From 56d90565e16a6d4ae37af6324a3f85d3bdeb5d39 Mon Sep 17 00:00:00 2001
From: gspark <gspark@etri.re.kr>
Date: Fri, 23 Oct 2026 15:27:46 +0900
Subject: [PATCH 1/6] feat: add experimental Qwen3 RNGD prefill validation

---
 opt_prime/examples/pd/qwen3_rngd/README.md    |  63 +++++
 .../examples/pd/qwen3_rngd/check_atom.py      | 111 ++++++++
 opt_prime/examples/pd/qwen3_rngd/prefill.py   | 196 ++++++++++++++
 .../examples/pd/qwen3_rngd/run_prefill.py     |  47 ++++
 opt_prime/examples/pd/qwen3_rngd/spec.py      |  33 +++
 opt_prime/examples/pd/qwen3_rngd/test_spec.py |  43 ++++
 .../examples/pd/qwen3_rngd/validation.json    | 239 ++++++++++++++++++
 7 files changed, 732 insertions(+)
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/README.md
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/check_atom.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/prefill.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/run_prefill.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/spec.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/test_spec.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/validation.json

diff --git a/opt_prime/examples/pd/qwen3_rngd/README.md b/opt_prime/examples/pd/qwen3_rngd/README.md
new file mode 100644
index 00000000..4f98ba0d
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/README.md
@@ -0,0 +1,63 @@
+# Experimental Qwen3 RNGD prefill → ATOM validation (#197)
+
+Qwen3-0.6B requires Q/K normalization, unbiased projections, and a2048-wide query
+projection despite hidden_size1024. This separate graph implements those semantics;
+it does not loosen the existing Qwen2-only DenseModel contract or register a new
+production serving backend. Quantized weights, scaled RoPE, sliding attention and
+non-Qwen3 configs are rejected. Only Qwen3-0.6B has been hardware-validated here.
+
+## Reproduce
+
+Use the isolated single-card device-plugin Pod setup from PR195's
+`ATOM_CONTAINER_EXPERIMENT.md`, including validated Python environments, HOME,
+offline model cache, device leases, and the fusion4/Mesh2 Furiosa launcher. Requires
+existing `rngd_pd` and optimus-rm modules on PYTHONPATH, plus PR195's
+`atom_stream_upload.py` and `check_atom_container.py`. Do not install into or mutate
+another workload's environment. Work from a writable scratch directory containing
+these files. Use only the allocated devices.
+
+```bash
+python3 test_spec.py
+# Export FURIOSA_PYTHON to the validated Furiosa2026.4 interpreter.
+./furiosa_prefill_python.sh -m run_prefill
+"$ATOM_PYTHON" check_atom_container.py check_atom.py
+```
+
+`run_prefill.py` writes trusted tensor data to qwen3_197_kv.pt. `check_atom.py`
+compiles/loads its own B1/context512/3KV-block artifact under qwen3-197-atom, compares
+CPU FP32 prefill, native ATOM prefill and imported RNGD KV, and writes
+qwen3_197_atom.json. It uses both the existing and portable importers and reads back
+every imported slot. These scripts run a finite fixture, not an HTTP service.
+
+## Hardware evidence (2026-10-23)
+
+RNGD: Furiosa2026.4,fusion4/Mesh2,width128,28layers,8KV heads,head_dim128.
+ATOM: rebel-compiler0.11.0,optimum-rbln0.11.0.post1,torch2.11.0+cpu,
+B1/context512/3KV blocks. Checkpoint Qwen/Qwen3-0.6B; three chat prompts with thinking
+disabled. Exact records in validation.json; compilation provenance in the local
+research experiment2026-10-23-gemma-fix-qwen3-koni.md.
+
+-3/3first tokens match CPU FP32 and native ATOM.
+-3/3fixed16-token PD outputs match native ATOM.
+-3/3full decode-logit sequences match between original and portable importers.
+-336/336slot readbacks exact:3prompts ×2importers ×56slots.
+-RNGD-vs-CPU relative RMS:K0.470–0.491%,V1.757–1.911%,last logits1.164–1.604%.
+-Prefill wall time including host KV/logits:1.321s first,0.723/0.714s subsequent.
+  This is a correctness bring-up, not vendor-native performance parity.
+
+Fixed-length diagnostic generation continues beyondEOS; service termination semantics
+are not tested. The Korean-capital answer is incorrectlyIncheon on both CPU/ATOM in
+this fixture; matching tokens does not establish factual quality. No DMA, zero-copy,
+streamed source export or overlap claim. Host transfer occurs after complete prefill.
+
+The initial compile failed because epsilon/exponent inputs wereBF16 while rms_dsl
+requiresFP32. They now remainFP32; weights/activation carriers remainBF16. Failed
+compile log is retained with the experiment. CPU contract tests reject unsupported
+semantics; hardware records are essential evidence beyond those tests.
+
+## Limits
+
+Other sizes, longer prompts, concurrent requests, prefix reuse, cancellation and
+production recovery/lifetime behavior are unvalidated. The handoff has not been
+wired into npu_pd_bench/serve_batched. Keep model-specific math in this prefiller;
+the portable56-slot import succeeded without communication-library changes.
diff --git a/opt_prime/examples/pd/qwen3_rngd/check_atom.py b/opt_prime/examples/pd/qwen3_rngd/check_atom.py
new file mode 100644
index 00000000..61edc38d
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/check_atom.py
@@ -0,0 +1,111 @@
+import json
+from pathlib import Path
+
+import numpy as np
+import torch
+from atom_stream_upload import import_atom_span
+from opt_prime.accel.platforms.atom.engine import AtomRblnEngine
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+mid = "Qwen/Qwen3-0.6B"
+physical = int(next(Path("/dev").glob("rbln[0-9]*")).name[4:])
+e = AtomRblnEngine(
+    devices=[physical],
+    max_seq_len=512,
+    batch_size=1,
+    kv_blocks=3,
+    cache_dir=str(Path("qwen3-197-atom").resolve()),
+)
+e.start(mid)
+records = torch.load("qwen3_197_kv.pt", weights_only=True)
+ref = AutoModelForCausalLM.from_pretrained(mid, dtype=torch.float32).eval()
+tok = AutoTokenizer.from_pretrained(mid)
+b = e.batcher(prefix_cache=False)
+rows = []
+
+
+def decode(first, n):
+    out = [first]
+    vectors = []
+    for j in range(15):
+        z = (
+            b.decoder(
+                input_ids=torch.tensor([[out[-1]]]),
+                cache_position=torch.tensor([[n + j]], dtype=torch.int32),
+                block_tables=b.blocks.table([0]),
+            )
+            .logits[0, -1]
+            .float()
+        )
+        assert torch.isfinite(z).all()
+        vectors.append(z.clone())
+        out.append(int(z.argmax()))
+    return out, torch.stack(vectors)
+
+
+with torch.inference_mode():
+    for r in records:
+        ids = r["ids"]
+        n = len(ids)
+        b.blocks.admit(0, ids)
+        block = b.blocks.lane_block[0]
+        nl = b._prefill(0, ids).logits[0, -1].float()
+        native, _ = decode(int(nl.argmax()), n)
+        hf = ref(torch.tensor([ids]), use_cache=True)
+        hl = hf.logits[0, -1].float()
+        ll = r["logits"].float()
+        ck = torch.stack([l.keys[0] for l in hf.past_key_values.layers])
+        cv = torch.stack([l.values[0] for l in hf.past_key_values.layers])
+        kr = float(
+            (r["keys"].float() - ck).square().mean().sqrt() / ck.square().mean().sqrt()
+        )
+        vr = float(
+            (r["values"].float() - cv).square().mean().sqrt()
+            / cv.square().mean().sqrt()
+        )
+        variants = {}
+        saved = []
+        for name in ["original", "portable"]:
+            if name == "original":
+                b.blocks.import_span(0, 0, r["keys"], r["values"])
+            else:
+                import_atom_span(b.blocks, 0, 0, r["keys"], r["values"])
+            exact = 0
+            for meta, buf in zip(b.blocks.metas, b.blocks._staging()):
+                expected = buf[block, :, :n].copy()
+                buf.fill(0)
+                b.blocks.io_rt.get_kv_cache_for_layer(meta["name"], block, 0, n, buf)
+                assert np.array_equal(expected, buf[block, :, :n])
+                exact += 1
+            tokens, vectors = decode(int(ll.argmax()), n)
+            saved.append(vectors)
+            variants[name] = {
+                "tokens": tokens,
+                "text": tok.decode(tokens),
+                "readback_exact_slots": exact,
+                "native_equal": tokens == native,
+            }
+        row = {
+            "prompt": r["prompt"],
+            "prompt_tokens": n,
+            "first_cpu": int(hl.argmax()),
+            "first_rngd": int(ll.argmax()),
+            "first_atom": int(nl.argmax()),
+            "key_relative_rms": kr,
+            "value_relative_rms": vr,
+            "logits_relative_rms": float(
+                (ll - hl).square().mean().sqrt() / hl.square().mean().sqrt()
+            ),
+            "prefill_s": r["prefill_s"],
+            "native_tokens": native,
+            "native_text": tok.decode(native),
+            "variants": variants,
+            "importer_logits_exact": bool(torch.equal(*saved)),
+        }
+        rows.append(row)
+        print(json.dumps(row, ensure_ascii=False), flush=True)
+        b.blocks.release(0, None)
+Path("qwen3_197_atom.json").write_text(
+    json.dumps(rows, ensure_ascii=False, indent=2) + "\n"
+)
+e.close()
diff --git a/opt_prime/examples/pd/qwen3_rngd/prefill.py b/opt_prime/examples/pd/qwen3_rngd/prefill.py
new file mode 100644
index 00000000..ad45b85f
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/prefill.py
@@ -0,0 +1,196 @@
+"""Experimental Qwen3 full-prefix graph; requires validated Furiosa environment."""
+
+import math
+
+import torch
+from rngd_pd.engines.fsdk import (
+    CompileModule,
+    FCConfig,
+    TacticKernelModule,
+    dev_str,
+    hint,
+    sync,
+)
+from rngd_pd.kernels.prefill_chunk import rms_dsl, softmax_dsl, swiglu_dsl, zipadd_dsl
+from spec import validate
+
+BF = torch.bfloat16
+
+
+class Layer(torch.nn.Module):
+    def __init__(self, s, width):
+        super().__init__()
+        self.s = s
+        self.width = width
+        h = s["hidden_size"]
+        d = s["head_dim"]
+        nh = s["num_attention_heads"]
+        nk = s["num_key_value_heads"]
+        self.norm = TacticKernelModule(rms_dsl(width, h))
+        self.add = TacticKernelModule(zipadd_dsl(width, h))
+        self.qnorm = TacticKernelModule(rms_dsl(width * nh, d))
+        self.knorm = TacticKernelModule(rms_dsl(width * nk, d))
+        self.act = TacticKernelModule(swiglu_dsl(width, s["intermediate_size"]))
+        self.softmax = TacticKernelModule(
+            softmax_dsl(nk, (nh // nk) * width, width).replace("/bf16", "/f32")
+        )
+
+    def forward(
+        self,
+        x,
+        eps,
+        exponent,
+        qeps,
+        qexp,
+        keps,
+        kexp,
+        n1,
+        n2,
+        qn,
+        kn,
+        wq,
+        wk,
+        wv,
+        wo,
+        wg,
+        wu,
+        wd,
+        rotation,
+        mask,
+    ):
+        s = self.s
+        c = self.width
+        d = s["head_dim"]
+        nh = s["num_attention_heads"]
+        nk = s["num_key_value_heads"]
+        g = nh // nk
+        h = self.norm(x, n1, eps, exponent)[0]
+        q = self.qnorm((h @ wq).reshape(c * nh, d), qn, qeps, qexp)[0].reshape(c, nh, d)
+        k = self.knorm((h @ wk).reshape(c * nk, d), kn, keps, kexp)[0].reshape(c, nk, d)
+        q = torch.matmul(q, rotation).to(BF)
+        k = torch.matmul(k, rotation).to(BF)
+        keys = k.permute(1, 2, 0)
+        values = (h @ wv).reshape(c, nk, d).permute(1, 0, 2)
+        query = q.reshape(c, nk, g, d).permute(1, 2, 0, 3).reshape(nk, g * c, d)
+        probs = self.softmax((query.float() @ keys.float()) * (d**-0.5) + mask.float())[
+            0
+        ].to(BF)
+        att = (
+            (probs @ values).reshape(nk, g, c, d).permute(2, 0, 1, 3).reshape(c, nh * d)
+        )
+        res = self.add(att @ wo, x)[0]
+        h = self.norm(res, n2, eps, exponent)[0]
+        out = self.add(self.act(h @ wg, h @ wu)[0] @ wd, res)[0]
+        return out, torch.stack((keys, values.transpose(-1, -2)))
+
+
+class Prefill:
+    def __init__(self, config, weights, width=128, chip=0):
+        s = self.s = validate(config)
+        self.width = width
+        self.chip = chip
+        if width <= 0 or width > 512 or width % 32:
+            raise ValueError("width must be a multiple of32 within512")
+        device = dev_str(chip)
+        self.device = device
+
+        def put(x):
+            return x.contiguous().to(BF).to(device)
+
+        self.embedding = weights["model.embed_tokens.weight"]
+        h = s["hidden_size"]
+        d = s["head_dim"]
+        nh = s["num_attention_heads"]
+        nk = s["num_key_value_heads"]
+        eps = s["rms_norm_eps"]
+        c = width
+        self.fixed = [
+            torch.full((r, col), v, dtype=torch.float32).to(device)
+            for r, col, v in [
+                (c, h, eps * h),
+                (c, h, -0.5),
+                (c * nh, d, eps * d),
+                (c * nh, d, -0.5),
+                (c * nk, d, eps * d),
+                (c * nk, d, -0.5),
+            ]
+        ]
+        self.layers = []
+        for i in range(s["num_hidden_layers"]):
+
+            def w(k, i=i):
+                return weights[f"model.layers.{i}." + k]
+
+            norms = [
+                put(w(k).repeat(n, 1) * math.sqrt(dim))
+                for k, n, dim in [
+                    ("input_layernorm.weight", c, h),
+                    ("post_attention_layernorm.weight", c, h),
+                    ("self_attn.q_norm.weight", c * nh, d),
+                    ("self_attn.k_norm.weight", c * nk, d),
+                ]
+            ]
+            mats = [
+                put(w(k).T)
+                for k in [
+                    "self_attn.q_proj.weight",
+                    "self_attn.k_proj.weight",
+                    "self_attn.v_proj.weight",
+                    "self_attn.o_proj.weight",
+                    "mlp.gate_proj.weight",
+                    "mlp.up_proj.weight",
+                    "mlp.down_proj.weight",
+                ]
+            ]
+            self.layers.append(norms + mats)
+        inv = 1 / (s["rope_theta"] ** (torch.arange(0, d, 2).float() / d))
+        angle = torch.arange(c)[:, None] * inv[None]
+        rot = torch.zeros(c, d, d)
+        idx = torch.arange(d // 2)
+        rot[:, idx, idx] = angle.cos()
+        rot[:, idx + d // 2, idx + d // 2] = angle.cos()
+        rot[:, idx, idx + d // 2] = angle.sin()
+        rot[:, idx + d // 2, idx] = -angle.sin()
+        self.rotation = put(rot)
+        self.mask = put(
+            torch.triu(torch.full((c, c), -1e4), 1).repeat(nh // nk, 1)[None]
+        )
+        tensors = [
+            torch.zeros(c, h, dtype=BF),
+            *self.fixed,
+            *self.layers[0],
+            self.rotation,
+            self.mask,
+        ]
+        sample = tuple(torch.zeros(tuple(t.shape), dtype=t.dtype) for t in tensors)
+        self.module = CompileModule.from_module(
+            Layer(s, c),
+            sample,
+            compiler_config=FCConfig(tactic_hint=hint("ForLlmModelPrefill")),
+        )
+        self.module.to(device)
+        sync(chip)
+        self.norm = weights["model.norm.weight"].float()
+        self.head = weights.get("lm_head.weight", self.embedding).float()
+
+    def prefill(self, tokens):
+        if not 0 < len(tokens) <= self.width:
+            raise ValueError("prompt exceeds compiled capacity")
+        x = (
+            self.embedding[tokens + [tokens[-1]] * (self.width - len(tokens))]
+            .to(BF)
+            .to(self.device)
+        )
+        keys = []
+        values = []
+        for w in self.layers:
+            x, kv = self.module(x, *self.fixed, *w, self.rotation, self.mask)
+            keys.append(kv[0])
+            values.append(kv[1].transpose(-1, -2))
+        sync(self.chip)
+        n = len(tokens)
+        k = torch.stack(keys).cpu()[:, :, :, :n].transpose(-1, -2)
+        v = torch.stack(values).cpu()[:, :, :n]
+        z = x.cpu()[n - 1].float()
+        z = z * torch.rsqrt(z.square().mean() + self.s["rms_norm_eps"]) * self.norm
+        return k, v, z @ self.head.T
diff --git a/opt_prime/examples/pd/qwen3_rngd/run_prefill.py b/opt_prime/examples/pd/qwen3_rngd/run_prefill.py
new file mode 100644
index 00000000..337fd4ed
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/run_prefill.py
@@ -0,0 +1,47 @@
+import json
+import time
+from pathlib import Path
+
+import torch
+from huggingface_hub import snapshot_download
+from prefill import Prefill
+from safetensors import safe_open
+from transformers import AutoTokenizer
+
+mid = "Qwen/Qwen3-0.6B"
+path = Path(snapshot_download(mid, local_files_only=True))
+c = json.loads((path / "config.json").read_text())
+w = {}
+for shard in sorted(path.glob("*.safetensors")):
+    with safe_open(shard, framework="pt") as f:
+        for key in f:
+            w[key] = f.get_tensor(key).float()
+e = Prefill(c, w, width=128, chip=0)
+print("READY", flush=True)
+tok = AutoTokenizer.from_pretrained(mid)
+records = []
+for prompt in [
+    "What is the capital of France?",
+    "Compute 2 + 3.",
+    "한국의 수도는 어디인가요?",
+]:
+    ids = tok.apply_chat_template(
+        [{"role": "user", "content": prompt}],
+        tokenize=True,
+        add_generation_prompt=True,
+        return_dict=False,
+        enable_thinking=False,
+    )
+    t = time.perf_counter()
+    k, v, l = e.prefill(ids)
+    row = {
+        "prompt": prompt,
+        "ids": ids,
+        "keys": k,
+        "values": v,
+        "logits": l,
+        "prefill_s": time.perf_counter() - t,
+    }
+    records.append(row)
+    print(prompt, "first", int(l.argmax()), "seconds", row["prefill_s"], flush=True)
+torch.save(records, "qwen3_197_kv.pt")
diff --git a/opt_prime/examples/pd/qwen3_rngd/spec.py b/opt_prime/examples/pd/qwen3_rngd/spec.py
new file mode 100644
index 00000000..903c30fe
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/spec.py
@@ -0,0 +1,33 @@
+"""Strict semantics for the experimental dense Qwen3 graph, SDK independent."""
+
+
+def validate(config):
+    c = dict(config)
+    if (
+        c.get("model_type") != "qwen3"
+        or c.get("hidden_act") != "silu"
+        or c.get("attention_bias", False)
+    ):
+        raise ValueError("requires unbiased dense Qwen3/SwiGLU")
+    if (
+        c.get("quantization_config")
+        or c.get("rope_scaling")
+        or c.get("use_sliding_window", False)
+    ):
+        raise ValueError(
+            "quantization, scaled RoPE and sliding attention need separate validation"
+        )
+    for k in [
+        "hidden_size",
+        "num_attention_heads",
+        "num_key_value_heads",
+        "num_hidden_layers",
+        "head_dim",
+    ]:
+        if type(c.get(k)) is not int or c[k] <= 0:
+            raise ValueError("invalid " + k)
+    if c["head_dim"] != 128 or c["num_attention_heads"] % c["num_key_value_heads"]:
+        raise ValueError("unsupported head geometry")
+    c["query_width"] = c["num_attention_heads"] * c["head_dim"]
+    c["slots"] = 2 * c["num_hidden_layers"]
+    return c
diff --git a/opt_prime/examples/pd/qwen3_rngd/test_spec.py b/opt_prime/examples/pd/qwen3_rngd/test_spec.py
new file mode 100644
index 00000000..93ba1be9
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/test_spec.py
@@ -0,0 +1,43 @@
+import unittest
+
+from spec import validate
+
+
+class SpecTests(unittest.TestCase):
+    def config(self, **kw):
+        c = {
+            "model_type": "qwen3",
+            "hidden_size": 1024,
+            "num_attention_heads": 16,
+            "num_key_value_heads": 8,
+            "head_dim": 128,
+            "num_hidden_layers": 28,
+            "hidden_act": "silu",
+            "attention_bias": False,
+            "rope_theta": 1e6,
+        }
+        c.update(kw)
+        return c
+
+    def test_query_width_is_independent(self):
+        s = validate(self.config())
+        self.assertEqual(s["query_width"], 2048)
+        self.assertEqual(s["hidden_size"], 1024)
+        self.assertEqual(s["slots"], 56)
+
+    def test_rejects_other_semantics(self):
+        for kw in [
+            {"model_type": "qwen2"},
+            {"quantization_config": {"quant_method": "fp8"}},
+            {"use_sliding_window": True},
+            {"rope_scaling": {"rope_type": "linear"}},
+            {"attention_bias": True},
+            {"num_key_value_heads": 3},
+            {"head_dim": 64},
+        ]:
+            with self.subTest(kw=kw), self.assertRaises(ValueError):
+                validate(self.config(**kw))
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/opt_prime/examples/pd/qwen3_rngd/validation.json b/opt_prime/examples/pd/qwen3_rngd/validation.json
new file mode 100644
index 00000000..3211a05c
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/validation.json
@@ -0,0 +1,239 @@
+[
+  {
+    "prompt": "What is the capital of France?",
+    "prompt_tokens": 19,
+    "first_cpu": 785,
+    "first_rngd": 785,
+    "first_atom": 785,
+    "key_relative_rms": 0.004695013631135225,
+    "value_relative_rms": 0.01862264797091484,
+    "logits_relative_rms": 0.011644327081739902,
+    "prefill_s": 1.3208831560332328,
+    "native_tokens": [
+      785,
+      6722,
+      315,
+      9625,
+      374,
+      3070,
+      59604,
+      334,
+      13,
+      151645,
+      198,
+      151643,
+      33975,
+      25,
+      3555,
+      374
+    ],
+    "native_text": "The capital of France is **Paris**.<|im_end|>\n<|endoftext|>Human: What is",
+    "variants": {
+      "original": {
+        "tokens": [
+          785,
+          6722,
+          315,
+          9625,
+          374,
+          3070,
+          59604,
+          334,
+          13,
+          151645,
+          198,
+          151643,
+          33975,
+          25,
+          3555,
+          374
+        ],
+        "text": "The capital of France is **Paris**.<|im_end|>\n<|endoftext|>Human: What is",
+        "readback_exact_slots": 56,
+        "native_equal": true
+      },
+      "portable": {
+        "tokens": [
+          785,
+          6722,
+          315,
+          9625,
+          374,
+          3070,
+          59604,
+          334,
+          13,
+          151645,
+          198,
+          151643,
+          33975,
+          25,
+          3555,
+          374
+        ],
+        "text": "The capital of France is **Paris**.<|im_end|>\n<|endoftext|>Human: What is",
+        "readback_exact_slots": 56,
+        "native_equal": true
+      }
+    },
+    "importer_logits_exact": true
+  },
+  {
+    "prompt": "Compute 2 + 3.",
+    "prompt_tokens": 19,
+    "first_cpu": 17,
+    "first_rngd": 17,
+    "first_atom": 17,
+    "key_relative_rms": 0.004911903291940689,
+    "value_relative_rms": 0.01911376416683197,
+    "logits_relative_rms": 0.01556027214974165,
+    "prefill_s": 0.723041116958484,
+    "native_tokens": [
+      17,
+      488,
+      220,
+      18,
+      284,
+      220,
+      20,
+      13,
+      151645,
+      198,
+      151643,
+      641,
+      279,
+      2266,
+      315,
+      279
+    ],
+    "native_text": "2 + 3 = 5.<|im_end|>\n<|endoftext|>In the context of the",
+    "variants": {
+      "original": {
+        "tokens": [
+          17,
+          488,
+          220,
+          18,
+          284,
+          220,
+          20,
+          13,
+          151645,
+          198,
+          151643,
+          641,
+          279,
+          2266,
+          315,
+          279
+        ],
+        "text": "2 + 3 = 5.<|im_end|>\n<|endoftext|>In the context of the",
+        "readback_exact_slots": 56,
+        "native_equal": true
+      },
+      "portable": {
+        "tokens": [
+          17,
+          488,
+          220,
+          18,
+          284,
+          220,
+          20,
+          13,
+          151645,
+          198,
+          151643,
+          641,
+          279,
+          2266,
+          315,
+          279
+        ],
+        "text": "2 + 3 = 5.<|im_end|>\n<|endoftext|>In the context of the",
+        "readback_exact_slots": 56,
+        "native_equal": true
+      }
+    },
+    "importer_logits_exact": true
+  },
+  {
+    "prompt": "한국의 수도는 어디인가요?",
+    "prompt_tokens": 22,
+    "first_cpu": 23573,
+    "first_rngd": 23573,
+    "first_atom": 23573,
+    "key_relative_rms": 0.004876873455941677,
+    "value_relative_rms": 0.01756548509001732,
+    "logits_relative_rms": 0.016039395704865456,
+    "prefill_s": 0.7141205350635573,
+    "native_tokens": [
+      23573,
+      124785,
+      20401,
+      134013,
+      16560,
+      3070,
+      31328,
+      129034,
+      334,
+      78952,
+      13,
+      151645,
+      198,
+      151643,
+      33975,
+      25
+    ],
+    "native_text": "한국의 수도는 **인천**입니다.<|im_end|>\n<|endoftext|>Human:",
+    "variants": {
+      "original": {
+        "tokens": [
+          23573,
+          124785,
+          20401,
+          134013,
+          16560,
+          3070,
+          31328,
+          129034,
+          334,
+          78952,
+          13,
+          151645,
+          198,
+          151643,
+          33975,
+          25
+        ],
+        "text": "한국의 수도는 **인천**입니다.<|im_end|>\n<|endoftext|>Human:",
+        "readback_exact_slots": 56,
+        "native_equal": true
+      },
+      "portable": {
+        "tokens": [
+          23573,
+          124785,
+          20401,
+          134013,
+          16560,
+          3070,
+          31328,
+          129034,
+          334,
+          78952,
+          13,
+          151645,
+          198,
+          151643,
+          33975,
+          25
+        ],
+        "text": "한국의 수도는 **인천**입니다.<|im_end|>\n<|endoftext|>Human:",
+        "readback_exact_slots": 56,
+        "native_equal": true
+      }
+    },
+    "importer_logits_exact": true
+  }
+]

From d4b6fe6ba764a43ec66538e448cff5eb98220965 Mon Sep 17 00:00:00 2001
From: gspark <gspark@etri.re.kr>
Date: Fri, 23 Oct 2026 16:02:29 +0900
Subject: [PATCH 2/6] perf: remove Qwen3 device-stack bottleneck and validate
 HTTP PD (#199)

---
 opt_prime/examples/pd/qwen3_rngd/README.md    |  53 +-
 .../pd/qwen3_rngd/SERVING_EXPERIMENT.md       |  61 ++
 .../examples/pd/qwen3_rngd/bench_export.py    |  52 ++
 .../examples/pd/qwen3_rngd/bench_http.py      |  68 ++
 .../examples/pd/qwen3_rngd/bench_stream.py    |  74 ++
 opt_prime/examples/pd/qwen3_rngd/prefill.py   |   5 +-
 .../examples/pd/qwen3_rngd/run_prefill.py     |   2 +-
 opt_prime/examples/pd/qwen3_rngd/serve.py     | 100 +++
 .../qwen3_197_export_profile.json             | 155 ++++
 .../serving_results/qwen3_197_profile.json    | 614 ++++++++++++++++
 .../serving_results/qwen3_199_export_ab.json  | 170 +++++
 .../serving_results/qwen3_199_http_atom.json  | 665 ++++++++++++++++++
 .../qwen3_199_http_baseline.json              | 665 ++++++++++++++++++
 .../serving_results/qwen3_199_http_pd.json    | 665 ++++++++++++++++++
 .../qwen3_199_stream_atom.json                |  97 +++
 .../qwen3_199_stream_baseline.json            |  97 +++
 .../serving_results/qwen3_199_stream_pd.json  |  97 +++
 .../examples/pd/qwen3_rngd/test_worker.py     |  48 ++
 opt_prime/examples/pd/qwen3_rngd/worker.py    | 121 ++++
 .../examples/pd/qwen3_rngd/worker_python.sh   |   9 +
 20 files changed, 3812 insertions(+), 6 deletions(-)
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/SERVING_EXPERIMENT.md
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/bench_export.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/bench_http.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/bench_stream.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serve.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_197_export_profile.json
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_197_profile.json
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_export_ab.json
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_atom.json
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_baseline.json
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_pd.json
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_atom.json
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_baseline.json
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_pd.json
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/test_worker.py
 create mode 100644 opt_prime/examples/pd/qwen3_rngd/worker.py
 create mode 100755 opt_prime/examples/pd/qwen3_rngd/worker_python.sh

diff --git a/opt_prime/examples/pd/qwen3_rngd/README.md b/opt_prime/examples/pd/qwen3_rngd/README.md
index 4f98ba0d..6c6c8f4e 100644
--- a/opt_prime/examples/pd/qwen3_rngd/README.md
+++ b/opt_prime/examples/pd/qwen3_rngd/README.md
@@ -27,7 +27,7 @@ python3 test_spec.py
 compiles/loads its own B1/context512/3KV-block artifact under qwen3-197-atom, compares
 CPU FP32 prefill, native ATOM prefill and imported RNGD KV, and writes
 qwen3_197_atom.json. It uses both the existing and portable importers and reads back
-every imported slot. These scripts run a finite fixture, not an HTTP service.
+every imported slot. These two scripts run a finite fixture; the experimental HTTP adapter is described below.
 
 ## Hardware evidence (2026-10-23)
 
@@ -58,6 +58,53 @@ semantics; hardware records are essential evidence beyond those tests.
 ## Limits
 
 Other sizes, longer prompts, concurrent requests, prefix reuse, cancellation and
-production recovery/lifetime behavior are unvalidated. The handoff has not been
-wired into npu_pd_bench/serve_batched. Keep model-specific math in this prefiller;
+production recovery/lifetime behavior are unvalidated. The handoff is exposed by the explicit experimental `serve.py` adapter below;
+it is not automatically registered in npu_pd_bench/serve_batched. Keep model-specific math in this prefiller;
 the portable56-slot import succeeded without communication-library changes.
+
+
+## Experimental HTTP adapter and export improvement (#199)
+
+`serve.py` reuses the existing optimus-rm `SubprocessPrefillWorker`, ATOM batcher,
+`BatchedService` and `BatchedOwner`. `worker_python.sh` redirects only the expected
+prefill worker module launch to `worker.py`; model-specific math stays in
+`prefill.py`. This is an explicit experiment requiring the existing optimus-rm
+runtime on PYTHONPATH, not a new production backend in this checkout.
+
+Place PR195's `furiosa_prefill_python.sh` alongside `worker_python.sh` and use its
+single-card allocation/bootstrap instructions. Keep the task's own compiled
+artifact cache and validated interpreters. The HTTP listener binds only loopback.
+The service is fixed to Qwen3-0.6B, B1/context512/3KV blocks, no prefix caching.
+Prompts above `--width` (default128) are rejected before worker submission. Decode
+requests may queue; only one lane executes. SIGTERM drains queued work before
+closing the worker and slab. Disconnect cancellation and fault recovery have not
+been hardware-validated in this experiment.
+
+```bash
+python3 -m unittest discover -s . -p 'test_*.py'
+# In the writable scratch directory with the validated fixture/artifacts:
+"$ATOM_PYTHON" check_atom_container.py serve.py --cache-dir "$PWD/qwen3-197-atom"
+# Another process in the same allocated Pod:
+"$ATOM_PYTHON" bench_http.py --name pd
+"$ATOM_PYTHON" bench_stream.py --name pd
+# Stop/drain the service, then restart with --local-prefill for the same ATOM-only baseline.
+```
+
+The export fix replaces device-side `torch.stack(...).cpu()` with per-layer
+`.cpu()` followed by CPU stacking. The device graph and width remain unchanged.
+To reproduce the alternating A/B, save the old implementation in the scratch
+working directory (not over the new prefiller):
+
+```bash
+git show 353d15090aa7908e0bb2ea5422f047e59ea9273d:opt_prime/examples/pd/qwen3_rngd/prefill.py > prefill_baseline.py
+./furiosa_prefill_python.sh -m bench_export
+```
+
+Measured warm prefill: 728.435→55.823ms mean (nine pairs); all twelve paired K,V
+and full-logit comparisons exact. Detailed serving results and limitations are
+in `SERVING_EXPERIMENT.md` and `serving_results/`. The old roughly719ms measurement
+above is historical, before this export fix. Width32/64 were also measured but
+changed numerical values, so the verified improvement retains width128.
+
+`safe_open` must be iterated through `.keys()` in this SDK environment; the
+specific lint exemption preserves that behavior.
diff --git a/opt_prime/examples/pd/qwen3_rngd/SERVING_EXPERIMENT.md b/opt_prime/examples/pd/qwen3_rngd/SERVING_EXPERIMENT.md
new file mode 100644
index 00000000..dbfb3bf0
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/SERVING_EXPERIMENT.md
@@ -0,0 +1,61 @@
+# Qwen3 RNGD → ATOM: host stacking and experimental HTTP serving
+
+- **Date**: 2026-10-23
+- **Task**: Qwen3 serving and export profiling (aicomp #199)
+- **Host / device**: RNGD+ATOM host, dedicated single-card device-plugin Pod; Furiosa 2026.4 fusion4/Mesh2; ATOM rebel-compiler 0.11.0, optimum-rbln 0.11.0.post1, torch 2.11.0+cpu
+- **Commit**: baseline 353d15090aa7908e0bb2ea5422f047e59ea9273d; experiment branch exp/199-qwen3-serving (56d9056 plus working changes before final commit)
+
+## Question
+
+Is the roughly 719 ms Qwen3 research prefill dominated by model computation or KV materialization, and does its improvement survive the existing batched HTTP service?
+
+## Setup
+
+Checkpoint Qwen/Qwen3-0.6B revision c1899de289a04d12100db370d81485cdf75e47ca. Three fixed chat prompts, thinking disabled, 19/19/22 tokens, fixed 16 output tokens for parity diagnostics. RNGD width128, 28 layers, 8 KV heads, dimension128. ATOM own B1/context512/3KV-block artifact from the previous experiment. Prefix caching disabled; B1 serializes three concurrent requests through the existing queue.
+
+Baseline stacks each K/V tensor on RNGD before copying. Improvement copies each layer to CPU and stacks there; model math, widths and sampler unchanged. Existing SubprocessPrefillWorker, AtomRblnEngine, BatchedOwner and HTTP handlers are reused through an explicit experimental launcher. No shared SDK files or other workloads changed. Host-mediated KV, no DMA or producer overlap.
+
+## Command
+
+Use PR195 isolated-container environment and bootstrap; working directory contains the Qwen3 example files and the previous trusted fixture. Set HOME to the SDK owner's home, offline HF cache, OMP_NUM_THREADS=4, MKL_NUM_THREADS=4, RBLN_DEVICES=0, RBLN_NPUS_PER_DEVICE=1, and PYTHONPATH to the existing optimus-rm, rngd_pd and hetero_p2p sources. FURIOSA_PYTHON and ATOM_PYTHON select the previously validated interpreters.
+
+```bash
+# From the checkout, save the exact pre-change implementation into the scratch directory:
+git show 353d15090aa7908e0bb2ea5422f047e59ea9273d:opt_prime/examples/pd/qwen3_rngd/prefill.py > prefill_baseline.py
+./furiosa_prefill_python.sh -m bench_export
+# Service on a dedicated Pod, loopback only. Baseline adds --local-prefill.
+"$ATOM_PYTHON" check_atom_container.py serve.py --cache-dir "$PWD/qwen3-197-atom"
+"$ATOM_PYTHON" bench_http.py --name pd
+"$ATOM_PYTHON" bench_stream.py --name pd
+```
+
+A/B export alternates baseline/optimized order over four rounds and three prompts; round0 excluded from latency means. Whole K,V,logits are compared bit-for-bit each pair. HTTP harness performs twelve sequential requests, three concurrent requests, an over-capacity HTTP400 check and a successful follow-up. Streaming harness performs nine requests plus EOS termination; first round excluded from warm means.
+
+## Results
+
+- Alternating A/B: baseline 728.435 ms mean (688.581–829.348), host stacking 55.823 ms (44.516–67.473), nine warm pairs; 12/12 complete K,V,logit comparisons exact including first round.
+- Phase diagnostic: device stack/export roughly 650 ms versus per-layer CPU export/host stack roughly 11 ms. Layer execution plus final hidden materialization roughly 5 ms; CPU output head roughly 26–40 ms. These are diagnostic wall times with fences, not isolated DMA bandwidth.
+- Width32/64 reduced the original stack cost but changed KV/logits numerically; not adopted. Width128 stays unchanged.
+- Optimized PD HTTP: 16/16 full output sequences equal ATOM-only; 16 remote prefills, zero local fallback. Warm sequential nonstream mean201.439 ms vs ATOM121.831 ms.
+- Warm streaming (six requests): optimized PD TTFT69.567 ms/E2E167.535 ms; ATOM TTFT15.883 ms/E2E128.254 ms.
+- Oversize129-token prompt rejectedHTTP400 before device work; next request succeeds. EOS termination passes both routes. Fixed-length diagnostics intentionally ignoreEOS and are not quality tests.
+
+Failures retained: first profiling script attempted to iterate safe_open directly (TypeError); changed to .keys(), with targeted SIM118 exemption. Initial experimental worker launch used a relative executable path and failed before worker startup; changed to an absolute sibling path. Parent resource_tracker cleaned the abandoned slab. No shared package modified.
+
+Raw logs/results: analysis/local_prefill/native_dma/general/serving199/. Exact local launch scripts retained with this experiment's artifacts; public reproduction uses role/environment variable names.
+
+### Pre-change HTTP baseline and final comparison
+
+| Route | Warm streaming TTFT (ms) | Warm streaming E2E,16 outputs (ms) | Warm nonstream E2E (ms) |
+|---|---:|---:|---:|
+| PD, device-stack baseline |755.042|905.176|887.234|
+| PD, host-stack improvement |69.567|167.535|201.439|
+| ATOM-only |15.883|128.254|121.831|
+
+Streaming means use six requests after excluding the first round; nonstream means use nine requests after excluding the first round. Routes ran sequentially (optimized PD, ATOM, old PD), not a randomized serving trial. No quiet-host guarantee or confidence intervals; the alternating in-process A/B supports the export mechanism independently. All16 nonstream token sequences match across all three routes. All three pass oversized-request rejection/recovery and EOS termination. The pre-change service uses exactly the old prefiller file in the same task-owned scratch directory, restored to the improved file afterwards.
+
+The improved PD first-token gap to ATOM is53.684ms; E2E gap39.281ms in this small streaming sample. Source export and CPU LM head now merit optimization; no producer-side overlap was added. Server drains were successful and the dedicated validation namespace was deleted; other workloads untouched.
+
+## Conclusion
+
+The largest original delay was device-side KV stacking/materialization, not inherent PD disaggregation or transformer-layer compute. Moving the stack to CPU yields about13× prefill speedup without numerical changes and enables the existing HTTP service. ATOM-only remains faster for these short B1 requests; this does not establish native Furiosa parity, throughput at high concurrency, or an advantage for PD. Remaining work: output-head cost, ATOM import/IPC and longer prompts; source-streaming overlap and KONI HTTP integration remain separate validations.
diff --git a/opt_prime/examples/pd/qwen3_rngd/bench_export.py b/opt_prime/examples/pd/qwen3_rngd/bench_export.py
new file mode 100644
index 00000000..1eeedf81
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/bench_export.py
@@ -0,0 +1,52 @@
+import json
+import time
+from pathlib import Path
+
+import torch
+from huggingface_hub import snapshot_download
+from prefill import Prefill
+from safetensors import safe_open
+
+path = Path(snapshot_download("Qwen/Qwen3-0.6B", local_files_only=True))
+c = json.loads((path / "config.json").read_text())
+w = {}
+for shard in path.glob("*.safetensors"):
+    with safe_open(shard, framework="pt") as f:
+        for k in f.keys():  # noqa: SIM118 - safe_open is not iterable
+            w[k] = f.get_tensor(k).float()
+records = torch.load("qwen3_197_kv.pt", weights_only=True)
+from prefill_baseline import Prefill as Baseline
+
+rows = []
+e = Prefill(c, w, width=128, chip=0)
+for repeat in range(4):
+    for i, r in enumerate(records):
+        outputs = {}
+        times = {}
+        for name in (
+            ["baseline", "host_stack"]
+            if repeat % 2 == 0
+            else ["host_stack", "baseline"]
+        ):
+            t = time.perf_counter()
+            outputs[name] = (
+                Baseline.prefill(e, r["ids"])
+                if name == "baseline"
+                else e.prefill(r["ids"])
+            )
+            times[name] = time.perf_counter() - t
+        exact = [
+            torch.equal(a, b)
+            for a, b in zip(outputs["baseline"], outputs["host_stack"])
+        ]
+        assert all(exact), exact
+        row = {
+            "repeat": repeat,
+            "prompt": i,
+            "n": len(r["ids"]),
+            "seconds": times,
+            "kv_logits_exact": exact,
+        }
+        rows.append(row)
+        print(json.dumps(row), flush=True)
+Path("qwen3_199_export_ab.json").write_text(json.dumps(rows, indent=2) + "\n")
diff --git a/opt_prime/examples/pd/qwen3_rngd/bench_http.py b/opt_prime/examples/pd/qwen3_rngd/bench_http.py
new file mode 100644
index 00000000..5d28d1fb
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/bench_http.py
@@ -0,0 +1,68 @@
+import argparse
+import concurrent.futures
+import json
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+import torch
+
+ap = argparse.ArgumentParser()
+ap.add_argument("--name", required=True)
+a = ap.parse_args()
+base = "http://127.0.0.1:18197"
+records = torch.load("qwen3_197_kv.pt", weights_only=True)
+rows = []
+
+
+def request(i, repeat, extra=None):
+    body = {
+        "model": "Qwen/Qwen3-0.6B",
+        "prompt": records[i]["ids"],
+        "temperature": 0,
+        "max_tokens": 16,
+        "ignore_eos": True,
+    }
+    body.update(extra or {})
+    req = urllib.request.Request(
+        base + "/v1/completions",
+        json.dumps(body).encode(),
+        headers={
+            "Content-Type": "application/json",
+            "x-request-id": f"{a.name}-{repeat}-{i}",
+        },
+    )
+    t = time.perf_counter()
+    with urllib.request.urlopen(req, timeout=120) as r:
+        out = json.load(r)
+    row = {
+        "repeat": repeat,
+        "prompt": i,
+        "seconds": time.perf_counter() - t,
+        "response": out,
+    }
+    print(json.dumps(row), flush=True)
+    return row
+
+
+for repeat in range(4):
+    for i in range(3):
+        rows.append(request(i, repeat))
+with concurrent.futures.ThreadPoolExecutor(max_workers=3) as pool:
+    rows.extend(pool.map(lambda i: request(i, 4), range(3)))
+# Capacity must be rejected before submitting to the worker, then service must remain usable.
+try:
+    request(0, 5, {"prompt": [1] * 129})
+    raise AssertionError("oversize request unexpectedly accepted")
+except urllib.error.HTTPError as e:
+    assert e.code == 400, e.code
+    rejected = json.loads(e.read())
+rows.append(request(0, 6))
+with urllib.request.urlopen(base + "/metrics") as r:
+    metrics = r.read().decode()
+Path(f"qwen3_199_http_{a.name}.json").write_text(
+    json.dumps({"rows": rows, "oversize": rejected, "metrics": metrics}, indent=2)
+    + "\n"
+)
+print(metrics, flush=True)
diff --git a/opt_prime/examples/pd/qwen3_rngd/bench_stream.py b/opt_prime/examples/pd/qwen3_rngd/bench_stream.py
new file mode 100644
index 00000000..56818743
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/bench_stream.py
@@ -0,0 +1,74 @@
+import argparse
+import json
+import time
+import urllib.request
+from pathlib import Path
+
+import torch
+
+ap = argparse.ArgumentParser()
+ap.add_argument("--name", required=True)
+a = ap.parse_args()
+records = torch.load("qwen3_197_kv.pt", weights_only=True)
+rows = []
+for repeat in range(3):
+    for i, r in enumerate(records):
+        payload = {
+            "model": "Qwen/Qwen3-0.6B",
+            "prompt": r["ids"],
+            "max_tokens": 16,
+            "temperature": 0,
+            "ignore_eos": True,
+            "stream": True,
+        }
+        req = urllib.request.Request(
+            "http://127.0.0.1:18197/v1/completions",
+            json.dumps(payload).encode(),
+            headers={
+                "Content-Type": "application/json",
+                "x-request-id": f"{a.name}-stream-{repeat}-{i}",
+            },
+        )
+        t = time.perf_counter()
+        first = None
+        done = False
+        chunks = []
+        with urllib.request.urlopen(req, timeout=120) as f:
+            for line in f:
+                if not line.startswith(b"data: "):
+                    continue
+                if line.strip() == b"data: [DONE]":
+                    done = True
+                    break
+                chunk = json.loads(line[6:])
+                chunks.append(chunk)
+                if (
+                    chunk.get("choices")
+                    and chunk["choices"][0].get("text")
+                    and first is None
+                ):
+                    first = time.perf_counter() - t
+        assert done and first is not None
+        row = {
+            "repeat": repeat,
+            "prompt": i,
+            "ttft_s": first,
+            "e2e_s": time.perf_counter() - t,
+            "text": "".join(
+                c["choices"][0].get("text", "") for c in chunks if c.get("choices")
+            ),
+        }
+        rows.append(row)
+        print(json.dumps(row), flush=True)
+payload.update(stream=False, ignore_eos=False, prompt=records[0]["ids"])
+req = urllib.request.Request(
+    "http://127.0.0.1:18197/v1/completions",
+    json.dumps(payload).encode(),
+    headers={"Content-Type": "application/json", "x-request-id": a.name + "-eos"},
+)
+with urllib.request.urlopen(req, timeout=120) as f:
+    eos = json.load(f)
+assert eos["choices"][0]["finish_reason"] == "stop", eos
+Path(f"qwen3_199_stream_{a.name}.json").write_text(
+    json.dumps({"rows": rows, "eos": eos}, indent=2) + "\n"
+)
diff --git a/opt_prime/examples/pd/qwen3_rngd/prefill.py b/opt_prime/examples/pd/qwen3_rngd/prefill.py
index ad45b85f..0fc527d1 100644
--- a/opt_prime/examples/pd/qwen3_rngd/prefill.py
+++ b/opt_prime/examples/pd/qwen3_rngd/prefill.py
@@ -189,8 +189,9 @@ def prefill(self, tokens):
             values.append(kv[1].transpose(-1, -2))
         sync(self.chip)
         n = len(tokens)
-        k = torch.stack(keys).cpu()[:, :, :, :n].transpose(-1, -2)
-        v = torch.stack(values).cpu()[:, :, :n]
+        # Host stacking avoids the expensive device-side concatenation path.
+        k = torch.stack([key.cpu() for key in keys])[:, :, :, :n].transpose(-1, -2)
+        v = torch.stack([value.cpu() for value in values])[:, :, :n]
         z = x.cpu()[n - 1].float()
         z = z * torch.rsqrt(z.square().mean() + self.s["rms_norm_eps"]) * self.norm
         return k, v, z @ self.head.T
diff --git a/opt_prime/examples/pd/qwen3_rngd/run_prefill.py b/opt_prime/examples/pd/qwen3_rngd/run_prefill.py
index 337fd4ed..90b53efe 100644
--- a/opt_prime/examples/pd/qwen3_rngd/run_prefill.py
+++ b/opt_prime/examples/pd/qwen3_rngd/run_prefill.py
@@ -14,7 +14,7 @@
 w = {}
 for shard in sorted(path.glob("*.safetensors")):
     with safe_open(shard, framework="pt") as f:
-        for key in f:
+        for key in f.keys():  # noqa: SIM118 - safe_open is not iterable
             w[key] = f.get_tensor(key).float()
 e = Prefill(c, w, width=128, chip=0)
 print("READY", flush=True)
diff --git a/opt_prime/examples/pd/qwen3_rngd/serve.py b/opt_prime/examples/pd/qwen3_rngd/serve.py
new file mode 100644
index 00000000..f6ba6738
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serve.py
@@ -0,0 +1,100 @@
+"""Experimental B1 Qwen3 HTTP service using existing optimus-rm ownership and batching."""
+
+import argparse
+import os
+import signal
+from pathlib import Path
+
+from opt_prime.accel.pd.pd_batching import SubprocessPrefillWorker
+from opt_prime.accel.platforms.atom.engine import AtomRblnEngine
+from opt_prime.accel.serve.batched import BatchedOwner, BatchedService
+from opt_prime.accel.serve.llmd import StageError, serve_stage
+from transformers import AutoTokenizer
+
+
+class BoundedService(BatchedService):
+    def __init__(self, *args, prompt_capacity, **kwargs):
+        self.prompt_capacity = prompt_capacity
+        super().__init__(*args, **kwargs)
+
+    def _request(self, rid, body, chat=False):
+        request, stops = super()._request(rid, body, chat)
+        if len(request.input_ids) > self.prompt_capacity:
+            raise StageError(400, "prompt exceeds compiled Qwen3 prefill width")
+        return request, stops
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--port", type=int, default=18197)
+    ap.add_argument("--width", type=int, default=128)
+    ap.add_argument("--cache-dir", required=True)
+    ap.add_argument(
+        "--local-prefill",
+        action="store_true",
+        help="ATOM-only baseline using the same service",
+    )
+    a = ap.parse_args()
+    devices = list(Path("/dev").glob("rbln[0-9]*"))
+    if len(devices) != 1:
+        raise ValueError("requires an isolated single-ATOM allocation")
+    mid = "Qwen/Qwen3-0.6B"
+    tok = AutoTokenizer.from_pretrained(mid)
+    engine = AtomRblnEngine(
+        devices=[int(devices[0].name[4:])],
+        max_seq_len=512,
+        batch_size=1,
+        kv_blocks=3,
+        cache_dir=a.cache_dir,
+    )
+    worker = server = None
+    try:
+        engine.start(mid)
+        c = engine.model.config.get_text_config()
+        if not a.local_prefill:
+            worker = SubprocessPrefillWorker(
+                str(Path(__file__).resolve().with_name("worker_python.sh")),
+                ["--model", mid, "--width", str(a.width)],
+                slots=1,
+                layers=c.num_hidden_layers,
+                kv_heads=c.num_key_value_heads,
+                head_dim=c.head_dim,
+                max_context=512,
+                env=dict(os.environ),
+            )
+        b = engine.batcher(
+            prefix_cache=False, prefill_worker=worker, local_prefill=a.local_prefill
+        )
+        eos = engine.model.generation_config.eos_token_id
+        svc = BoundedService(
+            b,
+            prompt_capacity=a.width,
+            model=mid,
+            vocab_size=c.vocab_size,
+            max_output=128,
+            render=lambda t: tok.decode(t, skip_special_tokens=True),
+            tokenizer=tok,
+            eos_token_ids=tuple(eos) if isinstance(eos, (tuple, list)) else (eos,),
+        )
+        owner = BatchedOwner(svc)
+
+        # Drain in-flight requests before releasing the shared-memory slab and devices.
+        def stop(*_):
+            owner.submit_async("stop")
+
+        signal.signal(signal.SIGTERM, stop)
+        signal.signal(signal.SIGINT, stop)
+        server, _ = serve_stage(owner, "decode", a.port)
+        print("HTTP_READY", flush=True)
+        owner.run()
+    finally:
+        if server is not None:
+            server.shutdown()
+            server.server_close()
+        if worker is not None:
+            worker.close()
+        engine.close()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_197_export_profile.json b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_197_export_profile.json
new file mode 100644
index 00000000..209ee0fc
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_197_export_profile.json
@@ -0,0 +1,155 @@
+[
+  {
+    "width": 128,
+    "repeat": 0,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0006219040369614959,
+    "layers_and_hidden_s": 0.004504419979639351,
+    "kv_export_s": 0.01061674696393311,
+    "hidden_export_s": 0.0002370449947193265,
+    "head_s": 0.030969161074608564,
+    "total_s": 0.04694927704986185,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 0,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.000629443908110261,
+    "layers_and_hidden_s": 0.00464651407673955,
+    "kv_export_s": 0.010553185013122857,
+    "hidden_export_s": 0.00026791589334607124,
+    "head_s": 0.03083170810714364,
+    "total_s": 0.04692876699846238,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 0,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0005821130471304059,
+    "layers_and_hidden_s": 0.004609663039445877,
+    "kv_export_s": 0.011402474017813802,
+    "hidden_export_s": 0.0002546359319239855,
+    "head_s": 0.030859099002555013,
+    "total_s": 0.04770798503886908,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 1,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0006276339991018176,
+    "layers_and_hidden_s": 0.004648193949833512,
+    "kv_export_s": 0.010822981013916433,
+    "hidden_export_s": 0.00021584494970738888,
+    "head_s": 0.03149873309303075,
+    "total_s": 0.0478133870055899,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 1,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0008148879278451204,
+    "layers_and_hidden_s": 0.004738376010209322,
+    "kv_export_s": 0.011202770052477717,
+    "hidden_export_s": 0.0002653059782460332,
+    "head_s": 0.031153884949162602,
+    "total_s": 0.048175224917940795,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 1,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0005853730253875256,
+    "layers_and_hidden_s": 0.0046637540217489,
+    "kv_export_s": 0.011821603984571993,
+    "hidden_export_s": 0.00026384496595710516,
+    "head_s": 0.030933211091905832,
+    "total_s": 0.04826778708957136,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 2,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0007110660662874579,
+    "layers_and_hidden_s": 0.0046412929659709334,
+    "kv_export_s": 0.011320922989398241,
+    "hidden_export_s": 0.00027665600646287203,
+    "head_s": 0.03126502805389464,
+    "total_s": 0.04821496608201414,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 2,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0008229190716519952,
+    "layers_and_hidden_s": 0.004791516927070916,
+    "kv_export_s": 0.011187429074198008,
+    "hidden_export_s": 0.00023271492682397366,
+    "head_s": 0.03093738097231835,
+    "total_s": 0.04797196097206324,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 2,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0006468149367719889,
+    "layers_and_hidden_s": 0.00464962306432426,
+    "kv_export_s": 0.011294952011667192,
+    "hidden_export_s": 0.00025847600772976875,
+    "head_s": 0.031100824009627104,
+    "total_s": 0.04795069003012031,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  }
+]
diff --git a/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_197_profile.json b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_197_profile.json
new file mode 100644
index 00000000..a07b921b
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_197_profile.json
@@ -0,0 +1,614 @@
+[
+  {
+    "width": 128,
+    "repeat": 0,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0008347210241481662,
+    "layers_s": 0.004783479031175375,
+    "kv_export_s": 0.6512959250248969,
+    "hidden_export_s": 0.0005497429519891739,
+    "head_s": 0.37443894799798727,
+    "total_s": 1.0319028160301968,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 0,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0008467710576951504,
+    "layers_s": 0.004424049984663725,
+    "kv_export_s": 0.6519521150039509,
+    "hidden_export_s": 0.0005255630239844322,
+    "head_s": 0.03527562797535211,
+    "total_s": 0.6930241270456463,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 0,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.001019895076751709,
+    "layers_s": 0.00468721694778651,
+    "kv_export_s": 0.6547371220076457,
+    "hidden_export_s": 0.0005087120225653052,
+    "head_s": 0.03990906197577715,
+    "total_s": 0.7008620080305263,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 1,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0009103730553761125,
+    "layers_s": 0.004774917964823544,
+    "kv_export_s": 0.6529824699973688,
+    "hidden_export_s": 0.0004310410004109144,
+    "head_s": 0.035557662951759994,
+    "total_s": 0.6946564649697393,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 1,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0008877119980752468,
+    "layers_s": 0.004763347911648452,
+    "kv_export_s": 0.6488819480873644,
+    "hidden_export_s": 0.00042821094393730164,
+    "head_s": 0.03430014196783304,
+    "total_s": 0.6892613609088585,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 1,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0010040350025519729,
+    "layers_s": 0.006587793934158981,
+    "kv_export_s": 0.6493192040361464,
+    "hidden_export_s": 0.00040609994903206825,
+    "head_s": 0.0340559349860996,
+    "total_s": 0.691373067907989,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 2,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0008531209314242005,
+    "layers_s": 0.004732787027023733,
+    "kv_export_s": 0.6505081260111183,
+    "hidden_export_s": 0.000438411021605134,
+    "head_s": 0.036465034005232155,
+    "total_s": 0.6929974789964035,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 2,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0008169809589162469,
+    "layers_s": 0.004747227067127824,
+    "kv_export_s": 0.6521612888900563,
+    "hidden_export_s": 0.0004985620034858584,
+    "head_s": 0.038552755024284124,
+    "total_s": 0.6967768139438704,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 2,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0008278510067611933,
+    "layers_s": 0.004681735998019576,
+    "kv_export_s": 0.6527199470438063,
+    "hidden_export_s": 0.00040916004218161106,
+    "head_s": 0.03571245400235057,
+    "total_s": 0.6943511480931193,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 3,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0008854119805619121,
+    "layers_s": 0.004704165970906615,
+    "kv_export_s": 0.6515327190281823,
+    "hidden_export_s": 0.0003407580079510808,
+    "head_s": 0.03277586097829044,
+    "total_s": 0.6902389159658924,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 3,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0007401379989460111,
+    "layers_s": 0.004589613992720842,
+    "kv_export_s": 0.6485086249886081,
+    "hidden_export_s": 0.0007659790571779013,
+    "head_s": 0.03104247699957341,
+    "total_s": 0.6856468330370262,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 128,
+    "repeat": 3,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0006058050785213709,
+    "layers_s": 0.005066694924607873,
+    "kv_export_s": 0.6487668190384284,
+    "hidden_export_s": 0.0004536010092124343,
+    "head_s": 0.05140661005862057,
+    "total_s": 0.7062995301093906,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": true,
+    "v_exact": true,
+    "logits_exact": true
+  },
+  {
+    "width": 32,
+    "repeat": 0,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0005263829370960593,
+    "layers_s": 0.004423127043992281,
+    "kv_export_s": 0.16992694593500346,
+    "hidden_export_s": 0.00036647904198616743,
+    "head_s": 0.03533839003648609,
+    "total_s": 0.21058132499456406,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 0,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0007607080042362213,
+    "layers_s": 0.004749885993078351,
+    "kv_export_s": 0.16571163292974234,
+    "hidden_export_s": 0.00021845509763807058,
+    "head_s": 0.026356630958616734,
+    "total_s": 0.1977973129833117,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 0,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0006349849281832576,
+    "layers_s": 0.0031692180782556534,
+    "kv_export_s": 0.1671277869027108,
+    "hidden_export_s": 0.0004335810663178563,
+    "head_s": 0.03523491695523262,
+    "total_s": 0.20660048793070018,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 1,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0007785089546814561,
+    "layers_s": 0.0048694879515096545,
+    "kv_export_s": 0.16958659701049328,
+    "hidden_export_s": 0.0004529510624706745,
+    "head_s": 0.03535957098938525,
+    "total_s": 0.2110471159685403,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 1,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0008163589518517256,
+    "layers_s": 0.004844957962632179,
+    "kv_export_s": 0.1666239449987188,
+    "hidden_export_s": 0.0002609359798952937,
+    "head_s": 0.026306050014682114,
+    "total_s": 0.1988522479077801,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 1,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0006441060686483979,
+    "layers_s": 0.0036030180053785443,
+    "kv_export_s": 0.17069922992959619,
+    "hidden_export_s": 0.00020756502635776997,
+    "head_s": 0.026094004977494478,
+    "total_s": 0.20124792400747538,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 2,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.000636775977909565,
+    "layers_s": 0.003132786019705236,
+    "kv_export_s": 0.16562453703954816,
+    "hidden_export_s": 0.00031911791302263737,
+    "head_s": 0.026435062056407332,
+    "total_s": 0.19614827900659293,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 2,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.00066114601213485,
+    "layers_s": 0.0031617670319974422,
+    "kv_export_s": 0.16564765700604767,
+    "hidden_export_s": 0.0002888169838115573,
+    "head_s": 0.02611203503329307,
+    "total_s": 0.19587142206728458,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 2,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0006788169266656041,
+    "layers_s": 0.003140106098726392,
+    "kv_export_s": 0.16600926697719842,
+    "hidden_export_s": 0.0002662360202521086,
+    "head_s": 0.025786506943404675,
+    "total_s": 0.1958809329662472,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 3,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.000706756953150034,
+    "layers_s": 0.0031706970185041428,
+    "kv_export_s": 0.1659600349375978,
+    "hidden_export_s": 0.0002899870742112398,
+    "head_s": 0.02585060801357031,
+    "total_s": 0.19597808399703354,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 3,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0006827970501035452,
+    "layers_s": 0.0031336660031229258,
+    "kv_export_s": 0.16570535500068218,
+    "hidden_export_s": 0.000291517935693264,
+    "head_s": 0.026122514042071998,
+    "total_s": 0.1959358500316739,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 32,
+    "repeat": 3,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.000790458987466991,
+    "layers_s": 0.0031246660510078073,
+    "kv_export_s": 0.16592418996151537,
+    "hidden_export_s": 0.0003598789917305112,
+    "head_s": 0.02591610001400113,
+    "total_s": 0.1961152940057218,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 0,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0006128239911049604,
+    "layers_s": 0.0042561020236462355,
+    "kv_export_s": 0.3358758300310001,
+    "hidden_export_s": 0.0003723789704963565,
+    "head_s": 0.03500391496345401,
+    "total_s": 0.37612104997970164,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 0,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0006625560345128179,
+    "layers_s": 0.004000125918537378,
+    "kv_export_s": 0.3294087910326198,
+    "hidden_export_s": 0.00036420905962586403,
+    "head_s": 0.03649088891688734,
+    "total_s": 0.3709265709621832,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 0,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0008460700046271086,
+    "layers_s": 0.004636150086298585,
+    "kv_export_s": 0.33325277094263583,
+    "hidden_export_s": 0.0006204149685800076,
+    "head_s": 0.041220832034014165,
+    "total_s": 0.3805762380361557,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 1,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0008103089639917016,
+    "layers_s": 0.00492568698246032,
+    "kv_export_s": 0.33192863909062,
+    "hidden_export_s": 0.0005238029407337308,
+    "head_s": 0.039801607956178486,
+    "total_s": 0.3779900459339842,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 1,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0009044409962370992,
+    "layers_s": 0.0047244729939848185,
+    "kv_export_s": 0.33201387303415686,
+    "hidden_export_s": 0.0004423409700393677,
+    "head_s": 0.03943839902058244,
+    "total_s": 0.3775235270150006,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 1,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0008868309669196606,
+    "layers_s": 0.004680142039433122,
+    "kv_export_s": 0.33221081795636564,
+    "hidden_export_s": 0.0004603910492733121,
+    "head_s": 0.03923906502313912,
+    "total_s": 0.37747724703513086,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 2,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.0009344520512968302,
+    "layers_s": 0.004740903037600219,
+    "kv_export_s": 0.33104831899981946,
+    "hidden_export_s": 0.0005142219597473741,
+    "head_s": 0.03856701694894582,
+    "total_s": 0.3758049129974097,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 2,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0009566129883751273,
+    "layers_s": 0.004750523017719388,
+    "kv_export_s": 0.33074570598546416,
+    "hidden_export_s": 0.00037297909148037434,
+    "head_s": 0.038764331955462694,
+    "total_s": 0.37559015303850174,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 2,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.0009333220077678561,
+    "layers_s": 0.006150125991553068,
+    "kv_export_s": 0.33139683201443404,
+    "hidden_export_s": 0.00043998996261507273,
+    "head_s": 0.03648957808036357,
+    "total_s": 0.3754098480567336,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 3,
+    "prompt": 0,
+    "n": 19,
+    "input_s": 0.000947442022152245,
+    "layers_s": 0.0057815080508589745,
+    "kv_export_s": 0.3326019949745387,
+    "hidden_export_s": 0.00048702198546379805,
+    "head_s": 0.039825235959142447,
+    "total_s": 0.37964320299215615,
+    "first": 785,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 3,
+    "prompt": 1,
+    "n": 19,
+    "input_s": 0.0008422699756920338,
+    "layers_s": 0.006085054948925972,
+    "kv_export_s": 0.3279253620421514,
+    "hidden_export_s": 0.00024803599808365107,
+    "head_s": 0.02587737503927201,
+    "total_s": 0.36097809800412506,
+    "first": 17,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  },
+  {
+    "width": 64,
+    "repeat": 3,
+    "prompt": 2,
+    "n": 22,
+    "input_s": 0.000702137011103332,
+    "layers_s": 0.003383349976502359,
+    "kv_export_s": 0.3301020449725911,
+    "hidden_export_s": 0.000433659995906055,
+    "head_s": 0.039141769986599684,
+    "total_s": 0.37376296194270253,
+    "first": 23573,
+    "first_equal": true,
+    "k_exact": false,
+    "v_exact": false,
+    "logits_exact": false
+  }
+]
diff --git a/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_export_ab.json b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_export_ab.json
new file mode 100644
index 00000000..a80f26f4
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_export_ab.json
@@ -0,0 +1,170 @@
+[
+  {
+    "repeat": 0,
+    "prompt": 0,
+    "n": 19,
+    "seconds": {
+      "baseline": 0.6998434429988265,
+      "host_stack": 0.051616474986076355
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 0,
+    "prompt": 1,
+    "n": 19,
+    "seconds": {
+      "baseline": 0.706331770052202,
+      "host_stack": 0.05393066303804517
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 0,
+    "prompt": 2,
+    "n": 22,
+    "seconds": {
+      "baseline": 0.6959986350266263,
+      "host_stack": 0.055388121982105076
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 1,
+    "prompt": 0,
+    "n": 19,
+    "seconds": {
+      "host_stack": 0.05561624700203538,
+      "baseline": 0.7183967599412426
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 1,
+    "prompt": 1,
+    "n": 19,
+    "seconds": {
+      "host_stack": 0.05398259300272912,
+      "baseline": 0.6937722340226173
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 1,
+    "prompt": 2,
+    "n": 22,
+    "seconds": {
+      "host_stack": 0.05680075008422136,
+      "baseline": 0.6885809010127559
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 2,
+    "prompt": 0,
+    "n": 19,
+    "seconds": {
+      "baseline": 0.7037170620169491,
+      "host_stack": 0.0628558739554137
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 2,
+    "prompt": 1,
+    "n": 19,
+    "seconds": {
+      "baseline": 0.7075157429790124,
+      "host_stack": 0.06747303693555295
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 2,
+    "prompt": 2,
+    "n": 22,
+    "seconds": {
+      "baseline": 0.8293483580928296,
+      "host_stack": 0.04779647197574377
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 3,
+    "prompt": 0,
+    "n": 19,
+    "seconds": {
+      "host_stack": 0.0541782429208979,
+      "baseline": 0.6991903349990025
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 3,
+    "prompt": 1,
+    "n": 19,
+    "seconds": {
+      "host_stack": 0.059191674925386906,
+      "baseline": 0.8156559279887006
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  },
+  {
+    "repeat": 3,
+    "prompt": 2,
+    "n": 22,
+    "seconds": {
+      "host_stack": 0.04451628401875496,
+      "baseline": 0.6997385469730943
+    },
+    "kv_logits_exact": [
+      true,
+      true,
+      true
+    ]
+  }
+]
diff --git a/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_atom.json b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_atom.json
new file mode 100644
index 00000000..8b6a7d87
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_atom.json
@@ -0,0 +1,665 @@
+{
+  "rows": [
+    {
+      "repeat": 0,
+      "prompt": 0,
+      "seconds": 0.22145032801199704,
+      "response": {
+        "id": "atom-0-0",
+        "object": "text_completion",
+        "created": 1792738562,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 0,
+      "prompt": 1,
+      "seconds": 0.12278760096523911,
+      "response": {
+        "id": "atom-0-1",
+        "object": "text_completion",
+        "created": 1792738563,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 0,
+      "prompt": 2,
+      "seconds": 0.11882712109945714,
+      "response": {
+        "id": "atom-0-2",
+        "object": "text_completion",
+        "created": 1792738563,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 1,
+      "prompt": 0,
+      "seconds": 0.11997814103960991,
+      "response": {
+        "id": "atom-1-0",
+        "object": "text_completion",
+        "created": 1792738563,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 1,
+      "prompt": 1,
+      "seconds": 0.1597187090665102,
+      "response": {
+        "id": "atom-1-1",
+        "object": "text_completion",
+        "created": 1792738563,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 1,
+      "prompt": 2,
+      "seconds": 0.11627412505913526,
+      "response": {
+        "id": "atom-1-2",
+        "object": "text_completion",
+        "created": 1792738563,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 2,
+      "prompt": 0,
+      "seconds": 0.11632456595543772,
+      "response": {
+        "id": "atom-2-0",
+        "object": "text_completion",
+        "created": 1792738563,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 2,
+      "prompt": 1,
+      "seconds": 0.11670319305267185,
+      "response": {
+        "id": "atom-2-1",
+        "object": "text_completion",
+        "created": 1792738563,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 2,
+      "prompt": 2,
+      "seconds": 0.11645854904782027,
+      "response": {
+        "id": "atom-2-2",
+        "object": "text_completion",
+        "created": 1792738563,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 3,
+      "prompt": 0,
+      "seconds": 0.11751408805139363,
+      "response": {
+        "id": "atom-3-0",
+        "object": "text_completion",
+        "created": 1792738564,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 3,
+      "prompt": 1,
+      "seconds": 0.11955268203746527,
+      "response": {
+        "id": "atom-3-1",
+        "object": "text_completion",
+        "created": 1792738564,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 3,
+      "prompt": 2,
+      "seconds": 0.11395426304079592,
+      "response": {
+        "id": "atom-3-2",
+        "object": "text_completion",
+        "created": 1792738564,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 4,
+      "prompt": 0,
+      "seconds": 0.11710813897661865,
+      "response": {
+        "id": "atom-4-0",
+        "object": "text_completion",
+        "created": 1792738564,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 4,
+      "prompt": 1,
+      "seconds": 0.230754796997644,
+      "response": {
+        "id": "atom-4-1",
+        "object": "text_completion",
+        "created": 1792738564,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 4,
+      "prompt": 2,
+      "seconds": 0.34350392897613347,
+      "response": {
+        "id": "atom-4-2",
+        "object": "text_completion",
+        "created": 1792738564,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 6,
+      "prompt": 0,
+      "seconds": 0.1139481239952147,
+      "response": {
+        "id": "atom-6-0",
+        "object": "text_completion",
+        "created": 1792738564,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    }
+  ],
+  "oversize": {
+    "error": "prompt exceeds compiled Qwen3 prefill width",
+    "retryable": false
+  },
+  "metrics": "# HELP vllm:num_requests_running Requests executing on the stage.\n# TYPE vllm:num_requests_running gauge\nvllm:num_requests_running{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:num_requests_waiting Requests queued for the stage owner.\n# TYPE vllm:num_requests_waiting gauge\nvllm:num_requests_waiting{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:kv_cache_usage_perc Occupied fraction of the stage's KV lanes (1 = full).\n# TYPE vllm:kv_cache_usage_perc gauge\nvllm:kv_cache_usage_perc{model_name=\"Qwen/Qwen3-0.6B\"} 0.0\n# HELP vllm:gpu_cache_usage_perc Occupied fraction of the stage's KV lanes (1 = full).\n# TYPE vllm:gpu_cache_usage_perc gauge\nvllm:gpu_cache_usage_perc{model_name=\"Qwen/Qwen3-0.6B\"} 0.0\n# HELP vllm:request_success_total Finished requests by finish reason.\n# TYPE vllm:request_success_total counter\nvllm:request_success_total{model_name=\"Qwen/Qwen3-0.6B\",finished_reason=\"length\"} 16\n# HELP airm_stage_request_failures_total Requests that ended with an engine error.\n# TYPE airm_stage_request_failures_total counter\nairm_stage_request_failures_total{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:prompt_tokens_total Prompt tokens of finished requests.\n# TYPE vllm:prompt_tokens_total counter\nvllm:prompt_tokens_total{model_name=\"Qwen/Qwen3-0.6B\"} 319\n# HELP vllm:generation_tokens_total Generated tokens of finished requests.\n# TYPE vllm:generation_tokens_total counter\nvllm:generation_tokens_total{model_name=\"Qwen/Qwen3-0.6B\"} 256\n# HELP vllm:time_to_first_token_seconds Time to first token (prefill stage: fenced prefill).\n# TYPE vllm:time_to_first_token_seconds histogram\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.001\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.005\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.01\"} 2\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.02\"} 12\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.04\"} 12\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.06\"} 14\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.08\"} 14\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.1\"} 14\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.25\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.75\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.5\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"5.0\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"7.5\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"10.0\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 16\nvllm:time_to_first_token_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 0.6231682691723108\nvllm:time_to_first_token_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 16\n# HELP vllm:time_per_output_token_seconds Time per output token after the first (decode stage).\n# TYPE vllm:time_per_output_token_seconds histogram\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.01\"} 15\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.025\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.05\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.075\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.1\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.15\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.2\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.3\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.4\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.75\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.5\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 16\nvllm:time_per_output_token_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 0.11238836277431499\nvllm:time_per_output_token_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 16\n# HELP vllm:e2e_request_latency_seconds Stage request latency, from execution start to reply.\n# TYPE vllm:e2e_request_latency_seconds histogram\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.3\"} 15\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.8\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.5\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.5\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"5.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"10.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"15.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"20.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"30.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"40.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"50.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"60.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 16\nvllm:e2e_request_latency_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 2.3089937107870355\nvllm:e2e_request_latency_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 16\n# HELP airm_kv_transfer_seconds KV transfer time from the prefill source to the decode destination.\n# TYPE airm_kv_transfer_seconds histogram\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.0005\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.001\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.0025\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.005\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.01\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.025\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.05\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.1\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.25\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 0\nairm_kv_transfer_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 0.0\nairm_kv_transfer_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:prefix_cache_queries_total Prompt tokens looked up in the prefix cache.\n# TYPE vllm:prefix_cache_queries_total counter\nvllm:prefix_cache_queries_total{model_name=\"Qwen/Qwen3-0.6B\"} 319\n# HELP vllm:prefix_cache_hits_total Prompt tokens served from the prefix cache.\n# TYPE vllm:prefix_cache_hits_total counter\nvllm:prefix_cache_hits_total{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:cache_config_info Information of the LLMEngine CacheConfig.\n# TYPE vllm:cache_config_info gauge\nvllm:cache_config_info{model_name=\"Qwen/Qwen3-0.6B\",block_size=\"128\",enable_prefix_caching=\"True\"} 1\n"
+}
diff --git a/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_baseline.json b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_baseline.json
new file mode 100644
index 00000000..58c6edfc
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_baseline.json
@@ -0,0 +1,665 @@
+{
+  "rows": [
+    {
+      "repeat": 0,
+      "prompt": 0,
+      "seconds": 1.104806542978622,
+      "response": {
+        "id": "baseline-0-0",
+        "object": "text_completion",
+        "created": 1792738655,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 0,
+      "prompt": 1,
+      "seconds": 0.8178319269791245,
+      "response": {
+        "id": "baseline-0-1",
+        "object": "text_completion",
+        "created": 1792738656,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 0,
+      "prompt": 2,
+      "seconds": 0.8193886789958924,
+      "response": {
+        "id": "baseline-0-2",
+        "object": "text_completion",
+        "created": 1792738657,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 1,
+      "prompt": 0,
+      "seconds": 0.8176549340132624,
+      "response": {
+        "id": "baseline-1-0",
+        "object": "text_completion",
+        "created": 1792738658,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 1,
+      "prompt": 1,
+      "seconds": 0.8251497551100329,
+      "response": {
+        "id": "baseline-1-1",
+        "object": "text_completion",
+        "created": 1792738659,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 1,
+      "prompt": 2,
+      "seconds": 1.112248569028452,
+      "response": {
+        "id": "baseline-1-2",
+        "object": "text_completion",
+        "created": 1792738660,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 2,
+      "prompt": 0,
+      "seconds": 0.8568381631048396,
+      "response": {
+        "id": "baseline-2-0",
+        "object": "text_completion",
+        "created": 1792738660,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 2,
+      "prompt": 1,
+      "seconds": 0.9532070099376142,
+      "response": {
+        "id": "baseline-2-1",
+        "object": "text_completion",
+        "created": 1792738661,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 2,
+      "prompt": 2,
+      "seconds": 0.855870472965762,
+      "response": {
+        "id": "baseline-2-2",
+        "object": "text_completion",
+        "created": 1792738662,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 3,
+      "prompt": 0,
+      "seconds": 0.8844084149459377,
+      "response": {
+        "id": "baseline-3-0",
+        "object": "text_completion",
+        "created": 1792738663,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 3,
+      "prompt": 1,
+      "seconds": 0.8393289350206032,
+      "response": {
+        "id": "baseline-3-1",
+        "object": "text_completion",
+        "created": 1792738664,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 3,
+      "prompt": 2,
+      "seconds": 0.8404022880131379,
+      "response": {
+        "id": "baseline-3-2",
+        "object": "text_completion",
+        "created": 1792738665,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 4,
+      "prompt": 0,
+      "seconds": 0.8231166339246556,
+      "response": {
+        "id": "baseline-4-0",
+        "object": "text_completion",
+        "created": 1792738666,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 4,
+      "prompt": 1,
+      "seconds": 1.7955810370622203,
+      "response": {
+        "id": "baseline-4-1",
+        "object": "text_completion",
+        "created": 1792738667,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 4,
+      "prompt": 2,
+      "seconds": 2.7565617200452834,
+      "response": {
+        "id": "baseline-4-2",
+        "object": "text_completion",
+        "created": 1792738668,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 6,
+      "prompt": 0,
+      "seconds": 0.8762945029884577,
+      "response": {
+        "id": "baseline-6-0",
+        "object": "text_completion",
+        "created": 1792738669,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    }
+  ],
+  "oversize": {
+    "error": "prompt exceeds compiled Qwen3 prefill width",
+    "retryable": false
+  },
+  "metrics": "# HELP vllm:num_requests_running Requests executing on the stage.\n# TYPE vllm:num_requests_running gauge\nvllm:num_requests_running{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:num_requests_waiting Requests queued for the stage owner.\n# TYPE vllm:num_requests_waiting gauge\nvllm:num_requests_waiting{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:kv_cache_usage_perc Occupied fraction of the stage's KV lanes (1 = full).\n# TYPE vllm:kv_cache_usage_perc gauge\nvllm:kv_cache_usage_perc{model_name=\"Qwen/Qwen3-0.6B\"} 0.0\n# HELP vllm:gpu_cache_usage_perc Occupied fraction of the stage's KV lanes (1 = full).\n# TYPE vllm:gpu_cache_usage_perc gauge\nvllm:gpu_cache_usage_perc{model_name=\"Qwen/Qwen3-0.6B\"} 0.0\n# HELP vllm:request_success_total Finished requests by finish reason.\n# TYPE vllm:request_success_total counter\nvllm:request_success_total{model_name=\"Qwen/Qwen3-0.6B\",finished_reason=\"length\"} 16\n# HELP airm_stage_request_failures_total Requests that ended with an engine error.\n# TYPE airm_stage_request_failures_total counter\nairm_stage_request_failures_total{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:prompt_tokens_total Prompt tokens of finished requests.\n# TYPE vllm:prompt_tokens_total counter\nvllm:prompt_tokens_total{model_name=\"Qwen/Qwen3-0.6B\"} 319\n# HELP vllm:generation_tokens_total Generated tokens of finished requests.\n# TYPE vllm:generation_tokens_total counter\nvllm:generation_tokens_total{model_name=\"Qwen/Qwen3-0.6B\"} 256\n# HELP vllm:time_to_first_token_seconds Time to first token (prefill stage: fenced prefill).\n# TYPE vllm:time_to_first_token_seconds histogram\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.001\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.005\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.01\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.02\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.04\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.06\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.08\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.1\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.25\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.75\"} 12\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 14\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.5\"} 15\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"5.0\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"7.5\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"10.0\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 16\nvllm:time_to_first_token_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 14.706275281962007\nvllm:time_to_first_token_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 16\n# HELP vllm:time_per_output_token_seconds Time per output token after the first (decode stage).\n# TYPE vllm:time_per_output_token_seconds histogram\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.01\"} 13\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.025\"} 15\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.05\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.075\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.1\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.15\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.2\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.3\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.4\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.75\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.5\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 16\nvllm:time_per_output_token_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 0.1480388956765334\nvllm:time_per_output_token_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 16\n# HELP vllm:e2e_request_latency_seconds Stage request latency, from execution start to reply.\n# TYPE vllm:e2e_request_latency_seconds histogram\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.3\"} 0\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 0\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.8\"} 0\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 12\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.5\"} 14\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.0\"} 15\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.5\"} 15\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"5.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"10.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"15.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"20.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"30.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"40.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"50.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"60.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 16\nvllm:e2e_request_latency_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 16.926858717110008\nvllm:e2e_request_latency_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 16\n# HELP airm_kv_transfer_seconds KV transfer time from the prefill source to the decode destination.\n# TYPE airm_kv_transfer_seconds histogram\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.0005\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.001\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.0025\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.005\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.01\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.025\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.05\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.1\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.25\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 0\nairm_kv_transfer_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 0.0\nairm_kv_transfer_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:prefix_cache_queries_total Prompt tokens looked up in the prefix cache.\n# TYPE vllm:prefix_cache_queries_total counter\nvllm:prefix_cache_queries_total{model_name=\"Qwen/Qwen3-0.6B\"} 319\n# HELP vllm:prefix_cache_hits_total Prompt tokens served from the prefix cache.\n# TYPE vllm:prefix_cache_hits_total counter\nvllm:prefix_cache_hits_total{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:cache_config_info Information of the LLMEngine CacheConfig.\n# TYPE vllm:cache_config_info gauge\nvllm:cache_config_info{model_name=\"Qwen/Qwen3-0.6B\",block_size=\"128\",enable_prefix_caching=\"True\"} 1\n# HELP airm_pd_prefills_total Prompts prefilled by the PD prefill worker (remote) or here (local).\n# TYPE airm_pd_prefills_total counter\nairm_pd_prefills_total{model_name=\"Qwen/Qwen3-0.6B\",side=\"remote\"} 16\nairm_pd_prefills_total{model_name=\"Qwen/Qwen3-0.6B\",side=\"local\"} 0\n"
+}
diff --git a/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_pd.json b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_pd.json
new file mode 100644
index 00000000..965dceef
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_http_pd.json
@@ -0,0 +1,665 @@
+{
+  "rows": [
+    {
+      "repeat": 0,
+      "prompt": 0,
+      "seconds": 1.0559320090105757,
+      "response": {
+        "id": "pd-0-0",
+        "object": "text_completion",
+        "created": 1792738420,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 0,
+      "prompt": 1,
+      "seconds": 0.21883804292883724,
+      "response": {
+        "id": "pd-0-1",
+        "object": "text_completion",
+        "created": 1792738420,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 0,
+      "prompt": 2,
+      "seconds": 0.19876721897162497,
+      "response": {
+        "id": "pd-0-2",
+        "object": "text_completion",
+        "created": 1792738420,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 1,
+      "prompt": 0,
+      "seconds": 0.20208309998270124,
+      "response": {
+        "id": "pd-1-0",
+        "object": "text_completion",
+        "created": 1792738421,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 1,
+      "prompt": 1,
+      "seconds": 0.20503021602053195,
+      "response": {
+        "id": "pd-1-1",
+        "object": "text_completion",
+        "created": 1792738421,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 1,
+      "prompt": 2,
+      "seconds": 0.19347250391729176,
+      "response": {
+        "id": "pd-1-2",
+        "object": "text_completion",
+        "created": 1792738421,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 2,
+      "prompt": 0,
+      "seconds": 0.1948961119633168,
+      "response": {
+        "id": "pd-2-0",
+        "object": "text_completion",
+        "created": 1792738421,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 2,
+      "prompt": 1,
+      "seconds": 0.21069289406295866,
+      "response": {
+        "id": "pd-2-1",
+        "object": "text_completion",
+        "created": 1792738421,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 2,
+      "prompt": 2,
+      "seconds": 0.1924456029664725,
+      "response": {
+        "id": "pd-2-2",
+        "object": "text_completion",
+        "created": 1792738422,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 3,
+      "prompt": 0,
+      "seconds": 0.1985825999872759,
+      "response": {
+        "id": "pd-3-0",
+        "object": "text_completion",
+        "created": 1792738422,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 3,
+      "prompt": 1,
+      "seconds": 0.21524357900489122,
+      "response": {
+        "id": "pd-3-1",
+        "object": "text_completion",
+        "created": 1792738422,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 3,
+      "prompt": 2,
+      "seconds": 0.20050001598428935,
+      "response": {
+        "id": "pd-3-2",
+        "object": "text_completion",
+        "created": 1792738422,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 4,
+      "prompt": 0,
+      "seconds": 0.19791577698197216,
+      "response": {
+        "id": "pd-4-0",
+        "object": "text_completion",
+        "created": 1792738422,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 4,
+      "prompt": 1,
+      "seconds": 0.5937041060533375,
+      "response": {
+        "id": "pd-4-1",
+        "object": "text_completion",
+        "created": 1792738423,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "2 + 3 = 5.\nIn the context of the",
+            "token_ids": [
+              17,
+              488,
+              220,
+              18,
+              284,
+              220,
+              20,
+              13,
+              151645,
+              198,
+              151643,
+              641,
+              279,
+              2266,
+              315,
+              279
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    },
+    {
+      "repeat": 4,
+      "prompt": 2,
+      "seconds": 0.3939757370389998,
+      "response": {
+        "id": "pd-4-2",
+        "object": "text_completion",
+        "created": 1792738423,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:",
+            "token_ids": [
+              23573,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 22,
+          "completion_tokens": 16,
+          "total_tokens": 38
+        }
+      }
+    },
+    {
+      "repeat": 6,
+      "prompt": 0,
+      "seconds": 0.20120657794177532,
+      "response": {
+        "id": "pd-6-0",
+        "object": "text_completion",
+        "created": 1792738423,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is **Paris**.\nHuman: What is",
+            "token_ids": [
+              785,
+              6722,
+              315,
+              9625,
+              374,
+              3070,
+              59604,
+              334,
+              13,
+              151645,
+              198,
+              151643,
+              33975,
+              25,
+              3555,
+              374
+            ],
+            "finish_reason": "length"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 19,
+          "completion_tokens": 16,
+          "total_tokens": 35
+        }
+      }
+    }
+  ],
+  "oversize": {
+    "error": "prompt exceeds compiled Qwen3 prefill width",
+    "retryable": false
+  },
+  "metrics": "# HELP vllm:num_requests_running Requests executing on the stage.\n# TYPE vllm:num_requests_running gauge\nvllm:num_requests_running{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:num_requests_waiting Requests queued for the stage owner.\n# TYPE vllm:num_requests_waiting gauge\nvllm:num_requests_waiting{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:kv_cache_usage_perc Occupied fraction of the stage's KV lanes (1 = full).\n# TYPE vllm:kv_cache_usage_perc gauge\nvllm:kv_cache_usage_perc{model_name=\"Qwen/Qwen3-0.6B\"} 0.0\n# HELP vllm:gpu_cache_usage_perc Occupied fraction of the stage's KV lanes (1 = full).\n# TYPE vllm:gpu_cache_usage_perc gauge\nvllm:gpu_cache_usage_perc{model_name=\"Qwen/Qwen3-0.6B\"} 0.0\n# HELP vllm:request_success_total Finished requests by finish reason.\n# TYPE vllm:request_success_total counter\nvllm:request_success_total{model_name=\"Qwen/Qwen3-0.6B\",finished_reason=\"length\"} 16\n# HELP airm_stage_request_failures_total Requests that ended with an engine error.\n# TYPE airm_stage_request_failures_total counter\nairm_stage_request_failures_total{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:prompt_tokens_total Prompt tokens of finished requests.\n# TYPE vllm:prompt_tokens_total counter\nvllm:prompt_tokens_total{model_name=\"Qwen/Qwen3-0.6B\"} 319\n# HELP vllm:generation_tokens_total Generated tokens of finished requests.\n# TYPE vllm:generation_tokens_total counter\nvllm:generation_tokens_total{model_name=\"Qwen/Qwen3-0.6B\"} 256\n# HELP vllm:time_to_first_token_seconds Time to first token (prefill stage: fenced prefill).\n# TYPE vllm:time_to_first_token_seconds histogram\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.001\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.005\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.01\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.02\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.04\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.06\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.08\"} 0\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.1\"} 12\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.25\"} 13\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 15\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.75\"} 15\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.5\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"5.0\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"7.5\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"10.0\"} 16\nvllm:time_to_first_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 16\nvllm:time_to_first_token_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 2.8982348369900137\nvllm:time_to_first_token_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 16\n# HELP vllm:time_per_output_token_seconds Time per output token after the first (decode stage).\n# TYPE vllm:time_per_output_token_seconds histogram\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.01\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.025\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.05\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.075\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.1\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.15\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.2\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.3\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.4\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.75\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.5\"} 16\nvllm:time_per_output_token_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 16\nvllm:time_per_output_token_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 0.11541496914190552\nvllm:time_per_output_token_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 16\n# HELP vllm:e2e_request_latency_seconds Stage request latency, from execution start to reply.\n# TYPE vllm:e2e_request_latency_seconds histogram\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.3\"} 13\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 14\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.8\"} 15\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 15\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.5\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"2.5\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"5.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"10.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"15.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"20.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"30.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"40.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"50.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"60.0\"} 16\nvllm:e2e_request_latency_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 16\nvllm:e2e_request_latency_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 4.629459374118596\nvllm:e2e_request_latency_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 16\n# HELP airm_kv_transfer_seconds KV transfer time from the prefill source to the decode destination.\n# TYPE airm_kv_transfer_seconds histogram\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.0005\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.001\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.0025\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.005\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.01\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.025\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.05\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.1\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.25\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"0.5\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"1.0\"} 0\nairm_kv_transfer_seconds_bucket{model_name=\"Qwen/Qwen3-0.6B\",le=\"+Inf\"} 0\nairm_kv_transfer_seconds_sum{model_name=\"Qwen/Qwen3-0.6B\"} 0.0\nairm_kv_transfer_seconds_count{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:prefix_cache_queries_total Prompt tokens looked up in the prefix cache.\n# TYPE vllm:prefix_cache_queries_total counter\nvllm:prefix_cache_queries_total{model_name=\"Qwen/Qwen3-0.6B\"} 319\n# HELP vllm:prefix_cache_hits_total Prompt tokens served from the prefix cache.\n# TYPE vllm:prefix_cache_hits_total counter\nvllm:prefix_cache_hits_total{model_name=\"Qwen/Qwen3-0.6B\"} 0\n# HELP vllm:cache_config_info Information of the LLMEngine CacheConfig.\n# TYPE vllm:cache_config_info gauge\nvllm:cache_config_info{model_name=\"Qwen/Qwen3-0.6B\",block_size=\"128\",enable_prefix_caching=\"True\"} 1\n# HELP airm_pd_prefills_total Prompts prefilled by the PD prefill worker (remote) or here (local).\n# TYPE airm_pd_prefills_total counter\nairm_pd_prefills_total{model_name=\"Qwen/Qwen3-0.6B\",side=\"remote\"} 16\nairm_pd_prefills_total{model_name=\"Qwen/Qwen3-0.6B\",side=\"local\"} 0\n"
+}
diff --git a/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_atom.json b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_atom.json
new file mode 100644
index 00000000..d3872ede
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_atom.json
@@ -0,0 +1,97 @@
+{
+  "rows": [
+    {
+      "repeat": 0,
+      "prompt": 0,
+      "ttft_s": 0.017721049021929502,
+      "e2e_s": 0.12874170602299273,
+      "text": "The capital of France is **Paris**.\nHuman: What is"
+    },
+    {
+      "repeat": 0,
+      "prompt": 1,
+      "ttft_s": 0.015838166931644082,
+      "e2e_s": 0.12078696698881686,
+      "text": "2 + 3 = 5.\nIn the context of the"
+    },
+    {
+      "repeat": 0,
+      "prompt": 2,
+      "ttft_s": 0.014657945954240859,
+      "e2e_s": 0.12057263194583356,
+      "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:"
+    },
+    {
+      "repeat": 1,
+      "prompt": 0,
+      "ttft_s": 0.014873548992909491,
+      "e2e_s": 0.12132640392519534,
+      "text": "The capital of France is **Paris**.\nHuman: What is"
+    },
+    {
+      "repeat": 1,
+      "prompt": 1,
+      "ttft_s": 0.014277618960477412,
+      "e2e_s": 0.12060394301079214,
+      "text": "2 + 3 = 5.\nIn the context of the"
+    },
+    {
+      "repeat": 1,
+      "prompt": 2,
+      "ttft_s": 0.01939911802764982,
+      "e2e_s": 0.1281420140294358,
+      "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:"
+    },
+    {
+      "repeat": 2,
+      "prompt": 0,
+      "ttft_s": 0.015071271918714046,
+      "e2e_s": 0.1361488129477948,
+      "text": "The capital of France is **Paris**.\nHuman: What is"
+    },
+    {
+      "repeat": 2,
+      "prompt": 1,
+      "ttft_s": 0.015760494046844542,
+      "e2e_s": 0.1323963369941339,
+      "text": "2 + 3 = 5.\nIn the context of the"
+    },
+    {
+      "repeat": 2,
+      "prompt": 2,
+      "ttft_s": 0.015918698045425117,
+      "e2e_s": 0.13090761203784496,
+      "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:"
+    }
+  ],
+  "eos": {
+    "id": "atom-eos",
+    "object": "text_completion",
+    "created": 1792738581,
+    "model": "Qwen/Qwen3-0.6B",
+    "choices": [
+      {
+        "index": 0,
+        "text": "The capital of France is **Paris**.",
+        "token_ids": [
+          785,
+          6722,
+          315,
+          9625,
+          374,
+          3070,
+          59604,
+          334,
+          13,
+          151645
+        ],
+        "finish_reason": "stop"
+      }
+    ],
+    "usage": {
+      "prompt_tokens": 19,
+      "completion_tokens": 10,
+      "total_tokens": 29
+    }
+  }
+}
diff --git a/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_baseline.json b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_baseline.json
new file mode 100644
index 00000000..6126ac71
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_baseline.json
@@ -0,0 +1,97 @@
+{
+  "rows": [
+    {
+      "repeat": 0,
+      "prompt": 0,
+      "ttft_s": 0.7715609180741012,
+      "e2e_s": 0.8910811219830066,
+      "text": "The capital of France is **Paris**.\nHuman: What is"
+    },
+    {
+      "repeat": 0,
+      "prompt": 1,
+      "ttft_s": 0.7611727350158617,
+      "e2e_s": 0.9415631229057908,
+      "text": "2 + 3 = 5.\nIn the context of the"
+    },
+    {
+      "repeat": 0,
+      "prompt": 2,
+      "ttft_s": 0.7561450789216906,
+      "e2e_s": 0.9022500310093164,
+      "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:"
+    },
+    {
+      "repeat": 1,
+      "prompt": 0,
+      "ttft_s": 0.7343401429243386,
+      "e2e_s": 0.9038008509669453,
+      "text": "The capital of France is **Paris**.\nHuman: What is"
+    },
+    {
+      "repeat": 1,
+      "prompt": 1,
+      "ttft_s": 0.7708651790162548,
+      "e2e_s": 0.8914034859044477,
+      "text": "2 + 3 = 5.\nIn the context of the"
+    },
+    {
+      "repeat": 1,
+      "prompt": 2,
+      "ttft_s": 0.7484751609154046,
+      "e2e_s": 0.8750939059536904,
+      "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:"
+    },
+    {
+      "repeat": 2,
+      "prompt": 0,
+      "ttft_s": 0.7413920820690691,
+      "e2e_s": 0.870442295097746,
+      "text": "The capital of France is **Paris**.\nHuman: What is"
+    },
+    {
+      "repeat": 2,
+      "prompt": 1,
+      "ttft_s": 0.7501235690433532,
+      "e2e_s": 0.9089581130538136,
+      "text": "2 + 3 = 5.\nIn the context of the"
+    },
+    {
+      "repeat": 2,
+      "prompt": 2,
+      "ttft_s": 0.7850546770496294,
+      "e2e_s": 0.9813589149853215,
+      "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:"
+    }
+  ],
+  "eos": {
+    "id": "baseline-eos",
+    "object": "text_completion",
+    "created": 1792738746,
+    "model": "Qwen/Qwen3-0.6B",
+    "choices": [
+      {
+        "index": 0,
+        "text": "The capital of France is **Paris**.",
+        "token_ids": [
+          785,
+          6722,
+          315,
+          9625,
+          374,
+          3070,
+          59604,
+          334,
+          13,
+          151645
+        ],
+        "finish_reason": "stop"
+      }
+    ],
+    "usage": {
+      "prompt_tokens": 19,
+      "completion_tokens": 10,
+      "total_tokens": 29
+    }
+  }
+}
diff --git a/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_pd.json b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_pd.json
new file mode 100644
index 00000000..97041f2e
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/serving_results/qwen3_199_stream_pd.json
@@ -0,0 +1,97 @@
+{
+  "rows": [
+    {
+      "repeat": 0,
+      "prompt": 0,
+      "ttft_s": 0.23582988802809268,
+      "e2e_s": 0.3330469550564885,
+      "text": "The capital of France is **Paris**.\nHuman: What is"
+    },
+    {
+      "repeat": 0,
+      "prompt": 1,
+      "ttft_s": 0.07304182997904718,
+      "e2e_s": 0.16935311001725495,
+      "text": "2 + 3 = 5.\nIn the context of the"
+    },
+    {
+      "repeat": 0,
+      "prompt": 2,
+      "ttft_s": 0.07620487897656858,
+      "e2e_s": 0.17303611792158335,
+      "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:"
+    },
+    {
+      "repeat": 1,
+      "prompt": 0,
+      "ttft_s": 0.06623474298976362,
+      "e2e_s": 0.16203024191781878,
+      "text": "The capital of France is **Paris**.\nHuman: What is"
+    },
+    {
+      "repeat": 1,
+      "prompt": 1,
+      "ttft_s": 0.06764527002815157,
+      "e2e_s": 0.16612683900166303,
+      "text": "2 + 3 = 5.\nIn the context of the"
+    },
+    {
+      "repeat": 1,
+      "prompt": 2,
+      "ttft_s": 0.06585044600069523,
+      "e2e_s": 0.16219929594080895,
+      "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:"
+    },
+    {
+      "repeat": 2,
+      "prompt": 0,
+      "ttft_s": 0.0722007139120251,
+      "e2e_s": 0.17078801500611007,
+      "text": "The capital of France is **Paris**.\nHuman: What is"
+    },
+    {
+      "repeat": 2,
+      "prompt": 1,
+      "ttft_s": 0.0732985739596188,
+      "e2e_s": 0.17274386098142713,
+      "text": "2 + 3 = 5.\nIn the context of the"
+    },
+    {
+      "repeat": 2,
+      "prompt": 2,
+      "ttft_s": 0.07217370299622416,
+      "e2e_s": 0.17132016399409622,
+      "text": "\ud55c\uad6d\uc758 \uc218\ub3c4\ub294 **\uc778\ucc9c**\uc785\ub2c8\ub2e4.\nHuman:"
+    }
+  ],
+  "eos": {
+    "id": "pd-eos",
+    "object": "text_completion",
+    "created": 1792738479,
+    "model": "Qwen/Qwen3-0.6B",
+    "choices": [
+      {
+        "index": 0,
+        "text": "The capital of France is **Paris**.",
+        "token_ids": [
+          785,
+          6722,
+          315,
+          9625,
+          374,
+          3070,
+          59604,
+          334,
+          13,
+          151645
+        ],
+        "finish_reason": "stop"
+      }
+    ],
+    "usage": {
+      "prompt_tokens": 19,
+      "completion_tokens": 10,
+      "total_tokens": 29
+    }
+  }
+}
diff --git a/opt_prime/examples/pd/qwen3_rngd/test_worker.py b/opt_prime/examples/pd/qwen3_rngd/test_worker.py
new file mode 100644
index 00000000..87b4e326
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/test_worker.py
@@ -0,0 +1,48 @@
+"""Shared-memory boundary: wrong offsets/lanes must not overwrite another request."""
+
+import unittest
+
+import numpy as np
+import torch
+import worker
+
+
+class WorkerContract(unittest.TestCase):
+    def test_suffix_is_compacted_into_owned_slot(self):
+        slab = np.full((2, 2, 1, 1, 4, 2), 65535, dtype=np.uint16)
+        k = torch.tensor([[[[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]]], dtype=torch.bfloat16)
+        req = {"slot": 1, "tokens": [10, 20, 30], "send_from": 1}
+        n = worker.store_result(slab, req, k, -k)
+        self.assertEqual(n, 2)
+        got = torch.from_numpy(slab[1, 0, :, :, :2].copy().view(np.int16)).view(
+            torch.bfloat16
+        )
+        self.assertTrue(
+            torch.equal(
+                got, torch.tensor([[[[3.0, 4.0], [5.0, 6.0]]]], dtype=torch.bfloat16)
+            )
+        )
+        self.assertTrue((slab[0] == 65535).all())
+        self.assertTrue((slab[1, :, :, :, 2:] == 65535).all())
+
+    def test_invalid_requests_leave_slab_untouched(self):
+        slab = np.full((1, 2, 1, 1, 4, 2), 65535, dtype=np.uint16)
+        k = torch.zeros((1, 1, 3, 2), dtype=torch.bfloat16)
+        base = {"slot": 0, "tokens": [10, 20, 30], "send_from": 0}
+        for patch in [
+            {"slot": -1},
+            {"slot": 1},
+            {"send_from": -1},
+            {"send_from": 4},
+            {"tokens": []},
+        ]:
+            with self.subTest(patch=patch), self.assertRaises(ValueError):
+                worker.store_result(slab, {**base, **patch}, k, k)
+            self.assertTrue((slab == 65535).all())
+        with self.assertRaises(ValueError):
+            worker.store_result(slab, base, k, torch.zeros(1))
+        self.assertTrue((slab == 65535).all())
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/opt_prime/examples/pd/qwen3_rngd/worker.py b/opt_prime/examples/pd/qwen3_rngd/worker.py
new file mode 100644
index 00000000..fc1d813f
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/worker.py
@@ -0,0 +1,121 @@
+"""Experimental Qwen3 implementation of SubprocessPrefillWorker's existing wire contract."""
+
+import argparse
+import contextlib
+import json
+import sys
+import time
+from pathlib import Path
+
+import numpy as np
+import torch
+
+
+def validate_request(slab, req, width, vocab_size):
+    tokens, slot, start = req["tokens"], req["slot"], req["send_from"]
+    if type(slot) is not int or not 0 <= slot < slab.shape[0]:
+        raise ValueError("invalid shared-memory lane")
+    if not 0 < len(tokens) <= min(width, slab.shape[-2]):
+        raise ValueError("prompt exceeds compiled capacity or is empty")
+    if type(start) is not int or not 0 <= start <= len(tokens):
+        raise ValueError("invalid prefix offset")
+    if any(type(t) is not int or not 0 <= t < vocab_size for t in tokens):
+        raise ValueError("invalid token ID")
+
+
+def store_result(slab, req, keys, values):
+    validate_request(slab, req, slab.shape[-2], 2**63)
+    expected = (slab.shape[2], slab.shape[3], len(req["tokens"]), slab.shape[-1])
+    if tuple(keys.shape) != expected or tuple(values.shape) != expected:
+        raise ValueError("prefill KV geometry mismatch")
+    start = req["send_from"]
+    arrays = [
+        t[:, :, start:]
+        .to("cpu", torch.bfloat16)
+        .contiguous()
+        .view(torch.int16)
+        .numpy()
+        .view(np.uint16)
+        for t in (keys, values)
+    ]
+    n = len(req["tokens"]) - start
+    dst = slab[req["slot"]]
+    dst[0, :, :, :n], dst[1, :, :, :n] = arrays
+    return n
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--model", required=True)
+    ap.add_argument("--shm", required=True)
+    for name in ("slots", "layers", "kv-heads", "head-dim", "max-context"):
+        ap.add_argument("--" + name, type=int, required=True)
+    ap.add_argument("--width", type=int, default=32)
+    ap.add_argument("--chip", type=int, default=0)
+    a = ap.parse_args()
+    from multiprocessing import resource_tracker, shared_memory
+
+    from huggingface_hub import snapshot_download
+    from opt_prime.accel.pd import SamplingParams
+    from opt_prime.accel.pd.sampling import select_token
+    from prefill import Prefill
+    from safetensors import safe_open
+    from spec import validate
+
+    out = sys.stdout
+    with contextlib.redirect_stdout(sys.stderr):
+        path = (
+            Path(a.model)
+            if Path(a.model).is_dir()
+            else Path(snapshot_download(a.model, local_files_only=True))
+        )
+        config = validate(json.loads((path / "config.json").read_text()))
+        shape = (a.slots, 2, a.layers, a.kv_heads, a.max_context, a.head_dim)
+        if min(shape) <= 0 or (a.layers, a.kv_heads, a.head_dim) != (
+            config["num_hidden_layers"],
+            config["num_key_value_heads"],
+            config["head_dim"],
+        ):
+            raise ValueError("shared-memory geometry does not match checkpoint")
+        weights = {}
+        for shard in sorted(path.glob("*.safetensors")):
+            with safe_open(shard, framework="pt") as f:
+                for key in f.keys():  # noqa: SIM118 - safe_open is not iterable
+                    weights[key] = f.get_tensor(key).float()
+        engine = Prefill(config, weights, width=a.width, chip=a.chip)
+        shm = shared_memory.SharedMemory(name=a.shm)
+        # The parent owns unlinking; this child must not unlink another process's slab.
+        resource_tracker.unregister(shm._name, "shared_memory")
+        try:
+            slab = np.ndarray(shape, dtype=np.uint16, buffer=shm.buf)
+            print(json.dumps({"ready": True}), file=out, flush=True)
+            for line in sys.stdin:
+                req = {}
+                try:
+                    req = json.loads(line)
+                    validate_request(slab, req, a.width, config["vocab_size"])
+                    sampling = dict(req.get("sampling") or {})
+                    sampling["stop_token_ids"] = tuple(
+                        sampling.get("stop_token_ids", ())
+                    )
+                    sampling = SamplingParams(**sampling)
+                    t = time.perf_counter()
+                    k, v, logits = engine.prefill(list(req["tokens"]))
+                    token = select_token(logits, sampling, len(req["tokens"]))
+                    n = store_result(slab, req, k, v)
+                    result = {
+                        "id": req["id"],
+                        "token": int(token),
+                        "n": n,
+                        "cached": 0,
+                        "prefill_s": time.perf_counter() - t,
+                    }
+                except Exception as error:  # noqa: BLE001 - report errors through the worker protocol
+                    result = {"id": req.get("id"), "error": repr(error)}
+                print(json.dumps(result), file=out, flush=True)
+        finally:
+            shm.close()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/opt_prime/examples/pd/qwen3_rngd/worker_python.sh b/opt_prime/examples/pd/qwen3_rngd/worker_python.sh
new file mode 100755
index 00000000..fcf2071a
--- /dev/null
+++ b/opt_prime/examples/pd/qwen3_rngd/worker_python.sh
@@ -0,0 +1,9 @@
+#!/bin/bash
+# Adapt the existing worker launch without modifying the installed serving package.
+set -eu
+if [ "${1:-}" != -m ] || [ "${2:-}" != opt_prime.accel.pd.prefill_server ]; then
+    echo 'expected SubprocessPrefillWorker module launch' >&2
+    exit 2
+fi
+shift 2
+exec "$(dirname "$0")/furiosa_prefill_python.sh" -m worker "$@"

From 3855050a23cfc9f866d959e2b2ae147ef0273a9a Mon Sep 17 00:00:00 2001
From: gspark <gspark@etri.re.kr>
Date: Fri, 23 Oct 2026 22:32:12 +0900
Subject: [PATCH 3/6] feat: expose reusable Gemma prefill through the live PD
 worker

---
 .../examples/pd/qwen3_rngd/test_worker.py     |  23 ++++
 opt_prime/examples/pd/qwen3_rngd/worker.py    |  28 ++++-
 .../examples/pd/serving_matrix/README.md      |   8 ++
 .../examples/pd/serving_matrix/STATUS.md      |   4 +
 .../pd/serving_matrix/check_live_worker.py    |  64 +++++++++++
 .../serving_matrix/evidence234/events.jsonl   |   1 +
 .../live234-gemma-refactor-parity.json        |  30 ++++++
 .../evidence234/live234-gemma-rngd128.jsonl   |  12 +++
 .../evidence234/live234-gemma-worker.json     |  41 +++++++
 .../2026-10-23-live-model-pd234.md            |  42 ++++++++
 .../pd/serving_matrix/gemma_prefill.py        | 100 ++++++++++++++++++
 .../pd/serving_matrix/rngd_gemma_sweep.py     |  59 +----------
 .../pd/serving_matrix/test_gemma_prefill.py   |  75 +++++++++++++
 13 files changed, 428 insertions(+), 59 deletions(-)
 create mode 100644 opt_prime/examples/pd/serving_matrix/check_live_worker.py
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-refactor-parity.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-rngd128.jsonl
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-worker.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-model-pd234.md
 create mode 100644 opt_prime/examples/pd/serving_matrix/gemma_prefill.py
 create mode 100644 opt_prime/examples/pd/serving_matrix/test_gemma_prefill.py

diff --git a/opt_prime/examples/pd/qwen3_rngd/test_worker.py b/opt_prime/examples/pd/qwen3_rngd/test_worker.py
index 87b4e326..d7fdd2bf 100644
--- a/opt_prime/examples/pd/qwen3_rngd/test_worker.py
+++ b/opt_prime/examples/pd/qwen3_rngd/test_worker.py
@@ -3,6 +3,7 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 import worker
 
@@ -46,3 +47,25 @@ def test_invalid_requests_leave_slab_untouched(self):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+def test_explicit_engine_factory_is_local_callable():
+    import sys
+    import types
+    from unittest.mock import patch
+
+    from worker import engine_class
+
+    module = types.ModuleType("test_local_prefill")
+
+    class Engine:
+        def __init__(self, config, weights, width, chip):
+            self.args = config, weights, width, chip
+
+    module.Prefill = Engine
+    with patch.dict(sys.modules, {"test_local_prefill": module}):
+        cls = engine_class("test_local_prefill:Prefill")
+        assert cls({}, {}, width=128, chip=0).args == ({}, {}, 128, 0)
+        for invalid in ["test_local_prefill", ":Prefill", "test_local_prefill:"]:
+            with pytest.raises(ValueError):
+                engine_class(invalid)
diff --git a/opt_prime/examples/pd/qwen3_rngd/worker.py b/opt_prime/examples/pd/qwen3_rngd/worker.py
index fc1d813f..104375bb 100644
--- a/opt_prime/examples/pd/qwen3_rngd/worker.py
+++ b/opt_prime/examples/pd/qwen3_rngd/worker.py
@@ -44,6 +44,19 @@ def store_result(slab, req, keys, values):
     return n
 
 
+def engine_class(factory):
+    """Load an explicit trusted local CLI factory; never select this from request JSON."""
+    from importlib import import_module
+
+    module, separator, name = factory.partition(":")
+    if not separator or not module or not name:
+        raise ValueError("engine factory must be module:callable")
+    cls = getattr(import_module(module), name)
+    if not callable(cls):
+        raise TypeError("engine factory must be callable")
+    return cls
+
+
 def main():
     ap = argparse.ArgumentParser(description=__doc__)
     ap.add_argument("--model", required=True)
@@ -52,15 +65,17 @@ def main():
         ap.add_argument("--" + name, type=int, required=True)
     ap.add_argument("--width", type=int, default=32)
     ap.add_argument("--chip", type=int, default=0)
+    ap.add_argument(
+        "--engine-factory",
+        help="trusted local module:callable accepting config, weights, width, chip",
+    )
     a = ap.parse_args()
     from multiprocessing import resource_tracker, shared_memory
 
     from huggingface_hub import snapshot_download
     from opt_prime.accel.pd import SamplingParams
     from opt_prime.accel.pd.sampling import select_token
-    from prefill import Prefill
     from safetensors import safe_open
-    from spec import validate
 
     out = sys.stdout
     with contextlib.redirect_stdout(sys.stderr):
@@ -69,7 +84,14 @@ def main():
             if Path(a.model).is_dir()
             else Path(snapshot_download(a.model, local_files_only=True))
         )
-        config = validate(json.loads((path / "config.json").read_text()))
+        config = json.loads((path / "config.json").read_text())
+        if a.engine_factory:
+            Prefill = engine_class(a.engine_factory)
+        else:
+            from prefill import Prefill
+            from spec import validate
+
+            config = validate(config)
         shape = (a.slots, 2, a.layers, a.kv_heads, a.max_context, a.head_dim)
         if min(shape) <= 0 or (a.layers, a.kv_heads, a.head_dim) != (
             config["num_hidden_layers"],
diff --git a/opt_prime/examples/pd/serving_matrix/README.md b/opt_prime/examples/pd/serving_matrix/README.md
index 4ae571c6..daa0cc2d 100644
--- a/opt_prime/examples/pd/serving_matrix/README.md
+++ b/opt_prime/examples/pd/serving_matrix/README.md
@@ -166,3 +166,11 @@ python check_rngd_gpu.py --model "$MODEL_SNAPSHOT" \
 ```
 
 For Qwen use `rngd_qwen_sweep --model MODEL --experimental-long-width`; for KONI use `rngd_koni_sweep --partitioned --group 1 --rms-tile 512`. Pass the same fixture/case/width/output options. Each selected case saves a numbered export containing its exact IDs and EOS set. Both `check_rngd_atom.py` and `check_gemma_atom_handoff.py` accept `--eos-aware --max-output 64`; legacy fixed-length checks remain the default. Raw full-prefix host exports are not streaming overlap. Shared wrong native answers remain wrong after a successful handoff; inspect the semantic records separately from transport parity.
+
+## Callable Gemma and real worker (#234)
+
+`gemma_prefill.Prefill` is shared by `rngd_gemma_sweep.py` and the experimental worker's explicit `--engine-factory gemma_prefill:Prefill` choice. Add this directory and `../qwen3_rngd` to the isolated worker PYTHONPATH, with the existing optimus-rm runtime and validated SDK launcher. Factories are trusted local CLI code, not request parameters. The default Qwen worker remains unchanged. The shared-memory layout and JSON request/reply fields remain compatible with SubprocessPrefillWorker.
+
+`check_live_worker.py --python "$WORKER_PYTHON" --model unsloth/gemma-3-1b-it --engine-factory gemma_prefill:Prefill --width 128 --exports "$EXPORT0" "$EXPORT1" "$EXPORT2" "$EXPORT3" --output live-worker.json` runs real prefill per request, verifies returned KV and first tokens against trusted exports, and checks rejection/recovery for an oversized prompt. Run under the decoder-side environment with its existing core package; the worker interpreter owns the RNGD SDK.
+
+Short-input producer/full-logit parity and real shared-memory requests are verified. This is not yet full HTTP PD or long live integration. Numerical mismatch #233 remains, and no compute-transfer overlap or serving speedup is claimed. See the live-model-pd234 experiment and evidence234.
diff --git a/opt_prime/examples/pd/serving_matrix/STATUS.md b/opt_prime/examples/pd/serving_matrix/STATUS.md
index 857ec49a..b5f538c2 100644
--- a/opt_prime/examples/pd/serving_matrix/STATUS.md
+++ b/opt_prime/examples/pd/serving_matrix/STATUS.md
@@ -55,3 +55,7 @@ KONI (6137–6143 tokens) and Qwen3-0.6B (6026–6031 tokens) pass16/16 GPU/ATOM
 Gemma long-context update:6423–6428tokens,12producer runs, GPU3/4 and ATOM4/4 EOS matches; cumulative23/24. Arithmetic nativeGPU30 vs RNGD/ATOM25, both wrong; nativeGPU FP32 also30.52slot readbacks and native ATOM roundtrip exact. Same layer equations on GPU choose3 at BF16/FP32; accumulating RNGD numerical differences tracked in #233. Qwen3-4B queue remains live.
 
 Final6K matrix:48producer runs,31/32 EOS comparisons across four models/four questions/two decoders. Qwen3-4B4/4 on each decoder,72slots exact; the original queue exited0. Gemma arithmetic mismatch remains#233. Real live PD integration starts under#234; no overall-goal completion claim.
+
+## Live worker integration (#234)
+
+Reusable Gemma producer matches original RNGD K/V/full logits4/4 at width128. Real SubprocessPrefillWorker5requests match K/V/token and recover after an oversized request;15CPU tests plus5subtests pass. No communication-protocol change. Full live HTTP, other adapters and long/lifecycle conditions remain open; #233 numerical limits retained.
diff --git a/opt_prime/examples/pd/serving_matrix/check_live_worker.py b/opt_prime/examples/pd/serving_matrix/check_live_worker.py
new file mode 100644
index 00000000..9c84b920
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/check_live_worker.py
@@ -0,0 +1,64 @@
+"""Validate actual subprocess prefill results and recovery against trusted exports."""
+
+import argparse
+import json
+import os
+from pathlib import Path
+
+import torch
+from opt_prime.accel.pd import PDRequest
+from opt_prime.accel.pd.pd_batching import SubprocessPrefillWorker
+
+p = argparse.ArgumentParser(description=__doc__)
+p.add_argument("--python", required=True)
+p.add_argument("--model", required=True)
+p.add_argument("--engine-factory", required=True)
+p.add_argument("--width", type=int, required=True)
+p.add_argument("--exports", nargs="+", required=True)
+p.add_argument("--output", required=True)
+a = p.parse_args()
+exports = [torch.load(path, weights_only=True) for path in a.exports]
+layers, heads, _, dim = exports[0]["keys"].shape
+rows = []
+worker = SubprocessPrefillWorker(
+    a.python,
+    ["--model", a.model, "--width", str(a.width), "--engine-factory", a.engine_factory],
+    slots=1,
+    layers=layers,
+    kv_heads=heads,
+    head_dim=dim,
+    max_context=a.width,
+    env=dict(os.environ),
+)
+try:
+    for i, record in enumerate(exports + exports[:1]):
+        if i == len(exports):
+            bad = worker.submit(
+                PDRequest("oversize", tuple([1] * (a.width + 1)), 1), slot=0
+            )
+            try:
+                bad.result(timeout=120)
+            except RuntimeError as error:
+                assert "capacity" in str(error)
+            else:
+                raise AssertionError("oversize request was accepted")
+        request = PDRequest(str(i), tuple(record["ids"]), 1)
+        out = worker.submit(request, slot=0).result(timeout=120)
+        assert out.start == 0 and out.token == int(record["logits"].argmax())
+        assert torch.equal(out.keys, record["keys"]) and torch.equal(
+            out.values, record["values"]
+        )
+        rows.append(
+            {
+                "case": record["metadata"]["case"],
+                "keys_exact": True,
+                "values_exact": True,
+                "first_token_exact": True,
+                "prefill_s": out.prefill_s,
+            }
+        )
+    result = {"worker": "live_rngd_subprocess", "rows": rows, "oversize_recovery": True}
+    Path(a.output).write_text(json.dumps(result, indent=2) + "\n")
+    print(json.dumps(result), flush=True)
+finally:
+    worker.close()
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl b/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
new file mode 100644
index 00000000..bcd190fe
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
@@ -0,0 +1 @@
+{"id": "819c75b3957c4f0b980915096014c36f", "ts": "2026-10-23T22:30:07+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Callable Gemma refactor:4/4 full K/V/logits bit-exact on RNGD; actual subprocess/shared-memory worker5/5 exact and oversize recovery pass.15CPU tests+5subtests. No live HTTP claim yet.", "next": "Connect live worker to production ATOM HTTP; extend Qwen/KONI adapters and long-prefix/lifecycle conditions.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"model": "Gemma3-1B", "width": "128"}, "artifacts": ["opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-worker.json"], "experiments": ["opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-model-pd234.md"], "labels": ["experiment"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "validation host", "branch": "main", "commit": "351d39d-dirty"}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-refactor-parity.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-refactor-parity.json
new file mode 100644
index 00000000..7a33da60
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-refactor-parity.json
@@ -0,0 +1,30 @@
+[
+  {
+    "case": "capital-0",
+    "ids_equal": true,
+    "keys_equal": true,
+    "values_equal": true,
+    "logits_equal": true
+  },
+  {
+    "case": "arithmetic-0",
+    "ids_equal": true,
+    "keys_equal": true,
+    "values_equal": true,
+    "logits_equal": true
+  },
+  {
+    "case": "korean-0",
+    "ids_equal": true,
+    "keys_equal": true,
+    "values_equal": true,
+    "logits_equal": true
+  },
+  {
+    "case": "retrieval-0",
+    "ids_equal": true,
+    "keys_equal": true,
+    "values_equal": true,
+    "logits_equal": true
+  }
+]
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-rngd128.jsonl b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-rngd128.jsonl
new file mode 100644
index 00000000..e70ece24
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-rngd128.jsonl
@@ -0,0 +1,12 @@
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 23, "case": "capital-0", "repeat": 0, "prefill_s": 0.060643895994871855, "repeat_exact": true, "first_token": 50429, "kv_shape": [26, 1, 23, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 23, "case": "capital-0", "repeat": 1, "prefill_s": 0.04894804093055427, "repeat_exact": true, "first_token": 50429, "kv_shape": [26, 1, 23, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 23, "case": "capital-0", "repeat": 2, "prefill_s": 0.04832607100252062, "repeat_exact": true, "first_token": 50429, "kv_shape": [26, 1, 23, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 24, "case": "arithmetic-0", "repeat": 0, "prefill_s": 0.04934179107658565, "repeat_exact": true, "first_token": 236812, "kv_shape": [26, 1, 24, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 24, "case": "arithmetic-0", "repeat": 1, "prefill_s": 0.048956381040625274, "repeat_exact": true, "first_token": 236812, "kv_shape": [26, 1, 24, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 24, "case": "arithmetic-0", "repeat": 2, "prefill_s": 0.04907395096961409, "repeat_exact": true, "first_token": 236812, "kv_shape": [26, 1, 24, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 24, "case": "korean-0", "repeat": 0, "prefill_s": 0.04923301108647138, "repeat_exact": true, "first_token": 222340, "kv_shape": [26, 1, 24, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 24, "case": "korean-0", "repeat": 1, "prefill_s": 0.04606894007883966, "repeat_exact": true, "first_token": 222340, "kv_shape": [26, 1, 24, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 24, "case": "korean-0", "repeat": 2, "prefill_s": 0.04869454097934067, "repeat_exact": true, "first_token": 222340, "kv_shape": [26, 1, 24, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 28, "case": "retrieval-0", "repeat": 0, "prefill_s": 0.04887848102953285, "repeat_exact": true, "first_token": 28487, "kv_shape": [26, 1, 28, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 28, "case": "retrieval-0", "repeat": 1, "prefill_s": 0.0483387210406363, "repeat_exact": true, "first_token": 28487, "kv_shape": [26, 1, 28, 256]}
+{"model": "unsloth/gemma-3-1b-it", "revision": "5b11413a10db4e486ef16a20101fd028f8f2499c", "width": 128, "input_tokens": 28, "case": "retrieval-0", "repeat": 2, "prefill_s": 0.04836872196756303, "repeat_exact": true, "first_token": 28487, "kv_shape": [26, 1, 28, 256]}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-worker.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-worker.json
new file mode 100644
index 00000000..b0282d83
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-worker.json
@@ -0,0 +1,41 @@
+{
+  "worker": "live_rngd_subprocess",
+  "rows": [
+    {
+      "case": "capital-0",
+      "keys_exact": true,
+      "values_exact": true,
+      "first_token_exact": true,
+      "prefill_s": 0.04225390707142651
+    },
+    {
+      "case": "arithmetic-0",
+      "keys_exact": true,
+      "values_exact": true,
+      "first_token_exact": true,
+      "prefill_s": 0.040302990935742855
+    },
+    {
+      "case": "korean-0",
+      "keys_exact": true,
+      "values_exact": true,
+      "first_token_exact": true,
+      "prefill_s": 0.040417241980321705
+    },
+    {
+      "case": "retrieval-0",
+      "keys_exact": true,
+      "values_exact": true,
+      "first_token_exact": true,
+      "prefill_s": 0.04033046099357307
+    },
+    {
+      "case": "capital-0",
+      "keys_exact": true,
+      "values_exact": true,
+      "first_token_exact": true,
+      "prefill_s": 0.04012918099761009
+    }
+  ],
+  "oversize_recovery": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-model-pd234.md b/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-model-pd234.md
new file mode 100644
index 00000000..853e5f74
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-model-pd234.md
@@ -0,0 +1,42 @@
+# Callable Gemma producer and live worker integration
+
+- **Date**: 2026-10-23
+- **Task**: aicomp #234
+- **Host / device**: CPU tests on validation host; owned RNGD+ATOM allocation for pending hardware refactor check.
+- **Commit**: feat/234-live-model-pd working tree, based on67fc47d and merged dependency exp/199-qwen3-serving(d4b6fe6).
+
+## Question
+
+Can the validated Gemma prefill computation be called repeatedly through the existing shared-memory worker without changing its wire protocol?
+
+## Setup
+
+Extract gemma_prefill.Prefill(config,weights,width,chip) from the existing sweep and route the sweep through it. It retains the exact compiled layer, normalization/rotary/mask construction, BF16 carriers, true-length KV export and CPU final head. Add an explicit trusted local module:callable factory option to the existing #199 worker; default Qwen behavior and request/slab protocol remain unchanged. The factory is a CLI choice, never request data.
+
+## Command
+
+```bash
+PYTHONPATH=opt_prime/examples/pd/qwen3_rngd:opt_prime/examples/pd/serving_matrix python -m pytest -q opt_prime/examples/pd/qwen3_rngd/test_worker.py opt_prime/examples/pd/serving_matrix/test_gemma_prefill.py opt_prime/examples/pd/serving_matrix/test_gemma_parts.py
+bash run-gemma234.sh --width 128 --fixtures semantic227-gemma-fixtures.json --cases capital-0 arithmetic-0 korean-0 retrieval-0 --output live234-gemma-rngd128
+```
+
+## Results
+
+- New callable test initially fails because the class is absent; factory test initially1failed/2passed before implementation.
+- A pointwise BF16 KV comparison initially failed at one near-zero element (absolute0.0327); use a documented aggregate relative RMS criterion below3% for the tiny BF16 HF comparison. This is approximate numerical validation, not exact HF parity. Existing FP32 layer equation tests remain strict and unchanged. Repeated identical prompts separated by a longer prompt must return bit-identical K/V/logits; invalid tokens and capacity are rejected.
+- CPU result:15passed and5subtests passed; scoped Ruff passes. CPU SDK stub executes the layer equations, so it does not prove hardware compilation.
+- Hardware refactor run launched in the owned paired allocation, comparing four short cases against the retained #227 exports. Completed:12producer runs; all four cases have bit-identical K,V and full logits versus the original #227 exports. No live HTTP claim yet.
+
+Raw logs: analysis/serving-matrix-201/live234-{gemma-red,gemma-green,gemma-tests,worker-red,tests,gemma-rngd128}.log. No shared SDK installation modified.
+
+## Conclusion
+
+The callable producer and worker hook are implemented locally; live worker/HTTP validation and Qwen/KONI integration remain open. The full four-model goal is active, including numerical issue#233 and real overlap/cache/performance work. Do not publish success based solely on the CPU stub.
+
+## Actual subprocess-worker validation
+
+The new worker factory was exercised through the existing production SubprocessPrefillWorker, in an owned paired RNGD+ATOM allocation (only RNGD used by this check). The child compiles once, then performs real prefill for four questions and a repeated first question. Parent receives KV via the original shared-memory slab and JSON reply protocol. All5requests match retained raw K/V and first-token argmax exactly. An oversized request fails through the worker error reply; a following valid request succeeds. Measured worker-prefill fields40.13–42.25ms include the protocol's reported producer/export work; they are not HTTP TTFT or serving speedup.
+
+Command: `bash run-live-worker234.sh --python "$WORKER_PYTHON" --model unsloth/gemma-3-1b-it --engine-factory gemma_prefill:Prefill --width 128 --exports semantic227-gemma-rngd128-0.pt semantic227-gemma-rngd128-1.pt semantic227-gemma-rngd128-2.pt semantic227-gemma-rngd128-3.pt --output live234-gemma-worker.json`. The isolated launcher redirects the expected prefill_server module to the candidate worker and uses the validated Furiosa interpreter. CPU/ATOM environment hosts the parent; no communication-library code changed.
+
+Evidence:live234-gemma-refactor-parity.json, live234-gemma-worker.json, live234-gemma-rngd128.jsonl. Short-prompt standalone producer and actual worker are now verified; HTTP decode, long live prefixes, cancellation and the other model adapters remain open.
diff --git a/opt_prime/examples/pd/serving_matrix/gemma_prefill.py b/opt_prime/examples/pd/serving_matrix/gemma_prefill.py
new file mode 100644
index 00000000..ea0d5008
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/gemma_prefill.py
@@ -0,0 +1,100 @@
+"""Reusable Gemma3 full-prefix producer for the existing PD worker contract."""
+
+import torch
+from gemma_parts import (
+    FlatGemmaLayer,
+    attention_inputs,
+    final_logits,
+    initial_hidden,
+    layer_parameters,
+)
+from semantic_export import valid_prefix
+
+
+class Prefill:
+    def __init__(self, config, weights, width, chip=0):
+        if type(width) is not int or width <= 0:
+            raise ValueError("compiled width must be positive")
+        from rngd_pd.engines.fsdk import CompileModule, FCConfig, dev_str, hint, sync
+
+        self.config, self.width, self.chip, self.sync = config, width, chip, sync
+        self.weights = {
+            name: value.detach().to("cpu", torch.bfloat16)
+            for name, value in weights.items()
+        }
+        self.device = dev_str(chip)
+        module = FlatGemmaLayer(config)
+        params = [
+            layer_parameters(self.weights, i)
+            for i in range(config["num_hidden_layers"])
+        ]
+        self.local = [
+            kind == "sliding_attention" for kind in config.get("layer_types", [])
+        ]
+        if not self.local:
+            self.local = [
+                (i + 1) % config.get("sliding_window_pattern", 6) != 0
+                for i in range(len(params))
+            ]
+        if len(self.local) != len(params):
+            raise ValueError("layer types do not match layer count")
+        positions = {
+            local: attention_inputs(config, width, self.local.index(local))
+            for local in set(self.local)
+        }
+        rot, mask = positions[self.local[0]]
+        sample = (
+            torch.zeros(width, config["hidden_size"], dtype=torch.bfloat16),
+            *params[0][0],
+            *params[0][1],
+            rot.to(torch.bfloat16),
+            mask.to(torch.bfloat16),
+        )
+        self.compiled = CompileModule.from_module(
+            module,
+            tuple(torch.zeros(t.shape, dtype=t.dtype) for t in sample),
+            compiler_config=FCConfig(tactic_hint=hint("ForLlmModelPrefill")),
+        )
+        self.compiled.to(self.device)
+        self.params = [
+            tuple(t.contiguous().to(self.device) for t in (*norms, *mats))
+            for norms, mats in params
+        ]
+        self.positions = {
+            local: tuple(t.to(torch.bfloat16).to(self.device) for t in pair)
+            for local, pair in positions.items()
+        }
+        sync(chip)
+
+    @torch.inference_mode()
+    def prefill(self, tokens):
+        ids = list(tokens)
+        if not 0 < len(ids) <= self.width or any(
+            type(t) is not int or not 0 <= t < self.config["vocab_size"] for t in ids
+        ):
+            raise ValueError("invalid tokens or prompt exceeds compiled width")
+        pad = self.config.get("pad_token_id")
+        if pad is None:
+            pad = 0
+        x = initial_hidden(
+            self.weights["model.embed_tokens.weight"],
+            ids + [pad] * (self.width - len(ids)),
+        ).to(self.device)
+        keys, values = [], []
+        for i, layer in enumerate(self.params):
+            x, k, v = self.compiled(x, *layer, *self.positions[self.local[i]])
+            keys.append(k)
+            values.append(v)
+        self.sync(self.chip)
+        k = torch.stack([t.cpu().permute(1, 0, 2) for t in keys])
+        v = torch.stack([t.cpu().permute(1, 0, 2) for t in values])
+        k, v, last_hidden = valid_prefix(k, v, x.cpu(), len(ids))
+        logits = final_logits(
+            last_hidden,
+            self.weights["model.norm.weight"],
+            self.weights.get(
+                "lm_head.weight", self.weights["model.embed_tokens.weight"]
+            ),
+            self.config["rms_norm_eps"],
+        )
+        return k, v, logits
diff --git a/opt_prime/examples/pd/serving_matrix/rngd_gemma_sweep.py b/opt_prime/examples/pd/serving_matrix/rngd_gemma_sweep.py
index c585930c..8b4d1af5 100644
--- a/opt_prime/examples/pd/serving_matrix/rngd_gemma_sweep.py
+++ b/opt_prime/examples/pd/serving_matrix/rngd_gemma_sweep.py
@@ -6,18 +6,10 @@
 from pathlib import Path
 
 import torch
-from gemma_parts import (
-    FlatGemmaLayer,
-    attention_inputs,
-    final_logits,
-    initial_hidden,
-    layer_parameters,
-)
+from gemma_prefill import Prefill
 from huggingface_hub import snapshot_download
 from inputs import exact_prompt
-from rngd_pd.engines.fsdk import CompileModule, FCConfig, dev_str, hint, sync
 from safetensors import safe_open
-from semantic_export import valid_prefix
 from transformers import AutoTokenizer
 
 MODEL = "unsloth/gemma-3-1b-it"
@@ -30,7 +22,6 @@
 a = p.parse_args()
 path = Path(snapshot_download(MODEL, revision=REVISION, local_files_only=True))
 c = json.loads((path / "config.json").read_text())
-module = FlatGemmaLayer(c)
 weights = {}
 for shard in sorted(path.glob("*.safetensors")):
     with safe_open(shard, framework="pt") as f:
@@ -60,35 +51,9 @@
         a.width,
     )
     cases = [{"case": "fixed", "ids": ids}]
-dev = dev_str(0)
 start = time.perf_counter()
-params = [layer_parameters(weights, i) for i in range(c["num_hidden_layers"])]
-rot, mask = attention_inputs(c, a.width, 0)
-sample = (
-    torch.zeros(a.width, c["hidden_size"], dtype=torch.bfloat16),
-    *params[0][0],
-    *params[0][1],
-    rot.to(torch.bfloat16),
-    mask.to(torch.bfloat16),
-)
 print(json.dumps({"compile_start": True, "width": a.width}), flush=True)
-# Match the contiguous runtime tensors; zeros_like preserves transposed weight strides.
-compiled = CompileModule.from_module(
-    module,
-    tuple(torch.zeros(t.shape, dtype=t.dtype) for t in sample),
-    compiler_config=FCConfig(tactic_hint=hint("ForLlmModelPrefill")),
-)
-compiled.to(dev)
-params = [
-    tuple(t.contiguous().to(dev) for t in (*norms, *mats)) for norms, mats in params
-]
-# Only two distinct positional/mask inputs, reused across the layer schedule.
-positions = {}
-for index in [0, 5]:
-    positions[index] = tuple(
-        t.to(torch.bfloat16).to(dev) for t in attention_inputs(c, a.width, index)
-    )
-sync(0)
+engine = Prefill(c, weights, width=a.width)
 print(
     json.dumps(
         {"ready": True, "init_s": time.perf_counter() - start, "width": a.width}
@@ -98,28 +63,10 @@
 with torch.inference_mode(), open(a.output + ".jsonl", "w", buffering=1) as f:
     for case_index, case in enumerate(cases):
         ids = case["ids"]
-        padded_ids = ids + [tok.pad_token_id] * (a.width - len(ids))
         previous = None
         for repeat in range(3):
             start = time.perf_counter()
-            x = initial_hidden(weights["model.embed_tokens.weight"], padded_ids).to(dev)
-            keys = []
-            values = []
-            for i, layer in enumerate(params):
-                local = (i + 1) % c.get("sliding_window_pattern", 6) != 0
-                x, k, v = compiled(x, *layer, *positions[0 if local else 5])
-                keys.append(k)
-                values.append(v)
-            sync(0)
-            k = torch.stack([t.cpu().permute(1, 0, 2) for t in keys])
-            v = torch.stack([t.cpu().permute(1, 0, 2) for t in values])
-            k, v, last_hidden = valid_prefix(k, v, x.cpu(), len(ids))
-            logits = final_logits(
-                last_hidden,
-                weights["model.norm.weight"],
-                weights.get("lm_head.weight", weights["model.embed_tokens.weight"]),
-                c["rms_norm_eps"],
-            )
+            k, v, logits = engine.prefill(ids)
             elapsed = time.perf_counter() - start
             assert k.shape == v.shape == (26, 1, len(ids), 256)
             assert all(bool(torch.isfinite(t).all()) for t in (k, v, logits))
diff --git a/opt_prime/examples/pd/serving_matrix/test_gemma_prefill.py b/opt_prime/examples/pd/serving_matrix/test_gemma_prefill.py
new file mode 100644
index 00000000..60d7c19b
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/test_gemma_prefill.py
@@ -0,0 +1,75 @@
+import sys
+import types
+
+import pytest
+import torch
+from gemma_prefill import Prefill
+from transformers import Gemma3ForCausalLM, Gemma3TextConfig
+
+
+@pytest.fixture
+def cpu_sdk(monkeypatch):
+    sdk = types.ModuleType("rngd_pd.engines.fsdk")
+
+    class Compile:
+        @staticmethod
+        def from_module(module, sample, compiler_config):
+            return module
+
+    sdk.CompileModule = Compile
+    sdk.FCConfig = lambda **kw: kw
+    sdk.dev_str = lambda chip: "cpu"
+    sdk.hint = lambda name: name
+    sdk.sync = lambda chip: None
+    for name in ["rngd_pd", "rngd_pd.engines"]:
+        monkeypatch.setitem(sys.modules, name, types.ModuleType(name))
+    monkeypatch.setitem(sys.modules, "rngd_pd.engines.fsdk", sdk)
+
+
+def test_reusable_gemma_prefill_matches_hf_and_reuses_without_stale_prefix(cpu_sdk):
+    torch.manual_seed(234)
+    config = Gemma3TextConfig(
+        hidden_size=16,
+        intermediate_size=32,
+        num_hidden_layers=6,
+        num_attention_heads=4,
+        num_key_value_heads=1,
+        head_dim=4,
+        sliding_window=4,
+        sliding_window_pattern=6,
+        query_pre_attn_scalar=4,
+        vocab_size=32,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+    )
+    config._attn_implementation = "eager"
+    ref = Gemma3ForCausalLM(config).to(torch.bfloat16).eval()
+    engine = Prefill(config.to_dict(), ref.state_dict(), width=12)
+    first = None
+    with torch.inference_mode():
+        for ids in [[1, 4, 5], [1, 7, 8, 9, 10, 11, 12], [1, 4, 5]]:
+            k, v, logits = engine.prefill(ids)
+            out = ref(input_ids=torch.tensor([ids]), use_cache=True)
+            assert k.shape == v.shape == (6, 1, len(ids), 4)
+            torch.testing.assert_close(
+                logits, out.logits[0, -1].float(), atol=0.02, rtol=0.08
+            )
+            for i, layer in enumerate(out.past_key_values.layers):
+                kept = layer.keys.shape[2]
+                for actual, expected in [
+                    (k[i, :, -kept:], layer.keys[0]),
+                    (v[i, :, -kept:], layer.values[0]),
+                ]:
+                    relative = (
+                        actual.float() - expected.float()
+                    ).square().mean().sqrt() / expected.float().square().mean().sqrt()
+                    assert float(relative) < 0.03, (i, float(relative))
+            if len(ids) == 3:
+                if first is None:
+                    first = (k.clone(), v.clone(), logits.clone())
+                else:
+                    assert all(torch.equal(a, b) for a, b in zip(first, (k, v, logits)))
+        for ids in [[], list(range(13)), [-1], [32]]:
+            with pytest.raises(ValueError):
+                engine.prefill(ids)

From 74d0d3b934861265f175dea5c02e33cc383ea85f Mon Sep 17 00:00:00 2001
From: gspark <gspark@etri.re.kr>
Date: Fri, 23 Oct 2026 22:57:39 +0900
Subject: [PATCH 4/6] feat: validate live RNGD to ATOM HTTP across four models

---
 .../examples/pd/qwen3_rngd/test_worker.py     |  38 ++
 opt_prime/examples/pd/qwen3_rngd/worker.py    |  19 +-
 .../examples/pd/serving_matrix/README.md      |   4 +
 .../examples/pd/serving_matrix/STATUS.md      |   4 +
 .../pd/serving_matrix/atom_single_block.py    |   2 +
 .../pd/serving_matrix/check_live_http.py      | 194 +++++++
 .../serving_matrix/evidence234/events.jsonl   |   1 +
 .../evidence234/live234-gemma-http.json       | 400 +++++++++++++++
 .../evidence234/live234-http-summary.json     |  47 ++
 .../evidence234/live234-koni-http.json        | 415 +++++++++++++++
 .../evidence234/live234-qwen06b-http.json     | 484 ++++++++++++++++++
 .../evidence234/live234-qwen4b-http.json      | 400 +++++++++++++++
 .../experiments/2026-10-23-live-http234.md    |  44 ++
 .../pd/serving_matrix/koni_prefill.py         |  45 ++
 .../pd/serving_matrix/long_width_variant.py   |   5 +
 .../pd/serving_matrix/test_koni_prefill.py    |  23 +
 16 files changed, 2121 insertions(+), 4 deletions(-)
 create mode 100644 opt_prime/examples/pd/serving_matrix/check_live_http.py
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-http-summary.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http234.md
 create mode 100644 opt_prime/examples/pd/serving_matrix/koni_prefill.py
 create mode 100644 opt_prime/examples/pd/serving_matrix/test_koni_prefill.py

diff --git a/opt_prime/examples/pd/qwen3_rngd/test_worker.py b/opt_prime/examples/pd/qwen3_rngd/test_worker.py
index d7fdd2bf..8b382b0f 100644
--- a/opt_prime/examples/pd/qwen3_rngd/test_worker.py
+++ b/opt_prime/examples/pd/qwen3_rngd/test_worker.py
@@ -69,3 +69,41 @@ def __init__(self, config, weights, width, chip):
         for invalid in ["test_local_prefill", ":Prefill", "test_local_prefill:"]:
             with pytest.raises(ValueError):
                 engine_class(invalid)
+
+
+@pytest.mark.parametrize(
+    "config,expected",
+    [
+        (
+            {
+                "num_hidden_layers": 32,
+                "num_key_value_heads": 8,
+                "hidden_size": 4096,
+                "num_attention_heads": 32,
+            },
+            (32, 8, 128),
+        ),
+        (
+            {
+                "num_hidden_layers": 36,
+                "num_key_value_heads": 8,
+                "hidden_size": 2560,
+                "num_attention_heads": 32,
+                "head_dim": 128,
+            },
+            (36, 8, 128),
+        ),
+        (
+            {
+                "num_hidden_layers": 26,
+                "num_key_value_heads": 1,
+                "hidden_size": 1152,
+                "num_attention_heads": 4,
+                "head_dim": 256,
+            },
+            (26, 1, 256),
+        ),
+    ],
+)
+def test_cache_geometry_preserves_explicit_head_dim(config, expected):
+    assert worker.cache_geometry(config) == expected
diff --git a/opt_prime/examples/pd/qwen3_rngd/worker.py b/opt_prime/examples/pd/qwen3_rngd/worker.py
index 104375bb..3b21c3f8 100644
--- a/opt_prime/examples/pd/qwen3_rngd/worker.py
+++ b/opt_prime/examples/pd/qwen3_rngd/worker.py
@@ -44,6 +44,19 @@ def store_result(slab, req, keys, values):
     return n
 
 
+def cache_geometry(config):
+    head_dim = config.get("head_dim")
+    if head_dim is None:
+        heads, hidden = config["num_attention_heads"], config["hidden_size"]
+        if type(heads) is not int or heads <= 0 or hidden % heads:
+            raise ValueError("invalid implicit head dimension")
+        head_dim = hidden // heads
+    shape = (config["num_hidden_layers"], config["num_key_value_heads"], head_dim)
+    if any(type(n) is not int or n <= 0 for n in shape):
+        raise ValueError("invalid cache geometry")
+    return shape
+
+
 def engine_class(factory):
     """Load an explicit trusted local CLI factory; never select this from request JSON."""
     from importlib import import_module
@@ -93,10 +106,8 @@ def main():
 
             config = validate(config)
         shape = (a.slots, 2, a.layers, a.kv_heads, a.max_context, a.head_dim)
-        if min(shape) <= 0 or (a.layers, a.kv_heads, a.head_dim) != (
-            config["num_hidden_layers"],
-            config["num_key_value_heads"],
-            config["head_dim"],
+        if min(shape) <= 0 or (a.layers, a.kv_heads, a.head_dim) != cache_geometry(
+            config
         ):
             raise ValueError("shared-memory geometry does not match checkpoint")
         weights = {}
diff --git a/opt_prime/examples/pd/serving_matrix/README.md b/opt_prime/examples/pd/serving_matrix/README.md
index daa0cc2d..53695d9f 100644
--- a/opt_prime/examples/pd/serving_matrix/README.md
+++ b/opt_prime/examples/pd/serving_matrix/README.md
@@ -174,3 +174,7 @@ For Qwen use `rngd_qwen_sweep --model MODEL --experimental-long-width`; for KONI
 `check_live_worker.py --python "$WORKER_PYTHON" --model unsloth/gemma-3-1b-it --engine-factory gemma_prefill:Prefill --width 128 --exports "$EXPORT0" "$EXPORT1" "$EXPORT2" "$EXPORT3" --output live-worker.json` runs real prefill per request, verifies returned KV and first tokens against trusted exports, and checks rejection/recovery for an oversized prompt. Run under the decoder-side environment with its existing core package; the worker interpreter owns the RNGD SDK.
 
 Short-input producer/full-logit parity and real shared-memory requests are verified. This is not yet full HTTP PD or long live integration. Numerical mismatch #233 remains, and no compute-transfer overlap or serving speedup is claimed. See the live-model-pd234 experiment and evidence234.
+
+`check_live_http.py` runs actual RNGD prefill and ATOM HTTP decode from token fixtures (it does not load saved KV). Pass `--batching` with the reviewed ATOM scheduler, `--http-module` with the #214 SIGPIPE-fixed server, the model/artifact cache, `--worker-python`, factory/width and fixture/baseline JSON. It checks two serial passes, four queued clients, cancel/disconnect and reuse. Factories: `prefill:Prefill` for short Qwen3, `long_width_variant:make_prefill` for the explicit experimental long Qwen variant, `gemma_prefill:Prefill`, and `koni_prefill:Prefill` for partitioned KONI widths divisible by512. KONI uses `--kv-blocks 1` with #236; other measured artifacts use3. Explicit head_dim is preserved; Llama's missing field is inferred from hidden size/attention heads.
+
+All four short-input live HTTP paths pass56normal/reuse response checks and64remote prefills including cancellations. Native-output parity does not imply answer quality. Width2048 live validation is pending; hardware batch remains1 and no matched performance benefit is established. See live-http234 record.
diff --git a/opt_prime/examples/pd/serving_matrix/STATUS.md b/opt_prime/examples/pd/serving_matrix/STATUS.md
index b5f538c2..d1d148d4 100644
--- a/opt_prime/examples/pd/serving_matrix/STATUS.md
+++ b/opt_prime/examples/pd/serving_matrix/STATUS.md
@@ -59,3 +59,7 @@ Final6K matrix:48producer runs,31/32 EOS comparisons across four models/four que
 ## Live worker integration (#234)
 
 Reusable Gemma producer matches original RNGD K/V/full logits4/4 at width128. Real SubprocessPrefillWorker5requests match K/V/token and recover after an oversized request;15CPU tests plus5subtests pass. No communication-protocol change. Full live HTTP, other adapters and long/lifecycle conditions remain open; #233 numerical limits retained.
+
+## Four-model actual HTTP PD (#234/#236)
+
+All four models:56/56 normal/concurrent/reuse responses equal native ATOM,64RNGD prefills and0local fallbacks. Cancellation/disconnect reuse passes. B1, cache-off, short prompts; KONI uses existingk1 artifact with#236. HTTP uses#214 SIGPIPE fix. No live-transfer-overlap or throughput gain claim. Width2048 live queue pending; long/cache/hardware-batch/performance scope remains.
diff --git a/opt_prime/examples/pd/serving_matrix/atom_single_block.py b/opt_prime/examples/pd/serving_matrix/atom_single_block.py
index 279dd435..0095521a 100644
--- a/opt_prime/examples/pd/serving_matrix/atom_single_block.py
+++ b/opt_prime/examples/pd/serving_matrix/atom_single_block.py
@@ -28,6 +28,8 @@ def __init__(self, model):
                 or m.get("layer_type", "full_attention") != "full_attention"
             ):
                 raise ValueError("requires uniform full-attention cache geometry")
+        self.hybrid_cache = False
+        self.single_block = True
         self.io_rt = model.model[1]
         self.lane_block = [0]
         self._stage = None
diff --git a/opt_prime/examples/pd/serving_matrix/check_live_http.py b/opt_prime/examples/pd/serving_matrix/check_live_http.py
new file mode 100644
index 00000000..ffecd42d
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/check_live_http.py
@@ -0,0 +1,194 @@
+"""Actual RNGD subprocess prefill to ATOM HTTP, with native-output and lifecycle checks."""
+
+import argparse
+import importlib.util
+import json
+import os
+import sys
+import threading
+import time
+import urllib.request
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+p = argparse.ArgumentParser(description=__doc__)
+p.add_argument("--batching", required=True)
+p.add_argument("--cache-dir", required=True)
+p.add_argument("--fixtures", required=True)
+p.add_argument("--cases", nargs="+", required=True)
+p.add_argument("--worker-python", required=True)
+p.add_argument("--engine-factory", required=True)
+p.add_argument("--width", type=int, required=True)
+p.add_argument("--kv-blocks", type=int, default=3)
+p.add_argument("--model", default="unsloth/gemma-3-1b-it")
+p.add_argument("--http-module", required=True)
+p.add_argument("--baseline", required=True)
+p.add_argument("--output", required=True)
+a = p.parse_args()
+for name, path in [
+    ("opt_prime.accel.platforms.atom.batching", a.batching),
+    ("opt_prime.accel.serve.llmd", a.http_module),
+]:
+    spec = importlib.util.spec_from_file_location(name, path)
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[name] = module
+    spec.loader.exec_module(module)
+from opt_prime.accel.pd.pd_batching import SubprocessPrefillWorker
+from opt_prime.accel.platforms.atom.engine import AtomRblnEngine
+from opt_prime.accel.serve.batched import BatchedOwner, BatchedService
+from opt_prime.accel.serve.llmd import serve_stage
+from transformers import AutoTokenizer
+
+fixtures = json.loads(Path(a.fixtures).read_text())
+cases = [r for r in fixtures["cases"] if r["case"] in a.cases]
+assert {r["case"] for r in cases} == set(a.cases)
+assert cases and all(0 < len(r["ids"]) <= a.width for r in cases)
+baseline = {
+    r["case"]: r["tokens"]
+    for r in map(json.loads, Path(a.baseline).read_text().splitlines())
+}
+
+
+nodes = list(Path("/dev").glob("rbln[0-9]*"))
+assert len(nodes) == 1
+engine = AtomRblnEngine(
+    devices=[int(nodes[0].name[4:])],
+    max_seq_len=9216,
+    batch_size=1,
+    kv_blocks=a.kv_blocks,
+    cache_dir=a.cache_dir,
+)
+worker = None
+owner = thread = server = http = None
+rows = []
+try:
+    model = a.model
+    engine.start(model)
+    c = engine.model.config.get_text_config()
+    worker = SubprocessPrefillWorker(
+        a.worker_python,
+        [
+            "--model",
+            model,
+            "--width",
+            str(a.width),
+            "--engine-factory",
+            a.engine_factory,
+        ],
+        slots=1,
+        layers=c.num_hidden_layers,
+        kv_heads=c.num_key_value_heads,
+        head_dim=getattr(c, "head_dim", None) or c.hidden_size // c.num_attention_heads,
+        max_context=a.width,
+        env=dict(os.environ),
+    )
+    tok = AutoTokenizer.from_pretrained(model, local_files_only=True)
+    b = engine.batcher(prefix_cache=False, prefill_worker=worker, local_prefill=False)
+    service = BatchedService(
+        b,
+        model=model,
+        vocab_size=engine.model.config.vocab_size,
+        max_output=64,
+        render=lambda ids: tok.decode(ids, skip_special_tokens=True),
+        tokenizer=tok,
+        eos_token_ids=tuple(fixtures["eos_ids"]),
+    )
+    owner = BatchedOwner(service)
+    thread = threading.Thread(target=owner.run, daemon=True)
+    thread.start()
+    server, http = serve_stage(owner, "decode", 0)
+
+    def request(path, body, rid):
+        req = urllib.request.Request(
+            f"http://127.0.0.1:{server.server_port}{path}",
+            data=json.dumps(body).encode(),
+            headers={"x-request-id": rid},
+        )
+        return urllib.request.urlopen(req, timeout=60)
+
+    def complete(r, rid):
+        started = time.perf_counter()
+        with request(
+            "/v1/completions",
+            {"model": model, "prompt": r["ids"], "max_tokens": 64},
+            rid,
+        ) as response:
+            out = json.load(response)
+        assert out["choices"][0]["finish_reason"] == "stop"
+        assert out["choices"][0]["token_ids"] == baseline[r["case"]], out
+        rows.append(
+            {
+                "case": r["case"],
+                "request": rid,
+                "response": out,
+                "latency_s": time.perf_counter() - started,
+            }
+        )
+
+    # Alternate long and short prefixes while blocks rotate under the same local lane.
+    for repeat in range(2):
+        for i, r in enumerate(cases):
+            complete(r, f"normal-{repeat}-{i}")
+    with ThreadPoolExecutor(max_workers=4) as clients:
+        list(
+            clients.map(
+                lambda pair: complete(pair[1], f"concurrent-{pair[0]}"),
+                enumerate(cases),
+            )
+        )
+    for mode in ["cancel", "disconnect"]:
+        r = cases[-1]
+        with request(
+            "/v1/completions",
+            {
+                "model": model,
+                "prompt": r["ids"],
+                "max_tokens": 64,
+                "stream": True,
+                "ignore_eos": True,
+            },
+            mode,
+        ) as response:
+            for raw in response:
+                if raw.startswith(b"data: ") and b"[DONE]" not in raw:
+                    event = json.loads(raw[6:])
+                    if event.get("choices") and event["choices"][0].get("text"):
+                        break
+            else:
+                raise AssertionError("no streamed text")
+            if mode == "cancel":
+                with request("/v1/cancel", {}, mode) as cancelled:
+                    assert cancelled.status == 202
+        end = time.monotonic() + 10
+        while mode in b.by_id and time.monotonic() < end:
+            time.sleep(0.01)
+        assert mode not in b.by_id
+        complete(cases[0], "after-" + mode)
+    assert b.stats_pd["local"] == 0 and sorted(b.free) == [0]
+    assert b.stats_pd["remote"] == len(rows) + 2
+    result = {
+        "worker": "live_rngd_subprocess_prefill",
+        "rows": rows,
+        "stats_pd": b.stats_pd,
+        "free": b.free,
+        "cancellation_reuse": True,
+    }
+    Path(a.output).write_text(json.dumps(result, ensure_ascii=False, indent=2) + "\n")
+    print(
+        json.dumps(
+            {"requests": len(rows), "stats_pd": b.stats_pd, "cancellation_reuse": True}
+        ),
+        flush=True,
+    )
+finally:
+    if owner:
+        owner.submit("stop")
+    if thread:
+        thread.join(15)
+    if server:
+        server.shutdown()
+        server.server_close()
+        http.join(5)
+    if worker:
+        worker.close()
+    engine.close()
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl b/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
index bcd190fe..86b05c78 100644
--- a/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
@@ -1 +1,2 @@
 {"id": "819c75b3957c4f0b980915096014c36f", "ts": "2026-10-23T22:30:07+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Callable Gemma refactor:4/4 full K/V/logits bit-exact on RNGD; actual subprocess/shared-memory worker5/5 exact and oversize recovery pass.15CPU tests+5subtests. No live HTTP claim yet.", "next": "Connect live worker to production ATOM HTTP; extend Qwen/KONI adapters and long-prefix/lifecycle conditions.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"model": "Gemma3-1B", "width": "128"}, "artifacts": ["opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-worker.json"], "experiments": ["opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-model-pd234.md"], "labels": ["experiment"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "validation host", "branch": "main", "commit": "351d39d-dirty"}
+{"id": "fdd4a316d5cb433e8b2b56028ac53379", "ts": "2026-10-23T22:52:06+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Real RNGD->ATOM HTTP four models:56/56 normal/reuse outputs exact,64remote/0local prefills; concurrent clients and cancel/disconnect recovery pass. KONI k1 enabled by#236;22adapter tests+5subtests and43scheduler tests. Width2048 queue started.", "next": "Collect live medium-context queue; publish #236 and adapter updates, then controlled native/PD performance and remaining cache/overlap.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"hardware_batch": "1", "mode": "live_rngd_atom_http"}, "artifacts": ["opt_prime/examples/pd/serving_matrix/evidence234/live234-http-summary.json"], "experiments": ["opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http234.md"], "labels": ["experiment"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "RNGD+ATOM host", "branch": "main", "commit": "351d39d-dirty"}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http.json
new file mode 100644
index 00000000..b822177b
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http.json
@@ -0,0 +1,400 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-0",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792762562,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 23,
+          "completion_tokens": 2,
+          "total_tokens": 25
+        }
+      },
+      "latency_s": 0.17397043097298592
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792762562,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42\n",
+            "token_ids": [
+              236812,
+              236778,
+              107,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 24,
+          "completion_tokens": 4,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.08812864299397916
+    },
+    {
+      "case": "korean-0",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792762562,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울특별시",
+            "token_ids": [
+              222340,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 24,
+          "completion_tokens": 2,
+          "total_tokens": 26
+        }
+      },
+      "latency_s": 0.06688711000606418
+    },
+    {
+      "case": "retrieval-0",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792762562,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              28487,
+              1618,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 28,
+          "completion_tokens": 3,
+          "total_tokens": 31
+        }
+      },
+      "latency_s": 0.07794601400382817
+    },
+    {
+      "case": "capital-0",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792762562,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 23,
+          "completion_tokens": 2,
+          "total_tokens": 25
+        }
+      },
+      "latency_s": 0.06611597491428256
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792762563,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42\n",
+            "token_ids": [
+              236812,
+              236778,
+              107,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 24,
+          "completion_tokens": 4,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.08932010200805962
+    },
+    {
+      "case": "korean-0",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792762563,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울특별시",
+            "token_ids": [
+              222340,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 24,
+          "completion_tokens": 2,
+          "total_tokens": 26
+        }
+      },
+      "latency_s": 0.06798817799426615
+    },
+    {
+      "case": "retrieval-0",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792762563,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              28487,
+              1618,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 28,
+          "completion_tokens": 3,
+          "total_tokens": 31
+        }
+      },
+      "latency_s": 0.07829866698011756
+    },
+    {
+      "case": "capital-0",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792762563,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 23,
+          "completion_tokens": 2,
+          "total_tokens": 25
+        }
+      },
+      "latency_s": 0.06778034707531333
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792762563,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42\n",
+            "token_ids": [
+              236812,
+              236778,
+              107,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 24,
+          "completion_tokens": 4,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.15490090299863368
+    },
+    {
+      "case": "korean-0",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792762563,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울특별시",
+            "token_ids": [
+              222340,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 24,
+          "completion_tokens": 2,
+          "total_tokens": 26
+        }
+      },
+      "latency_s": 0.21974424994550645
+    },
+    {
+      "case": "retrieval-0",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792762563,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              28487,
+              1618,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 28,
+          "completion_tokens": 3,
+          "total_tokens": 31
+        }
+      },
+      "latency_s": 0.29593874304555357
+    },
+    {
+      "case": "capital-0",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792762563,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 23,
+          "completion_tokens": 2,
+          "total_tokens": 25
+        }
+      },
+      "latency_s": 0.06744227407034487
+    },
+    {
+      "case": "capital-0",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792762563,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 23,
+          "completion_tokens": 2,
+          "total_tokens": 25
+        }
+      },
+      "latency_s": 0.06791288801468909
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-http-summary.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-http-summary.json
new file mode 100644
index 00000000..78dbb18f
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-http-summary.json
@@ -0,0 +1,47 @@
+{
+  "gemma": {
+    "responses": 14,
+    "remote_prefills": 16,
+    "local_prefills": 0,
+    "cancellation_reuse": true,
+    "input_tokens": [
+      23,
+      24,
+      28
+    ]
+  },
+  "qwen06b": {
+    "responses": 14,
+    "remote_prefills": 16,
+    "local_prefills": 0,
+    "cancellation_reuse": true,
+    "input_tokens": [
+      26,
+      27,
+      31
+    ]
+  },
+  "qwen4b": {
+    "responses": 14,
+    "remote_prefills": 16,
+    "local_prefills": 0,
+    "cancellation_reuse": true,
+    "input_tokens": [
+      26,
+      27,
+      31
+    ]
+  },
+  "koni": {
+    "responses": 14,
+    "remote_prefills": 16,
+    "local_prefills": 0,
+    "cancellation_reuse": true,
+    "input_tokens": [
+      137,
+      138,
+      139,
+      143
+    ]
+  }
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http.json
new file mode 100644
index 00000000..b3a13143
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http.json
@@ -0,0 +1,415 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-0",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792763172,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 138,
+          "completion_tokens": 2,
+          "total_tokens": 140
+        }
+      },
+      "latency_s": 0.4229255300015211
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792763173,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42.",
+            "token_ids": [
+              1114,
+              489,
+              220,
+              914,
+              284,
+              220,
+              2983,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 137,
+          "completion_tokens": 9,
+          "total_tokens": 146
+        }
+      },
+      "latency_s": 0.7874407280469313
+    },
+    {
+      "case": "korean-0",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792763174,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              115978,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 139,
+          "completion_tokens": 2,
+          "total_tokens": 141
+        }
+      },
+      "latency_s": 0.3736360300099477
+    },
+    {
+      "case": "retrieval-0",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792763174,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              18082,
+              877,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 143,
+          "completion_tokens": 3,
+          "total_tokens": 146
+        }
+      },
+      "latency_s": 0.41646945395041257
+    },
+    {
+      "case": "capital-0",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792763174,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 138,
+          "completion_tokens": 2,
+          "total_tokens": 140
+        }
+      },
+      "latency_s": 0.3646502789342776
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792763175,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42.",
+            "token_ids": [
+              1114,
+              489,
+              220,
+              914,
+              284,
+              220,
+              2983,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 137,
+          "completion_tokens": 9,
+          "total_tokens": 146
+        }
+      },
+      "latency_s": 0.8413291039178148
+    },
+    {
+      "case": "korean-0",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792763176,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              115978,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 139,
+          "completion_tokens": 2,
+          "total_tokens": 141
+        }
+      },
+      "latency_s": 0.3697471059858799
+    },
+    {
+      "case": "retrieval-0",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792763176,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              18082,
+              877,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 143,
+          "completion_tokens": 3,
+          "total_tokens": 146
+        }
+      },
+      "latency_s": 0.44573811802547425
+    },
+    {
+      "case": "capital-0",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792763176,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 138,
+          "completion_tokens": 2,
+          "total_tokens": 140
+        }
+      },
+      "latency_s": 0.36879166797734797
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792763177,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42.",
+            "token_ids": [
+              1114,
+              489,
+              220,
+              914,
+              284,
+              220,
+              2983,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 137,
+          "completion_tokens": 9,
+          "total_tokens": 146
+        }
+      },
+      "latency_s": 1.2019858930725604
+    },
+    {
+      "case": "korean-0",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792763178,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              115978,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 139,
+          "completion_tokens": 2,
+          "total_tokens": 141
+        }
+      },
+      "latency_s": 1.5741613419959322
+    },
+    {
+      "case": "retrieval-0",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792763178,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              18082,
+              877,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 143,
+          "completion_tokens": 3,
+          "total_tokens": 146
+        }
+      },
+      "latency_s": 2.007744933012873
+    },
+    {
+      "case": "capital-0",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792763179,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 138,
+          "completion_tokens": 2,
+          "total_tokens": 140
+        }
+      },
+      "latency_s": 0.3682253049919382
+    },
+    {
+      "case": "capital-0",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792763180,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 138,
+          "completion_tokens": 2,
+          "total_tokens": 140
+        }
+      },
+      "latency_s": 0.3731464589945972
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http.json
new file mode 100644
index 00000000..380d2f71
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http.json
@@ -0,0 +1,484 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-0",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792762650,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.3779727650107816
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792762650,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42\n\nReply with only the integer: **42**.",
+            "token_ids": [
+              16,
+              22,
+              488,
+              220,
+              17,
+              20,
+              284,
+              220,
+              19,
+              17,
+              271,
+              20841,
+              448,
+              1172,
+              279,
+              7546,
+              25,
+              3070,
+              19,
+              17,
+              334,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 27,
+          "completion_tokens": 23,
+          "total_tokens": 50
+        }
+      },
+      "latency_s": 0.227519836043939
+    },
+    {
+      "case": "korean-0",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792762650,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 서울입니다.",
+            "token_ids": [
+              66845,
+              23573,
+              125496,
+              124785,
+              20401,
+              134013,
+              16560,
+              136905,
+              78952,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 11,
+          "total_tokens": 42
+        }
+      },
+      "latency_s": 0.14687276305630803
+    },
+    {
+      "case": "retrieval-0",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792762650,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 3,
+          "total_tokens": 34
+        }
+      },
+      "latency_s": 0.09232904796954244
+    },
+    {
+      "case": "capital-0",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792762650,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.08383574697654694
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792762651,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42\n\nReply with only the integer: **42**.",
+            "token_ids": [
+              16,
+              22,
+              488,
+              220,
+              17,
+              20,
+              284,
+              220,
+              19,
+              17,
+              271,
+              20841,
+              448,
+              1172,
+              279,
+              7546,
+              25,
+              3070,
+              19,
+              17,
+              334,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 27,
+          "completion_tokens": 23,
+          "total_tokens": 50
+        }
+      },
+      "latency_s": 0.22187654604204
+    },
+    {
+      "case": "korean-0",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792762651,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 서울입니다.",
+            "token_ids": [
+              66845,
+              23573,
+              125496,
+              124785,
+              20401,
+              134013,
+              16560,
+              136905,
+              78952,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 11,
+          "total_tokens": 42
+        }
+      },
+      "latency_s": 0.1449671690352261
+    },
+    {
+      "case": "retrieval-0",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792762651,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 3,
+          "total_tokens": 34
+        }
+      },
+      "latency_s": 0.09237896895501763
+    },
+    {
+      "case": "capital-0",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792762651,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.08503968501463532
+    },
+    {
+      "case": "korean-0",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792762651,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 서울입니다.",
+            "token_ids": [
+              66845,
+              23573,
+              125496,
+              124785,
+              20401,
+              134013,
+              16560,
+              136905,
+              78952,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 11,
+          "total_tokens": 42
+        }
+      },
+      "latency_s": 0.22497885895427316
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792762651,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42\n\nReply with only the integer: **42**.",
+            "token_ids": [
+              16,
+              22,
+              488,
+              220,
+              17,
+              20,
+              284,
+              220,
+              19,
+              17,
+              271,
+              20841,
+              448,
+              1172,
+              279,
+              7546,
+              25,
+              3070,
+              19,
+              17,
+              334,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 27,
+          "completion_tokens": 23,
+          "total_tokens": 50
+        }
+      },
+      "latency_s": 0.44751804997213185
+    },
+    {
+      "case": "retrieval-0",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792762651,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 3,
+          "total_tokens": 34
+        }
+      },
+      "latency_s": 0.5371290389448404
+    },
+    {
+      "case": "capital-0",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792762652,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.08413173002190888
+    },
+    {
+      "case": "capital-0",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792762652,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.0885854009538889
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http.json
new file mode 100644
index 00000000..0084d1d6
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http.json
@@ -0,0 +1,400 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-0",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792762779,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.6855041530216113
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792762779,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "52",
+            "token_ids": [
+              20,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 27,
+          "completion_tokens": 3,
+          "total_tokens": 30
+        }
+      },
+      "latency_s": 0.19621982797980309
+    },
+    {
+      "case": "korean-0",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792762780,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              26698,
+              126893,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 3,
+          "total_tokens": 34
+        }
+      },
+      "latency_s": 0.20220612501725554
+    },
+    {
+      "case": "retrieval-0",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792762780,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 3,
+          "total_tokens": 34
+        }
+      },
+      "latency_s": 0.21435286104679108
+    },
+    {
+      "case": "capital-0",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792762780,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.17381490301340818
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792762780,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "52",
+            "token_ids": [
+              20,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 27,
+          "completion_tokens": 3,
+          "total_tokens": 30
+        }
+      },
+      "latency_s": 0.20950887200888246
+    },
+    {
+      "case": "korean-0",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792762780,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              26698,
+              126893,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 3,
+          "total_tokens": 34
+        }
+      },
+      "latency_s": 0.21188442199490964
+    },
+    {
+      "case": "retrieval-0",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792762781,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 3,
+          "total_tokens": 34
+        }
+      },
+      "latency_s": 0.21210102399345487
+    },
+    {
+      "case": "capital-0",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792762781,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.18024980393238366
+    },
+    {
+      "case": "arithmetic-0",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792762781,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "52",
+            "token_ids": [
+              20,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 27,
+          "completion_tokens": 3,
+          "total_tokens": 30
+        }
+      },
+      "latency_s": 0.3912503090687096
+    },
+    {
+      "case": "korean-0",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792762781,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              26698,
+              126893,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 3,
+          "total_tokens": 34
+        }
+      },
+      "latency_s": 0.6002984389197081
+    },
+    {
+      "case": "retrieval-0",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792762781,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 31,
+          "completion_tokens": 3,
+          "total_tokens": 34
+        }
+      },
+      "latency_s": 0.8154195169918239
+    },
+    {
+      "case": "capital-0",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792762782,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.1792328970041126
+    },
+    {
+      "case": "capital-0",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792762782,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 26,
+          "completion_tokens": 2,
+          "total_tokens": 28
+        }
+      },
+      "latency_s": 0.21111171692609787
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http234.md b/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http234.md
new file mode 100644
index 00000000..6431c29c
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http234.md
@@ -0,0 +1,44 @@
+# Four-model live RNGD to ATOM HTTP PD
+
+- **Date**: 2026-10-23
+- **Task**: aicomp #234, single-block support #236
+- **Host / device**: owned paired RNGD+ATOM allocation, one of each accelerator; hardware batch1, CPU threads4.
+- **Commit**: adapter branch3855050 plus current HTTP/KONI changes; AtomBatcher1e54786 for Gemma/Qwen, #236 working tree based69bca73 for KONI; HTTP SIGPIPE fix9e442cd loaded process-locally.
+- **SDKs**: producer furiosa-torch2026.4.0 / torch2.10.0+cpu / transformers5.19.0; decoder optimum-rbln0.11.0.post1 / rebel-compiler0.11.0 / torch2.11.0+cpu / transformers5.8.1.
+
+## Question
+
+Can each validated producer handle real HTTP-triggered prefill through SubprocessPrefillWorker, shared-memory KV import and the production ATOM scheduler/owner/server, including queued clients and cancellation/reuse?
+
+## Setup
+
+No saved KV is loaded by check_live_http.py. It reads only input/EOS fixtures and native output baselines. Each request performs live RNGD prefill, returns KV through the existing shared-memory protocol, and decodes on ATOM. Four semantic questions, two sequential passes, one four-client concurrent pass, two streaming cancellation/disconnect requests and a successful reuse after each.14normal/reuse output checks and16remote prefills per model. Concurrency queues on a B1 lane; this is not hardware batch4.
+
+Gemma/Qwen artifacts use context9216/k3; KONI uses its existing static context9216/k1 artifact with the #236 scheduler fix. Prefix and host caches disabled. Corrected Gemma global-causal artifact from #225; pinned revisions are unchanged from #227. Token lists and EOS finish reason must equal native ATOM baselines. This criterion does not establish answer quality (known shared wrong answers remain in #227/#233).
+
+## Command
+
+```bash
+bash run-live-http234.sh --batching atom_batching231.py --http-module llmd214.py --cache-dir atom-artifacts-global225 --fixtures semantic227-gemma-fixtures.json --cases capital-0 arithmetic-0 korean-0 retrieval-0 --worker-python "$WORKER_PYTHON" --engine-factory gemma_prefill:Prefill --width 128 --baseline semantic227-gemma-atom.jsonl --output live234-gemma-http.json
+```
+
+The isolated launcher runs check_live_http.py under the decoder interpreter/bootstrap. For Qwen use --model Qwen/Qwen3-0.6B orQwen/Qwen3-4B, --cache-dir atom-artifacts, --engine-factory prefill:Prefill and matching qwen06b/qwen4b fixtures/baselines/output. For KONI use --model KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024, --batching atom_batching236.py, --cache-dir atom-artifacts --kv-blocks 1, --engine-factory koni_prefill:Prefill --width 512, semantic227-koni-fixtures-eos.json and semantic227-koni-atom.jsonl. No installed SDK/core files changed.
+
+## Results
+
+| Model | Prompt tokens | Compiled width | Exact normal/reuse responses | Remote/local prefills | Cancel/disconnect reuse |
+|---|---:|---:|---:|---:|---|
+| Gemma3-1B | 23–28 | 128 | 14/14 | 16/0 | pass |
+| Qwen3-0.6B | 26–31 | 128 | 14/14 | 16/0 | pass |
+| Qwen3-4B | 26–31 | 128 | 14/14 | 16/0 | pass |
+| KONI-8B | 137–143 | 512 | 14/14 | 16/0 | pass |
+
+Total56exact responses and64actual RNGD prefills, zero local fallbacks. Each final free-lane list is[0]. Client latency observations are included in JSON but this is not a matched native-vs-PD benchmark or throughput claim. The previous vendor SIGPIPE backtraces do not appear with the separately tested #214 module; the ordinary tokenizer cleanup warning remains on KONI.
+
+Supporting CPU tests: #234 adapter/geometry22tests plus5subtests; #236 scheduler43tests, one existing unfused attention warning. Before #236, single-block PD fails at construction (1failed/3passed); cache/multi-lane rejection tests remain green. KONI raw config omits head_dim: worker infers128 only when absent and preserves explicit Qwen/Gemma head dimensions.
+
+## Conclusion
+
+All four models now have short-input real RNGD->ATOM HTTP evidence, replacing replay-only coverage for this condition. Remaining: live medium/long prefixes, prefill-time cancellation/fault recovery, prefix/host-cache restoration, hardware batch coverage, true compute-transfer overlap and controlled serving performance comparison. A width2048 queue for the four *-100 fixtures has started in the owned paired allocation; results are pending. Overall goal remains active.
+
+Raw evidence: analysis/serving-matrix-201/live234-{gemma,qwen06b,qwen4b,koni}-http.json/.log and live234-http-summary.json. Pending queue: live234-mid-queue.log.
diff --git a/opt_prime/examples/pd/serving_matrix/koni_prefill.py b/opt_prime/examples/pd/serving_matrix/koni_prefill.py
new file mode 100644
index 00000000..ece376aa
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/koni_prefill.py
@@ -0,0 +1,45 @@
+"""Callable adapter for the validated partitioned Llama3.1/KONI prefill graph."""
+
+from types import SimpleNamespace
+
+import torch
+from koni_export import export_groups, validate
+
+
+class Prefill:
+    def __init__(self, config, weights, width, chip=0):
+        validate(config)
+        if type(width) is not int or not 0 < width <= 8192 or width % 512:
+            raise ValueError("partitioned KONI requires multiples of512 through8192")
+        from koni_partitioned import PartitionedPrefill
+        from rngd_pd.engines.fsdk import sync
+        from rngd_pd.engines.model import ModelConfig
+
+        arrays = {k: v.detach().cpu().float().numpy() for k, v in weights.items()}
+        self.engine = PartitionedPrefill(
+            ModelConfig(SimpleNamespace(**config)),
+            arrays,
+            C=width,
+            S=width,
+            group=1,
+            chip=chip,
+            cache_dir=None,
+            rms_tile=512,
+        )
+        self.width, self.chip, self.sync, self.config = width, chip, sync, config
+        self.norm = torch.from_numpy(arrays["model.norm.weight"])
+        self.head = torch.from_numpy(arrays["lm_head.weight"])
+        sync(chip)
+
+    @torch.inference_mode()
+    def prefill(self, tokens):
+        ids = list(tokens)
+        if not 0 < len(ids) <= self.width or any(
+            type(t) is not int or not 0 <= t < self.config["vocab_size"] for t in ids
+        ):
+            raise ValueError("invalid tokens or prompt exceeds compiled width")
+        stage = self.engine.begin_device_prefill(ids)
+        while not stage.advance():
+            pass
+        self.sync(self.chip)
+        return export_groups(stage, self.norm, self.head, self.config["rms_norm_eps"])
diff --git a/opt_prime/examples/pd/serving_matrix/long_width_variant.py b/opt_prime/examples/pd/serving_matrix/long_width_variant.py
index 91548f52..ac26a862 100644
--- a/opt_prime/examples/pd/serving_matrix/long_width_variant.py
+++ b/opt_prime/examples/pd/serving_matrix/long_width_variant.py
@@ -14,3 +14,8 @@ def load_variant():
     namespace = {"__name__": "prefill_width8192"}
     exec(compile(source, "prefill_width8192", "exec"), namespace)  # noqa: S102 - trusted local experimental source
     return namespace["Prefill"]
+
+
+def make_prefill(config, weights, width, chip=0):
+    """Explicit experimental worker factory using the already validated long-width variant."""
+    return load_variant()(config, weights, width=width, chip=chip)
diff --git a/opt_prime/examples/pd/serving_matrix/test_koni_prefill.py b/opt_prime/examples/pd/serving_matrix/test_koni_prefill.py
new file mode 100644
index 00000000..6ffc94bc
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/test_koni_prefill.py
@@ -0,0 +1,23 @@
+import pytest
+from koni_prefill import Prefill
+
+
+def test_rejects_wrong_geometry_before_loading_sdk():
+    with pytest.raises(ValueError, match="Llama"):
+        Prefill({}, {}, width=512)
+
+
+@pytest.mark.parametrize("width", [0, 128, 513])
+def test_rejects_unsupported_width_before_loading_sdk(width):
+    config = {
+        "model_type": "llama",
+        "hidden_act": "silu",
+        "hidden_size": 4096,
+        "intermediate_size": 14336,
+        "num_attention_heads": 32,
+        "num_key_value_heads": 8,
+        "num_hidden_layers": 32,
+        "rope_scaling": {"rope_type": "llama3"},
+    }
+    with pytest.raises(ValueError, match="512"):
+        Prefill(config, {}, width=width)

From 75ce140026d37a96ca9f74c157682e2a4684c569 Mon Sep 17 00:00:00 2001
From: gspark <gspark@etri.re.kr>
Date: Fri, 23 Oct 2026 23:07:02 +0900
Subject: [PATCH 5/6] test: verify medium-context live HTTP PD across four
 models

---
 .../examples/pd/serving_matrix/COVERAGE.md    |  20 +-
 .../examples/pd/serving_matrix/README.md      |   2 +-
 .../examples/pd/serving_matrix/STATUS.md      |   4 +-
 .../serving_matrix/evidence234/events.jsonl   |   1 +
 .../evidence234/live234-gemma-http-mid.json   | 400 ++++++++++++++
 .../evidence234/live234-http-mid-summary.json |  47 ++
 .../evidence234/live234-koni-http-mid.json    | 445 ++++++++++++++++
 .../evidence234/live234-qwen06b-http-mid.json | 490 ++++++++++++++++++
 .../evidence234/live234-qwen4b-http-mid.json  | 400 ++++++++++++++
 .../2026-10-23-live-http-mid234.md            |  42 ++
 10 files changed, 1844 insertions(+), 7 deletions(-)
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http-mid.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-http-mid-summary.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http-mid.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http-mid.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http-mid.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-mid234.md

diff --git a/opt_prime/examples/pd/serving_matrix/COVERAGE.md b/opt_prime/examples/pd/serving_matrix/COVERAGE.md
index 9983bef9..6e8dc403 100644
--- a/opt_prime/examples/pd/serving_matrix/COVERAGE.md
+++ b/opt_prime/examples/pd/serving_matrix/COVERAGE.md
@@ -13,9 +13,9 @@ All direct runs use fixed output lengths ignoring EOS. Repeats match within each
 
 ## Outstanding acceptance work
 
-- Standalone GPU HTTP covered by PR213 and ATOM HTTP by PR216: all four lifecycle checks pass through8K. Cache-off concurrent/serial equality48/48 on each backend. RNGD/heterogeneous HTTP remains incomplete.
+- Standalone GPU HTTP covered by PR213 and ATOM HTTP by PR216: all four lifecycle checks pass through8K. Cache-off concurrent/serial equality48/48 on each backend. Real RNGD->ATOM short-input HTTP now passes56/56 output checks (#235/#237); medium/long live and broader conditions remain incomplete.
 - Qwen ATOM device/host-cache cases pass12/12 concurrent equality each (PR216). GPU BF16 cache has numerical ties; FP32 controls match48/48 (PR213). Gemma sliding-state restoration is unimplemented, KONI k1 has no spare prefix-cache blocks.
-- Production integration of validated research KONI/Gemma long-prefix producers; genuine cache-aware producer chunking remains open.
+- Short live producer/scheduler integration is verified for all four models. Live long prefixes and genuine cache-aware producer chunking remain open.
 - Layer/logit diagnosis and broader semantic fixtures for cross-backend output differences; matching a repeated capital question is not a quality evaluation.
 - Actual producer/consumer overlap and end-to-end serving benefit, distinct from faster CPU staging.
 
@@ -23,12 +23,22 @@ Evidence: PR203 matrix results, PR206 staging optimization, PR209 Gemma probe, e
 
 ## EOS-aware native fixtures (#227)
 
-192 GPU/ATOM runs, all EOS-ended;96 repeat pairs exact. Cross-backend exact cases: Qwen0.6B12/12,Qwen4B12/12,KONI12/12,Gemma11/12. Four short questions with0/100/400 repeated context notes give23–6428 input tokens. Shared wrong factual answers are retained, so this is not a general quality pass. Research EOS policy includes both generation-config and tokenizer turn-end tokens; original KONI policy failure preserved. RNGD and heterogeneous HTTP coverage of these exact semantic inputs remains missing.
+192 GPU/ATOM runs, all EOS-ended;96 repeat pairs exact. Cross-backend exact cases: Qwen0.6B12/12,Qwen4B12/12,KONI12/12,Gemma11/12. Four short questions with0/100/400 repeated context notes give23–6428 input tokens. Shared wrong factual answers are retained, so this is not a general quality pass. Research EOS policy includes both generation-config and tokenizer turn-end tokens; original KONI policy failure preserved. Subsequent RNGD and live HTTP evidence is listed below; matching outputs still does not certify answer quality.
 
 ## RNGD semantic handoff follow-up (#227)
 
-All four models now execute the same four short chat cases through RNGD prefill and GPU/ATOM continuation.32/32 decoder comparisons match native tokens through EOS,48prefill runs finite/repeat-exact. Exact saved IDs and true KV lengths verified, excluding right padding; per-model52/56/72/64slots read back exactly. Qwen/KONI original/portable importer logits agree; Gemma raw native-state roundtrip also passes. Inputs23–143 tokens (compiled width128 or512), so long semantic fixtures remain pending. This is host-staged full-prefix handoff, not actual compute/transfer overlap or production PD serving. See semantic-rngd227 record and evidence227.
+All four models now execute the same four short chat cases through RNGD prefill and GPU/ATOM continuation.32/32 decoder comparisons match native tokens through EOS,48prefill runs finite/repeat-exact. Exact saved IDs and true KV lengths verified, excluding right padding; per-model52/56/72/64slots read back exactly. Qwen/KONI original/portable importer logits agree; Gemma raw native-state roundtrip also passes. Inputs23–143 tokens (compiled width128 or512), with subsequent long-context results below. This is host-staged full-prefix handoff, not actual compute/transfer overlap or production PD serving. See semantic-rngd227 record and evidence227.
 
 ## Medium semantic handoff (#227)
 
-Inputs1526–1643tokens in width2048 buckets now pass all32 model/case/decoder EOS comparisons (four models,four cases,GPU/ATOM).48prefill runs finite/repeat-exact; true IDs/KV lengths and52/56/72/64slot readbacks verified. Native GPU/ATOM sequences agree for these cases, including shared wrong answers documented earlier. Approximately6K semantic inputs and production heterogeneous serving/overlap remain open. See semantic-mid227 experiment and evidence227.
+Inputs1526–1643tokens in width2048 buckets now pass all32 model/case/decoder EOS comparisons (four models,four cases,GPU/ATOM).48prefill runs finite/repeat-exact; true IDs/KV lengths and52/56/72/64slot readbacks verified. Native GPU/ATOM sequences agree for these cases, including shared wrong answers documented earlier. Approximately6K semantic results and short live HTTP are now available below; true overlap and broader serving conditions remain open. See semantic-mid227 experiment and evidence227.
+
+## Long semantic handoff (#227/#233)
+
+Four models,48producer runs,31/32 EOS comparisons at6026–6428input tokens. Both Qwens and KONI match4/4cases on each decoder; Gemma matches3/4GPU and4/4ATOM. All KV readbacks pass. The retained Gemma arithmetic mismatch is30(nativeGPU BF16/FP32) vs25(RNGD/ATOM), both wrong. Same layer equations on GPU choose a different first token from RNGD; operator cause unresolved in#233. See semantic-long227 and evidence227.
+
+## Real RNGD to ATOM HTTP (#234/#236)
+
+All four short-input paths pass14normal/concurrent/reuse responses each:56total,64remote prefills including cancellation requests,0local fallbacks. Requests trigger actual RNGD subprocess computation; no KV replay. Each model passes explicit cancellation, socket disconnection and subsequent reuse. Gemma/Qwen usek3; static B1 KONI usesk1 with#236. SIGPIPE-safe HTTP is#214/#215. These are B1 hardware requests with queued client concurrency, not hardware batch4. Native parity includes shared wrong answers.
+
+Width2048 live checks now pass for all four models:56responses and64remote prefills. Short+medium total112responses/128prefills, with no local fallbacks. Width8192 live queue is pending. Full live6K, pending-prefill hardware cancellation/worker-failure recovery, GPU live PD, hybrid prefix/host restore, applicable hardware batch>1, genuine producer/transfer overlap and matched native-vs-PD latency/throughput comparisons remain unverified. Overall goal remains active.
diff --git a/opt_prime/examples/pd/serving_matrix/README.md b/opt_prime/examples/pd/serving_matrix/README.md
index 53695d9f..be922bfe 100644
--- a/opt_prime/examples/pd/serving_matrix/README.md
+++ b/opt_prime/examples/pd/serving_matrix/README.md
@@ -177,4 +177,4 @@ Short-input producer/full-logit parity and real shared-memory requests are verif
 
 `check_live_http.py` runs actual RNGD prefill and ATOM HTTP decode from token fixtures (it does not load saved KV). Pass `--batching` with the reviewed ATOM scheduler, `--http-module` with the #214 SIGPIPE-fixed server, the model/artifact cache, `--worker-python`, factory/width and fixture/baseline JSON. It checks two serial passes, four queued clients, cancel/disconnect and reuse. Factories: `prefill:Prefill` for short Qwen3, `long_width_variant:make_prefill` for the explicit experimental long Qwen variant, `gemma_prefill:Prefill`, and `koni_prefill:Prefill` for partitioned KONI widths divisible by512. KONI uses `--kv-blocks 1` with #236; other measured artifacts use3. Explicit head_dim is preserved; Llama's missing field is inferred from hidden size/attention heads.
 
-All four short-input live HTTP paths pass56normal/reuse response checks and64remote prefills including cancellations. Native-output parity does not imply answer quality. Width2048 live validation is pending; hardware batch remains1 and no matched performance benefit is established. See live-http234 record.
+All four short-input live HTTP paths pass56normal/reuse response checks and64remote prefills including cancellations. Native-output parity does not imply answer quality. Width2048 live validation also passes all four models; width8192 live queue is pending; hardware batch remains1 and no matched performance benefit is established. See live-http234 record.
diff --git a/opt_prime/examples/pd/serving_matrix/STATUS.md b/opt_prime/examples/pd/serving_matrix/STATUS.md
index d1d148d4..2a6d7004 100644
--- a/opt_prime/examples/pd/serving_matrix/STATUS.md
+++ b/opt_prime/examples/pd/serving_matrix/STATUS.md
@@ -62,4 +62,6 @@ Reusable Gemma producer matches original RNGD K/V/full logits4/4 at width128. Re
 
 ## Four-model actual HTTP PD (#234/#236)
 
-All four models:56/56 normal/concurrent/reuse responses equal native ATOM,64RNGD prefills and0local fallbacks. Cancellation/disconnect reuse passes. B1, cache-off, short prompts; KONI uses existingk1 artifact with#236. HTTP uses#214 SIGPIPE fix. No live-transfer-overlap or throughput gain claim. Width2048 live queue pending; long/cache/hardware-batch/performance scope remains.
+All four models:56/56 normal/concurrent/reuse responses equal native ATOM,64RNGD prefills and0local fallbacks. Cancellation/disconnect reuse passes. B1, cache-off, short prompts; KONI uses existingk1 artifact with#236. HTTP uses#214 SIGPIPE fix. No live-transfer-overlap or throughput gain claim. Width2048 live queue passes all four models; width8192 live queue pending; long/cache/hardware-batch/performance scope remains.
+
+Medium live HTTP:1526–1643tokens,56/56 exact outputs and64remote/0local; short+medium112responses/128remote. Same cancellation/disconnect reuse checks pass. Long live queue active; full goal remains open. See live-http-mid234.
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl b/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
index 86b05c78..9ab9df1b 100644
--- a/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
@@ -1,2 +1,3 @@
 {"id": "819c75b3957c4f0b980915096014c36f", "ts": "2026-10-23T22:30:07+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Callable Gemma refactor:4/4 full K/V/logits bit-exact on RNGD; actual subprocess/shared-memory worker5/5 exact and oversize recovery pass.15CPU tests+5subtests. No live HTTP claim yet.", "next": "Connect live worker to production ATOM HTTP; extend Qwen/KONI adapters and long-prefix/lifecycle conditions.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"model": "Gemma3-1B", "width": "128"}, "artifacts": ["opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-worker.json"], "experiments": ["opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-model-pd234.md"], "labels": ["experiment"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "validation host", "branch": "main", "commit": "351d39d-dirty"}
 {"id": "fdd4a316d5cb433e8b2b56028ac53379", "ts": "2026-10-23T22:52:06+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Real RNGD->ATOM HTTP four models:56/56 normal/reuse outputs exact,64remote/0local prefills; concurrent clients and cancel/disconnect recovery pass. KONI k1 enabled by#236;22adapter tests+5subtests and43scheduler tests. Width2048 queue started.", "next": "Collect live medium-context queue; publish #236 and adapter updates, then controlled native/PD performance and remaining cache/overlap.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"hardware_batch": "1", "mode": "live_rngd_atom_http"}, "artifacts": ["opt_prime/examples/pd/serving_matrix/evidence234/live234-http-summary.json"], "experiments": ["opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http234.md"], "labels": ["experiment"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "RNGD+ATOM host", "branch": "main", "commit": "351d39d-dirty"}
+{"id": "50c06e38ffe04b859420a115b72f5cc6", "ts": "2026-10-23T23:05:26+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Medium1526-1643token live RNGD->ATOM HTTP four models56/56 responses exact,64remote/0local and cancel/disconnect reuse pass. Short+medium112responses/128remote. PR237df02223 single-block support; long8192-width/*-400 queue started.", "next": "Collect long live queue; then pending-prefill/worker-failure recovery, matched native/PD performance, cache/overlap and#233 numerics.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"width": "2048", "hardware_batch": "1"}, "artifacts": ["opt_prime/examples/pd/serving_matrix/evidence234/live234-http-mid-summary.json"], "experiments": ["opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-mid234.md"], "labels": ["experiment"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "RNGD+ATOM host", "branch": "main", "commit": "351d39d-dirty"}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http-mid.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http-mid.json
new file mode 100644
index 00000000..4cb2342d
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http-mid.json
@@ -0,0 +1,400 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-100",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792763625,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1623,
+          "completion_tokens": 2,
+          "total_tokens": 1625
+        }
+      },
+      "latency_s": 2.158342153998092
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792763627,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "40\n",
+            "token_ids": [
+              236812,
+              236771,
+              107,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1624,
+          "completion_tokens": 4,
+          "total_tokens": 1628
+        }
+      },
+      "latency_s": 2.0099443329963833
+    },
+    {
+      "case": "korean-100",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792763629,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              98665,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1624,
+          "completion_tokens": 2,
+          "total_tokens": 1626
+        }
+      },
+      "latency_s": 1.9934004180831835
+    },
+    {
+      "case": "retrieval-100",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792763631,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              28487,
+              1618,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1628,
+          "completion_tokens": 3,
+          "total_tokens": 1631
+        }
+      },
+      "latency_s": 2.0018371921032667
+    },
+    {
+      "case": "capital-100",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792763633,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1623,
+          "completion_tokens": 2,
+          "total_tokens": 1625
+        }
+      },
+      "latency_s": 1.9967707360628992
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792763635,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "40\n",
+            "token_ids": [
+              236812,
+              236771,
+              107,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1624,
+          "completion_tokens": 4,
+          "total_tokens": 1628
+        }
+      },
+      "latency_s": 2.0195871680043638
+    },
+    {
+      "case": "korean-100",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792763637,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              98665,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1624,
+          "completion_tokens": 2,
+          "total_tokens": 1626
+        }
+      },
+      "latency_s": 1.9965247600339353
+    },
+    {
+      "case": "retrieval-100",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792763639,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              28487,
+              1618,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1628,
+          "completion_tokens": 3,
+          "total_tokens": 1631
+        }
+      },
+      "latency_s": 2.000015356927179
+    },
+    {
+      "case": "capital-100",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792763641,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1623,
+          "completion_tokens": 2,
+          "total_tokens": 1625
+        }
+      },
+      "latency_s": 2.00175220600795
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792763643,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "40\n",
+            "token_ids": [
+              236812,
+              236771,
+              107,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1624,
+          "completion_tokens": 4,
+          "total_tokens": 1628
+        }
+      },
+      "latency_s": 4.0120902369963005
+    },
+    {
+      "case": "korean-100",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792763645,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              98665,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1624,
+          "completion_tokens": 2,
+          "total_tokens": 1626
+        }
+      },
+      "latency_s": 5.998096699011512
+    },
+    {
+      "case": "retrieval-100",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792763647,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              28487,
+              1618,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1628,
+          "completion_tokens": 3,
+          "total_tokens": 1631
+        }
+      },
+      "latency_s": 7.989423095015809
+    },
+    {
+      "case": "capital-100",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792763651,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1623,
+          "completion_tokens": 2,
+          "total_tokens": 1625
+        }
+      },
+      "latency_s": 1.9823684650473297
+    },
+    {
+      "case": "capital-100",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792763655,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1623,
+          "completion_tokens": 2,
+          "total_tokens": 1625
+        }
+      },
+      "latency_s": 1.9897793129784986
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-http-mid-summary.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-http-mid-summary.json
new file mode 100644
index 00000000..678070d9
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-http-mid-summary.json
@@ -0,0 +1,47 @@
+{
+  "gemma": {
+    "responses": 14,
+    "remote_prefills": 16,
+    "local_prefills": 0,
+    "cancellation_reuse": true,
+    "input_tokens": [
+      1623,
+      1624,
+      1628
+    ]
+  },
+  "qwen06b": {
+    "responses": 14,
+    "remote_prefills": 16,
+    "local_prefills": 0,
+    "cancellation_reuse": true,
+    "input_tokens": [
+      1526,
+      1527,
+      1531
+    ]
+  },
+  "qwen4b": {
+    "responses": 14,
+    "remote_prefills": 16,
+    "local_prefills": 0,
+    "cancellation_reuse": true,
+    "input_tokens": [
+      1526,
+      1527,
+      1531
+    ]
+  },
+  "koni": {
+    "responses": 14,
+    "remote_prefills": 16,
+    "local_prefills": 0,
+    "cancellation_reuse": true,
+    "input_tokens": [
+      1637,
+      1638,
+      1639,
+      1643
+    ]
+  }
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http-mid.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http-mid.json
new file mode 100644
index 00000000..b0254943
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http-mid.json
@@ -0,0 +1,445 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-100",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792764074,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is Paris.",
+            "token_ids": [
+              791,
+              6864,
+              315,
+              9822,
+              374,
+              12366,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1638,
+          "completion_tokens": 8,
+          "total_tokens": 1646
+        }
+      },
+      "latency_s": 2.5393513729795814
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792764076,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              2983,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1637,
+          "completion_tokens": 2,
+          "total_tokens": 1639
+        }
+      },
+      "latency_s": 1.323795892065391
+    },
+    {
+      "case": "korean-100",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792764077,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 서울입니다.",
+            "token_ids": [
+              113760,
+              108709,
+              21028,
+              116992,
+              16969,
+              106010,
+              80052,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1639,
+          "completion_tokens": 9,
+          "total_tokens": 1648
+        }
+      },
+      "latency_s": 1.7511723199859262
+    },
+    {
+      "case": "retrieval-100",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792764079,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              18082,
+              877,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1643,
+          "completion_tokens": 3,
+          "total_tokens": 1646
+        }
+      },
+      "latency_s": 1.280155176995322
+    },
+    {
+      "case": "capital-100",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792764080,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is Paris.",
+            "token_ids": [
+              791,
+              6864,
+              315,
+              9822,
+              374,
+              12366,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1638,
+          "completion_tokens": 8,
+          "total_tokens": 1646
+        }
+      },
+      "latency_s": 1.5194910150021315
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792764081,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              2983,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1637,
+          "completion_tokens": 2,
+          "total_tokens": 1639
+        }
+      },
+      "latency_s": 1.0988063589902595
+    },
+    {
+      "case": "korean-100",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792764083,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 서울입니다.",
+            "token_ids": [
+              113760,
+              108709,
+              21028,
+              116992,
+              16969,
+              106010,
+              80052,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1639,
+          "completion_tokens": 9,
+          "total_tokens": 1648
+        }
+      },
+      "latency_s": 1.4861711430130526
+    },
+    {
+      "case": "retrieval-100",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792764084,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              18082,
+              877,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1643,
+          "completion_tokens": 3,
+          "total_tokens": 1646
+        }
+      },
+      "latency_s": 1.0849265129072592
+    },
+    {
+      "case": "capital-100",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792764085,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is Paris.",
+            "token_ids": [
+              791,
+              6864,
+              315,
+              9822,
+              374,
+              12366,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1638,
+          "completion_tokens": 8,
+          "total_tokens": 1646
+        }
+      },
+      "latency_s": 1.425872788997367
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792764086,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              2983,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1637,
+          "completion_tokens": 2,
+          "total_tokens": 1639
+        }
+      },
+      "latency_s": 2.4230585580226034
+    },
+    {
+      "case": "korean-100",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792764088,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 서울입니다.",
+            "token_ids": [
+              113760,
+              108709,
+              21028,
+              116992,
+              16969,
+              106010,
+              80052,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1639,
+          "completion_tokens": 9,
+          "total_tokens": 1648
+        }
+      },
+      "latency_s": 3.934952615061775
+    },
+    {
+      "case": "retrieval-100",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792764089,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              18082,
+              877,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1643,
+          "completion_tokens": 3,
+          "total_tokens": 1646
+        }
+      },
+      "latency_s": 5.026510157971643
+    },
+    {
+      "case": "capital-100",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792764091,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is Paris.",
+            "token_ids": [
+              791,
+              6864,
+              315,
+              9822,
+              374,
+              12366,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1638,
+          "completion_tokens": 8,
+          "total_tokens": 1646
+        }
+      },
+      "latency_s": 1.4203430189518258
+    },
+    {
+      "case": "capital-100",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792764094,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "The capital of France is Paris.",
+            "token_ids": [
+              791,
+              6864,
+              315,
+              9822,
+              374,
+              12366,
+              13,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1638,
+          "completion_tokens": 8,
+          "total_tokens": 1646
+        }
+      },
+      "latency_s": 1.435384924057871
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http-mid.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http-mid.json
new file mode 100644
index 00000000..6a5b9049
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http-mid.json
@@ -0,0 +1,490 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-100",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792763718,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 0.9969441100256518
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792763718,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42\n\nReply with only the integer: 42.",
+            "token_ids": [
+              16,
+              22,
+              488,
+              220,
+              17,
+              20,
+              284,
+              220,
+              19,
+              17,
+              271,
+              20841,
+              448,
+              1172,
+              279,
+              7546,
+              25,
+              220,
+              19,
+              17,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1527,
+          "completion_tokens": 22,
+          "total_tokens": 1549
+        }
+      },
+      "latency_s": 0.6752042719163001
+    },
+    {
+      "case": "korean-100",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792763719,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 **인천**입니다.",
+            "token_ids": [
+              66845,
+              23573,
+              125496,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 14,
+          "total_tokens": 1545
+        }
+      },
+      "latency_s": 0.6178859820356593
+    },
+    {
+      "case": "retrieval-100",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792763719,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 3,
+          "total_tokens": 1534
+        }
+      },
+      "latency_s": 0.5457541500218213
+    },
+    {
+      "case": "capital-100",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792763720,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 0.5465415481012315
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792763721,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42\n\nReply with only the integer: 42.",
+            "token_ids": [
+              16,
+              22,
+              488,
+              220,
+              17,
+              20,
+              284,
+              220,
+              19,
+              17,
+              271,
+              20841,
+              448,
+              1172,
+              279,
+              7546,
+              25,
+              220,
+              19,
+              17,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1527,
+          "completion_tokens": 22,
+          "total_tokens": 1549
+        }
+      },
+      "latency_s": 0.6642461880110204
+    },
+    {
+      "case": "korean-100",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792763721,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 **인천**입니다.",
+            "token_ids": [
+              66845,
+              23573,
+              125496,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 14,
+          "total_tokens": 1545
+        }
+      },
+      "latency_s": 0.6021285310853273
+    },
+    {
+      "case": "retrieval-100",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792763722,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 3,
+          "total_tokens": 1534
+        }
+      },
+      "latency_s": 0.5523455439833924
+    },
+    {
+      "case": "capital-100",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792763722,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 0.5414341390132904
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792763723,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42\n\nReply with only the integer: 42.",
+            "token_ids": [
+              16,
+              22,
+              488,
+              220,
+              17,
+              20,
+              284,
+              220,
+              19,
+              17,
+              271,
+              20841,
+              448,
+              1172,
+              279,
+              7546,
+              25,
+              220,
+              19,
+              17,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1527,
+          "completion_tokens": 22,
+          "total_tokens": 1549
+        }
+      },
+      "latency_s": 1.227440306916833
+    },
+    {
+      "case": "korean-100",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792763724,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 **인천**입니다.",
+            "token_ids": [
+              66845,
+              23573,
+              125496,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 14,
+          "total_tokens": 1545
+        }
+      },
+      "latency_s": 1.8558082920499146
+    },
+    {
+      "case": "retrieval-100",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792763724,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 3,
+          "total_tokens": 1534
+        }
+      },
+      "latency_s": 2.406756103038788
+    },
+    {
+      "case": "capital-100",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792763725,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 0.6081401809351519
+    },
+    {
+      "case": "capital-100",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792763727,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 0.550125173991546
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http-mid.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http-mid.json
new file mode 100644
index 00000000..0e13cc7e
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http-mid.json
@@ -0,0 +1,400 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-100",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792763846,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 1.5593066349392757
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792763847,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              19,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1527,
+          "completion_tokens": 3,
+          "total_tokens": 1530
+        }
+      },
+      "latency_s": 0.9446719089755788
+    },
+    {
+      "case": "korean-100",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792763848,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              26698,
+              126893,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 3,
+          "total_tokens": 1534
+        }
+      },
+      "latency_s": 1.038334756041877
+    },
+    {
+      "case": "retrieval-100",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792763849,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 3,
+          "total_tokens": 1534
+        }
+      },
+      "latency_s": 1.0008713440038264
+    },
+    {
+      "case": "capital-100",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792763850,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 1.0365961409406736
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792763851,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              19,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1527,
+          "completion_tokens": 3,
+          "total_tokens": 1530
+        }
+      },
+      "latency_s": 0.9256328369956464
+    },
+    {
+      "case": "korean-100",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792763852,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              26698,
+              126893,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 3,
+          "total_tokens": 1534
+        }
+      },
+      "latency_s": 0.9329322580015287
+    },
+    {
+      "case": "retrieval-100",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792763853,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 3,
+          "total_tokens": 1534
+        }
+      },
+      "latency_s": 0.9429352049482986
+    },
+    {
+      "case": "capital-100",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792763854,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 0.9035777349490672
+    },
+    {
+      "case": "arithmetic-100",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792763855,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              19,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1527,
+          "completion_tokens": 3,
+          "total_tokens": 1530
+        }
+      },
+      "latency_s": 1.836488423985429
+    },
+    {
+      "case": "korean-100",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792763856,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              26698,
+              126893,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 3,
+          "total_tokens": 1534
+        }
+      },
+      "latency_s": 2.7430692090420052
+    },
+    {
+      "case": "retrieval-100",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792763857,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1531,
+          "completion_tokens": 3,
+          "total_tokens": 1534
+        }
+      },
+      "latency_s": 3.6805838339496404
+    },
+    {
+      "case": "capital-100",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792763858,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 0.8944114779587835
+    },
+    {
+      "case": "capital-100",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792763860,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 1526,
+          "completion_tokens": 2,
+          "total_tokens": 1528
+        }
+      },
+      "latency_s": 0.9089197190478444
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-mid234.md b/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-mid234.md
new file mode 100644
index 00000000..703ea7e8
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-mid234.md
@@ -0,0 +1,42 @@
+# Four-model medium-input live HTTP PD
+
+- **Date**: 2026-10-23
+- **Task**: aicomp #234, #236
+- **Host / device**: owned paired RNGD+ATOM allocation; one of each, hardwareB1, context9216. SDKs/model revisions unchanged from live-http234.
+- **Commit**: live adapter74d0d3b; core1e54786 for Gemma/Qwen, df02223 behavior for KONI; HTTP9e442cd.
+
+## Question
+
+Does live producer->shared memory->ATOM HTTP preserve native tokens and lifecycle behavior with longer prompts?
+
+## Setup
+
+Same check_live_http.py as the short matrix, replacing four *-0 cases by *-100 and compiled width by2048. Four questions, two serial passes, four queued clients, cancellation/disconnection and reuse. No saved KV loaded. Prefix/host cache off. Existing Gemma/Qwen k3 and KONI k1 artifacts reused. Exact output/EOS comparisons use native ATOM baseline records. Shared wrong answers remain known; parity is not quality.
+
+## Command
+
+```bash
+bash run-live-http234.sh --batching atom_batching231.py --http-module llmd214.py --cache-dir atom-artifacts-global225 --kv-blocks 3 --model unsloth/gemma-3-1b-it --fixtures semantic227-gemma-fixtures.json --cases capital-100 arithmetic-100 korean-100 retrieval-100 --worker-python "$WORKER_PYTHON" --engine-factory gemma_prefill:Prefill --width 2048 --baseline semantic227-gemma-atom.jsonl --output live234-gemma-http-mid.json
+bash run-live-http234.sh --batching atom_batching231.py --http-module llmd214.py --cache-dir atom-artifacts --kv-blocks 3 --model Qwen/Qwen3-0.6B --fixtures semantic227-qwen06b-fixtures.json --cases capital-100 arithmetic-100 korean-100 retrieval-100 --worker-python "$WORKER_PYTHON" --engine-factory long_width_variant:make_prefill --width 2048 --baseline semantic227-qwen06b-atom.jsonl --output live234-qwen06b-http-mid.json
+bash run-live-http234.sh --batching atom_batching231.py --http-module llmd214.py --cache-dir atom-artifacts --kv-blocks 3 --model Qwen/Qwen3-4B --fixtures semantic227-qwen4b-fixtures.json --cases capital-100 arithmetic-100 korean-100 retrieval-100 --worker-python "$WORKER_PYTHON" --engine-factory long_width_variant:make_prefill --width 2048 --baseline semantic227-qwen4b-atom.jsonl --output live234-qwen4b-http-mid.json
+bash run-live-http234.sh --batching atom_batching236.py --http-module llmd214.py --cache-dir atom-artifacts --kv-blocks 1 --model KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024 --fixtures semantic227-koni-fixtures-eos.json --cases capital-100 arithmetic-100 korean-100 retrieval-100 --worker-python "$WORKER_PYTHON" --engine-factory koni_prefill:Prefill --width 2048 --baseline semantic227-koni-atom.jsonl --output live234-koni-http-mid.json
+```
+
+This sequential queue invokes the exact short-record command per model with --width 2048 --cases capital-100 arithmetic-100 korean-100 retrieval-100 and output live234-<model>-http-mid.json. Qwen uses --engine-factory long_width_variant:make_prefill (explicit existing experimental8192 guard variant), Gemma uses gemma_prefill:Prefill, KONI uses koni_prefill:Prefill. Model-specific artifact/batching/fixture/baseline arguments are unchanged from live-http234. Queue stops on any nonzero result.
+
+## Results
+
+| Model | Actual prompt tokens | Exact responses | Remote/local prefills | Cancel/disconnect reuse |
+|---|---:|---:|---:|---|
+| Gemma3-1B | 1623–1628 | 14/14 | 16/0 | pass |
+| Qwen3-0.6B | 1526–1531 | 14/14 | 16/0 | pass |
+| Qwen3-4B | 1526–1531 | 14/14 | 16/0 | pass |
+| KONI-8B | 1637–1643 | 14/14 | 16/0 | pass |
+
+All four queue stages exited0. Total56response checks and64live RNGD prefills, no local fallback. Together with the short matrix:112exact normal/concurrent/reuse responses and128remote prefills. Client concurrency is queued on B1, not hardware batch4. Recorded latency fields are observations; no native-vs-PD performance comparison was performed here.
+
+## Conclusion
+
+Medium-input live HTTP correctness/lifecycle passes for all four selected models. Original whole-goal scope remains open: live6K, prefill-time hardware cancellation/worker crashes, cache restoration, applicable hardware batch>1, GPU live PD, numerical fidelity and real overlap/serving benefit. After the medium queue terminated, the same owned allocation started run-live-long234.sh with width8192 and *-400 cases. That new queue is pending; do not restart while its processes are live.
+
+Raw evidence: analysis/serving-matrix-201/live234-<model>-http-mid.json, live234-mid-queue.log and live234-http-mid-summary.json. New queue log: live234-long-queue.log.

From 04dab7cb6961103f9404c99e9fcbcaadac14b098 Mon Sep 17 00:00:00 2001
From: gspark <gspark@etri.re.kr>
Date: Fri, 23 Oct 2026 23:52:03 +0900
Subject: [PATCH 6/6] test: record four-model long live RNGD to ATOM HTTP
 results (#234)

---
 .../examples/pd/serving_matrix/COVERAGE.md    |   6 +-
 .../examples/pd/serving_matrix/STATUS.md      |   2 +
 .../serving_matrix/evidence234/events.jsonl   |   1 +
 .../evidence234/live234-gemma-http-long.json  | 397 +++++++++++++++
 .../live234-http-long-summary.json            |  44 ++
 .../evidence234/live234-koni-http-long.json   | 394 +++++++++++++++
 .../live234-qwen06b-http-long.json            | 457 ++++++++++++++++++
 .../evidence234/live234-qwen4b-http-long.json | 400 +++++++++++++++
 .../2026-10-23-live-http-long234.md           |  34 ++
 9 files changed, 1732 insertions(+), 3 deletions(-)
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http-long.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-http-long-summary.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http-long.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http-long.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http-long.json
 create mode 100644 opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-long234.md

diff --git a/opt_prime/examples/pd/serving_matrix/COVERAGE.md b/opt_prime/examples/pd/serving_matrix/COVERAGE.md
index 6e8dc403..9eb3bc2b 100644
--- a/opt_prime/examples/pd/serving_matrix/COVERAGE.md
+++ b/opt_prime/examples/pd/serving_matrix/COVERAGE.md
@@ -13,9 +13,9 @@ All direct runs use fixed output lengths ignoring EOS. Repeats match within each
 
 ## Outstanding acceptance work
 
-- Standalone GPU HTTP covered by PR213 and ATOM HTTP by PR216: all four lifecycle checks pass through8K. Cache-off concurrent/serial equality48/48 on each backend. Real RNGD->ATOM short-input HTTP now passes56/56 output checks (#235/#237); medium/long live and broader conditions remain incomplete.
+- Standalone GPU HTTP covered by PR213 and ATOM HTTP by PR216: all four lifecycle checks pass through8K. Cache-off concurrent/serial equality48/48 on each backend. Real RNGD->ATOM short-input HTTP now passes56/56 output checks (#235/#237); short/medium/long live checks pass168/168responses; broader conditions remain incomplete.
 - Qwen ATOM device/host-cache cases pass12/12 concurrent equality each (PR216). GPU BF16 cache has numerical ties; FP32 controls match48/48 (PR213). Gemma sliding-state restoration is unimplemented, KONI k1 has no spare prefix-cache blocks.
-- Short live producer/scheduler integration is verified for all four models. Live long prefixes and genuine cache-aware producer chunking remain open.
+- Short live producer/scheduler integration is verified for all four models. Live long prefixes now pass for all four; genuine cache-aware producer chunking remains open.
 - Layer/logit diagnosis and broader semantic fixtures for cross-backend output differences; matching a repeated capital question is not a quality evaluation.
 - Actual producer/consumer overlap and end-to-end serving benefit, distinct from faster CPU staging.
 
@@ -41,4 +41,4 @@ Four models,48producer runs,31/32 EOS comparisons at6026–6428input tokens. Bot
 
 All four short-input paths pass14normal/concurrent/reuse responses each:56total,64remote prefills including cancellation requests,0local fallbacks. Requests trigger actual RNGD subprocess computation; no KV replay. Each model passes explicit cancellation, socket disconnection and subsequent reuse. Gemma/Qwen usek3; static B1 KONI usesk1 with#236. SIGPIPE-safe HTTP is#214/#215. These are B1 hardware requests with queued client concurrency, not hardware batch4. Native parity includes shared wrong answers.
 
-Width2048 live checks now pass for all four models:56responses and64remote prefills. Short+medium total112responses/128prefills, with no local fallbacks. Width8192 live queue is pending. Full live6K, pending-prefill hardware cancellation/worker-failure recovery, GPU live PD, hybrid prefix/host restore, applicable hardware batch>1, genuine producer/transfer overlap and matched native-vs-PD latency/throughput comparisons remain unverified. Overall goal remains active.
+Width2048 live checks now pass for all four models:56responses and64remote prefills. Short+medium total112responses/128prefills, with no local fallbacks. Width8192 live checks also pass all four models:56responses/64remote prefills at6026–6428input tokens. Combined short+medium+long:168responses/192remote prefills,0fallback. Pending-prefill hardware cancellation/worker-failure recovery, cross-host RNGD-to-GPU live PD, hybrid prefix/host restore, applicable hardware batch>1, genuine producer/transfer overlap and matched native-vs-PD latency/throughput remain unverified. GPU-to-GPU live HTTP is tested separately under#240: default BF16 concurrent35/48; explicit uncached/chunk256 producer controls#241 recover48/48concurrent and12/12native/PD serial outputs. Gemma numerical#233 remains unresolved. Overall goal remains active.
diff --git a/opt_prime/examples/pd/serving_matrix/STATUS.md b/opt_prime/examples/pd/serving_matrix/STATUS.md
index 2a6d7004..cc1c7e9e 100644
--- a/opt_prime/examples/pd/serving_matrix/STATUS.md
+++ b/opt_prime/examples/pd/serving_matrix/STATUS.md
@@ -65,3 +65,5 @@ Reusable Gemma producer matches original RNGD K/V/full logits4/4 at width128. Re
 All four models:56/56 normal/concurrent/reuse responses equal native ATOM,64RNGD prefills and0local fallbacks. Cancellation/disconnect reuse passes. B1, cache-off, short prompts; KONI uses existingk1 artifact with#236. HTTP uses#214 SIGPIPE fix. No live-transfer-overlap or throughput gain claim. Width2048 live queue passes all four models; width8192 live queue pending; long/cache/hardware-batch/performance scope remains.
 
 Medium live HTTP:1526–1643tokens,56/56 exact outputs and64remote/0local; short+medium112responses/128remote. Same cancellation/disconnect reuse checks pass. Long live queue active; full goal remains open. See live-http-mid234.
+
+Long live HTTP follow-up: all four width8192 routes pass56/56responses and64remote/0local prefills, cancellation/disconnect reuse included. Short+medium+long168/168responses/192remote. All queue stages exited0. HardwareB1, nativeATOM equality only; no quality, performance or overlap claim. See live-http-long234 record/evidence234.
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl b/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
index 9ab9df1b..a1efcedb 100644
--- a/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/events.jsonl
@@ -1,3 +1,4 @@
 {"id": "819c75b3957c4f0b980915096014c36f", "ts": "2026-10-23T22:30:07+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Callable Gemma refactor:4/4 full K/V/logits bit-exact on RNGD; actual subprocess/shared-memory worker5/5 exact and oversize recovery pass.15CPU tests+5subtests. No live HTTP claim yet.", "next": "Connect live worker to production ATOM HTTP; extend Qwen/KONI adapters and long-prefix/lifecycle conditions.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"model": "Gemma3-1B", "width": "128"}, "artifacts": ["opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-worker.json"], "experiments": ["opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-model-pd234.md"], "labels": ["experiment"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "validation host", "branch": "main", "commit": "351d39d-dirty"}
 {"id": "fdd4a316d5cb433e8b2b56028ac53379", "ts": "2026-10-23T22:52:06+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Real RNGD->ATOM HTTP four models:56/56 normal/reuse outputs exact,64remote/0local prefills; concurrent clients and cancel/disconnect recovery pass. KONI k1 enabled by#236;22adapter tests+5subtests and43scheduler tests. Width2048 queue started.", "next": "Collect live medium-context queue; publish #236 and adapter updates, then controlled native/PD performance and remaining cache/overlap.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"hardware_batch": "1", "mode": "live_rngd_atom_http"}, "artifacts": ["opt_prime/examples/pd/serving_matrix/evidence234/live234-http-summary.json"], "experiments": ["opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http234.md"], "labels": ["experiment"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "RNGD+ATOM host", "branch": "main", "commit": "351d39d-dirty"}
 {"id": "50c06e38ffe04b859420a115b72f5cc6", "ts": "2026-10-23T23:05:26+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Medium1526-1643token live RNGD->ATOM HTTP four models56/56 responses exact,64remote/0local and cancel/disconnect reuse pass. Short+medium112responses/128remote. PR237df02223 single-block support; long8192-width/*-400 queue started.", "next": "Collect long live queue; then pending-prefill/worker-failure recovery, matched native/PD performance, cache/overlap and#233 numerics.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"width": "2048", "hardware_batch": "1"}, "artifacts": ["opt_prime/examples/pd/serving_matrix/evidence234/live234-http-mid-summary.json"], "experiments": ["opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-mid234.md"], "labels": ["experiment"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "RNGD+ATOM host", "branch": "main", "commit": "351d39d-dirty"}
+{"id": "41fd11829a7f4bd2a868afd0f12e194d", "ts": "2026-10-23T23:50:41+09:00", "task": "Live four-model PD integration (aicomp #234)", "result": "Long all4models56/56 nativeATOM output checks,64remote/0local,cancel/disconnect reuse pass. Short+medium+long168/168responses,192remote prefills. Queue70411 terminal0; no active benchmark in owned paired allocation.", "next": "Publish long evidence; next hardware pending-prefill/fault checks and matched performance/overlap, retain Gemma#233.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"width": "8192", "hardware_batch": "1"}, "artifacts": [{"path": "analysis/serving-matrix-201/live234-http-long-summary.json", "size": 844, "sha256": "2aa497c94edb6e28"}], "experiments": ["experiments/2026-10-23-live-http-long234.md"], "labels": ["experiment", "rngd"], "review": true, "todos": [], "note": "", "agent": "codex", "host": "xeon-max", "branch": "main", "commit": "351d39d-dirty"}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http-long.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http-long.json
new file mode 100644
index 00000000..8c856ee9
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-gemma-http-long.json
@@ -0,0 +1,397 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-400",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792765083,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6423,
+          "completion_tokens": 2,
+          "total_tokens": 6425
+        }
+      },
+      "latency_s": 2.059294487000443
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792765084,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "25",
+            "token_ids": [
+              236778,
+              236810,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6424,
+          "completion_tokens": 3,
+          "total_tokens": 6427
+        }
+      },
+      "latency_s": 1.7813942750217393
+    },
+    {
+      "case": "korean-400",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792765086,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              98665,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6424,
+          "completion_tokens": 2,
+          "total_tokens": 6426
+        }
+      },
+      "latency_s": 1.7830313210142776
+    },
+    {
+      "case": "retrieval-400",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792765088,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              28487,
+              1618,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6428,
+          "completion_tokens": 3,
+          "total_tokens": 6431
+        }
+      },
+      "latency_s": 1.8204669379629195
+    },
+    {
+      "case": "capital-400",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792765090,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6423,
+          "completion_tokens": 2,
+          "total_tokens": 6425
+        }
+      },
+      "latency_s": 1.8128947280347347
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792765092,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "25",
+            "token_ids": [
+              236778,
+              236810,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6424,
+          "completion_tokens": 3,
+          "total_tokens": 6427
+        }
+      },
+      "latency_s": 1.8391432100906968
+    },
+    {
+      "case": "korean-400",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792765094,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              98665,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6424,
+          "completion_tokens": 2,
+          "total_tokens": 6426
+        }
+      },
+      "latency_s": 1.8177044560434297
+    },
+    {
+      "case": "retrieval-400",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792765095,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              28487,
+              1618,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6428,
+          "completion_tokens": 3,
+          "total_tokens": 6431
+        }
+      },
+      "latency_s": 1.8186313869664446
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792765097,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "25",
+            "token_ids": [
+              236778,
+              236810,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6424,
+          "completion_tokens": 3,
+          "total_tokens": 6427
+        }
+      },
+      "latency_s": 1.8222813030006364
+    },
+    {
+      "case": "capital-400",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792765099,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6423,
+          "completion_tokens": 2,
+          "total_tokens": 6425
+        }
+      },
+      "latency_s": 3.6308326860889792
+    },
+    {
+      "case": "retrieval-400",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792765101,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              28487,
+              1618,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6428,
+          "completion_tokens": 3,
+          "total_tokens": 6431
+        }
+      },
+      "latency_s": 5.447109714965336
+    },
+    {
+      "case": "korean-400",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792765103,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              98665,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6424,
+          "completion_tokens": 2,
+          "total_tokens": 6426
+        }
+      },
+      "latency_s": 7.257134675048292
+    },
+    {
+      "case": "capital-400",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792765106,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6423,
+          "completion_tokens": 2,
+          "total_tokens": 6425
+        }
+      },
+      "latency_s": 1.8154899249784648
+    },
+    {
+      "case": "capital-400",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792765110,
+        "model": "unsloth/gemma-3-1b-it",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              50429,
+              106
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6423,
+          "completion_tokens": 2,
+          "total_tokens": 6425
+        }
+      },
+      "latency_s": 1.791391227976419
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-http-long-summary.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-http-long-summary.json
new file mode 100644
index 00000000..da51cb0b
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-http-long-summary.json
@@ -0,0 +1,44 @@
+{
+  "models": [
+    {
+      "model": "gemma",
+      "responses": 14,
+      "remote": 16,
+      "local": 0,
+      "cancellation_reuse": true,
+      "input_min": 6423,
+      "input_max": 6428
+    },
+    {
+      "model": "qwen06b",
+      "responses": 14,
+      "remote": 16,
+      "local": 0,
+      "cancellation_reuse": true,
+      "input_min": 6026,
+      "input_max": 6031
+    },
+    {
+      "model": "qwen4b",
+      "responses": 14,
+      "remote": 16,
+      "local": 0,
+      "cancellation_reuse": true,
+      "input_min": 6026,
+      "input_max": 6031
+    },
+    {
+      "model": "koni",
+      "responses": 14,
+      "remote": 16,
+      "local": 0,
+      "cancellation_reuse": true,
+      "input_min": 6137,
+      "input_max": 6143
+    }
+  ],
+  "total_responses": 56,
+  "remote_prefills": 64,
+  "local_fallbacks": 0,
+  "hardware_batch": 1
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http-long.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http-long.json
new file mode 100644
index 00000000..41262463
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-koni-http-long.json
@@ -0,0 +1,394 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-400",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792766816,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6138,
+          "completion_tokens": 2,
+          "total_tokens": 6140
+        }
+      },
+      "latency_s": 5.335317382006906
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792766820,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              2983,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6137,
+          "completion_tokens": 2,
+          "total_tokens": 6139
+        }
+      },
+      "latency_s": 4.379292714060284
+    },
+    {
+      "case": "korean-400",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792766824,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              115978,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6139,
+          "completion_tokens": 2,
+          "total_tokens": 6141
+        }
+      },
+      "latency_s": 4.36221725505311
+    },
+    {
+      "case": "retrieval-400",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792766829,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              18082,
+              877,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6143,
+          "completion_tokens": 3,
+          "total_tokens": 6146
+        }
+      },
+      "latency_s": 4.450131780933589
+    },
+    {
+      "case": "capital-400",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792766833,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6138,
+          "completion_tokens": 2,
+          "total_tokens": 6140
+        }
+      },
+      "latency_s": 4.333481329958886
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792766837,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              2983,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6137,
+          "completion_tokens": 2,
+          "total_tokens": 6139
+        }
+      },
+      "latency_s": 4.402466899016872
+    },
+    {
+      "case": "korean-400",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792766842,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              115978,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6139,
+          "completion_tokens": 2,
+          "total_tokens": 6141
+        }
+      },
+      "latency_s": 4.322912700008601
+    },
+    {
+      "case": "retrieval-400",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792766846,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              18082,
+              877,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6143,
+          "completion_tokens": 3,
+          "total_tokens": 6146
+        }
+      },
+      "latency_s": 4.414992478094064
+    },
+    {
+      "case": "capital-400",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792766851,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6138,
+          "completion_tokens": 2,
+          "total_tokens": 6140
+        }
+      },
+      "latency_s": 4.283409978961572
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792766855,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              2983,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6137,
+          "completion_tokens": 2,
+          "total_tokens": 6139
+        }
+      },
+      "latency_s": 8.67299879703205
+    },
+    {
+      "case": "korean-400",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792766859,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              115978,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6139,
+          "completion_tokens": 2,
+          "total_tokens": 6141
+        }
+      },
+      "latency_s": 13.047355814022012
+    },
+    {
+      "case": "retrieval-400",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792766864,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              18082,
+              877,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6143,
+          "completion_tokens": 3,
+          "total_tokens": 6146
+        }
+      },
+      "latency_s": 17.474600348970853
+    },
+    {
+      "case": "capital-400",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792766873,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6138,
+          "completion_tokens": 2,
+          "total_tokens": 6140
+        }
+      },
+      "latency_s": 4.395555546972901
+    },
+    {
+      "case": "capital-400",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792766881,
+        "model": "KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              60704,
+              128009
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6138,
+          "completion_tokens": 2,
+          "total_tokens": 6140
+        }
+      },
+      "latency_s": 4.39396666304674
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http-long.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http-long.json
new file mode 100644
index 00000000..e9d48283
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen06b-http-long.json
@@ -0,0 +1,457 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-400",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792765440,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 3.1041212170384824
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792765442,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42",
+            "token_ids": [
+              16,
+              22,
+              488,
+              220,
+              17,
+              20,
+              284,
+              220,
+              19,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6027,
+          "completion_tokens": 11,
+          "total_tokens": 6038
+        }
+      },
+      "latency_s": 2.572454360080883
+    },
+    {
+      "case": "korean-400",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792765445,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 **인천**입니다.",
+            "token_ids": [
+              66845,
+              23573,
+              125496,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 14,
+          "total_tokens": 6045
+        }
+      },
+      "latency_s": 2.60976382100489
+    },
+    {
+      "case": "retrieval-400",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792765447,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 3,
+          "total_tokens": 6034
+        }
+      },
+      "latency_s": 2.5075355779845268
+    },
+    {
+      "case": "capital-400",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792765450,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 2.431598277995363
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792765452,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42",
+            "token_ids": [
+              16,
+              22,
+              488,
+              220,
+              17,
+              20,
+              284,
+              220,
+              19,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6027,
+          "completion_tokens": 11,
+          "total_tokens": 6038
+        }
+      },
+      "latency_s": 2.365722585003823
+    },
+    {
+      "case": "korean-400",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792765455,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 **인천**입니다.",
+            "token_ids": [
+              66845,
+              23573,
+              125496,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 14,
+          "total_tokens": 6045
+        }
+      },
+      "latency_s": 2.4915975880576298
+    },
+    {
+      "case": "retrieval-400",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792765457,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 3,
+          "total_tokens": 6034
+        }
+      },
+      "latency_s": 2.4608081240439788
+    },
+    {
+      "case": "capital-400",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792765459,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 2.436914485064335
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792765462,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "17 + 25 = 42",
+            "token_ids": [
+              16,
+              22,
+              488,
+              220,
+              17,
+              20,
+              284,
+              220,
+              19,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6027,
+          "completion_tokens": 11,
+          "total_tokens": 6038
+        }
+      },
+      "latency_s": 4.775390743045136
+    },
+    {
+      "case": "korean-400",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792765464,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "대한민국의 수도는 **인천**입니다.",
+            "token_ids": [
+              66845,
+              23573,
+              125496,
+              124785,
+              20401,
+              134013,
+              16560,
+              3070,
+              31328,
+              129034,
+              334,
+              78952,
+              13,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 14,
+          "total_tokens": 6045
+        }
+      },
+      "latency_s": 7.070237320964225
+    },
+    {
+      "case": "retrieval-400",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792765467,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 3,
+          "total_tokens": 6034
+        }
+      },
+      "latency_s": 9.64104292797856
+    },
+    {
+      "case": "capital-400",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792765472,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 2.510781381977722
+    },
+    {
+      "case": "capital-400",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792765476,
+        "model": "Qwen/Qwen3-0.6B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 2.521972832037136
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http-long.json b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http-long.json
new file mode 100644
index 00000000..da36214f
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/evidence234/live234-qwen4b-http-long.json
@@ -0,0 +1,400 @@
+{
+  "worker": "live_rngd_subprocess_prefill",
+  "rows": [
+    {
+      "case": "capital-400",
+      "request": "normal-0-0",
+      "response": {
+        "id": "normal-0-0",
+        "object": "text_completion",
+        "created": 1792766310,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 5.043774194899015
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "normal-0-1",
+      "response": {
+        "id": "normal-0-1",
+        "object": "text_completion",
+        "created": 1792766314,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              19,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6027,
+          "completion_tokens": 3,
+          "total_tokens": 6030
+        }
+      },
+      "latency_s": 4.239778788993135
+    },
+    {
+      "case": "korean-400",
+      "request": "normal-0-2",
+      "response": {
+        "id": "normal-0-2",
+        "object": "text_completion",
+        "created": 1792766319,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              26698,
+              126893,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 3,
+          "total_tokens": 6034
+        }
+      },
+      "latency_s": 4.6180367700289935
+    },
+    {
+      "case": "retrieval-400",
+      "request": "normal-0-3",
+      "response": {
+        "id": "normal-0-3",
+        "object": "text_completion",
+        "created": 1792766323,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 3,
+          "total_tokens": 6034
+        }
+      },
+      "latency_s": 4.293596021016128
+    },
+    {
+      "case": "capital-400",
+      "request": "normal-1-0",
+      "response": {
+        "id": "normal-1-0",
+        "object": "text_completion",
+        "created": 1792766328,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 4.356954317074269
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "normal-1-1",
+      "response": {
+        "id": "normal-1-1",
+        "object": "text_completion",
+        "created": 1792766332,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              19,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6027,
+          "completion_tokens": 3,
+          "total_tokens": 6030
+        }
+      },
+      "latency_s": 4.4063200840028
+    },
+    {
+      "case": "korean-400",
+      "request": "normal-1-2",
+      "response": {
+        "id": "normal-1-2",
+        "object": "text_completion",
+        "created": 1792766336,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              26698,
+              126893,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 3,
+          "total_tokens": 6034
+        }
+      },
+      "latency_s": 4.39458605600521
+    },
+    {
+      "case": "retrieval-400",
+      "request": "normal-1-3",
+      "response": {
+        "id": "normal-1-3",
+        "object": "text_completion",
+        "created": 1792766341,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 3,
+          "total_tokens": 6034
+        }
+      },
+      "latency_s": 4.390069373999722
+    },
+    {
+      "case": "capital-400",
+      "request": "concurrent-0",
+      "response": {
+        "id": "concurrent-0",
+        "object": "text_completion",
+        "created": 1792766345,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 4.352759285015054
+    },
+    {
+      "case": "arithmetic-400",
+      "request": "concurrent-1",
+      "response": {
+        "id": "concurrent-1",
+        "object": "text_completion",
+        "created": 1792766350,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "42",
+            "token_ids": [
+              19,
+              17,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6027,
+          "completion_tokens": 3,
+          "total_tokens": 6030
+        }
+      },
+      "latency_s": 8.775955583085306
+    },
+    {
+      "case": "korean-400",
+      "request": "concurrent-2",
+      "response": {
+        "id": "concurrent-2",
+        "object": "text_completion",
+        "created": 1792766354,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "서울",
+            "token_ids": [
+              26698,
+              126893,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 3,
+          "total_tokens": 6034
+        }
+      },
+      "latency_s": 13.207480930024758
+    },
+    {
+      "case": "retrieval-400",
+      "request": "concurrent-3",
+      "response": {
+        "id": "concurrent-3",
+        "object": "text_completion",
+        "created": 1792766358,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "MAPLE",
+            "token_ids": [
+              17617,
+              867,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6031,
+          "completion_tokens": 3,
+          "total_tokens": 6034
+        }
+      },
+      "latency_s": 17.583430355996825
+    },
+    {
+      "case": "capital-400",
+      "request": "after-cancel",
+      "response": {
+        "id": "after-cancel",
+        "object": "text_completion",
+        "created": 1792766367,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 4.211048972909339
+    },
+    {
+      "case": "capital-400",
+      "request": "after-disconnect",
+      "response": {
+        "id": "after-disconnect",
+        "object": "text_completion",
+        "created": 1792766376,
+        "model": "Qwen/Qwen3-4B",
+        "choices": [
+          {
+            "index": 0,
+            "text": "Paris",
+            "token_ids": [
+              59604,
+              151645
+            ],
+            "finish_reason": "stop"
+          }
+        ],
+        "usage": {
+          "prompt_tokens": 6026,
+          "completion_tokens": 2,
+          "total_tokens": 6028
+        }
+      },
+      "latency_s": 4.349409435992129
+    }
+  ],
+  "stats_pd": {
+    "remote": 16,
+    "local": 0
+  },
+  "free": [
+    0
+  ],
+  "cancellation_reuse": true
+}
diff --git a/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-long234.md b/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-long234.md
new file mode 100644
index 00000000..4d829c71
--- /dev/null
+++ b/opt_prime/examples/pd/serving_matrix/experiments/2026-10-23-live-http-long234.md
@@ -0,0 +1,34 @@
+# Long-input live RNGD to ATOM HTTP PD (complete)
+
+- **Date**: 2026-10-23
+- **Task**: aicomp #234
+- **Host / device**: existing owned paired RNGD+ATOM allocation; hardwareB1, context9216, same SDK/artifacts as live-http-mid234.
+- **Commit**: adapter74d0d3b (evidence/docs75ce140); core1e54786 for Gemma/Qwen, df02223 behavior for KONI; HTTP9e442cd.
+
+## Question
+
+Do real HTTP-triggered prefill and decoder lifecycle checks remain correct for approximately6K token inputs?
+
+## Setup and command
+
+After run-live-mid234.sh exited0 for all models, start `bash run-live-long234.sh` in the same allocation. It uses the four exact medium-record commands with --width8192 (CLI syntax: --width 8192), *-400 instead of *-100 cases, and output live234-<model>-http-long.json. Model/factory/artifact/core module arguments are unchanged. Order:Gemma,Qwen0.6B,Qwen4B,KONI. Two sequential passes, four queued clients, cancellation/disconnect and reuse;14normal/reuse output checks per model. B1, cache-off, no saved-KV replay.
+
+## Results
+
+All four stages exited0. Each model has14/14 normal/concurrent/reuse responses matching native ATOM tokens/EOS,16remote prefills,0local fallbacks, cancellation/disconnect reuse passing and lane0 released.
+
+| Model | Input tokens | Responses | Remote prefills |
+|---|---:|---:|---:|
+| Gemma3-1B | 6423–6428 | 14/14 | 16 |
+| Qwen3-0.6B | 6026–6031 | 14/14 | 16 |
+| Qwen3-4B | 6026–6031 | 14/14 | 16 |
+| KONI8B | 6137–6143 | 14/14 | 16 |
+
+Total56responses/64remote prefills. Across short+medium+long:168responses/192remote prefills,0fallback. These are B1 requests with queued client concurrency. No controlled performance result is claimed. The earlier offline31/32 long comparison includes an unresolved Gemma GPU-vs-RNGD numerical mismatch (#233); this live checker compares against nativeATOM and does not erase that discrepancy or certify answer quality.
+
+Local queue session70411 and all child benchmark processes completed. Queue stdout:analysis/serving-matrix-201/live234-long-queue.log. Per-model JSON and aggregate live234-http-long-summary.json are preserved in the same directory.
+
+
+## Conclusion
+
+The four-model long live matrix is complete. Keep the overall goal active. Hardware batch>1, pending-prefill/worker-failure recovery, GPU live PD, cache restoration, numerical fidelity, compute-transfer overlap and matched serving performance remain open.