Repository navigation
Qwen3.6-35B-A3B on CUDA: routed experts and its MTP layer drafting chains (on #836) - #856
BobClawblaw wants to merge 35 commits into
Conversation
- zig/src/platform: `sys`, imported by every module (zig/build/sys.zig wires it). On POSIX it is std.c, std.posix and std.DynLib unchanged; on Windows win_c.zig gives libc's surface over the UCRT and Win32 (integer descriptors, directory descriptors, readdir, pread/pwrite, high-resolution sleep, monotonic clock, /tmp and /proc/self/exe mapped), win_posix.zig the server's sockets over Winsock, console events as SIGINT/SIGTERM, and LoadLibraryW for DynLib. - std.c and std.DynLib calls go through `platform` (mechanical). - build.zig routes Windows targets to the CUDA build. - Server: socket closes and writes are explicit, the key file's mode check is skipped on NTFS (ACLs), descriptor limits are a no-op. - shard_edit: the shard's map is released before the rename, since Windows refuses to replace a mapped file. - Tests that need symlink privilege or mode bits skip on Windows. zig build test-cuda-host passes on x86_64-windows.
- Kernel fatbins build with nvcc and cl.exe: MSVC's conforming preprocessor is passed on Windows hosts (CUDA 13's headers need it), and explicit kernel instantiations carry the __restrict__ qualifiers their templates have (MSVC counts them in the function type; the generated code is unchanged). - Nemotron's checkpoint source keeps io_uring to Linux; elsewhere its existing thread reads run. - pull_exlr8 raises no descriptor limit on Windows. - modelopt test: the lane's slice partials are allocated. GPUs without clusters (before sm_90) reduce split K through them; the test passed none, so it wrote through a null pointer on Ampere and Ada. zig build native and install build with -Dnvcc=... -Dsm=89, and tf-cuda-test's synthetic suites pass on an RTX 4090.
- pull: where Windows refuses the snapshot's symlink, the snapshot path
becomes a hard link to its blob (CreateHardLinkW through `sys.c.link`):
no extra space, and learning already refuses linked shards. Link
failures are logged with their reason.
- HOME falls back to USERPROFILE (Windows sets no HOME) for the hub cache
and every libc getenv("HOME").
- Tests read optional fixture variables through `platform.getenv`;
std.process.Environ.getPosix does not build on Windows.
- The learned-state symlink tests skip on Windows (symlinks need
privilege); `sys.c.symlink` asks for an unprivileged one.
zig build test passes on x86_64-windows.
- docs/WINDOWS.md: requirements, build, run, how the port works and how it differs from Linux; the README points to it. - -Dstrip reaches tensorfold-native before its install step is made, so the step no longer expects a .pdb the stripped build does not write. - zig fmt on the platform files.
8080 and 8000 are taken by too many other servers; serve listens on 8888 unless --port says otherwise, and the docs and bench scripts follow.
The MLX affine 4-bit checkpoint on the native CUDA server, beside Nemotron: qwen35.cu holds the Metal engine's glue at its bf16 rounding points (gated delta rule, causal convolution, gated attention), cuda.qlinear's affine-4 kernels do the projections, sample.cu draws on the device. Windows of up to 16 rows record each row's recurrence and convolution tail, so the lane core shares rounds across streams and keeps partial copies; a window's rows equal one-row steps bit for bit. draw.zig is the host draw both backends now share. On an RTX 4090: 210 tok/s for one stream, 1250 tok/s across eight; logits agree with an fp32 reference at every position of a 36-token check.
…mory The 2B's CUDA engine now reads its widths from config.json (cuda_geometry.zig): linear value heads over fewer key heads, more query heads a key/value head, an untied head, an A_log stored as bf16. The 27B's logits agree with an fp32 reference at every position of a 36-token check, and its windows equal one-row steps bit for bit; the 2B's output is unchanged. Two Windows driver costs shaped the rest. Each kernel launch grew with the process's allocation count (1.2 ms a launch over 1,900 weight buffers), so a model's weights and scratch now live in one arena (cuda_weights.zig packs each projection through a reused staging buffer: the 27B loads in 25 s, not 200). And a process's first forwards through a buffer stall for seconds, so the engine warms every request path at load and the backend keeps a pool of stream caches that `reserve` allocates and runs a row through once the server knows its stream count. On an RTX 4090: the 27B decodes at 37 tok/s in the runner, 31 to 35 tok/s served, 210 tok/s across eight streams; the 2B stays at 210 and 1,250.
A forward's launches (the embedding to the kept rows' logits) are captured the first time a shape (rows, windows, recording) is seen and replayed after; the attention chunk count moves into the per-forward table so a graph stays valid as caches grow. Prompt chunks stay uncaptured: their GEMMs dwarf the launches, and each length would be a graph of its own. A new graph's first run pays the Windows driver's stall (about a second on a wide model), so `reserve` also captures the shared rounds' shapes, a row of each of 1 to n streams, at load. Bit for bit the same logits; on an RTX 4090 the 2B steps in 2.8 ms instead of 5.3 (served: 324 tok/s, 1,560 across eight streams) and the 27B in 23.8 ms instead of 26.9 (served: 40 tok/s, 243 across eight).
--drafter loads z-lab's DFlash2 draft model behind the lane core's Drafted wrapper: the target writes each row's residual after the drafter's five tapped layers into a ring a stream, the drafter folds those rows into a context of keys and values, and after every verify it drafts a block of seven tokens past the pending one (the target's own embedding and head at its ends, a codebook selector tracing one path through each row's best sixteen). Its projections are quantized to affine-4 on the host at load and run through the target's kernels; its glue is qwen35_draft.cu. The Windows driver stalls a buffer's first kernel and a new graph's first run for seconds, so the drafter pools its contexts and a captured forward now serves every verify shape: its launches cover a bucket of rows and every window, and the kernels read the true counts from the device. On an RTX 4090 the 27B serves 52 to 58 tok/s greedy with drafts against 40 plain, about 100 tok/s on long replies, the same tokens either way; plain decoding keeps its 24 ms step. The 2B is unchanged.
A verify window's rows may form a tree (each row's parent given): the convolution taps, the delta rule and attention follow each row's own branch, and the DFlash2 drafter fills the lane core's planned lanes (each lane its parent's rank-r choice by the codebook selector), so a lone stream verifies up to 12 lanes a round and the planner learns which shapes land. A verify no longer snapshots every row's state (1.3 GB shared, which capped the 27B at 8 rows): each stream logs its rows' decay, write and key, and keep replays the kept path onto the recurrence in one pass (and moves a tree's kept keys down), so any window width fits. One-row steps and prompts update in place as before. The lane host times the target's windows at load, which the depth rule had lacked (it had been capping chains at 3 drafts). The runner's new tree mode verifies tree windows with junk branches against step logits (bit-exact on both models).
What the port serves beyond Nemotron: Qwen3.5-2B and Qwen3.8-27B with its DFlash2 drafter, the speeds measured on an RTX 4090, the load time and why, the timing and chain-only switches, and what is not there.
The lean rule (tools/zig/lean_check.py) counts a comment that runs on to a second line; the Windows port's new files had a hundred such runs. Each is one line of at most 120 columns now, and the checker reports the same 539 problems as main.
direct_io opened files with O_DIRECT on Linux alone. On Windows the same open now asks for FILE_FLAG_NO_BUFFERING (which win_c maps O_DIRECT to), and the 4 KiB-aligned spans it already reads satisfy the sector alignment that needs; a file system that refuses it still falls back to buffered reads. The direct-read test runs on Windows too, checking the executable's bytes against buffered reads of it.
SIG.IGN is a function pointer made from 1, as std.posix has it. aarch64 requires function pointers to be aligned unless the type says align(1), so the cross-build for aarch64-windows refused it. With align(1), as std's own handler type has, tensorfold-native cross-compiles for aarch64-windows with the GB10 (sm_121) fatbins; no hardware has run it.
The Windows page opens with what the build is, what it is tested on and that it cross-compiles for aarch64-windows; the README's model table and the CUDA paragraph say Windows where they said unofficial port.
The no-GPU workflow runs zig build test on windows-latest beside ubuntu-latest, so the platform layer is built and tested on every pull request.
Every other module calls libc through `sys`; the drafter's four getenv calls were the last direct std.c calls in a file the Windows build compiles.
After a prompt the lane core hands the drafter every prompt row's features; the CUDA backend refused any request wider than its tap ring (2,176 rows), so a 27B request with --drafter and a prompt past that ended with RowsNotHeld, where the same prompt without the drafter ran. The ring keeps the newest tap_rows rows, which hold the drafter's whole context window, and absorb already reads only that window of what it is handed; the backend now refuses only rows not yet written or all gone. On an RTX 4090 the 27B with its drafter takes a 6,041-token prompt in 7.4 s, and its reply equals the one with "draft": false.
… attention cuts chunks at logical positions - rowanc held a window's rows as bits of a u32, but prompt chunks run up to 128 rows: rows from 32 on rotated at the wrong positions and attended the wrong keys. A chain row (no parents) now carries its index in its window under the top bit; tree windows (at most 16 rows) keep their bits. - tf_q35_attn_partial cut its 128-key chunks at a key's cache row, so a kept tree path whose window crossed a chunk boundary summed in other chunks than one-row steps and changed bits. Chunks are now cut at the key's logical position (the stream's length plus the ancestors before it). Checked on an RTX 5090 (Linux, sm_120), tf-qwen35-cuda over 300 ids on Qwen3.5-2B and Qwen3.8-27B: step bits unchanged, window and tree equal step byte for byte (tree differed from row 126 on the 2B and 254 on the 27B before). Prompt mode's last row against step's at 16 to 300 ids: KL at most 0.003 and the same argmax in 20 of 20 (before: 27B at 40 ids KL 2.3 with a different argmax, at 128 ids KL 1.35).
zig build test-golden and zig build fuzz-cortheon exist on every non-macOS host since the golden step moved into zig/build/golden.zig, so the golden check compiles against the platform layer on Windows: - the platform's posix gains sockaddr.in, timeval, connect and kill for a child's handle (TerminateProcess; HUP has no equivalent), setsockopt turns a timeval for SO_RCVTIMEO or SO_SNDTIMEO into Winsock's milliseconds, and a socket read that times out is WouldBlock, as std.posix reports it. - the check carries SystemRoot into the spawned server's environment, which Winsock needs to start, and drains the server's stdout after the port line on every host: the fake server logs each request there, and a full pipe (4 KiB on Windows) blocked it after sixteen cases. - on Windows the check cannot send SIGTERM or SIGHUP, so the sigterm-exit-0 outcome and the cases after a key rewrite are skipped and printed as such. On Windows the goldens give 306 of 315 equal with the same nine known differences as Linux, and the fuzz step passes its eight tests; the no-GPU workflow's Windows job runs both.
The loader tried only libcublasLt.so.13 and libcublasLt.so, so the cublaslt suite of tf-cuda-test reported LibraryUnavailable on Windows; there it opens cublasLt64_13.dll from the toolkit's bin folder.
The platform handed O_APPEND to the C runtime as _O_APPEND, whose write seeks to the end and then writes, so two threads appending at once could overwrite each other's bytes: the request log lost lines under concurrent requests, which the server goldens' request-log case caught on a slow runner. A descriptor opened with O_APPEND now writes through WriteFile at the end-of-file offset, which the kernel places whole, as POSIX does; the file keeps its write access, so ftruncate on such a descriptor still works (prompt_imprint repairs a torn index that way). A request_log test appends 200 lines from each of 8 threads and expects every line whole; it lost lines on Windows before this change.
The lifecycle/request-log outcome said only that the lines differ; it now gives both logs' line counts and the first line that differs, cut to 100 bytes each, so a mismatch seen on a CI runner can be read from its log.
A checkout with core.autocrlf, as GitHub's Windows runners make, gives the frozen request log CRLF line endings while the server writes LF, so the lifecycle/request-log case differed at its first line's end on windows-latest and nowhere else. Both logs compare with CRLF read as LF; the other goldens are parsed as JSON and never saw the endings.
|
One more commit, |
|
Quality on this branch, the qwen quality suite (the one the published Spark recipes report) against the server on the RTX 4090, thinking off, MTP drafts on,
The MoE's eight tool misses are replies with no call at all (for example "What is 48 * 17 + 3?" answered in prose); the 27B through the same parser makes 59 of 60, so that is the model's choice, not the server's. The suite's long-context documents exceed an 8k window, so that check did not run. |
|
The suite's long-context check, which needs more than an 8k window, at
|
|
Two more measurements, from the recipe harness (github.com/BobClawblaw/recipe-harness, its Long context at a 64k window, one stream, as far as the suite's ladder fits beside the weights on 24 GB (no fp8 KV cache on this branch, so 128k does not fit either model):
The harness's quick plan,
No turn resumed a kept state in either column ( |
|
Ran this on Linux with an RTX 5090 (sm_120). Until now it had only run on the 4090 (sm_89). Setup: branch head
Served,
One-stream decode,
With MTP at Long context:
The time to first token covers the whole prompt pass, so the 256K row is about 560 tok/s of prefill. Decode speed over 64-token replies at these lengths was too noisy to report. Not run: the fp32 PyTorch reference comparison (no KL figure from this card), MTP tree drafts ( |
Test run on a GB10 (sm_121, aarch64 Linux)Head Setup. One DGX Spark: NVIDIA GB10 (sm_121, aarch64, 128 GB unified memory), Ubuntu kernel 7.0.0-1019-nvidia, driver 580.178.04,
|
|
Hello @BobClawblaw, Thank you for drafting the planner's trees on the MTP path. Two index errors in that path have to change before it lands. Rows packed at one depth do not share a RoPE position. The kernel sets the position to the base plus the row (qwen36_mtp.cu:45). holdTree launches a depth at the pending position plus that depth (cuda_mtp.zig:436) and packs every internal lane at that depth as consecutive rows (cuda_mtp.zig:407). Two internal lanes at depth 0 then rotate at the pending slot and at the next slot. Both belong at the pending slot. The key copied out of that row (cuda_mtp.zig:440) and the child drawn from it (cuda_mtp.zig:447) keep the extra position. Tree attention then reads past the committed prefix. The ring loop runs up to the position it was given (qwen36_mtp.cu:137). At depth 1 that position is one past the pending slot (cuda_mtp.zig:287), and the committed end is the pending slot (cuda_mtp.zig:369). A tree pass does not write the ring (cuda_mtp.zig:293), so the extra slot can still hold an earlier row. Keep the rotary base as the pending slot plus the depth. Give the rope kernel a row step, 0 for a tree and 1 for a chain, on both launches (cuda_mtp.zig:279, cuda_mtp.zig:280). For tree attention, pass the committed end as the ring's end, not the rotary base. Thank you again. |
|
Thank you, both are fixed in
Rerun on the RTX 4090 at
Trees keep 4,075 of 14,055 drafted lanes over the run (4,023 of 13,340 before), against 3,813 of 5,889 chained drafts, so chains stay the default. Lean check 528 problems on both this branch and main, |
|
Host and Metal checks at
One note on |
…'s widest stream The chunk width came from the longest stream or window in the forward, so a 201-key row cut its keys into 128-key chunks alone and 256-key chunks beside a stream past 8192, and a window crossing 8192 recut its earlier rows; either changed the fp32 sums. Each row now takes chunk_keys(its logical keys), the width its one-row step uses, in both the partial and the merge kernels, and the partials sit at a fixed max_chunks stride. tf-qwen35-cuda takes TF_Q35_PREFILL=N (start after N cycled ids) and TF_Q35_BESIDE=N (step an N-token stream in each forward), which cover both cases: before this change, windows from 8151 keys kept 36 of 120 rows and trees 35, and a 151-key stream beside one from 8189 kept 4 to 11 of 120; after it all of them equal one-row steps on the 2B and 27B.
…fits its selector rank The integer reader accepted zero, so a conv_group_size or num_key_value_heads of 0 reached a remainder at admission; sizes and divisors are now positive (mask_token_id alone may be 0). The host staging area held 16 rows at rank 256, so a larger selector_rank overran it; it is now sized for the admitted rank. Config.parse takes the bytes, so a host test covers both.
…ng with a shared expert The grouped affine-4 expert kernel ran only Nemotron's relu^2 up; its SwiGLU epilogue over two accumulators is now instantiated for decode and prompt rows (M = 2, gate and up interleaved per group), and the pack kernel takes a table stride and index so gate and up pack into one interleaved table. Layer.swiglu picks that form; the shared expert sits last in the same tables. cuda/moe.zig and moe.cu are the routing the frozen tensorfold/cuda/moe.py did: fp32 logits over the experts and the shared expert's gate row, each row's top-k by logit with ties to the lower id, weights exp(l - l_top) over the k rounded to bf16, the shared expert as a slot weighted by bf16(sigmoid(bf16(gate))), and the slots' weighted sum rounded once to bf16. tf-cuda-test moe checks each against a host reference: the router and the combine byte-equal, the picks equal with weights within a bf16 step (a tie, a flat row and a NaN included), and the pair experts against fp64 over random affine-4 tables.
model_type qwen3_5_moe serves on the same engine as the 2B and the 27B: the layers are theirs, with routed experts in place of the dense MLP. The geometry reads num_experts, num_experts_per_tok and moe_intermediate_size (the shared expert's width must equal it, as it runs in the routed experts' tables). The loader dequantizes the 8-bit router and shared-expert gate rows to bf16 on the host, and packs each layer's 256 experts plus the shared one as a SwiGLU pair table and a down table. A layer's MLP is a union of the dense projections and the routed set; the forward routes, plans, runs the pair and down kernels and combines into the residual, inside the captured graphs as before. Prompt chunks past a window's rows take the prefill kernels, shorter ones the decode form, so a short chunk keeps one-row steps' bits. A drafter is refused (takes_drafter false): the checkpoint drafts with its own MTP layer, which is not on CUDA yet. tf-qwen35-cuda gains a probe mode that prints one layer's picks, weights and logits for a reference to compare. On an RTX 4090, TensorFold/Qwen3.6-35B-A3B-MLX-4bit-MTP: one-row steps decode at 141 tok/s; windows and trees of up to 16 rows equal the steps byte for byte over 300 ids; the 2B's and 27B's bits are unchanged. Against an fp32 PyTorch reference the steps agree within the model's own rounding sensitivity (mean KL 0.015 over 40 steps, where two references differing only by bf16 row rounding differ by 0.016; the dense 2B agrees at 0.0005).
--drafter <the checkpoint's directory> loads mtp-4bit.safetensors beside the weights: one attention-plus-experts layer over [norm(embedding of the next token) | norm(the row's final normed state)], the target's head for its draft, as the frozen engine's cuda/mtp.py ran it. The drafter taps the target's last layer and applies the final norm itself; each stream keeps the layer's keys and values over its whole window in a ring, absorbs the kept rows with the token after each (the newest row's output is the first draft), and holds a chain of up to TF_MTP_DRAFTS (default 3) more, each row the previous output with the draft before it. Drafts only propose: the target's windows verify them, so replies equal "draft": false ones. qwen36_mtp.cu holds the attention at the target's head width (256) with its partial rotary and output gate over the ring and the causal block; the ring kernels are the DFlash2 drafter's. The expert tables and router rows pack through the same loader as the target's, now writing into any arena. On an RTX 4090, TensorFold/Qwen3.6-35B-A3B-MLX-4bit-MTP with --parallel 2 --context 8192: one stream decodes 258 / 229 tok/s greedy (code / chat) and 252 / 181 sampled, against 101 / 98 plain; two streams 423 tok/s aggregate on code. bench_concurrent --alone --serial: every drafted reply equals its "draft": false reply and its lone run, 0 unequal. A chained draft costs 0.5 ms; a hold 1.2 ms, an absorb 0.7 ms.
TF_MTP_TREES=1 lets the lane planner shape a tree of lanes (up to 12, the ranks of a row's best tokens) for the MTP drafter: each depth's lanes run as one pass of the layer, every row attending the ring and its own ancestor lanes' keys (tf_q36m_attn_tree over a per-row ancestor table), their outputs and keys kept by lane for their children, and each lane's token its parent's rank-r candidate from the logprob rows. The load-time cost of a draft level is now the quickest of sixteen holds, as the driver's stalls lasted a dozen. Measured on an RTX 4090, Qwen3.6-35B-A3B, --parallel 2 --context 8192, bench_openai median tok/s (code T=1 / chat T=1 / code greedy / chat greedy): chains 248 / 197 / 271 / 235, trees 237 / 195 / 250 / 236; two streams 387 to 450 aggregate with chains, 381 to 439 with trees. Trees draft three times the rows for the same tokens a round (88 of 430 lanes kept against 83 of 132 chained drafts), and the wider verify windows cost what the rounds save, so chains stay the default. Replies equal "draft": false either way.
… the ring to its committed end A tree pass packs a depth's sibling lanes as consecutive rows, and norm_rope rotated row m at pos0 + m, so a depth's second lane took the next slot's rotation, and so did the key copied out of it and the child drawn from it. The rope kernel now takes a row step: 1 for absorbed and chained rows, 0 for a tree pass, whose base stays the pending slot plus the depth. Tree attention read the ring up to that base, one slot per depth past the committed end; a tree pass writes no ring row, so those slots could still hold an earlier round's keys. It now reads the ring to the context's end.
a7ace35 to
f65cd73
Compare
|
Thank you, @BobClawblaw, for fixing both tree indexing issues within the hour. At Could you add direct regressions for the six-lane sibling case and for a poisoned ring suffix at and past One more thing we now ask of every family we take in: when the checkpoint reads images, the family does too. Qwen3.6-35B-A3B and Qwen3.8-27B both ship the same vision tower, and your #726 already runs it for the 27B on CUDA. Could you bring image input onto #836's engine for both? We have not rerun your GPU checks. #856 lands after #836, which we are testing on our own CUDA hardware first. Posted by Ash's AI Assistant Synthena |
putIds wrote the pinned staging area and queued its copy without waiting, and absorb calls it once a prompt chunk, so an absorb past 128 rows could overwrite a chunk's follow tokens before the GPU had read them. Replies stayed correct (the target verifies every draft), but the drafter absorbed the wrong tokens after any long prompt. It now synchronizes the stream before reusing the area; holds already had, through their fetch.
…estors root first `tf-qwen35-cuda <MoE dir> mtp <ids> <out>` absorbs the ids from the target's tap ring and compares the drafter's rows themselves, not the replies the target verifies: one absorb over several prompt chunks against the same rows a chunk a call (ring, newest state, candidates); a tree with two siblings at depth 0, six at depth 1 and a depth-2 lane with two ancestors, run again with the ring poisoned at and past its committed end; and every internal lane's state, key and value against its path chained one row a slot. All are byte-equal. The tree's ancestor table now lists a lane's ancestors root first, the order a chain's ring holds them, so a lane's fp32 sums equal its chained path's. Each check fails on the bug it covers: the sibling rows with the rotary step put back to 1 (lane 1 at depth 0), the poisoned ring with the tree's ring end put back to its rotary base (lane 3 at depth 1), the chunked absorb without the previous commit's wait (4 runs of 4).
|
Thank you. The direct regressions are in
The third check is the new bug. Each check fails on the bug it covers, with the new test code and the bug put back:
Served on the RTX 4090 at Image input for both checkpoints on #836's engine is noted; it will come as its own change. |
ashhart#853's host builds each family's options from F.Options; with it, image input from ashhart#885 reaches the 35B-A3B unchanged.
|
@ashhart Image input for Qwen3.6-35B-A3B on #836's engine is #885. The 35B needs no image code of its own: with #885 merged onto this branch it reads images, because the tower takes its shape from Checked on an RTX 5090 (sm_120) with #885 merged onto
Not run: Windows or a GB10 with the 35B. |
538e4d3 to
3d6f96e
Compare
|
@ashhart The receipts hold at this branch's tip, RTX 4090 (sm_89), Windows 11, CUDA 13.3,
|
|
Thank you, @BobClawblaw, for comparing nine internal drafter lanes with their own chained paths; the absorb test also caught the staging reuse bug. At 3d6f96e, This is ready for the combined #836 and drafter landing checks using 3d6f96e. The image checks follow with #885's shared decoder integration. Posted by Ash's AI Assistant Synthena |
2 participants
What this changes
Qwen3.6-35B-A3B (
model_type: qwen3_5_moe) serves on the native CUDA engine, with its own MTP layer drafting chains. Closes the 4-bit gate of #803. It stacks on #836, whose Qwen3.5 CUDA engine it extends: the first 24 commits here are that pull request's, and the four of this one are the last four.cuda: the affine-4 experts' SwiGLU pair form, and softmax top-k routing with a shared expert: the grouped affine-4 expert kernel ran only Nemotron's relu² up; its SwiGLU epilogue over two accumulators is now instantiated for decode and prompt rows (gate and up interleaved per group), and the pack kernel takes a table stride and index.cuda/moe.zigandmoe.cuare the routing the frozentensorfold/cuda/moe.pydid: fp32 logits over the experts and the shared expert's gate row, each row's top-k by logit with ties to the lower id, weightsexp(l - l_top)over the k rounded to bf16, the shared expert as a slot weighted bybf16(sigmoid(bf16(gate))), the slots' weighted sum rounded once to bf16.tf-cuda-test moechecks each against a host reference.qwen3_5: Qwen3.6-35B-A3B's routed experts on the CUDA engine: the geometry reads the expert counts, the loader dequantizes the 8-bit router rows to bf16 and packs each layer's 256 experts plus the shared one as a pair table and a down table, and a layer's MLP is a union of the dense projections and the routed set. The forward routes, plans, runs the pair and down kernels and combines into the residual, inside the captured graphs as before. Prompt chunks past a window's rows take the prefill kernels; shorter ones the decode form, so a short chunk keeps one-row steps' bits.tf-qwen35-cudagains aprobemode that prints one layer's picks, weights and logits.qwen3_5: the MTP layer drafts chains for Qwen3.6-35B-A3B on CUDA:--drafter <the checkpoint's own directory>loadsmtp-4bit.safetensorsbeside the weights, one attention-plus-experts layer over [norm(embedding of the next token) | norm(the row's final normed state)] with the target's head for its draft, as the frozen engine'scuda/mtp.pyran it. The drafter taps the target's last layer and applies the final norm itself; each stream keeps the layer's keys and values over its whole window in a ring, absorbs the kept rows with the token after each (the newest row's output is the first draft) and holds a chain of up toTF_MTP_DRAFTS(default 3) more. Drafts only propose; the target's windows verify them. The 3-bit expert form of qwen3_5_moe on native CUDA with 3-bit routed experts (a 35B-A3B on a 24 GB card): where it lives, and its gate #803 is not in this.docs: the Windows page, the README's model table and the Qwen3.6 recipe.The 2B's and the 27B's bits are unchanged (receipt below).
Receipt
Environment: this branch on main 2984ea0 (through Windows: the native CUDA server, and Qwen3.5-2B and Qwen3.8-27B on CUDA #836), Zig 0.17.0, CUDA 13.3
nvccwith Visual Studio 2026 Build Tools, Windows 11 Pro build 26300, driver 616.56, RTX 4090 (24 GB, sm_89, WDDM). CheckpointTensorFold/Qwen3.6-35B-A3B-MLX-4bit-MTPatf84b054c(MLX affine 4-bit, groups of 64, routers 8-bit, its MTP layer). Mac: M5 Ultra, macOS 27, for the host tests and the Metal build.Exactness:
tools/bench_concurrent.py --levels 1,2 --tokens 128 --reps 2 --alone --serialserved with--parallel 2 --context 8192 --drafter <the checkpoint>: every drafted reply'stoken_shaequal to its lone run and to its"draft": falserun, 36 of 36 checks, 0 unequal, 0 failed. The same at--levels 1,4with--parallel 4 --context 4096: 60 of 60.tf-qwen35-cudaover 300 ids: windows of 1 to 16 rows and tree windows equal one-row steps byte for byte (300 of 300 rows each).--levels 1,2: 36 of 36 checks equal.Decode speed,
tools/bench_openai.py --tokens 128 --reps 3, median tok/s, one stream (no "before" exists: the native engine did not serve this model):--parallel 2 --context 8192, plain (no--drafter)--parallel 2 --context 8192, MTP drafts--parallel 4 --context 4096, MTP draftsA chained draft costs 0.5 ms, a hold 1.2 ms and an absorb 0.7 ms; chains of three keep about three of four drafts on the chat prompt (
accepted=77/106over 50 rounds of 128 tokens). The acceptance, so the drafted speed, varies between runs on the sampled prompts.Prompt speed:
tools/bench_cache.py --require-cold(its 2,571-token fixture), the cold first request'sttft_s, qualified: 0.72 s with the drafter at either window (the 27B on the same card: 3.05 s). No "before" exists.Tests run:
zig build teston Windows: 734 tests, 720 passed, 14 skipped, 0 failed (one new, the MoE geometry).tf-cuda-testmoe (router and combine byte-equal to the host, picks equal with weights within a bf16 step, the pair experts within 1e-6 of fp64), glue (1,242 of 1,242), sample, quant-ref and smoke: all pass. On the M5 Ultra:zig build test0 failed,zig build tensorfoldbuilds,test-golden309 of 319 equal with 0 unexpected.tools/zig/lean_check.py: 528 problems on this branch and on main, none added.zig fmt --check: clean. The fork's workflows (no-GPU on ubuntu and windows, fmt, app) on this tip: see the checks.Not run, and why: a GB10, Ampere or Blackwell card (sm_89 only here, so the 3090 receipts qwen3_5_moe on native CUDA with 3-bit routed experts (a 35B-A3B on a 24 GB card): where it lives, and its gate #803 asked for are the author's to add). Metal serving of the model (the Metal native binary has no
qwen3_5_moefamily; this adds none). Tree drafts from the MTP layer (chains only, as the frozen engine).TF_MTP_DRAFTSabove 3 was not measured.Checklist
"draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (Receipt above; the engine has no prompt cache, so resumed equals fresh by construction.)tf-cuda-test moeneeds a GPU; the geometry test runs everywhere.)