[OMNIML-5899] Export Q8_0 checkpoints and add recipes

CHANGELOG.rst

@@ -13,6 +13,7 @@ Changelog
 
 *Quantization*
 
+- Add Q8_0 weight-only quantization with 32-value GGML blocks, packed unified HF and Megatron export, and a built-in ``q8_0`` PTQ recipe.
 - Add Hugging Face PTQ calibration and export support for Nemotron-H MTP modules, preserving calibrated expert input scales.
 - Backfill checkpoint aliases for nine more published NVFP4 releases, so each is reachable from its source model's hub path: ``zai-org/GLM-5.1`` and ``GLM-5.2``, ``MiniMaxAI/MiniMax-M2.5`` and ``MiniMax-M3``, ``deepseek-ai/DeepSeek-V3.1`` and ``DeepSeek-V3.2``, ``Qwen/Qwen3-235B-A22B-Instruct-2507`` and ``-Thinking-2507``, and ``Qwen/Qwen3.6-27B``. Each imports an existing general or architecture recipe wholesale rather than copying its body.
 - Add composed Hugging Face AutoQuantize recipes that run fixed PTQ or weight AutoQuantize before

docs/source/deployment/3_unified_hf.rst

@@ -52,39 +52,41 @@ The unified HF export API supports the following quantization formats:
 6. W4A8_AWQ - 4-bit weights and 8-bit activations with AWQ optimization
 7. IQ1_S - 1-bit codebook quantization using the GGML block layout
 8. IQ2_XS - 2-bit codebook quantization using the GGML block layout
+9. Q8_0 - 8-bit symmetric integer quantization using the GGML block layout
 
 .. note::
    GGML has no equivalent for ModelOpt's per-tensor FP8 weight-and-activation format. In particular,
    GGML does not define a first-class FP8 tensor type with the corresponding per-tensor weight and
    activation scale semantics. Converting a ModelOpt FP8 checkpoint to GGUF therefore requires
    conversion to another GGML-supported tensor type rather than a lossless FP8 encoding.
 
-IQ weight representation
-~~~~~~~~~~~~~~~~~~~~~~~~
+GGML weight representation
+~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-For IQ1_S and IQ2_XS, unified export replaces each floating-point ``<module>.weight`` with a
-``uint8`` tensor containing byte-exact GGML blocks. Its shape is
-``[*logical_shape[:-1], logical_shape[-1] // 256, payload_bytes]``, where ``payload_bytes`` is 50
-for IQ1_S and 74 for IQ2_XS. No separate shape tensor is stored: a loader recovers the logical
-shape as ``[*weight.shape[:-2], weight.shape[-2] * 256]``. This is unambiguous because IQ export
-requires the logical last dimension to be divisible by 256.
+For IQ1_S, IQ2_XS, and Q8_0, unified export replaces each floating-point
+``<module>.weight`` with a ``uint8`` tensor containing byte-exact GGML blocks. Its shape is
+``[*logical_shape[:-1], logical_shape[-1] // block_size, payload_bytes]``. The block size and
+payload size are 256 and 50 for IQ1_S, 256 and 74 for IQ2_XS, and 32 and 34 for Q8_0. No separate
+shape tensor is stored: a loader recovers the logical shape as
+``[*weight.shape[:-2], weight.shape[-2] * block_size]``. This is unambiguous because export
+requires each logical row to contain complete blocks.
 
 .. note::
-   Megatron IQ export currently requires tensor and pipeline model parallel sizes of 1. Packing
+   Megatron GGML export currently requires tensor and pipeline model parallel sizes of 1. Packing
    happens during export, so a tensor-parallel shard would be packed as if it were a whole
-   weight, and a pipeline stage holding no IQ layer would not reach the same rejection as its
+   weight, and a pipeline stage holding no GGML layer would not reach the same rejection as its
    peers. Expert parallelism is supported, assuming every expert uses the same format.
 
 .. warning::
-   Megatron fused-MoE IQ export is not currently supported. Its packed tensor would require the
+   Megatron fused-MoE GGML export is not currently supported. Its packed tensor would require the
    deployment consumer to understand
-   ``[num_experts, out_features, in_features // 256, payload_bytes]`` rather than the ordinary HF
-   fused-expert order. The exporter raises ``NotImplementedError`` until a deployment loader owns
-   this layout and is covered by an integration test. Dense and individually named expert weights
-   continue to use the representation above.
+   ``[num_experts, out_features, in_features // block_size, payload_bytes]`` rather than the
+   ordinary HF fused-expert order. The exporter raises ``NotImplementedError`` until a
+   deployment loader owns this layout and is covered by an integration test. Dense and
+   individually named expert weights continue to use the representation above.
 
-The generated configuration records ``quant_method: modelopt``, ``packing: ggml``, the 256-value
-block size, and the payload byte count. IQ payloads are not represented as compressed-tensors
+The generated configuration records ``quant_method: modelopt``, ``packing: ggml``, the format's
+block size, and the payload byte count. GGML payloads are not represented as compressed-tensors
 integer ``weights`` groups because all scales and indices are embedded in each packed block.
 
 Each 74-byte IQ2_XS block represents 256 logical weights:
@@ -99,6 +101,10 @@ Each 74-byte IQ2_XS block represents 256 logical weights:
 The canonical 512-by-8 IQ2_XS codebook is part of the implementation rather than the checkpoint.
 The complete block therefore costs ``74 * 8 / 256 = 2.3125`` bits per logical weight.
 
+Each 34-byte Q8_0 block represents 32 logical weights: bytes 0--1 hold the little-endian FP16
+scale, and bytes 2--33 hold 32 signed int8 quants. The block costs
+``34 * 8 / 32 = 8.5`` bits per logical weight.
+
 Minimum Framework Versions
 --------------------------
 

modelopt/torch/export/convert_hf_config.py

@@ -19,9 +19,9 @@
 from collections import defaultdict
 from typing import Any
 
-from modelopt.torch.quantization.ggml import IQ_FORMAT_REGISTRY
+from modelopt.torch.quantization.ggml import GGML_FORMAT_REGISTRY
 
-from .quant_format import IQ_FORMATS
+from .quant_format import GGML_FORMATS
 
 
 def _quant_algo_to_group_config(quant_algo: str, group_size: int | None = None) -> dict[str, Any]:
@@ -34,7 +34,7 @@ def _quant_algo_to_group_config(quant_algo: str, group_size: int | None = None)
     Returns:
         Dictionary with ``input_activations`` and ``weights`` entries suitable for
         a compressed-tensors ``config_groups`` entry, or ModelOpt-owned metadata for
-        self-contained IQ payloads.
+        self-contained GGML payloads.
     """
     if quant_algo == "FP8":
         return {
@@ -122,13 +122,13 @@ def _quant_algo_to_group_config(quant_algo: str, group_size: int | None = None)
             },
             "weights": {"dynamic": False, "num_bits": 8, "type": "float", "group_size": gs},
         }
-    elif quant_algo.lower() in IQ_FORMATS:
-        iq_format = IQ_FORMAT_REGISTRY[quant_algo.lower()]
-        block_size, payload_bytes = iq_format.block_size, iq_format.block_bytes
-        effective_bits = iq_format.effective_bits
+    elif quant_algo.lower() in GGML_FORMATS:
+        ggml_format = GGML_FORMAT_REGISTRY[quant_algo.lower()]
+        block_size, payload_bytes = ggml_format.block_size, ggml_format.block_bytes
+        effective_bits = ggml_format.effective_bits
         if group_size not in (None, block_size):
             raise ValueError(f"{quant_algo} requires group size {block_size}, got {group_size}")
-        # IQ payloads are self-contained blocks, not compressed-tensors integer groups.
+        # GGML payloads are self-contained blocks, not compressed-tensors integer groups.
         # Keep their format marker outside a ``weights`` quantization scheme.
         return {
             "quant_algo": quant_algo,
@@ -229,13 +229,13 @@ def convert_hf_quant_config_format(input_config: dict[str, Any]) -> dict[str, An
             "targets": ["Linear"],
         }
         new_config["config_groups"] = {"group_0": config_group_details}
-    elif str(quant_algo_value).lower() in IQ_FORMATS:
+    elif str(quant_algo_value).lower() in GGML_FORMATS:
         # Forward the caller's group size so a mismatched one is rejected rather than rewritten
         # to the format's block size.
-        iq_metadata = _quant_algo_to_group_config(
+        ggml_metadata = _quant_algo_to_group_config(
             quant_algo_value, original_quantization_details.get("group_size")
         )
-        new_config.update(iq_metadata)
+        new_config.update(ggml_metadata)
     elif quant_algo_value == "NVFP4_SVD":
         # NVFP4 + SVDQuant: NVFP4 weights/activations plus an AWQ-style
         # pre_quant_scale and a low-rank residual (svdquant_lora_a/b) stored as

modelopt/torch/export/quant_format.py

@@ -19,7 +19,7 @@
 constants, for example, are in :mod:`modelopt.torch.export.trtllm.model_config`.
 """
 
-from modelopt.torch.quantization.ggml import IQ_FORMAT_REGISTRY
+from modelopt.torch.quantization.ggml import GGML_FORMAT_REGISTRY, IQ_FORMAT_REGISTRY
 
 QUANTIZATION_NONE = None
 QUANTIZATION_FP8 = "fp8"
@@ -43,22 +43,24 @@
 QUANTIZATION_IQ2_XXS = "iq2_xxs"
 QUANTIZATION_IQ2_XS = "iq2_xs"
 QUANTIZATION_IQ2_S = "iq2_s"
+QUANTIZATION_Q8_0 = "q8_0"
 
-# Every GGML IQ format, derived from the registry the quantization backend dispatches through, so
-# export and dispatch cannot disagree about which formats exist. They share the weight-only,
-# 256-value-block, per-module-scale shape, so export treats them as one family. A format's block
-# geometry and packer are read from IQ_FORMAT_REGISTRY directly.
+# Every GGML format is derived from the registry the quantization backend dispatches through, so
+# export and dispatch cannot disagree about which formats exist. A format's block geometry and
+# packer are read from GGML_FORMAT_REGISTRY directly. IQ_FORMATS remains as the compatibility
+# subset used by IQ-specific conformance tests.
 #
 # Registering a format therefore declares it exportable, and that is intended rather than a side
 # effect: fake quant is dequantize(quantize(w)), so a format cannot be dispatched without the
 # packer and block geometry that are all export reads.
 IQ_FORMATS = frozenset(IQ_FORMAT_REGISTRY)
+GGML_FORMATS = frozenset(GGML_FORMAT_REGISTRY)
 
 
 # Formats whose scales are purely per-module, so export never merges them across the q/k/v
 # and gate/up groups that share an input. Every other format unifies input_amax (and, for
 # NVFP4, weight_scale_2) across such a group, which only a whole-model forward can discover.
-FUSION_FREE_FORMATS = IQ_FORMATS | frozenset(
+FUSION_FREE_FORMATS = GGML_FORMATS | frozenset(
     {
         QUANTIZATION_FP8,
         QUANTIZATION_NONE,

modelopt/torch/export/quant_utils.py

@@ -27,7 +27,7 @@
 
 from modelopt import __version__
 from modelopt.torch.models import get_spec, list_all_possible
-from modelopt.torch.quantization.ggml import IQ_FORMAT_REGISTRY
+from modelopt.torch.quantization.ggml import GGML_FORMAT_REGISTRY
 from modelopt.torch.quantization.model_calib import (
     enable_stats_collection,
     finish_stats_collection,
@@ -52,7 +52,7 @@
 from ..quantization.nn import NVFP4StaticQuantizer, SequentialQuantizer, TensorQuantizer
 from .model_utils import TiedWeightMap, get_language_model_from_vl
 from .quant_format import (
-    IQ_FORMATS,
+    GGML_FORMATS,
     KV_CACHE_FP8,
     KV_CACHE_FP8_K_NVFP4_V,
     KV_CACHE_INT8,
@@ -444,23 +444,23 @@ def get_weight_block_size(module: nn.Module, weight_name: str = "weight") -> int
 
 
 def uses_iq_quantization(module) -> bool:
-    """Whether any weight quantizer in ``module`` or its children targets an IQ format.
+    """Whether any weight quantizer in ``module`` or its children targets a GGML format.
 
     ``get_quantization_format`` returns the *first* non-``NONE`` format it finds, so in a
-    mixed-format model IQ layers sitting behind, say, an FP8 layer are invisible to it. Callers
-    that must reject IQ specifically need to see every layer.
+    mixed-format model GGML layers sitting behind, say, an FP8 layer are invisible to it. Callers
+    that must reject GGML specifically need to see every layer.
 
     This reads ``num_bits`` directly rather than resolving each layer's full format, so an
     unrelated unsupported quantizer elsewhere in the model cannot turn the check into an error.
     """
     for weight_name in weight_attr_names(module):
         weight_quantizer = representative_weight_quantizer(module, weight_name)
-        # getattr: a SequentialQuantizer has is_enabled but no num_bits, and is never IQ --
-        # IQ is a single quantizer with backend="ggml".
+        # getattr: a SequentialQuantizer has is_enabled but no num_bits, and is never GGML --
+        # GGML is a single quantizer with backend="ggml".
         if (
             weight_quantizer is not None
             and weight_quantizer.is_enabled
-            and getattr(weight_quantizer, "num_bits", None) in IQ_FORMATS
+            and getattr(weight_quantizer, "num_bits", None) in GGML_FORMATS
         ):
             return True
     return any(uses_iq_quantization(child) for _, child in module.named_children())
@@ -500,21 +500,21 @@ def _get_quantization_from_layer(layer, quantizer_attr_names: QuantizerAttrNames
             return QUANTIZATION_W4A8_AWQ
 
         # Handle individual num_bits cases
-        if weight_quantizer.num_bits in IQ_FORMATS:
+        if weight_quantizer.num_bits in GGML_FORMATS:
             if weight_quantizer.backend != "ggml":
-                raise ValueError("IQ formats require the built-in 'ggml' quantization backend")
+                raise ValueError("GGML formats require the built-in 'ggml' quantization backend")
             # Both exporters return before collecting input_scale and before the pre_quant_scale
             # handling below, so an enabled activation quantizer would be dropped without a trace
             # and the checkpoint would load as weight-only. Refuse instead.
             if input_quantizer is not None and input_quantizer.is_enabled:
                 raise NotImplementedError(
-                    "IQ1_S/IQ2_XS export is weight-only, but this layer has an enabled input "
+                    "GGML export is weight-only, but this layer has an enabled input "
                     "quantizer. The GGML block payload carries no activation scale, so the "
                     "activation quantization would be silently lost."
                 )
             if input_quantizer is not None and hasattr(input_quantizer, "_pre_quant_scale"):
                 raise NotImplementedError(
-                    "IQ1_S/IQ2_XS export does not support an AWQ-style pre_quant_scale."
+                    "GGML export does not support an AWQ-style pre_quant_scale."
                 )
             return weight_quantizer.num_bits
 
@@ -766,10 +766,10 @@ def process_layer_quant_config(layer_config_dict):
                 "quant_algo": "MXFP8",
                 "group_size": block_size_value,
             }
-        elif v in IQ_FORMATS:
-            iq_format = IQ_FORMAT_REGISTRY[v]
-            block_size, payload_bytes = iq_format.block_size, iq_format.block_bytes
-            effective_bits = iq_format.effective_bits
+        elif v in GGML_FORMATS:
+            ggml_format = GGML_FORMAT_REGISTRY[v]
+            block_size, payload_bytes = ggml_format.block_size, ggml_format.block_bytes
+            effective_bits = ggml_format.effective_bits
             if block_size_value != block_size:
                 raise ValueError(
                     f"{v.upper()} requires block size {block_size}, got {block_size_value}"

modelopt/torch/export/unified_export_hf.py

@@ -68,7 +68,7 @@
 from modelopt.torch.opt.conversion import ModeloptStateManager, modelopt_state
 from modelopt.torch.opt.plugins.huggingface import _MODELOPT_STATE_SAVE_NAME
 from modelopt.torch.quantization import set_quantizer_by_cfg_context
-from modelopt.torch.quantization.ggml import IQ_FORMAT_REGISTRY
+from modelopt.torch.quantization.ggml import GGML_FORMAT_REGISTRY
 from modelopt.torch.quantization.nn import SequentialQuantizer, TensorQuantizer
 from modelopt.torch.quantization.qtensor import MXFP8QTensor, NVFP4QTensor
 from modelopt.torch.quantization.qtensor.base_qtensor import QTensorWrapper
@@ -102,7 +102,7 @@
 )
 from .quant_format import (
     FUSION_FREE_FORMATS,
-    IQ_FORMATS,
+    GGML_FORMATS,
     QUANTIZATION_FP8,
     QUANTIZATION_FP8_PB_REAL,
     QUANTIZATION_FP8_PC_PT,
@@ -634,13 +634,13 @@ def _export_quantized_weight(
             "which dispatches to the streaming writer that materialises weights layer-by-layer."
         )
 
-    if quantization_format in IQ_FORMATS:
+    if quantization_format in GGML_FORMATS:
         if weight_name != "weight":
             raise NotImplementedError(
-                "IQ unified export currently supports modules with a standard 'weight' "
+                "GGML unified export currently supports modules with a standard 'weight' "
                 f"attribute, got {weight_name!r} on {type(sub_module).__name__}"
             )
-        packed_weight = IQ_FORMAT_REGISTRY[quantization_format].pack(
+        packed_weight = GGML_FORMAT_REGISTRY[quantization_format].pack(
             weight.to(dtype), getattr(sub_module, quantizer_attrs.weight_quantizer, None)
         )
         setattr(sub_module, weight_name, nn.Parameter(packed_weight, requires_grad=False))