[Bugfix] Add deepseek_v32 to Quark dynamic MXFP4 model type check (#39498)

Signed-off-by: Shantipriya Parida <shantipriya.parida@amd.com>
This commit is contained in:
Shantipriya Parida
2026-06-10 02:52:33 -07:00
committed by GitHub
parent fdfb2566c0
commit a1ec011a83
@@ -5,6 +5,7 @@ import fnmatch
from typing import TYPE_CHECKING, Any, cast
import torch
from transformers import PretrainedConfig
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
@@ -45,6 +46,10 @@ __all__ = ["QuarkLinearMethod"]
logger = init_logger(__name__)
# model_type values that use dynamic MXFP4 re-quantization for
# OCP MX fp4 Quark checkpoints
_DEEPSEEK_V3_FAMILY_MODEL_TYPES = frozenset({"deepseek_v3", "deepseek_v32"})
class QuarkConfig(QuantizationConfig):
def __init__(
@@ -67,6 +72,33 @@ class QuarkConfig(QuantizationConfig):
# we want to re-enable it in the future.
self.dynamic_mxfp4_quant = False
def maybe_update_config(
self,
model_name: str,
hf_config: PretrainedConfig | None = None,
revision: str | None = None,
):
"""Enable dynamic MXFP4 only for DeepSeek-V3-family fp4 checkpoints."""
if hf_config is None:
return
if (
getattr(hf_config, "model_type", None)
not in _DEEPSEEK_V3_FAMILY_MODEL_TYPES
):
return
quant_config = getattr(hf_config, "quantization_config", None)
if isinstance(quant_config, dict):
quant_dtype = (
quant_config.get("global_quant_config", {})
.get("weight", {})
.get("dtype")
)
if quant_dtype == "fp4":
self.dynamic_mxfp4_quant = True
def get_linear_method(self) -> "QuarkLinearMethod":
return QuarkLinearMethod(self)