[Quantization] [Refactor] Create special "GptOssMxfp4MoeMethod" (#39604)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Yongye Zhu
2026-04-13 12:53:58 -04:00
committed by GitHub
parent 4d042ed85f
commit 739e5945dc
18 changed files with 129 additions and 34 deletions
+17
View File
@@ -108,6 +108,23 @@ class Gemma4Config(VerifyAndUpdateConfig):
class GptOssForCausalLMConfig(VerifyAndUpdateConfig):
@staticmethod
def verify_and_update_model_config(model_config: "ModelConfig") -> None:
quant_config = getattr(model_config.hf_config, "quantization_config", None)
if quant_config is not None and quant_config.get("quant_method") == "mxfp4":
model_config.hf_config.quantization_config["quant_method"] = "gpt_oss_mxfp4"
hf_text_quant_config = getattr(
model_config.hf_text_config, "quantization_config", None
)
if (
hf_text_quant_config is not None
and hf_text_quant_config.get("quant_method") == "mxfp4"
):
model_config.hf_text_config.quantization_config["quant_method"] = (
"gpt_oss_mxfp4"
)
@staticmethod
def verify_and_update_config(vllm_config: "VllmConfig") -> None:
structured_outputs_config = vllm_config.structured_outputs_config