biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Bump Flashinfer Version and Re-enable DeepSeek NVFP4 AR+Norm Fusion (#34899)

Browse Source 
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Wei Zhao
2026-02-20 16:37:31 -05:00
committed by GitHub
parent 0632ed8778
commit ea5f903f80
5 changed files with 6 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
vllm/model_executor/models/config.py
View File

@@ -536,34 +536,12 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
)
class DeepseekV3ForCausalLM(VerifyAndUpdateConfig):
@classmethod
def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
"""Disable AR-RMS-Quant fusion for DeepSeekV3 in NVFP4"""
# TODO: https://github.com/vllm-project/vllm/issues/34395
# disable AR-rms-fp4 fusion for DSv3+
ar_rms_enabled = vllm_config.compilation_config.pass_config.fuse_allreduce_rms
nvfp4 = vllm_config.model_config.is_nvfp4_quantized()
# Disable by default, warn if manually enabled:
if ar_rms_enabled is None and nvfp4:
vllm_config.compilation_config.pass_config.fuse_allreduce_rms = False
if ar_rms_enabled and nvfp4:
logger.warning(
"Allreduce-rms fusion broken for DeepSeekV3 with NVFP4 quant,"
"see https://github.com/vllm-project/vllm/issues/34395."
)
class DeepseekV32ForCausalLM(DeepseekV3ForCausalLM):
class DeepseekV32ForCausalLM(VerifyAndUpdateConfig):
@classmethod
def verify_and_update_config(cls, vllm_config: "VllmConfig") -> None:
"""
Updated fp8 cache to custom "fp8_ds_mla" format for DeepSeekV32
"""
super().verify_and_update_config(vllm_config)
hf_config = vllm_config.model_config.hf_config
# Mirror the check in vllm/model_executor/models/deepseek_v2.py
@@ -654,7 +632,6 @@ MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
"MambaForCausalLM": MambaModelConfig,
"Mamba2ForCausalLM": MambaModelConfig,
"FalconMambaForCausalLM": MambaModelConfig,
"DeepseekV3ForCausalLM": DeepseekV3ForCausalLM,
"DeepseekV32ForCausalLM": DeepseekV32ForCausalLM,
"NemotronHForCausalLM": NemotronHForCausalLMConfig,
"NemotronHPuzzleForCausalLM": NemotronHForCausalLMConfig,