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

[WideEP] Fix nvfp4 DeepEP High Throughput All2All backend (#33738)

Browse Source 
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
Tyler Michael Smith
2026-02-10 22:15:43 -05:00
committed by GitHub
parent e30cedd44b
commit 066c6da6a0
1 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py
View File

@@ -82,8 +82,12 @@ def _supports_routing_method(
def _supports_parallel_config(moe_parallel_config: FusedMoEParallelConfig) -> bool: def _supports_parallel_config(moe_parallel_config: FusedMoEParallelConfig) -> bool:
"""Supports EP.""" """
return True TRTLLM is a monolithic kernel that requires dispatch_router_logits() for
the naive dispatch/combine path. DeepEP HT only implements dispatch() for
the modular kernel path, so TRTLLM is incompatible with DeepEP HT.
"""
return not moe_parallel_config.use_deepep_ht_kernels
def is_supported_config_trtllm( def is_supported_config_trtllm(