A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 12:30 UTC
[WideEP] Fix nvfp4 DeepEP High Throughput All2All backend (#33738)
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Tyler Michael Smith committed
Feb 10, 2026 at 22:15 UTC
066c6da6a04906a89739fb7e6874ceb6cf714364
1 file changed
+6
-2
vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py
+6
-2
@@ -82,8 +82,12 @@ def _supports_routing_method(
82
83
84
def _supports_parallel_config(moe_parallel_config: FusedMoEParallelConfig) -> bool:
85
- """Supports EP."""
86
- return True
85
+ """
86
+ TRTLLM is a monolithic kernel that requires dispatch_router_logits() for
87
+ the naive dispatch/combine path. DeepEP HT only implements dispatch() for
88
+ the modular kernel path, so TRTLLM is incompatible with DeepEP HT.
89
+ """
90
+ return not moe_parallel_config.use_deepep_ht_kernels
91
92
93
def is_supported_config_trtllm(