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 7, 00:15 UTC
Fix: Re-Enable EP for trtllm MoE FP8 backend (#36494)
Signed-off-by: Amir Klein <203507526+amirkl94@users.noreply.github.com>
amirkl94 committed
Mar 10, 2026 at 08:11 UTC
156e33553ccdba940fec83a720290b30d2686ee8
1 file changed
-6
vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py
-6
@@ -35,12 +35,6 @@ class TrtLlmFp8Experts(mk.FusedMoEExpertsMonolithic):
35
):
36
super().__init__(moe_config, quant_config)
37
38
- if moe_config.moe_parallel_config.use_ep and quant_config.is_per_tensor:
39
- raise NotImplementedError(
40
- "EP parallelism is not supported with TRTLLM"
41
- "per-tensor FP8 quantization."
42
- )
43
-
38
self.routing_method_type = moe_config.routing_method
39
self.topk = moe_config.experts_per_token
40
self.intermediate_size_per_partition = (