[Performance] Dual stream execution of "shared_experts" and "selected_experts" inside FusedMoE (vllm-project#26440)

alexm-redhat · ilmarkov · commit 26ec33546773 · 2025-11-10T20:32:35.000Z
Signed-off-by: Alexander Matveev &lt;amatveev@redhat.com&gt;
diff --git a/vllm/envs.py b/vllm/envs.py
@@ -1481,7 +1481,6 @@ def get_vllm_port() -> int | None:
     # After enabled, PromptLogprobs and SampleLogprobs would populated as
     # FlattenLogprobs.
     "VLLM_FLATTEN_LOGPROBS": lambda: bool(int(os.getenv("VLLM_FLATTEN_LOGPROBS", "0"))),
-    # (EPLB-related envs moved to config: use ParallelConfig.eplb_config)
 }
 
 # --8<-- [end:env-vars-definition]

Original file line number	Diff line number	Diff line change
`@@ -1481,7 +1481,6 @@ def get_vllm_port() -> int \| None:`
`1481`	`1481`	`# After enabled, PromptLogprobs and SampleLogprobs would populated as`
`1482`	`1482`	`# FlattenLogprobs.`
`1483`	`1483`	`"VLLM_FLATTEN_LOGPROBS": lambda: bool(int(os.getenv("VLLM_FLATTEN_LOGPROBS", "0"))),`
`1484`		`- # (EPLB-related envs moved to config: use ParallelConfig.eplb_config)`
`1485`	`1484`	`}`
`1486`	`1485`
`1487`	`1486`	`# --8<-- [end:env-vars-definition]`