[Feat]support sequence parallelism by pass for VL models (#5632)

2026-02-27 08:27:41 +08:00
parent ed175d6d92
commit 5def28dcd3
22 changed files with 460 additions and 101 deletions
--- a/vllm_ascend/ops/fused_moe/fused_moe.py
+++ b/vllm_ascend/ops/fused_moe/fused_moe.py
@@ -440,7 +440,7 @@ class AscendFusedMoE(FusedMoE):
        hidden_states, router_logits, mc2_mask, context_metadata = forward_context.moe_comm_method.prepare(
            hidden_states=hidden_states,
            router_logits=router_logits,
-            replace_allreduce=forward_context.sp_enabled,
+            replace_allreduce=forward_context.flash_comm_v1_enabled,
            enable_shared_expert_dp=self.enable_shared_expert_dp,
            quant_type=self.quant_type,
        )