Merge branch 'v0.9.2-dev-ds-wm-1223' into 'v0.9.2-dev-ds'

[feat]低延迟模式采用int8 dispatch

See merge request dcutoolkit/deeplearing/vllm!314

vllm/model_executor/layers/fused_moe/deepep_ll_prepare_finalize.py

@@ -185,7 +185,10 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
         a1_dtype: torch.dtype,
         quant_config: FusedMoEQuantConfig,
     ) -> mk.PrepareResultType:
-        expert_x, expert_x_scale = self._do_quant(expert_x, a1_scale, a1_dtype, quant_config, expert_num_tokens)
+        if not self.use_int8_dispatch:
+            expert_x, expert_x_scale = self._do_quant(expert_x, a1_scale, a1_dtype, quant_config, expert_num_tokens)
+        else:
+            expert_x, expert_x_scale = expert_x
 
         expert_tokens_meta = mk.ExpertTokensMetadata(
             expert_num_tokens=expert_num_tokens, expert_num_tokens_cpu=None

vllm/model_executor/layers/fused_moe/layer.py

@@ -171,7 +171,7 @@ class FusedMoEMethodBase(QuantizeMethodBase):
                                 and moe.quant_config.block_shape
                                 == DEEPEP_QUANT_BLOCK_SHAPE)
             
-            use_int8_dispatch = False#moe.quant_config.quant_dtype == torch.int8
+            use_int8_dispatch = moe.quant_config.quant_dtype == torch.int8
 
             # Note (varun): Whether to use FP8 dispatch or not needs some
             # profiling. Turning it off for now.