Merge branch 'v0.15.1-dev_MOE_add_RQ_push1' into 'v0.15.1-dev'

perf: DS V2模型MOE部分增加rmsQuant

See merge request dcutoolkit/deeplearing/vllm!492

vllm/attention/layer.py

@@ -667,7 +667,6 @@ class MLAAttention(nn.Module, AttentionLayerBase):
             use_mla=True,
             use_sparse=use_sparse,
         )
-
         if (
             cache_config is not None
             and cache_config.enable_prefix_caching

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -1507,6 +1507,8 @@ def outplace_fused_experts(
     w1_bias: torch.Tensor | None = None,
     w2_bias: torch.Tensor | None = None,
     use_nn_moe: bool | None = False,
+    i_q: torch.Tensor | None = None,
+    i_s: torch.Tensor | None = None
 ) -> torch.Tensor:
     return fused_experts_impl(
         hidden_states,
@@ -1536,6 +1538,8 @@ def outplace_fused_experts(
         w1_bias,
         w2_bias,
         use_nn_moe,
+        i_q=i_q,
+        i_s=i_s,
     )
 
 
@@ -1614,7 +1618,7 @@ def fused_experts(
     use_nn_moe: bool | None = False,
     use_fused_gate: bool | None = False,
     i_q: torch.Tensor | None = None,
-    i_s: torch.Tensor | None = None # TODO:wjl
+    i_s: torch.Tensor | None = None
 ) -> torch.Tensor:
     if quant_config is None:
         quant_config = FUSED_MOE_UNQUANTIZED_CONFIG
@@ -1646,6 +1650,8 @@ def fused_experts(
         w1_bias=quant_config.w1_bias,
         w2_bias=quant_config.w2_bias,
         use_nn_moe=use_nn_moe,
+        i_q=i_q,
+        i_s=i_s,
     )
 
 
@@ -1705,6 +1711,8 @@ def fused_experts_impl(
     w1_bias: torch.Tensor | None = None,
     w2_bias: torch.Tensor | None = None,
     use_nn_moe: bool | None = False,
+    i_q: torch.Tensor | None = None,
+    i_s: torch.Tensor | None = None
 ) -> torch.Tensor:
     # Check constraints.
     num_tokens = hidden_states.size(0)
@@ -1851,7 +1859,9 @@ def fused_experts_impl(
                                        a1_scale=a1_scale,
                                        a2_scale=a2_scale,
                                        block_shape=block_shape,
-                                       use_nn_moe=False
+                                       use_nn_moe=False,
+                                       i_q=i_q,
+                                       i_s=i_s
                                        )
     elif use_int4_w4a8 is True:
         return fused_experts_impl_w4a8(hidden_states=hidden_states,

vllm/model_executor/layers/fused_moe/layer.py

@@ -1976,7 +1976,7 @@ class FusedMoE(CustomOp):
             # because matrix multiply maybe modify the hidden_states.
             if has_separate_shared_experts and not use_shared_experts_stream:
                 assert self.shared_experts is not None
-                shared_output = self.shared_experts(hidden_states)
+                shared_output = self.shared_experts(hidden_states, iqis=(i_q, i_s))
 
             # NOTE: Similar with DP, PCP also needs dispatch and combine. For
             # simplicity, AgRsAll2All was added separately for PCP here. Maybe
@@ -2014,7 +2014,6 @@ class FusedMoE(CustomOp):
                     self.capture(topk_ids)
                 
                 if envs.USE_FUSED_RMS_QUANT:
-
                     final_hidden_states = self.quant_method.apply(
                             layer=self,
                             x=x,  # The type signture of this is wrong due to the hack.

vllm/model_executor/models/deepseek_v2.py

@@ -370,7 +370,8 @@ class DeepseekV2MoE(nn.Module):
         if self.experts.is_internal_router:
             # In this case, the gate/router runs inside the FusedMoE class
             fused_moe_out = self.experts(
-                hidden_states=hidden_states, router_logits=hidden_states
+                hidden_states=hidden_states, router_logits=hidden_states,
+                iqis=iqis
             )
         else:
             # router_logits: (num_tokens, n_experts)
@@ -1087,7 +1088,7 @@ class DeepseekV2DecoderLayer(nn.Module):
                                                              )
         new_resi = residual
         hidden_states = self.mlp(hidden_states,
-                                #  iqis=(_i_q, _i_s) # TODO:wjl
+                                 iqis=(_i_q, _i_s)
                                  )
 
         if isinstance(self.mlp, DeepseekV2MLP) and hidden_states.dtype == torch.float16: