[BugFix] 1D query fix for MoE models (#3597)

tests/kernels/test_moe.py

@@ -81,11 +81,13 @@ def test_mixtral_moe(dtype: torch.dtype):
         vllm_moe.w2s[i][:] = hf_moe.experts[i].w2.weight.data
 
     # Generate input batch of dimensions [batch_size, seq_len, hidden_dim]
-    inputs = torch.randn((1, 64, config.hidden_size)).to(dtype).to("cuda")
+    hf_inputs = torch.randn((1, 64, config.hidden_size)).to(dtype).to("cuda")
+    # vLLM uses 1D query [num_tokens, hidden_dim]
+    vllm_inputs = hf_inputs.flatten(0, 1)
 
     # Run forward passes for both MoE blocks
-    hf_states, _ = hf_moe.forward(inputs)
-    vllm_states = vllm_moe.forward(inputs)
+    hf_states, _ = hf_moe.forward(hf_inputs)
+    vllm_states = vllm_moe.forward(vllm_inputs)
 
     mixtral_moe_tol = {
         torch.float32: 1e-3,
@@ -93,7 +95,7 @@ def test_mixtral_moe(dtype: torch.dtype):
         torch.bfloat16: 1e-2,
     }
 
-    assert torch.allclose(hf_states,
+    assert torch.allclose(hf_states.flatten(0, 1),
                           vllm_states,
                           rtol=mixtral_moe_tol[dtype],
                           atol=mixtral_moe_tol[dtype])

vllm/model_executor/models/deepseek.py

@@ -150,11 +150,11 @@ class DeepseekMoE(nn.Module):
         self.w2 = self.w2.view(len(w2), *w2s[0].shape)
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        num_tokens, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
         if self.config.n_shared_experts is not None:
             shared_output = self.shared_experts(hidden_states)
-        # router_logits: (batch * sequence_length, n_experts)
+        # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
         final_hidden_states = fused_moe(hidden_states,
                                         self.w1,
@@ -169,8 +169,7 @@ class DeepseekMoE(nn.Module):
         final_hidden_states = tensor_model_parallel_all_reduce(
             final_hidden_states)
 
-        return final_hidden_states.view(batch_size, sequence_length,
-                                        hidden_dim)
+        return final_hidden_states.view(num_tokens, hidden_dim)
 
 
 class DeepseekAttention(nn.Module):

vllm/model_executor/models/mixtral.py

@@ -124,9 +124,9 @@ class MixtralMoE(nn.Module):
             param_data[expert_id, :, :] = loaded_weight[:, shard]
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        batch_size, sequence_length, hidden_size = hidden_states.shape
+        num_tokens, hidden_size = hidden_states.shape
         hidden_states = hidden_states.view(-1, self.hidden_size)
-        # router_logits: (batch * sequence_length, n_experts)
+        # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
         final_hidden_states = fused_moe(hidden_states,
                                         self.ws,
@@ -140,8 +140,7 @@ class MixtralMoE(nn.Module):
             final_hidden_states = tensor_model_parallel_all_reduce(
                 final_hidden_states)
 
-        return final_hidden_states.view(batch_size, sequence_length,
-                                        hidden_size)
+        return final_hidden_states.view(num_tokens, hidden_size)
 
 
 class MixtralAttention(nn.Module):

vllm/model_executor/models/mixtral_quant.py

@@ -132,9 +132,9 @@ class MixtralMoE(nn.Module):
                                      linear_method=None)
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        num_tokens, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
-        # router_logits: (batch * sequence_length, n_experts)
+        # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
 
         routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
@@ -158,7 +158,7 @@ class MixtralMoE(nn.Module):
                 final_hidden_states.add_(current_hidden_states)
 
         return tensor_model_parallel_all_reduce(final_hidden_states).view(
-            batch_size, sequence_length, hidden_dim)
+            num_tokens, hidden_dim)
 
 
 class MixtralAttention(nn.Module):