[Bugfix] fix IMA issue in certain cases of the moe marlin kernel (#28619)

Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>

csrc/moe/marlin_moe_wna16/marlin_template.h

@@ -489,10 +489,11 @@ __global__ void Marlin(
   #pragma unroll
         for (int i = 0; i < 4; i++) {
           int idx = tid4 * 4 + i;
-          idx = idx < block_num_valid_tokens ? idx : 0;
+          if (idx < block_num_valid_tokens) {
             if constexpr (w_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn) {
-            sh_block_topk_weights[idx] = __hmul2(
-                global_scale, Dtype::num2num2(Dtype::float2num(
+              sh_block_topk_weights[idx] =
+                  __hmul2(global_scale,
+                          Dtype::num2num2(Dtype::float2num(
                               topk_weights_ptr[sh_block_sorted_ids[idx]])));
             } else {
               sh_block_topk_weights[idx] = Dtype::num2num2(
@@ -501,6 +502,7 @@ __global__ void Marlin(
           }
         }
       }
+    }
     __syncthreads();
   };
 

vllm/model_executor/layers/fused_moe/shared_fused_moe.py

@@ -38,7 +38,6 @@ class SharedFusedMoE(FusedMoE):
                 # TODO(wentao): find the root cause and remove this condition
                 self.enable_eplb
                 or (self.moe_config.use_flashinfer_cutlass_kernels and self.dp_size > 1)
-                or self.use_marlin_kernels
             )
             and self._shared_experts is not None
         )