Merge branch 'v0.11.0-dev-wm' into 'v0.11.0-dev'

[fix]解决moe_fused_gate编译错误

See merge request dcutoolkit/deeplearing/vllm!254

csrc/moe/moe_fused_gate.cu

@@ -331,10 +331,10 @@ __global__ void moe_fused_gate_kernel(
 // Macro to compute compile-time constants and launch the kernel.
 #define LAUNCH_MOE_GATE_CONFIG(T, EXPERTS, EXPERT_GROUP)                                                 \
   do {                                                                                                   \
-    int VPT = (EXPERTS) / (EXPERT_GROUP);                                                      \
+    constexpr int VPT = (EXPERTS) / (EXPERT_GROUP);                                                      \
     /* If EXPERT_GROUP > WARP_SIZE, fall back to 1 row per warp */                                       \
-    int ROWS_PER_WARP = ((EXPERT_GROUP) <= SIZE_WARP) ? (SIZE_WARP / (EXPERT_GROUP)) : 1;      \
-    int ROWS_PER_CTA = WARPS_PER_CTA * ROWS_PER_WARP;                                          \
+    constexpr int ROWS_PER_WARP = ((EXPERT_GROUP) <= SIZE_WARP) ? (SIZE_WARP / (EXPERT_GROUP)) : 1;      \
+    constexpr int ROWS_PER_CTA = WARPS_PER_CTA * ROWS_PER_WARP;                                          \
     moe_fused_gate_kernel<T, VPT, (EXPERTS), (EXPERT_GROUP), ROWS_PER_WARP, ROWS_PER_CTA, WARPS_PER_CTA> \
         <<<num_blocks, block_dim, 0, stream>>>(                                                          \
             input.data_ptr(),                                                                            \