@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 5b19b93082
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 00:45 UTC

[ROCm][Kernel] Using the correct warp_size value

Gregory Shtrasberg committed Feb 5, 2025 at 22:15 UTC 5b19b93082fc5ad0ce33752d8467337cbe93de21
1 file changed +2 -2
csrc/moe/moe_align_sum_kernels.cu
+2 -2
@@ -207,8 +207,8 @@ __global__ void sgl_moe_align_block_size_kernel(
207 __shared__ int32_t shared_counts[32][8];
208 __shared__ int32_t local_offsets[256];
209
210 - const int warp_id = threadIdx.x / WARP_SIZE;
211 - const int lane_id = threadIdx.x % WARP_SIZE;
210 + const int warp_id = threadIdx.x / 32;
211 + const int lane_id = threadIdx.x % 32;
212 const int experts_per_warp = 8;
213 const int my_expert_start = warp_id * experts_per_warp;
214