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