EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Bug] Fix Issue#10215 (#10572)

Browse Source
This commit is contained in:
Yuhao Yao
2025-09-25 09:51:50 +08:00
committed by GitHub
parent c4e314f986
commit fe531d6f4e
2 changed files with 9 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
sgl-kernel/csrc/cutlass_extensions/gemm/collective/sm90_mma_array_tma_gmma_rs_warpspecialized_mixed_input_.hpp
View File

@@ -1025,8 +1025,6 @@ struct CollectiveMmaArrayMixedInput<
// src: tCrA_load, dst: tCrA_mma
Utils::convert_A_kblock(tCrA_load, tCrA_mma, 0);
tiled_mma.accumulate_ = GMMA::ScaleOut::Zero;
// Unroll the K mode manually to set scale D to 1
CUTLASS_PRAGMA_UNROLL
for (int chunk_id = 0; chunk_id < NumChunksPerTileK; ++chunk_id) {
@@ -1060,6 +1058,8 @@ struct CollectiveMmaArrayMixedInput<
}
}
warpgroup_wait<0>();
CUTLASS_PRAGMA_UNROLL
for (int chunk_id_ = 0; chunk_id_ < NumChunksPerTileK; ++chunk_id_) {
warpgroup_fence_operand(intermediate_array[chunk_id_]);
@@ -1114,7 +1114,6 @@ struct CollectiveMmaArrayMixedInput<
1,
smem_pipe_read.index());
warpgroup_wait<K_WAIT_MAX>();
Utils::convert_A_kblock(tCrA_load, tCrA_mma, 0);
}
}
@@ -1148,8 +1147,6 @@ struct CollectiveMmaArrayMixedInput<
tiled_mma.accumulate_ = GMMA::ScaleOut::One;
warpgroup_commit_batch();
warpgroup_wait<K_WAIT_MAX>(); // We have K_BLOCK_MAX - 1 GMMA instructions pending for this stage, so we can
// release prior barrier
if (k_block == K_BLOCK_MAX - 1) {
pipeline.consumer_release(smem_pipe_release); // UNLOCK smem_pipe_release, done _computing_ on it
++smem_pipe_release;
@@ -1162,6 +1159,8 @@ struct CollectiveMmaArrayMixedInput<
if (k_block == K_BLOCK_MAX - 1) {
// The last k_block
warpgroup_wait<0>();
CUTLASS_PRAGMA_UNROLL
for (int chunk_id_ = 0; chunk_id_ < NumChunksPerTileK; ++chunk_id_) {
warpgroup_fence_operand(intermediate_array[chunk_id_]);
@@ -1241,7 +1240,6 @@ struct CollectiveMmaArrayMixedInput<
tiled_mma.accumulate_ = GMMA::ScaleOut::One;
warpgroup_commit_batch();
warpgroup_wait<K_WAIT_MAX>();
if (k_block == K_BLOCK_MAX - 1) {
// release prior barrier
pipeline.consumer_release(smem_pipe_release); // UNLOCK smem_pipe_release, done _computing_ on it
@@ -1264,6 +1262,8 @@ struct CollectiveMmaArrayMixedInput<
if ((k_block + 1) % NumMMAsPerChunk == 0) {
tiled_mma.accumulate_ = GMMA::ScaleOut::Zero;
warpgroup_wait<0>();
warpgroup_fence_operand(intermediate);
// Apply the group-wise scaling
@@ -1296,7 +1296,7 @@ struct CollectiveMmaArrayMixedInput<
smem_pipe_release.advance(k_tile_count);
// Wait on all GMMAs to complete
warpgroup_wait<0>();
// warpgroup_wait<0>();
for (int count = 0; count < prologue_mma_count; ++count) {
pipeline.consumer_release(smem_pipe_release); // UNLOCK smem_pipe_release, done _computing_ on it