Skip to content

[Bug Fix] Fix some bugs #7221

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ming1753 wants to merge 1 commit into
base: develop
Choose a base branch
from
+31 −2
Open

[Bug Fix] Fix some bugs #7221

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 14 additions & 2 deletions custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -293,10 +293,16 @@ void GetBlockShapeAndSplitKVBlock(
// Note (sunxin): Skip capturing the DtoH copy (it's time-consuming); CPU data
// is only for branching in attention.
#ifndef PADDLE_WITH_CUSTOM_DEVICE_METAX_GPU
if (!phi::backends::gpu::IsCUDAGraphCapturing())
#endif
if (!phi::backends::gpu::IsCUDAGraphCapturing()) {
max_len_tensor_cpu.copy_(
max_len_tensor_gpu, max_len_tensor_cpu.place(), false);
cudaStreamSynchronize(stream);
}
#else
max_len_tensor_cpu.copy_(
max_len_tensor_gpu, max_len_tensor_cpu.place(), false);
cudaStreamSynchronize(stream);
#endif

auto max_len_cpu_ptr = max_len_tensor_cpu.data<int>();
int max_len_this_time = max_len_cpu_ptr[0];
Expand Down Expand Up @@ -435,6 +441,12 @@ void GetBlockShapeAndSplitKVBlock(
encoder_num_blocks_x_cpu.copy_(
encoder_num_blocks_x, encoder_num_blocks_x_cpu.place(), false);
}
#ifndef PADDLE_WITH_CUSTOM_DEVICE_METAX_GPU
if (!phi::backends::gpu::IsCUDAGraphCapturing())
cudaStreamSynchronize(stream);
#else
cudaStreamSynchronize(stream);
#endif
}

std::vector<std::vector<int64_t>> GetBlockShapeAndSplitKVBlockInferShape(
Expand Down
17 changes: 17 additions & 0 deletions custom_ops/gpu_ops/flash_mask_attn/mainloop_attn.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -490,6 +490,23 @@ struct CollectiveMainloopAttn {

softmax.rescale_o(tOrO, scores_scale);
consumer_wait(pipeline_v, smem_pipe_read_v);
if (seq_len_k - n_block * kBlockN < kBlockN) {
int valid_k = seq_len_k - n_block * kBlockN;
auto sVt_this = sVt(_, _, smem_pipe_read_v.index());
constexpr int kHdLo = decltype(get<0, 0>(shape(sVt_this)))::value;
constexpr int kHdHi = decltype(get<0, 1>(shape(sVt_this)))::value;
if (thread_idx >= valid_k && thread_idx < kBlockN) {
#pragma unroll
for (int hd_hi = 0; hd_hi < kHdHi; ++hd_hi) {
#pragma unroll
for (int hd_lo = 0; hd_lo < kHdLo; ++hd_lo) {
sVt_this(make_coord(make_coord(hd_lo, hd_hi), thread_idx)) =
Element(0);
}
}
}
cutlass::arch::fence_view_async_shared();
}
gemm</*zero_init=*/false, /*wg_wait=*/-1>(
tiled_mma1, tOrP, tOrV(_, _, _, smem_pipe_read_v.index()), tOrO);
warp_scheduler_barrier_arrive();
Expand Down
Loading