Skip to content

bugfix: fix the accuracy error of NPU xattention. #1303

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

Merged
LMX-xin merged 1 commit into from
Apr 20, 2026
+21 −23
Merged

bugfix: fix the accuracy error of NPU xattention. #1303

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
44 changes: 21 additions & 23 deletions xllm/core/runtime/rec_worker_impl.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -929,17 +929,8 @@ std::optional<ForwardOutput> RecWorkerImpl::LlmRecMultiRoundPipeline::step(
<< ")";

#if defined(USE_NPU)
if (round == 0) {
// NPU beam_search_rec prefill only accepts top_k == 1. Flatten the
// sampler's per-request top-k output into one candidate per beam so
// the kernel can seed sequence_group without discarding beam diversity.
top_tokens =
sample_output.top_tokens.to(torch::kInt32).reshape({-1, 1});
top_logprobs = sample_output.top_logprobs.reshape({-1, 1});
} else {
top_tokens = sample_output.top_tokens.to(torch::kInt32);
top_logprobs = sample_output.top_logprobs;
}
top_tokens = sample_output.top_tokens.to(torch::kInt32);
top_logprobs = sample_output.top_logprobs;
#else
top_tokens = sample_output.top_tokens.to(torch::kInt32)
.reshape({-1, step_meta->beam_width});
Expand Down Expand Up @@ -997,18 +988,25 @@ void RecWorkerImpl::LlmRecMultiRoundPipeline::execute_beam_search(
int32_t round,
int32_t batch_size) {
#if defined(USE_NPU)
xllm::kernel::npu::beam_search_rec(
/*logprobs=*/beam_tensors.acc_logprob,
/*top_tokens=*/top_tokens.to(torch::kInt32),
/*top_logprobs=*/top_logprobs,
/*sequence_group=*/beam_tensors.sequence_group,
/*current_step=*/static_cast<int64_t>(round),
/*out_token_ids=*/beam_tensors.out_token_ids,
/*out_token_index=*/beam_tensors.out_token_index,
/*out_log_probs=*/beam_tensors.out_log_probs,
/*out_beam_count_prefix_sums=*/
beam_tensors.out_beam_count_prefix_sums,
/*out_sequence=*/beam_tensors.out_seqgroup);
if (round == 0) {
beam_tensors.out_token_ids.copy_(top_tokens.reshape({-1, 1}));
beam_tensors.out_log_probs.copy_(top_logprobs.reshape({-1, 1}));
beam_tensors.out_seqgroup.select(/*dim=*/2, /*index=*/0)
.copy_(top_tokens.reshape({-1}));
Comment thread
LMX-xin marked this conversation as resolved.
} else {
xllm::kernel::npu::beam_search_rec(
/*logprobs=*/beam_tensors.acc_logprob,
/*top_tokens=*/top_tokens,
/*top_logprobs=*/top_logprobs,
/*sequence_group=*/beam_tensors.sequence_group,
/*current_step=*/static_cast<int64_t>(round),
/*out_token_ids=*/beam_tensors.out_token_ids,
/*out_token_index=*/beam_tensors.out_token_index,
/*out_log_probs=*/beam_tensors.out_log_probs,
/*out_beam_count_prefix_sums=*/
beam_tensors.out_beam_count_prefix_sums,
/*out_sequence=*/beam_tensors.out_seqgroup);
}
#elif defined(USE_CUDA)
xllm::kernel::cuda::beam_search(beam_tensors.acc_logprob,
beam_tensors.sequence_group,
Expand Down
Loading