Skip to content

[bugfix]Correct chunk_end calculation in multi-CTA collaboration when max_len > length #2489

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
yzh119 merged 4 commits into from
Feb 20, 2026
Merged

[bugfix]Correct chunk_end calculation in multi-CTA collaboration when max_len > length #2489

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
68087b6
bugfix
Feb 4, 2026
307e7c6
Merge branch 'flashinfer-ai:main' into bugfix_2486
huangzhilin-hzl Feb 9, 2026
bd9b3fa
fix
Feb 9, 2026
8a78d18
Merge branch 'main' into bugfix_2486
yzh119 Feb 18, 2026
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
6 changes: 4 additions & 2 deletions include/flashinfer/topk.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -890,7 +890,9 @@ __global__ void __launch_bounds__(BLOCK_THREADS) RadixTopKKernel_Unified(
// k >= vocab_size: return all indices
const uint32_t chunk_start = cta_in_group * chunk_size;
const uint32_t chunk_end = min(chunk_start + chunk_size, length);
for (uint32_t i = tx; i < chunk_end - chunk_start; i += BLOCK_THREADS) {
const uint32_t actual_chunk_size = ((chunk_start < length) ? (chunk_end - chunk_start) : 0);

for (uint32_t i = tx; i < actual_chunk_size; i += BLOCK_THREADS) {
if (chunk_start + i < k) {
row_output[chunk_start + i] = static_cast<IdType>(chunk_start + i);
output_values[row_idx * top_k_val + chunk_start + i] =
Expand Down Expand Up @@ -950,7 +952,7 @@ __global__ void __launch_bounds__(BLOCK_THREADS) RadixTopKKernel_Unified(

const uint32_t chunk_start = cta_in_group * chunk_size;
const uint32_t chunk_end = min(chunk_start + chunk_size, length);
const uint32_t actual_chunk_size = chunk_end - chunk_start;
const uint32_t actual_chunk_size = ((chunk_start < length) ? (chunk_end - chunk_start) : 0);

// Stage 1: Load and convert to ordered representation
LoadToSharedOrdered<BLOCK_THREADS, VEC_SIZE, DType, Traits>(
Expand Down
2 changes: 1 addition & 1 deletion tests/utils/test_topk.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -592,7 +592,7 @@ def test_page_table_transform_with_row_to_batch(num_rows, max_len, k):


@pytest.mark.parametrize("num_rows", [4, 16])
@pytest.mark.parametrize("max_len", [4096])
@pytest.mark.parametrize("max_len", [4096, 131072])
@pytest.mark.parametrize("k", [256])
def test_page_table_transform_variable_lengths(num_rows, max_len, k):
"""Test page table transform with variable sequence lengths."""
Expand Down
Loading