Skip to content

[https://nvbugs/5970614][fix] Sync CTA before PDL trigger in quantize_with_block_size #14668

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
tianyuxbear wants to merge 2 commits into
base: main
Choose a base branch
from
+17 −2
Open

[https://nvbugs/5970614][fix] Sync CTA before PDL trigger in quantize_with_block_size #14668

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
61e4add
[https://nvbugs/6018046][fix] Lower throughput_pp4_mtp max_batch_size…
tianyuxbear May 26, 2026
f8d3481
[https://nvbugs/5970614][fix] Sync CTA before PDL trigger in quantize…
tianyuxbear May 26, 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
9 changes: 9 additions & 0 deletions cpp/tensorrt_llm/kernels/quantization.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -897,6 +897,15 @@ quantize_with_block_size(
}
}
}
// Fix for nvbugs/5970614 (https://nvbugspro.nvidia.com/bug/5970614).
// PDL completion is reported when every CTA has either exited or called
// this function at least once (per CUDA Programming Guide). Without a
// CTA-wide barrier, an early-finishing warp can trigger completion while
// other warps in the same CTA are still writing sf_out / out, allowing the
// downstream NVF4 GEMM consumer to read partial data once
// wait_on_dependent_grids returns. Drain the CTA's stores before trigger.
__syncthreads();
__threadfence();
cudaTriggerProgrammaticLaunchCompletion();
#endif
}
Expand Down
9 changes: 8 additions & 1 deletion tests/integration/defs/accuracy/test_llm_api_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2902,6 +2902,13 @@ class TestDeepSeekR1(LlmapiAccuracyTestHarness):
8,
"CUTLASS",
marks=pytest.mark.skip_less_mpi_world_size(8)),
# throughput_pp4_mtp (NVBug 6018046): on 4-GPU PP=4 + MTP, the
# per-bs CUDA-graph snapshot pool plus per-step activation and
# NCCL collective buffers consume the lazy cuBLAS Lt workspace
# headroom at bs=32, surfacing as mid-run
# CUBLAS_STATUS_EXECUTION_FAILED. Match the already-stable
# throughput_bs8_mtp configuration with bs=8; the default 0.70
# KV fraction is sufficient once bs is reduced.
pytest.param(1,
4,
1,
Expand All @@ -2911,7 +2918,7 @@ class TestDeepSeekR1(LlmapiAccuracyTestHarness):
False,
True,
True,
32,
8,
"CUTLASS",
marks=pytest.mark.skip_less_mpi_world_size(4)),
],
Expand Down
1 change: 0 additions & 1 deletion tests/integration/test_lists/waives.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,6 @@ accuracy/test_llm_api_autodeploy.py::TestQwen3_5_397B_MoE::test_bf16_small[4] SK
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale[throughput_mtp] SKIP (https://nvbugs/6215736)
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput] SKIP (https://nvbugs/6084775)
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_mtp] SKIP (https://nvbugs/6029882)
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_pp4_mtp] SKIP (https://nvbugs/6018046)
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_tp4] SKIP (https://nvbugs/6215793)
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1LongBenchV2::test_fp8_8gpus SKIP (https://nvbugs/6193778)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV32::test_dsa_host_cache_offload[host_cache_offload] SKIP (https://nvbugs/6185196)
Expand Down
Loading