Skip to content

[https://nvbugs/6426860][fix] Stabilize compressor BF16 tolerance #16376

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
lfr-0531 merged 1 commit into from
Jul 15, 2026
Merged

[https://nvbugs/6426860][fix] Stabilize compressor BF16 tolerance #16376

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
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 @@ -424,7 +424,6 @@ test_e2e.py::test_multi_nodes_eval[Kimi-K2-Thinking-NVFP4-tp16-mmlu] SKIP (https
test_e2e.py::test_multi_nodes_eval[MiniMax-M2-tp16-mmlu] SKIP (https://nvbugs/6373532)
test_e2e.py::test_multi_nodes_eval[MiniMax-M3-tp16-mmlu] SKIP (https://nvbugs/6373561)
test_e2e.py::test_ptp_quickstart_advanced_deepseek_r1_w4afp8_8gpus[DeepSeek-R1-W4AFP8-DeepSeek-R1/DeepSeek-R1-W4AFP8] SKIP (https://nvbugs/5836830)
unittest/_torch/attention/sparse/deepseek_v4/test_compressor_kernel.py::test_prefill_varlen[varlen_hd512_overlap] SKIP (https://nvbugs/6426860)
unittest/_torch/misc/test_share_tensor.py::TestShareTensor::test_share_tensor_different_dtypes SKIP (https://nvbugs/6418021)
unittest/_torch/modeling -k "modeling_qwen" SKIP (https://nvbugs/6433376)
unittest/_torch/modeling/test_modeling_qwen3_5_vl.py::test_qwen35_dense_vl_resolves_mamba_ssm_cache_dtype SKIP (https://nvbugs/6433376)
Expand Down
4 changes: 3 additions & 1 deletion tests/unittest/_torch/attention/sparse/deepseek_v4/test_compressor_kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1223,8 +1223,10 @@ def test_prefill_varlen(seq_lens_list, compress_ratio, head_dim, overlap):

if valid_outputs:
out_kernel_valid = torch.cat(valid_outputs, dim=0)
# The CUDA and PyTorch reductions can round to adjacent BF16 values near a tie.
output_rtol = max(2e-3, torch.finfo(out_kernel_valid.dtype).eps)
assert torch.allclose(
out_py.to(out_kernel_valid.dtype), out_kernel_valid, rtol=2e-3, atol=5e-3
out_py.to(out_kernel_valid.dtype), out_kernel_valid, rtol=output_rtol, atol=5e-3
), (
f"Output mismatch: max diff = {(out_py.to(out_kernel_valid.dtype) - out_kernel_valid).abs().max():.6f}"
)
Expand Down
Loading