Skip to content
Open
Changes from all commits
Commits
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
106 changes: 106 additions & 0 deletions reports/xfails_report.txt
Original file line number Diff line number Diff line change
@@ -0,0 +1,106 @@
Scanning for test files in /home/runner/work/flashinfer/flashinfer/tests...
Found 125 test files
Collecting xfail markers...
====================================================================================================
XFAILS REPORT
====================================================================================================

Total xfails: 11
Unique reasons: 9


----------------------------------------------------------------------------------------------------
Reason Count Type
----------------------------------------------------------------------------------------------------
Expected failure for SM120/121 for now since the tile s... 2 decorator
NOTE(Zihao): temporarily disable cuda graph until we fu... 2 runtime
NOTE(Zihao): attention sink with sliding window and non... 1 runtime
seq_len=514 is known to fail in full test suite occasio... 1 parameter
nvidia-cutlass-dsl has issue when l=1 1 runtime
str(e) 1 runtime
state_dtype=float32 not yet supported (only float16/bfl... 1 decorator
Note(jimmzhou): Make MxFP4xBf16 nonfunctional on SM103 ... 1 runtime
Numerical accuracy issue on SM 121 (Spark) 1 decorator
----------------------------------------------------------------------------------------------------


====================================================================================================
DETAILED BREAKDOWN BY REASON
====================================================================================================

[2 xfails] Expected failure for SM120/121 for now since the tile size/number of stages is too large.
----------------------------------------------------------------------------------------------------
• tests/attention/test_batch_attention.py:211
Test: test_batch_attention_with_noncontiguous_q
Type: decorator
Condition: get_compute_capability(torch.device(device='cuda'))[0] == 12

• tests/attention/test_batch_attention.py:260
Test: test_batch_attention_correctness
Type: decorator
Condition: get_compute_capability(torch.device(device='cuda'))[0] == 12


[2 xfails] NOTE(Zihao): temporarily disable cuda graph until we fully fix the workspace buffer overflow issue for prefill + cudagraph
----------------------------------------------------------------------------------------------------
• tests/attention/test_batch_prefill_kernels.py:81
Test: test_batch_prefill_with_paged_kv_cache
Type: runtime

• tests/attention/test_batch_prefill_kernels.py:321
Test: test_batch_prefill_with_tuple_paged_kv_cache
Type: runtime


[1 xfails] NOTE(Zihao): attention sink with sliding window and non-causal will fail after https://github.com/flashinfer-ai/flashinfer/pull/1661, temporarily xfail the test.
----------------------------------------------------------------------------------------------------
• tests/attention/test_attention_sink.py:643
Test: test_attention_sink_chunk_prefill
Type: runtime


[1 xfails] seq_len=514 is known to fail in full test suite occasionally
----------------------------------------------------------------------------------------------------
• tests/attention/test_xqa.py:138
Test: test_xqa
Type: parameter
Strict: False


[1 xfails] nvidia-cutlass-dsl has issue when l=1
----------------------------------------------------------------------------------------------------
• tests/gemm/test_cute_dsl_blockscaled_gemm.py:93
Test: test_blockscaled_gemm_python_interface
Type: runtime


[1 xfails] str(e)
----------------------------------------------------------------------------------------------------
• tests/gemm/test_mm_fp4.py:97
Test: _test_mm_fp4
Type: runtime
Comment on lines +77 to +81
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The xfail reason str(e) is not descriptive and makes the report less useful. This is caused by the use of pytest.xfail(str(e)) in tests/gemm/test_mm_fp4.py on line 92. The static analysis script that generates this report cannot resolve the value of str(e) and therefore uses the literal code as the reason.

To improve this, the pytest.xfail call should be updated to use a descriptive string literal. Based on the context in the test file, a more informative reason would be something like:

"cuDNN backend for FP4 GEMM with mxfp4 on SM120+ requires cuDNN 9.14.0+"

This change would make the xfail report more understandable and aid in tracking test failures.



[1 xfails] state_dtype=float32 not yet supported (only float16/bfloat16)
----------------------------------------------------------------------------------------------------
• tests/mamba/test_chunk_scan_combined.py:1989
Test: test_fp32_state_dtype
Type: decorator
Strict: True


[1 xfails] Note(jimmzhou): Make MxFP4xBf16 nonfunctional on SM103 to avoid B200 regression
----------------------------------------------------------------------------------------------------
• tests/moe/utils.py:158
Test: skip_checks
Type: runtime


[1 xfails] Numerical accuracy issue on SM 121 (Spark)
----------------------------------------------------------------------------------------------------
• tests/utils/test_jit_example.py:173
Test: test_dump_logits
Type: decorator
Condition: get_compute_capability(torch.device('cuda:0')) == (12, 1)

====================================================================================================
Loading