[FlexAttention] `PassManager run failed` during running FlexAttention UT

[hoshibara]

Describe the bug

After updating intel-xpu-backend-for-triton to 2be3060b74b56ee5c2a57c7170f37c5b737bbd36, we found some new errors in test_flex_attention.py that we haven't seen before.

Error: RuntimeError: PassManager::run failed
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_backward_stride_ordering_mode_inductor_permute_order0_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_backward_stride_ordering_mode_inductor_permute_order0_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_backward_stride_ordering_mode_inductor_permute_order1_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_backward_stride_ordering_mode_inductor_permute_order1_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_backward_stride_ordering_mode_inductor_permute_order2_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_backward_stride_ordering_mode_inductor_permute_order2_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_backward_stride_ordering_mode_inductor_permute_order3_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_backward_stride_ordering_mode_inductor_permute_order3_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_inductor_permute_order0_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_inductor_permute_order0_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_inductor_permute_order1_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_inductor_permute_order1_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_inductor_permute_order2_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_inductor_permute_order2_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_inductor_permute_order3_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_inductor_permute_order3_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_paged_attention_permute_order0_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_paged_attention_permute_order0_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_paged_attention_permute_order1_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_paged_attention_permute_order1_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_paged_attention_permute_order2_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_paged_attention_permute_order2_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_paged_attention_permute_order3_shape0
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_flex_attention_stride_ordering_mode_paged_attention_permute_order3_shape1
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_force_write_lse
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_load_from_bias_head_seq_batch_float16
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_load_from_bias_seq_batch_float16
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_load_from_bias_seq_only_float16
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_new_empty_mask_mod
  - python test/inductor/test_flex_attention.py TestFlexAttention.test_small_q_kv_len

Here's the log:

ERROR: test_new_empty_mask_mod (__main__.TestFlexAttention)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/testing/_internal/common_utils.py", line 3154, in wrapper
    method(*args, **kwargs)
  File "/workspace1/xingyuan/20250213-flexatt-enable/ut-test/../hoshibara-pytorch/test/inductor/test_flex_attention.py", line 2173, in test_new_empty_mask_mod
    torch.compile(flex_attention)(q, k, v, score_mod, block_mask=block_mask)
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_dynamo/eval_frame.py", line 671, in _fn
    raise e.remove_dynamo_frames() from None  # see TORCHDYNAMO_VERBOSE=1
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/compile_fx.py", line 768, in _compile_fx_inner
    raise InductorError(e, currentframe()).with_traceback(
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/compile_fx.py", line 753, in _compile_fx_inner
    mb_compiled_graph = fx_codegen_and_compile(
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/compile_fx.py", line 1357, in fx_codegen_and_compile
    return scheme.codegen_and_compile(gm, example_inputs, inputs_to_check, graph_kwargs)
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/compile_fx.py", line 1246, in codegen_and_compile
    compiled_module = graph.compile_to_module()
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/graph.py", line 2204, in compile_to_module
    return self._compile_to_module()
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/graph.py", line 2251, in _compile_to_module
    mod = PyCodeCache.load_by_key_path(
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/codecache.py", line 2893, in load_by_key_path
    mod = _reload_python_module(key, path, set_sys_modules=in_toplevel)
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/runtime/compile_tasks.py", line 31, in _reload_python_module
    exec(code, mod.__dict__, mod.__dict__)
  File "/tmp/tmpa4kow6_4/y2/cy2b4oaygkiyu6z65bqropo4q4ork45pp73yitw44piduxb3fdgo.py", line 519, in <module>
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/async_compile.py", line 446, in wait
    self._wait_futures(scope)
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/async_compile.py", line 466, in _wait_futures
    scope[key] = result.result()
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/codecache.py", line 3395, in result
    return self.result_fn()
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/async_compile.py", line 336, in get_result
    kernel, elapsed_us = task.result()
  File "/home/sdp/miniforge3/envs/xingyuan-flex-attention-enable-20250214/lib/python3.10/concurrent/futures/_base.py", line 458, in result
    return self.__get_result()
  File "/home/sdp/miniforge3/envs/xingyuan-flex-attention-enable-20250214/lib/python3.10/concurrent/futures/_base.py", line 403, in __get_result
    raise self._exception
torch._inductor.exc.InductorError: SubprocException: An exception occurred in a subprocess:

Traceback (most recent call last):
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/compile_worker/subproc_pool.py", line 323, in do_job
    result = job()
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/runtime/compile_tasks.py", line 57, in _worker_compile_triton
    kernel.precompile(warm_cache_only=True)
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/runtime/triton_heuristics.py", line 311, in precompile
    self._precompile_worker()
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/runtime/triton_heuristics.py", line 342, in _precompile_worker
    compile_results.append(self._precompile_config(c))
  File "/workspace1/xingyuan/20250213-flexatt-enable/hoshibara-pytorch/torch/_inductor/runtime/triton_heuristics.py", line 623, in _precompile_config
    binary = triton.compile(*compile_args, **compile_kwargs)
  File "/workspace1/xingyuan/20250213-flexatt-enable/intel-xpu-backend-for-triton/python/triton/compiler/compiler.py", line 287, in compile
    next_module = compile_ir(module, metadata)
  File "/workspace1/xingyuan/20250213-flexatt-enable/intel-xpu-backend-for-triton/python/triton/backends/intel/compiler.py", line 452, in <lambda>
    stages["llir"] = lambda src, metadata: self.make_llir(src, metadata, options)
  File "/workspace1/xingyuan/20250213-flexatt-enable/intel-xpu-backend-for-triton/python/triton/backends/intel/compiler.py", line 358, in make_llir
    pm.run(mod)
RuntimeError: PassManager::run failed


Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"


To execute this test, run the following from the base repo dir:
    python test/inductor/test_flex_attention.py TestFlexAttention.test_new_empty_mask_mod

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

Sometimes this error can be resolved by manually modifying the configs.
https://github.com/pytorch/pytorch/blob/24dc060a8ae5a57cce4c4a2c8b58ef5bb5303a11/torch/_inductor/kernel/flex_attention.py#L780

Environment details

You can setup test env according this issue:
#3518

PyTorch version: 2.8.0a0+git24dc060
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.31.6
Libc version: glibc-2.35

Python version: 3.10.16 | packaged by conda-forge | (main, Dec  5 2024, 14:16:10) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-5.15.50-051550-generic-x86_64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Address sizes:                   52 bits physical, 57 bits virtual
Byte Order:                      Little Endian
CPU(s):                          224
On-line CPU(s) list:             0-223
Vendor ID:                       GenuineIntel
Model name:                      Intel(R) Xeon(R) Platinum 8480+
CPU family:                      6
Model:                           143
Thread(s) per core:              2
Core(s) per socket:              56
Socket(s):                       2
Stepping:                        6
CPU max MHz:                     3800.0000
CPU min MHz:                     800.0000
BogoMIPS:                        4000.00
Flags:                           fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr avx512_fp16 flush_l1d arch_capabilities
Virtualization:                  VT-x
L1d cache:                       5.3 MiB (112 instances)
L1i cache:                       3.5 MiB (112 instances)
L2 cache:                        224 MiB (112 instances)
L3 cache:                        210 MiB (2 instances)
NUMA node(s):                    2
NUMA node0 CPU(s):               0-55,112-167
NUMA node1 CPU(s):               56-111,168-223
Vulnerability Itlb multihit:     Not affected
Vulnerability L1tf:              Not affected
Vulnerability Mds:               Not affected
Vulnerability Meltdown:          Not affected
Vulnerability Mmio stale data:   Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:        Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:        Mitigation; Enhanced IBRS, IBPB conditional, RSB filling
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected

Versions of relevant libraries:
[pip3] flake8==6.1.0
[pip3] flake8-bugbear==23.3.23
[pip3] flake8-comprehensions==3.15.0
[pip3] flake8-executable==2.1.3
[pip3] flake8-logging-format==0.9.0
[pip3] flake8-pyi==23.3.1
[pip3] flake8-simplify==0.19.3
[pip3] mypy==1.14.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] optree==0.13.0
[pip3] torch==2.8.0a0+git24dc060
[pip3] torchaudio==2.6.0.dev20250311+xpu
[pip3] torchinfo==1.8.0
[pip3] torchvision==0.22.0.dev20250311+xpu
[pip3] triton==3.3.0+git2be3060b
[conda] mkl-include               2025.0.1            hf2ce2f3_20    conda-forge
[conda] mkl-static                2025.0.1            ha770c72_20    conda-forge
[conda] numpy                     1.26.4                   pypi_0    pypi
[conda] optree                    0.13.0                   pypi_0    pypi
[conda] torch                     2.8.0a0+git24dc060           dev_0    <develop>
[conda] torchaudio                2.6.0.dev20250311+xpu          pypi_0    pypi
[conda] torchfix                  0.4.0                    pypi_0    pypi
[conda] torchinfo                 1.8.0                    pypi_0    pypi
[conda] torchvision               0.22.0.dev20250311+xpu          pypi_0    pypi
[conda] triton                    3.3.0+git2be3060b           dev_0    <develop>

[yudongsi]

Error: LLVM ERROR: Invalid basis 16 for in-dim 'iteration' and out-dim 'dim0'. Basis must be less than the out-dim size.

87b1723

[alexbaden]

@chengjunlu did you fix this? Looks like an issue with the linear layout code.

[alexbaden]

This doesn't reproduce for me using the pytorch flex attention patch -

» python ../pytorch/test/inductor/test_flex_attention.py TestFlexAttentionXPU.test_new_empty_mask_mod_xpu                                                                                                                                  
E0422 01:11:38.642000 694192 torch/_inductor/codegen/cuda/cuda_env.py:22] Error getting cuda arch: Torch not compiled with CUDA enabled
frames [('total', 1), ('ok', 1)]
inline_call []
stats [('calls_captured', 2), ('unique_graphs', 1)]
inductor [('triton_bundler_save_kernel', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1)]
aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
graph_break []
.
----------------------------------------------------------------------
Ran 1 test in 15.393s

OK

What am I doing wrong?

[yudongsi]

I repro this by build source torch commit 24dc060.

[alexbaden]

Can you run with TRITON_LLVM_DEBUG_ONLY=ttgpu_to_llvm and paste the output? And also provide the cached ttir/ttgir.

[alexbaden]

And note that as a temporary workaround you can run with TRITON_XPU_ENABLE_TILE_LOAD_LINEAR_LAYOUT=0.

[yudongsi]

utlog.txt

triton_tem_fused_4.ttgir.txt
triton_tem_fused_4.ttir.txt