[DEBUG] Try run FlexAttn in parallel
#4376
Draft
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Anatoly Myachev <[email protected]>
anmyachev
changed the title
FlexAttn in parallelFlexAttn in parallel
|
Good results! We can add a dedicated workflow for Flex Attention. The number of workers needs to be a parameter to adjust it for client GPUs. Also it would be nice to identify the root cause for failures, looks like accuracy errors in most cases, not sure if it is due to parallelism. |
|
Part of these problems I see in our usual run https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/15373802137/job/43256311378#step:8:14641: The following tests failed consistently: ['test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s0_v_s0_do_s0_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s0_v_s0_do_s1_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s0_v_s0_do_s2_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s1_v_s1_do_s0_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s1_v_s1_do_s1_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s1_v_s1_do_s2_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s2_v_s2_do_s0_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s2_v_s2_do_s1_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s2_v_s2_do_s2_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s3_v_s3_do_s0_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s3_v_s3_do_s1_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s0_k_s3_v_s3_do_s2_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s0_v_s0_do_s0_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s0_v_s0_do_s1_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s0_v_s0_do_s2_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s1_v_s1_do_s0_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s1_v_s1_do_s1_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s1_v_s1_do_s2_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s2_v_s2_do_s0_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s2_v_s2_do_s1_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s2_v_s2_do_s2_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s3_v_s3_do_s0_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s3_v_s3_do_s1_xpu_float16', 'test/inductor/test_flex_attention.py::TestFlexAttentionXPU::test_strided_inputs_q_s1_k_s3_v_s3_do_s2_xpu_float16', 'test/inductor/test_flex_attention.py::TestLearnableBiasesXPU::test_head_specific_gate_batch:2_head:4_seq_len:256_headdim:16_dtype:float16_mode_default_xpu', 'test/inductor/test_flex_attention.py::TestLearnableBiasesXPU::test_head_specific_gate_batch:2_head:4_seq_len:256_headdim:16_dtype:float16_mode_max-autotune-no-cudagraphs_xpu', 'test/inductor/test_flex_attention.py::TestLearnableBiasesXPU::test_head_specific_gate_batch:2_head:4_seq_len:277_headdim:16_dtype:float16_mode_default_xpu', 'test/inductor/test_flex_attention.py::TestLearnableBiasesXPU::test_head_specific_gate_batch:2_head:4_seq_len:277_headdim:16_dtype:float16_mode_max-autotune-no-cudagraphs_xpu', 'test/inductor/test_flex_attention.py::TestLearnableBiasesXPU::test_head_specific_gate_batch:2_head:4_seq_len:37_headdim:16_dtype:float16_mode_default_xpu', 'test/inductor/test_flex_attention.py::TestLearnableBiasesXPU::test_head_specific_gate_batch:2_head:4_seq_len:37_headdim:16_dtype:float16_mode_max-autotune-no-cudagraphs_xpu', 'test/inductor/test_flex_attention.py::TestLearnableBiasesXPU::test_relative_1d_bias_batch:2_head:4_seq_len:256_headdim:16_dtype:float32_mode_max-autotune-no-cudagraphs_xpu'] |
@chengjunlu investigated above failures before, and believe they are issues in SYCL. |
anmyachev
commented
Jun 2, 2025
|
Looks much better with last PyTorch pin update (https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/15416558343/job/43380600584): This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
====== 7 failed, 633 passed, 15 skipped, 1 xfailed in 2571.39s (0:42:51) ======= |
|
If you get intermittent failures you can try to rerun the failed tests (in parallel again, or perhaps sequentially). If you rerun in parallel you may have to rerun again... |
david-hls
pushed a commit
to david-hls/intel-xpu-backend-for-triton
that referenced
this pull request
Jun 18, 2025
…el#4376) There was no test coverage for this. I discovered this while implementing the CPU backend.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PyTorch CI:
-n 16)-n 8 --reruns 2, py3.9)n 8 --reruns 2, py3.10)-n 32then usual run)-n 16then usual run)-n 16 --reruns 2)n 16 --reruns 2with new PyTorch pin)Seems quite fast (without decoding). Before this, the tests were running in one process, apparently.
This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0 ===== 101 failed, 496 passed, 58 skipped, 1 xfailed in 2611.01s (0:43:31) ======Although I see a lot of errors, perhaps due to parallelism. We can experiment and choose the most successful combination. (It might also be good to enable restart in case of errors)
FYI @pbchekin @vlad-penkin @alexbaden @chengjunlu @whitneywhtsang @etiotto