mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-12 05:10:18 +08:00
KurisandClaude Opus 4.6
6e6295fb35
1. Add CI step to clean stale JIT temp files (/tmp/*.so, *.cu, *.cubin, tvm-debug-mode-tempdirs) before tests on self-hosted runners. These files accumulate across CI runs and can fill the disk, causing g++ to be killed (SIGTERM) during JIT compilation. 2. Skip CuTeDSL-incompatible example tests that use alloc_global (flash_decoding, deepseek_mla), since the CuTeDSL wrapper does not yet support alloc_global buffers. Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
30 lines
860 B
Python
30 lines
860 B
Python
import os
|
|
import pytest
|
|
import tilelang.testing
|
|
|
|
import example_gqa_decode
|
|
import example_mha_inference
|
|
import example_gqa_decode_varlen_logits
|
|
|
|
_is_cutedsl = os.environ.get("TILELANG_TARGET", "").lower() == "cutedsl"
|
|
|
|
|
|
@tilelang.testing.requires_cuda
|
|
@tilelang.testing.requires_cuda_compute_version_le(8, 9)
|
|
@pytest.mark.skipif(_is_cutedsl, reason="CuTeDSL backend does not support alloc_global yet")
|
|
def test_example_example_gqa_decode():
|
|
example_gqa_decode.main()
|
|
|
|
|
|
@pytest.mark.skipif(_is_cutedsl, reason="CuTeDSL backend does not support alloc_global yet")
|
|
def test_example_example_mha_inference():
|
|
example_mha_inference.main(BATCH=1, H=32, Q_CTX=128, KV_CTX=2048, D_HEAD=128, causal=False)
|
|
|
|
|
|
def test_example_example_gqa_decode_varlen_logits():
|
|
example_gqa_decode_varlen_logits.main()
|
|
|
|
|
|
if __name__ == "__main__":
|
|
tilelang.testing.main()
|