Add tests for custom operator implementation correctness #457
Conversation
Signed-off-by: Kacper Pietkun <[email protected]>
Kacper-Pietkun
requested review from
adobrzyn,
afierka-intel,
iboiko-habana,
kzawora-intel,
mgawarkiewicz-intel,
michalkuligowski,
mswiniarsk,
vivekgoe and
xuechendi
as code owners
|
/run-gaudi-tests |
Signed-off-by: Kacper Pietkun <[email protected]>
|
/run-gaudi-tests |
✅ CI PassedAll checks passed successfully against the following vllm commit: |
|
/run-gaudi-tests |
✅ CI PassedAll checks passed successfully against the following vllm commit: |
Signed-off-by: Kacper Pietkun <[email protected]>
🚧 CI BlockedThe main CI workflow was not started for the following reason:
|
|
/run-gaudi-tests |
✅ CI PassedAll checks passed successfully against the following vllm commit: |
✅ CI PassedAll checks passed successfully against the following vllm commit: |
There was a problem hiding this comment.
Pull Request Overview
This PR adds comprehensive unit tests for custom operators implemented in vllm_gaudi/ops. The tests verify correctness by comparing outputs between native VLLM operators and HPU-specific implementations. For operators using CUDA kernels, pre-computed reference tensors stored in safetensors format are used for validation.
Key changes:
- Native and HPU operator outputs are compared for operators compatible with both implementations
- Pre-computed reference data in safetensors format is used for CUDA kernel-based operators
- Test utilities added for temporary operator registry management and test data access
Reviewed Changes
Copilot reviewed 11 out of 19 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| vllm_gaudi/ops/hpu_fp8.py | Removed unused imports and decorator |
| tests/unit_tests/test_bucketing.py | Added config clearing before setup |
| tests/unit_tests/ops/utils.py | Added test utilities for operator registration and data loading |
| tests/unit_tests/ops/test_hpu_rotary_embedding.py | Tests for rotary embedding operator variants |
| tests/unit_tests/ops/test_hpu_multihead_attn.py | Tests for multi-head attention operator |
| tests/unit_tests/ops/test_hpu_layernorm.py | Tests for RMS normalization operator |
| tests/unit_tests/ops/test_hpu_gptq.py | Tests for GPTQ quantization operator |
| tests/unit_tests/ops/test_hpu_fused_moe.py | Tests for fused MoE operator |
| tests/unit_tests/ops/test_hpu_fp8.py | Tests for FP8 quantization operators |
| tests/unit_tests/ops/test_hpu_compressed_tensors.py | Tests for compressed tensor operators |
| tests/unit_tests/ops/test_hpu_awq.py | Tests for AWQ quantization operator |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Co-authored-by: Copilot <[email protected]> Signed-off-by: Kacper Pietkun <[email protected]>
Co-authored-by: Copilot <[email protected]> Signed-off-by: Kacper Pietkun <[email protected]>
Co-authored-by: Copilot <[email protected]> Signed-off-by: Kacper Pietkun <[email protected]>
Co-authored-by: Copilot <[email protected]> Signed-off-by: Kacper Pietkun <[email protected]>
Co-authored-by: Copilot <[email protected]> Signed-off-by: Kacper Pietkun <[email protected]>
|
All of the above changes are just corrections of spelling mistakes detected by copilot |
✅ CI PassedAll checks passed successfully against the following vllm commit: |
2 participants
I added tests for custom ops defined in
vllm_gaudi/ops: