A high-throughput and memory-efficient inference and serving engine for LLMs
[XPU] Fix test_spec_decode_logprobs: use FLASH_ATTN for XPU in GPU_DETERMINISM_KWARGS (#44468)
Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Chaojun Zhang committed
Jun 17, 2026 at 11:07 UTC
556b063e45a7f36d917ad6da03721e5f3fb78848
2 files changed
+9
-7
.buildkite/intel_jobs/misc_intel.yaml
+2
-1
@@ -63,7 +63,8 @@ steps:
63
pytest -v -s v1/test_oracle.py &&
64
pytest -v -s v1/test_request.py &&
65
pytest -v -s v1/test_outputs.py &&
66
- pytest -v -s v1/sample/test_topk_topp_sampler.py'
66
+ pytest -v -s v1/sample/test_topk_topp_sampler.py &&
67
+ pytest -v -s v1/sample/test_logprobs.py'
68
69
- label: XPU CPU Offload
70
timeout_in_minutes: 60
tests/v1/sample/test_logprobs.py
+7
-6
@@ -42,11 +42,12 @@ SAMPLE_PROMPT = BatchLogprobsComposition.SAMPLE_PROMPT
42
#
43
# Force LLM instances into an identical, deterministic execution
44
# mode so the test isolates spec-decode correctness only:
45
-ROCM_DETERMINISM_KWARGS: dict = (
46
- dict(max_num_seqs=1, attention_backend="TRITON_ATTN")
47
- if current_platform.is_rocm()
48
- else {}
49
-)
45
+if current_platform.is_rocm():
46
+ GPU_DETERMINISM_KWARGS: dict = dict(max_num_seqs=1, attention_backend="TRITON_ATTN")
47
+elif current_platform.is_xpu():
48
+ GPU_DETERMINISM_KWARGS = dict(max_num_seqs=1, attention_backend="FLASH_ATTN")
49
+else:
50
+ GPU_DETERMINISM_KWARGS = {}
51
52
53
@pytest.fixture(
@@ -1127,7 +1128,7 @@ def test_spec_decode_logprobs(
1128
enable_chunked_prefill=True,
1129
max_num_batched_tokens=32,
1130
enable_prefix_caching=False,
1130
- **ROCM_DETERMINISM_KWARGS,
1131
+ **GPU_DETERMINISM_KWARGS,
1132
)
1133
1134
# Run base LLM.