A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 07:15 UTC
[XPU] Fix spec-decode UTs under tests/v1/spec_decode (#38491)
Signed-off-by: Yan Ma <yan.ma@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Yan Ma committed
Apr 11, 2026 at 09:31 UTC
0f3ce4c74b1875791d6604e006b6e905fde9f698
3 files changed
+2
-14
tests/v1/spec_decode/test_eagle.py
-6
@@ -755,12 +755,6 @@ def test_load_model(
755
use_distinct_lm_head,
756
monkeypatch,
757
):
758
- if attn_backend == "TRITON_ATTN" and not current_platform.is_rocm():
759
- pytest.skip(
760
- "TRITON_ATTN does not support "
761
- "multi-token eagle spec decode on current platform"
762
- )
763
-
758
if attn_backend == "ROCM_AITER_FA" and current_platform.is_rocm():
759
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
760
tests/v1/spec_decode/test_eagle_step_kernel.py
+2
-2
@@ -15,8 +15,8 @@ DEVICE_TYPE = current_platform.device_type
15
16
# Skip if no CUDA - Triton kernel requires GPU
17
pytest.importorskip("triton")
18
-if not torch.cuda.is_available():
19
- pytest.skip("CUDA required for EAGLE kernel tests", allow_module_level=True)
18
+if not current_platform.is_cuda_alike() and not current_platform.is_xpu():
19
+ pytest.skip("CUDA/XPU required for EAGLE kernel tests", allow_module_level=True)
20
21
22
def _reference_eagle_step_slot_mapping(
tests/v1/spec_decode/test_max_len.py
-6
@@ -38,12 +38,6 @@ def test_ngram_max_len(num_speculative_tokens: int):
38
def test_eagle_max_len(
39
monkeypatch: pytest.MonkeyPatch, num_speculative_tokens: int, attn_backend: str
40
):
41
- if attn_backend == "TRITON_ATTN" and not current_platform.is_rocm():
42
- pytest.skip(
43
- "TRITON_ATTN does not support "
44
- "multi-token eagle spec decode on current platform"
45
- )
46
-
41
if attn_backend == "ROCM_AITER_FA" and current_platform.is_rocm():
42
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
43