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 15, 18:30 UTC
[Bugfix] Fix broken kernel test due to missing rename for v1 Triton backend (#15282)
Signed-off-by: Isotr0py <2037008807@qq.com>
Isotr0py committed
Mar 21, 2025 at 19:42 UTC
8afcd0f6336b076f1e8328d664f03f41319a65b8
2 files changed
+2
-2
tests/kernels/test_attention_selector.py
+1
-1
@@ -49,7 +49,7 @@ def test_env(
49
RocmPlatform()):
50
backend = get_attn_backend(16, torch.float16, torch.float16,
51
16, False)
52
- EXPECTED = "ROCM_ATTN_VLLM_V1" if use_v1 else "ROCM_FLASH"
52
+ EXPECTED = "TRITON_ATTN_VLLM_V1" if use_v1 else "ROCM_FLASH"
53
assert backend.get_name() == EXPECTED
54
elif device == "openvino":
55
with patch("vllm.attention.selector.current_platform",
tests/kernels/test_rocm_attention_selector.py
+1
-1
@@ -26,7 +26,7 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
26
# Test standard ROCm attention
27
backend = get_attn_backend(16, torch.float16, torch.float16, 16, False)
28
assert (backend.get_name() == "ROCM_FLASH"
29
- or backend.get_name() == "ROCM_ATTN_VLLM_V1")
29
+ or backend.get_name() == "TRITON_ATTN_VLLM_V1")
30
31
# mla test for deepseek related
32
backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False,