biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[V0 Deprecation] Remove V0 FlashInfer attention backend (#22776)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-08-18 19:54:16 -07:00
committed by GitHub
parent 6603288736
commit 14006840ea
8 changed files with 9 additions and 1133 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tests/core/block/e2e/test_correctness_sliding_window.py
View File

@@ -32,7 +32,7 @@ BLOCK_SIZE = 16
@pytest.mark.parametrize("test_llm_kwargs", [{}])
@pytest.mark.parametrize("batch_size", [5])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("backend", ["FLASH_ATTN", "FLASHINFER", "XFORMERS"])
@pytest.mark.parametrize("backend", ["FLASH_ATTN", "XFORMERS"])
def test_sliding_window_retrieval(baseline_llm_generator, test_llm_generator,
batch_size, seed, backend, monkeypatch):
"""
@@ -43,8 +43,6 @@ def test_sliding_window_retrieval(baseline_llm_generator, test_llm_generator,
Additionally, we compare the results of the v1 and v2 managers.
"""
if backend == "FLASHINFER" and current_platform.is_rocm():
pytest.skip("Flashinfer does not support ROCm/HIP.")
if backend == "XFORMERS" and current_platform.is_rocm():
pytest.skip("Xformers does not support ROCm/HIP.")
@@ -96,7 +94,7 @@ def test_sliding_window_retrieval(baseline_llm_generator, test_llm_generator,
@pytest.mark.parametrize("test_llm_kwargs", [{"enable_chunked_prefill": True}])
@pytest.mark.parametrize("batch_size", [5])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("backend", ["FLASH_ATTN", "FLASHINFER", "XFORMERS"])
@pytest.mark.parametrize("backend", ["FLASH_ATTN", "XFORMERS"])
def test_sliding_window_chunked_prefill(test_llm_generator, batch_size, seed,
backend, monkeypatch):
"""
@@ -107,8 +105,6 @@ def test_sliding_window_chunked_prefill(test_llm_generator, batch_size, seed,
The results with and without chunked prefill are not the same due to
numerical instabilities.
"""
if backend == "FLASHINFER" and current_platform.is_rocm():
pytest.skip("Flashinfer does not support ROCm/HIP.")
if backend == "XFORMERS" and current_platform.is_rocm():
pytest.skip("Xformers does not support ROCm/HIP.")
override_backend_env_variable(monkeypatch, backend)