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

[V0 deprecation] Remove V0 CPU/XPU/TPU backends (#20412)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
This commit is contained in:
Woosuk Kwon
2025-07-06 08:48:13 -07:00
committed by GitHub
parent 43813e6361
commit e202dd2736
20 changed files with 46 additions and 5034 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
tests/kernels/attention/test_attention_selector.py
View File

@@ -36,7 +36,8 @@ DEVICE_REGULAR_ATTN_BACKENDS = {
DEVICE_MLA_BLOCK_SIZES = {
"cuda": [16, 64], # CUDA supports both standard and extended block sizes
"hip": [16, 1], # HIP requires special handling for block_size=1
"cpu": [16] # CPU uses fixed block size from test cases
# "cpu": [16] # CPU uses fixed block size from test cases
"cpu": [] # FIXME(woosuk): Temporarily disable CPU tests
}
@@ -81,14 +82,14 @@ def test_env(
m.setenv("VLLM_MLA_DISABLE", "1" if use_mla else "0")
if device == "cpu":
if not use_v1:
pytest.skip("CPU backend only supports V1")
with patch("vllm.attention.selector.current_platform",
CpuPlatform()):
backend = get_attn_backend(16, torch.float16, torch.float16,
block_size, False)
if use_v1:
assert backend.get_name() == "TORCH_SDPA_VLLM_V1"
else:
assert backend.get_name() == "TORCH_SDPA"
assert backend.get_name() == "TORCH_SDPA_VLLM_V1"
elif device == "hip":
with patch("vllm.attention.selector.current_platform",
@@ -193,12 +194,14 @@ def test_fp32_fallback(
m.setenv("VLLM_USE_V1", "1" if use_v1 else "0")
if device == "cpu":
if not use_v1:
pytest.skip("CPU backend only supports V1")
with patch("vllm.attention.selector.current_platform",
CpuPlatform()):
backend = get_attn_backend(16, torch.float32, torch.float32,
16, False)
assert (backend.get_name() == "TORCH_SDPA_VLLM_V1"
if use_v1 else "TORCH_SDPA")
assert backend.get_name() == "TORCH_SDPA_VLLM_V1"
elif device == "cuda":
with patch("vllm.attention.selector.current_platform",