[V0 Deprecation] Deprecate BlockSparse Attention & Phi3-Small (#21217)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-07-19 13:53:17 -07:00
committed by GitHub
parent 881e3cbe3b
commit 752c6ade2e
38 changed files with 65 additions and 2435 deletions

View File

@@ -33,8 +33,12 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
# change the attention backend to triton MLA
m.setenv(STR_BACKEND_ENV_VAR, "TRITON_MLA")
backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False,
False, True)
backend = get_attn_backend(576,
torch.bfloat16,
"auto",
16,
False,
use_mla=True)
assert (backend.get_name() == "TRITON_MLA"
or backend.get_name() == "TRITON_MLA_VLLM_V1")
@@ -42,15 +46,23 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
# If use_mla is true
# The selected backend is triton MLA
m.setenv(STR_BACKEND_ENV_VAR, None)
backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False,
False, True)
backend = get_attn_backend(576,
torch.bfloat16,
"auto",
16,
False,
use_mla=True)
assert (backend.get_name() == "TRITON_MLA"
or backend.get_name() == "TRITON_MLA_VLLM_V1")
# change the attention backend to AITER MLA
m.setenv(STR_BACKEND_ENV_VAR, "ROCM_AITER_MLA")
backend = get_attn_backend(576, torch.bfloat16, "auto", 1, False,
False, True)
backend = get_attn_backend(576,
torch.bfloat16,
"auto",
1,
False,
use_mla=True)
assert (backend.get_name() == "ROCM_AITER_MLA"
or backend.get_name() == "ROCM_AITER_MLA_VLLM_V1")
@@ -60,7 +72,11 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
# The selected backend is ROCM_AITER_MLA
m.setenv(STR_BACKEND_ENV_VAR, None)
m.setenv("VLLM_ROCM_USE_AITER", "1")
backend = get_attn_backend(576, torch.bfloat16, "auto", 1, False,
False, True)
backend = get_attn_backend(576,
torch.bfloat16,
"auto",
1,
False,
use_mla=True)
assert (backend.get_name() == "ROCM_AITER_MLA"
or backend.get_name() == "ROCM_AITER_MLA_VLLM_V1")