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

[V0 deprecation] Remove _VLLM_V1 suffixes from attention backend names (#25489)

Browse Source 
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni001@gmail.com>
This commit is contained in:
Matthew Bonanni
2025-09-25 13:37:50 -04:00
committed by GitHub
parent 71b25b0d48
commit 3468f17ebe
42 changed files with 131 additions and 174 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
tests/v1/spec_decode/test_eagle.py
View File

@@ -314,12 +314,11 @@ def test_load_model(mock_get_model, mock_get_layers, mock_get_pp_group, method,
monkeypatch.setenv("VLLM_ATTENTION_BACKEND", attn_backend)
if (attn_backend == "TRITON_ATTN_VLLM_V1"
and not current_platform.is_rocm()):
pytest.skip("TRITON_ATTN_VLLM_V1 does not support "
if (attn_backend == "TRITON_ATTN" and not current_platform.is_rocm()):
pytest.skip("TRITON_ATTN does not support "
"multi-token eagle spec decode on current platform")
if attn_backend == "FLASH_ATTN_VLLM_V1" and current_platform.is_rocm():
if attn_backend == "FLASH_ATTN" and current_platform.is_rocm():
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
# Setup draft model mock
@@ -400,16 +399,15 @@ def test_propose(method, attn_backend, num_speculative_tokens, monkeypatch):
monkeypatch.setenv("VLLM_ATTENTION_BACKEND", attn_backend)
if (attn_backend == "TRITON_ATTN_VLLM_V1"
and not current_platform.is_rocm()):
pytest.skip("TRITON_ATTN_VLLM_V1 does not support "
if (attn_backend == "TRITON_ATTN" and not current_platform.is_rocm()):
pytest.skip("TRITON_ATTN does not support "
"multi-token eagle spec decode on current platform")
if (attn_backend == "TREE_ATTN"):
pytest.skip("TREE_ATTN is tested separately in test_propose_tree"
"because it requires special input mocking.")
if attn_backend == "FLASH_ATTN_VLLM_V1" and current_platform.is_rocm():
if attn_backend == "FLASH_ATTN" and current_platform.is_rocm():
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
# Use GPU device
@@ -510,12 +508,12 @@ def test_propose(method, attn_backend, num_speculative_tokens, monkeypatch):
device=device)
sampling_metadata = mock.MagicMock()
if attn_backend == "FLASH_ATTN_VLLM_V1":
if attn_backend == "FLASH_ATTN":
attn_metadata_builder_cls, _ = get_attention_backend(
_Backend.FLASH_ATTN_VLLM_V1)
elif attn_backend == "TRITON_ATTN_VLLM_V1":
_Backend.FLASH_ATTN)
elif attn_backend == "TRITON_ATTN":
attn_metadata_builder_cls, _ = get_attention_backend(
_Backend.TRITON_ATTN_VLLM_V1)
_Backend.TRITON_ATTN)
elif attn_backend == "TREE_ATTN":
attn_metadata_builder_cls, _ = get_attention_backend(
_Backend.TREE_ATTN)

7
tests/v1/spec_decode/test_max_len.py
View File

@@ -41,12 +41,11 @@ def test_eagle_max_len(monkeypatch: pytest.MonkeyPatch,
m.setenv("VLLM_USE_V1", "1")
m.setenv("VLLM_ATTENTION_BACKEND", attn_backend)
if (attn_backend == "TRITON_ATTN_VLLM_V1"
and not current_platform.is_rocm()):
pytest.skip("TRITON_ATTN_VLLM_V1 does not support "
if (attn_backend == "TRITON_ATTN" and not current_platform.is_rocm()):
pytest.skip("TRITON_ATTN does not support "
"multi-token eagle spec decode on current platform")
if attn_backend == "FLASH_ATTN_VLLM_V1" and current_platform.is_rocm():
if attn_backend == "FLASH_ATTN" and current_platform.is_rocm():
m.setenv("VLLM_ROCM_USE_AITER", "1")
llm = LLM(

2
tests/v1/spec_decode/test_tree_attention.py
View File

@@ -278,7 +278,7 @@ def test_tree_attn_correctness() -> None:
block_table=block_table,
slot_mapping=branch_slot_mapping,
seqlen_k=sequence_position + q_len,
backend=_Backend.FLASH_ATTN_VLLM_V1,
backend=_Backend.FLASH_ATTN,
).view(batch_size, -1, num_heads, dim_per_head)
# Compare the outputs.