[V0 deprecation] Remove _VLLM_V1 suffixes from attention backend names (#25489)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Signed-off-by: Matthew Bonanni <mbonanni001@gmail.com>
This commit is contained in:
@@ -21,16 +21,15 @@ from vllm.v1.attention.backends.utils import (CommonAttentionMetadata,
|
||||
from vllm.v1.kv_cache_interface import FullAttentionSpec
|
||||
|
||||
BACKENDS_TO_TEST = [
|
||||
_Backend.FLASH_ATTN_VLLM_V1, _Backend.FLASHINFER_VLLM_V1,
|
||||
_Backend.FLEX_ATTENTION, _Backend.TRITON_ATTN_VLLM_V1, _Backend.TREE_ATTN,
|
||||
"FLEX_ATTENTION_SLOW"
|
||||
_Backend.FLASH_ATTN, _Backend.FLASHINFER, _Backend.FLEX_ATTENTION,
|
||||
_Backend.TRITON_ATTN, _Backend.TREE_ATTN, "FLEX_ATTENTION_SLOW"
|
||||
]
|
||||
|
||||
# Remove flashinfer from the list if it's not available
|
||||
try:
|
||||
import flashinfer # noqa: F401
|
||||
except ImportError:
|
||||
BACKENDS_TO_TEST.remove(_Backend.FLASHINFER_VLLM_V1)
|
||||
BACKENDS_TO_TEST.remove(_Backend.FLASHINFER)
|
||||
|
||||
|
||||
def _convert_dtype_to_torch(dtype):
|
||||
@@ -214,7 +213,7 @@ def run_attention_backend(
|
||||
builder_cls, impl_cls = get_attention_backend(actual_backend)
|
||||
|
||||
# Mock flashinfer's get_per_layer_parameters if needed
|
||||
if actual_backend == _Backend.FLASHINFER_VLLM_V1:
|
||||
if actual_backend == _Backend.FLASHINFER:
|
||||
import unittest.mock
|
||||
|
||||
from vllm.v1.attention.backends.utils import PerLayerParameters
|
||||
@@ -434,7 +433,7 @@ def _test_backend_correctness(
|
||||
# [num_blocks, 2, block_size, num_kv_heads, head_size]
|
||||
# Select the appropriate KV cache format for each backend
|
||||
kv_cache_for_backend = kv_cache
|
||||
if backend_name == _Backend.FLASHINFER_VLLM_V1:
|
||||
if backend_name == _Backend.FLASHINFER:
|
||||
kv_cache_for_backend = kv_cache.transpose(0, 1)
|
||||
|
||||
# For FlashInfer default to HND layout and
|
||||
@@ -518,8 +517,8 @@ def test_causal_backend_correctness(batch_spec_name: str, model: str):
|
||||
|
||||
|
||||
SLIDING_WINDOW_BACKENDS_TO_TEST = [
|
||||
_Backend.FLASH_ATTN_VLLM_V1, _Backend.FLEX_ATTENTION,
|
||||
_Backend.TRITON_ATTN_VLLM_V1, "FLEX_ATTENTION_SLOW"
|
||||
_Backend.FLASH_ATTN, _Backend.FLEX_ATTENTION, _Backend.TRITON_ATTN,
|
||||
"FLEX_ATTENTION_SLOW"
|
||||
]
|
||||
|
||||
|
||||
|
||||
@@ -15,8 +15,8 @@ from vllm.v1.attention.backends.utils import CommonAttentionMetadata
|
||||
from vllm.v1.kv_cache_interface import FullAttentionSpec
|
||||
|
||||
BACKENDS_TO_TEST = [
|
||||
_Backend.CUTLASS_MLA, _Backend.FLASHMLA_VLLM_V1, _Backend.FLASH_ATTN_MLA,
|
||||
_Backend.TRITON_MLA_VLLM_V1
|
||||
_Backend.CUTLASS_MLA, _Backend.FLASHMLA, _Backend.FLASH_ATTN_MLA,
|
||||
_Backend.TRITON_MLA
|
||||
]
|
||||
|
||||
# Remove CUTLASS_MLA from the list if not using sm100
|
||||
|
||||
@@ -120,30 +120,30 @@ def get_attention_backend(backend_name: _Backend):
|
||||
Tuple of (backend_builder_class, backend_impl_class)
|
||||
"""
|
||||
backend_map = {
|
||||
_Backend.FLASH_ATTN_VLLM_V1:
|
||||
_Backend.FLASH_ATTN:
|
||||
("vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"
|
||||
if current_platform.is_cuda() else
|
||||
"vllm.v1.attention.backends.rocm_aiter_fa.AiterFlashAttentionBackend"
|
||||
),
|
||||
_Backend.FLASHINFER_VLLM_V1:
|
||||
_Backend.FLASHINFER:
|
||||
"vllm.v1.attention.backends.flashinfer.FlashInferBackend",
|
||||
_Backend.FLEX_ATTENTION:
|
||||
"vllm.v1.attention.backends.flex_attention.FlexAttentionBackend",
|
||||
_Backend.TRITON_ATTN_VLLM_V1:
|
||||
_Backend.TRITON_ATTN:
|
||||
"vllm.v1.attention.backends.triton_attn.TritonAttentionBackend",
|
||||
_Backend.TREE_ATTN:
|
||||
"vllm.v1.attention.backends.tree_attn.TreeAttentionBackend",
|
||||
_Backend.XFORMERS_VLLM_V1:
|
||||
_Backend.XFORMERS:
|
||||
"vllm.v1.attention.backends.xformers.XFormersAttentionBackend",
|
||||
_Backend.CUTLASS_MLA:
|
||||
"vllm.v1.attention.backends.mla.cutlass_mla.CutlassMLABackend",
|
||||
_Backend.FLASHMLA_VLLM_V1:
|
||||
_Backend.FLASHMLA:
|
||||
"vllm.v1.attention.backends.mla.flashmla.FlashMLABackend",
|
||||
_Backend.FLASH_ATTN_MLA:
|
||||
"vllm.v1.attention.backends.mla.flashattn_mla.FlashAttnMLABackend",
|
||||
_Backend.FLASHINFER_MLA:
|
||||
"vllm.v1.attention.backends.mla.flashinfer_mla.FlashInferMLABackend",
|
||||
_Backend.TRITON_MLA_VLLM_V1:
|
||||
_Backend.TRITON_MLA:
|
||||
"vllm.v1.attention.backends.mla.triton_mla.TritonMLABackend",
|
||||
}
|
||||
|
||||
|
||||
@@ -89,7 +89,7 @@ backend_configs = {
|
||||
# Triton Attention
|
||||
"TritonAttn":
|
||||
BackendConfig(name="TritonAttn",
|
||||
env_vars={"VLLM_ATTENTION_BACKEND": "TRITON_ATTN_VLLM_V1"},
|
||||
env_vars={"VLLM_ATTENTION_BACKEND": "TRITON_ATTN"},
|
||||
comp_config={
|
||||
"cudagraph_mode": "FULL_AND_PIECEWISE",
|
||||
}),
|
||||
|
||||
@@ -9,11 +9,14 @@ from ...utils import create_new_process_for_each_test
|
||||
|
||||
|
||||
@create_new_process_for_each_test()
|
||||
@pytest.mark.parametrize("attn_backend",
|
||||
["FLASH_ATTN_VLLM_V1", "FLASHINFER_VLLM_V1"])
|
||||
@pytest.mark.parametrize("attn_backend", ["FLASH_ATTN", "FLASHINFER"])
|
||||
def test_cascade_attention(example_system_message, monkeypatch, attn_backend):
|
||||
prompt = "\n<User>: Implement fibonacci sequence in Python.\n<Claude>:"
|
||||
|
||||
if attn_backend == "FLASHINFER":
|
||||
pytest.skip("This test is failing with FlashInfer backend and "
|
||||
"needs investigation. See issue #25679.")
|
||||
|
||||
with monkeypatch.context() as m:
|
||||
m.setenv("VLLM_USE_V1", "1")
|
||||
m.setenv("VLLM_ATTENTION_BACKEND", attn_backend)
|
||||
|
||||
@@ -176,12 +176,11 @@ def test_eagle_correctness(
|
||||
m.setenv("VLLM_MLA_DISABLE", "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")
|
||||
|
||||
method, model_name, spec_model_name, tp_size = model_setup
|
||||
|
||||
@@ -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)
|
||||
|
||||
@@ -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(
|
||||
|
||||
@@ -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.
|
||||
|
||||
@@ -54,26 +54,3 @@ def test_v1_llm_by_default(monkeypatch):
|
||||
print(llm.generate("Hello my name is"))
|
||||
assert hasattr(llm.llm_engine, "engine_core")
|
||||
m.delenv("VLLM_USE_V1")
|
||||
|
||||
|
||||
def test_v1_attn_backend(monkeypatch):
|
||||
with monkeypatch.context() as m:
|
||||
if os.getenv("VLLM_USE_V1", None):
|
||||
m.delenv("VLLM_USE_V1")
|
||||
m.setenv("VLLM_ATTENTION_BACKEND", "XFORMERS")
|
||||
|
||||
# Fall back to V0.
|
||||
_ = AsyncEngineArgs(model=MODEL).create_engine_config()
|
||||
assert not envs.VLLM_USE_V1
|
||||
m.delenv("VLLM_USE_V1")
|
||||
|
||||
# Reject if V1.
|
||||
m.setenv("VLLM_USE_V1", "1")
|
||||
with pytest.raises(NotImplementedError):
|
||||
AsyncEngineArgs(model=MODEL).create_engine_config()
|
||||
m.delenv("VLLM_USE_V1")
|
||||
|
||||
m.setenv("VLLM_ATTENTION_BACKEND", "FLASHMLA")
|
||||
_ = AsyncEngineArgs(model=MODEL).create_engine_config()
|
||||
assert envs.VLLM_USE_V1
|
||||
m.delenv("VLLM_USE_V1")
|
||||
|
||||
Reference in New Issue
Block a user