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

Remove V0 attention backends (#25351)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-09-21 16:03:28 -07:00
committed by GitHub
parent af7dfb0d1a
commit bc6e542d9f
28 changed files with 143 additions and 7376 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/engine/arg_utils.py
View File

@@ -44,8 +44,8 @@ from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
from vllm.transformers_utils.config import (get_model_path, is_interleaved,
maybe_override_with_speculators)
from vllm.transformers_utils.utils import check_gguf_file
from vllm.utils import (STR_DUAL_CHUNK_FLASH_ATTN_VAL, FlexibleArgumentParser,
GiB_bytes, get_ip, is_in_ray_actor)
from vllm.utils import (FlexibleArgumentParser, GiB_bytes, get_ip,
is_in_ray_actor)
from vllm.v1.sample.logits_processor import LogitsProcessor
# yapf: enable
@@ -1163,17 +1163,6 @@ class EngineArgs:
self._set_default_args_v0(model_config)
assert self.enable_chunked_prefill is not None
if envs.VLLM_ATTENTION_BACKEND in [STR_DUAL_CHUNK_FLASH_ATTN_VAL]:
assert self.enforce_eager, (
"Cuda graph is not supported with DualChunkFlashAttention. "
"To run the model in eager mode, set 'enforce_eager=True' "
"or use '--enforce-eager' in the CLI.")
assert current_platform.is_cuda(), (
"DualChunkFlashAttention is only supported on CUDA platform.")
assert not use_v1, (
"DualChunkFlashAttention is not supported on V1 engine. "
"To run the model in V0 engine, try set 'VLLM_USE_V1=0'")
sliding_window: Optional[int] = None
if not is_interleaved(model_config.hf_text_config):
# Only set CacheConfig.sliding_window if the model is all sliding