EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Drop 0.10.2 (#3284)

Browse Source 
Drop v0.10.2 support, we support vLLM 0.11.0rc3 now.
- vLLM version: v0.11.0rc3
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.0

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-10-09 10:28:38 +08:00
committed by GitHub
parent 2dde1268c7
commit f12f76d7ba
17 changed files with 202 additions and 653 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
vllm_ascend/patch/worker/patch_common/patch_attention_layer.py
View File

@@ -16,8 +16,6 @@ from vllm.model_executor.layers.quantization.base_config import \
from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod
from vllm.platforms import current_platform
from vllm_ascend.utils import vllm_version_is
class AscendAttention(Attention, nn.Module, AttentionLayerBase):
"""Attention layer.
@@ -69,12 +67,10 @@ class AscendAttention(Attention, nn.Module, AttentionLayerBase):
if cache_config is not None:
kv_cache_dtype = cache_config.cache_dtype
block_size = cache_config.block_size
is_attention_free = cache_config.is_attention_free
calculate_kv_scales = cache_config.calculate_kv_scales
else:
kv_cache_dtype = "auto"
block_size = 16
is_attention_free = False
calculate_kv_scales = False
if num_kv_heads is None:
num_kv_heads = num_heads
@@ -135,23 +131,13 @@ class AscendAttention(Attention, nn.Module, AttentionLayerBase):
# weight and activation dtype.
dtype = torch.get_default_dtype()
if attn_backend is None:
if vllm_version_is("0.10.2"):
self.attn_backend = get_attn_backend(head_size,
dtype,
kv_cache_dtype,
block_size,
is_attention_free,
use_mla=use_mla,
use_sfa=use_sfa,
has_sink=self.has_sink)
else:
self.attn_backend = get_attn_backend(head_size,
dtype,
kv_cache_dtype,
block_size,
use_mla=use_mla,
use_sfa=use_sfa,
has_sink=self.has_sink)
self.attn_backend = get_attn_backend(head_size,
dtype,
kv_cache_dtype,
block_size,
use_mla=use_mla,
use_sfa=use_sfa,
has_sink=self.has_sink)
else:
self.attn_backend = attn_backend

206
vllm_ascend/patch/worker/patch_common/patch_attention_selector.py
View File

@@ -27,154 +27,72 @@ from vllm.attention.selector import (backend_name_to_enum,
from vllm.platforms import _Backend, current_platform
from vllm.utils import resolve_obj_by_qualname
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("0.10.2"):
def get_attn_backend( # type: ignore[misc]
head_size: int,
dtype: torch.dtype,
kv_cache_dtype: Optional[str],
block_size: int,
use_mla: bool = False,
use_sfa: bool = False,
has_sink: bool = False,
) -> type[AttentionBackend]:
"""Selects which attention backend to use and lazily imports it."""
# Accessing envs.* behind an @lru_cache decorator can cause the wrong
# value to be returned from the cache if the value changes between calls.
# To avoid this, we read envs.VLLM_USE_V1 here and pass it explicitly to the
# private function.
return _cached_get_attn_backend(
head_size=head_size,
dtype=dtype,
kv_cache_dtype=kv_cache_dtype,
block_size=block_size,
use_v1=envs.VLLM_USE_V1,
use_mla=use_mla,
use_sfa=use_sfa,
has_sink=has_sink,
)
def get_attn_backend(
head_size: int,
dtype: torch.dtype,
kv_cache_dtype: Optional[str],
block_size: int,
is_attention_free: bool = False,
use_mla: bool = False,
use_sfa: bool = False,
has_sink: bool = False,
) -> type[AttentionBackend]:
"""Selects which attention backend to use and lazily imports it."""
# Accessing envs.* behind an @lru_cache decorator can cause the wrong
# value to be returned from the cache if the value changes between calls.
# To avoid this, we read envs.VLLM_USE_V1 here and pass it explicitly to the
# private function.
return _cached_get_attn_backend(
head_size=head_size,
dtype=dtype,
kv_cache_dtype=kv_cache_dtype,
block_size=block_size,
is_attention_free=is_attention_free,
use_v1=envs.VLLM_USE_V1,
use_mla=use_mla,
use_sfa=use_sfa,
has_sink=has_sink,
)
@cache
def _cached_get_attn_backend(
head_size: int,
dtype: torch.dtype,
kv_cache_dtype: Optional[str],
block_size: int,
is_attention_free: bool,
use_v1: bool = False,
use_mla: bool = False,
use_sfa: bool = False,
has_sink: bool = False,
) -> type[AttentionBackend]:
# If there are no attention layers (e.g. we are running Mamba),
# use the placeholder NO_ATTENTION
if is_attention_free:
from vllm.attention.backends.placeholder_attn import \
PlaceholderAttentionBackend
return PlaceholderAttentionBackend
@cache
def _cached_get_attn_backend(
head_size: int,
dtype: torch.dtype,
kv_cache_dtype: Optional[str],
block_size: int,
use_v1: bool = False,
use_mla: bool = False,
use_sfa: bool = False,
has_sink: bool = False,
) -> type[AttentionBackend]:
# Check whether a particular choice of backend was
# previously forced.
#
# THIS SELECTION OVERRIDES THE VLLM_ATTENTION_BACKEND
# ENVIRONMENT VARIABLE.
selected_backend = None
backend_by_global_setting: Optional[_Backend] = (
get_global_forced_attn_backend())
if backend_by_global_setting is not None:
selected_backend = backend_by_global_setting
else:
# Check the environment variable and override if specified
backend_by_env_var: Optional[str] = envs.VLLM_ATTENTION_BACKEND
if backend_by_env_var is not None:
selected_backend = backend_name_to_enum(backend_by_env_var)
if selected_backend is None:
raise ValueError(
f"Invalid attention backend: '{backend_by_env_var}'. "
f"Valid backends are: {list(_Backend.__members__.keys())}")
# Check whether a particular choice of backend was
# previously forced.
#
# THIS SELECTION OVERRIDES THE VLLM_ATTENTION_BACKEND
# ENVIRONMENT VARIABLE.
selected_backend = None
backend_by_global_setting: Optional[_Backend] = (
get_global_forced_attn_backend())
if backend_by_global_setting is not None:
selected_backend = backend_by_global_setting
else:
# Check the environment variable and override if specified
backend_by_env_var: Optional[str] = envs.VLLM_ATTENTION_BACKEND
if backend_by_env_var is not None:
selected_backend = backend_name_to_enum(backend_by_env_var)
if selected_backend is None:
raise ValueError(
f"Invalid attention backend: '{backend_by_env_var}'. "
f"Valid backends are: {list(_Backend.__members__.keys())}"
)
# get device-specific attn_backend
attention_cls = current_platform.get_attn_backend_cls(
selected_backend, head_size, dtype, kv_cache_dtype, block_size,
use_v1, use_mla, use_sfa, has_sink)
if not attention_cls:
raise ValueError(
f"Invalid attention backend for {current_platform.device_name}"
)
return resolve_obj_by_qualname(attention_cls)
else:
def get_attn_backend( # type: ignore[misc]
head_size: int,
dtype: torch.dtype,
kv_cache_dtype: Optional[str],
block_size: int,
use_mla: bool = False,
use_sfa: bool = False,
has_sink: bool = False,
) -> type[AttentionBackend]:
"""Selects which attention backend to use and lazily imports it."""
# Accessing envs.* behind an @lru_cache decorator can cause the wrong
# value to be returned from the cache if the value changes between calls.
# To avoid this, we read envs.VLLM_USE_V1 here and pass it explicitly to the
# private function.
return _cached_get_attn_backend(
head_size=head_size,
dtype=dtype,
kv_cache_dtype=kv_cache_dtype,
block_size=block_size,
use_v1=envs.VLLM_USE_V1,
use_mla=use_mla,
use_sfa=use_sfa,
has_sink=has_sink,
)
@cache
def _cached_get_attn_backend(
head_size: int,
dtype: torch.dtype,
kv_cache_dtype: Optional[str],
block_size: int,
use_v1: bool = False,
use_mla: bool = False,
use_sfa: bool = False,
has_sink: bool = False,
) -> type[AttentionBackend]:
# Check whether a particular choice of backend was
# previously forced.
#
# THIS SELECTION OVERRIDES THE VLLM_ATTENTION_BACKEND
# ENVIRONMENT VARIABLE.
selected_backend = None
backend_by_global_setting: Optional[_Backend] = (
get_global_forced_attn_backend())
if backend_by_global_setting is not None:
selected_backend = backend_by_global_setting
else:
# Check the environment variable and override if specified
backend_by_env_var: Optional[str] = envs.VLLM_ATTENTION_BACKEND
if backend_by_env_var is not None:
selected_backend = backend_name_to_enum(backend_by_env_var)
if selected_backend is None:
raise ValueError(
f"Invalid attention backend: '{backend_by_env_var}'. "
f"Valid backends are: {list(_Backend.__members__.keys())}"
)
# get device-specific attn_backend
attention_cls = current_platform.get_attn_backend_cls(
selected_backend, head_size, dtype, kv_cache_dtype, block_size,
use_v1, use_mla, use_sfa, has_sink)
if not attention_cls:
raise ValueError(
f"Invalid attention backend for {current_platform.device_name}"
)
return resolve_obj_by_qualname(attention_cls)
# get device-specific attn_backend
attention_cls = current_platform.get_attn_backend_cls(
selected_backend, head_size, dtype, kv_cache_dtype, block_size, use_v1,
use_mla, use_sfa, has_sink)
if not attention_cls:
raise ValueError(
f"Invalid attention backend for {current_platform.device_name}")
return resolve_obj_by_qualname(attention_cls)
vllm.attention.get_attn_backend = get_attn_backend

7
vllm_ascend/patch/worker/patch_common/patch_weight_loader.py
View File

@@ -1,11 +1,10 @@
import torch
from torch.nn.parameter import Parameter
from vllm.logger import init_logger
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
from vllm.model_executor.utils import set_weight_attrs
from vllm.utils import GiB_bytes
from vllm_ascend.utils import vllm_version_is
logger = init_logger(__name__)
@@ -39,6 +38,4 @@ def create_weights(self, layer: torch.nn.Module, input_size_per_partition: int,
set_weight_attrs(weight, extra_weight_attrs)
if not vllm_version_is("0.10.2"):
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
UnquantizedLinearMethod.create_weights = create_weights
UnquantizedLinearMethod.create_weights = create_weights