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

Upgrade vLLM to v0.10.0 (#1927)

Browse Source 
### What this PR does / why we need it?
- Upgrade to v0.10.0
- Drop v0.9.2 version compatibility
- Add patch for
`vllm_ascend/patch/worker/patch_common/patch_sampler_gather_logprobs.py`
as workaround of
f3a683b7c9
for v0.10.0 and also add e2e test `test_models_prompt_logprobs`
- Pin transformers<4.54.0 as workaround of
https://github.com/vllm-project/vllm-ascend/issues/2034

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
- Test locally:
`VLLM_USE_MODELSCOPE=true pytest -sv
tests/e2e/singlecard/test_offline_inference.py::test_models_prompt_logprobs`
- CI passed

- vLLM version: v0.9.2
- vLLM main:
7728dd77bb

---------

Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
This commit is contained in:
Yikun Jiang
2025-07-26 15:43:29 +08:00
committed by GitHub
parent 2f50304c19
commit 17a430f7b8
29 changed files with 198 additions and 251 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
vllm_ascend/attention/attention_v1.py
View File

@@ -17,7 +17,7 @@
from dataclasses import dataclass
from enum import Enum
from typing import Any, Dict, List, Optional, Tuple, Type
from typing import List, Optional, Tuple, Type
import torch
import torch_npu
@@ -31,7 +31,7 @@ from vllm.v1.worker.gpu_input_batch import InputBatch
from vllm_ascend.ops.attention import vanilla_chunked_prefill
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, aligned_16, is_310p,
nd_to_nz_2d, nd_to_nz_spec, vllm_version_is)
nd_to_nz_2d, nd_to_nz_spec)
class AscendAttentionBackend(AttentionBackend):
@@ -43,8 +43,6 @@ class AscendAttentionBackend(AttentionBackend):
@staticmethod
def get_impl_cls() -> Type["AscendAttentionBackendImpl"]:
if vllm_version_is("0.9.2"):
return AscendAttentionBackendImpl092
return AscendAttentionBackendImpl
@staticmethod
@@ -440,38 +438,6 @@ class AscendAttentionBackendImpl(AttentionImpl):
return output.view(num_tokens, self.hidden_size)
class AscendAttentionBackendImpl092(AscendAttentionBackendImpl):
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: int,
alibi_slopes: Optional[List[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[Dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: str = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[str] = None,
use_irope: bool = False,
) -> None:
super().__init__(
num_heads=num_heads,
head_size=head_size,
scale=scale,
num_kv_heads=num_kv_heads,
alibi_slopes=alibi_slopes,
sliding_window=sliding_window,
kv_cache_dtype=kv_cache_dtype,
logits_soft_cap=logits_soft_cap,
attn_type=attn_type,
kv_sharing_target_layer_name=kv_sharing_target_layer_name,
use_irope=use_irope,
)
def unified_ascend_attention_with_output(
query: torch.Tensor,
key: torch.Tensor,

39
vllm_ascend/attention/attention_v1_torchair.py
View File

@@ -16,7 +16,7 @@
#
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Type
from typing import List, Optional, Tuple, Type
import numpy as np
import torch
@@ -29,7 +29,7 @@ from vllm.v1.worker.gpu_input_batch import InputBatch
from vllm_ascend.attention.attention_v1 import AscendAttentionState
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, aligned_16, is_310p,
nd_to_nz_2d, vllm_version_is)
nd_to_nz_2d)
class AscendAttentionTorchairBackend(AttentionBackend):
@@ -41,8 +41,6 @@ class AscendAttentionTorchairBackend(AttentionBackend):
@staticmethod
def get_impl_cls() -> Type["AscendAttentionTorchairBackendImpl"]:
if vllm_version_is("0.9.2"):
return AscendAttentionTorchairBackendImpl092
return AscendAttentionTorchairBackendImpl
@staticmethod
@@ -489,36 +487,3 @@ class AscendAttentionTorchairBackendImpl(AttentionImpl):
"to use ascend scheduler.")
return output.view(num_tokens, self.hidden_size)
class AscendAttentionTorchairBackendImpl092(AscendAttentionTorchairBackendImpl
):
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: int,
alibi_slopes: Optional[List[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[Dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: str = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[str] = None,
use_irope: bool = False,
) -> None:
super().__init__(
num_heads=num_heads,
head_size=head_size,
scale=scale,
num_kv_heads=num_kv_heads,
alibi_slopes=alibi_slopes,
sliding_window=sliding_window,
kv_cache_dtype=kv_cache_dtype,
logits_soft_cap=logits_soft_cap,
attn_type=attn_type,
kv_sharing_target_layer_name=kv_sharing_target_layer_name,
use_irope=use_irope,
)

40
vllm_ascend/attention/mla_v1.py
View File

@@ -1,12 +1,11 @@
from dataclasses import dataclass
from typing import (TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type,
TypeVar)
from typing import TYPE_CHECKING, Optional, Tuple, Type, TypeVar
import numpy as np
import torch
import torch_npu
from vllm.attention.backends.abstract import (AttentionBackend, AttentionLayer,
AttentionMetadata, AttentionType,
AttentionMetadata,
MLAAttentionImpl)
from vllm.attention.backends.utils import PAD_SLOT_ID
from vllm.config import get_current_vllm_config
@@ -22,7 +21,7 @@ from vllm_ascend.multistream.context import get_multistream_comm_context
from vllm_ascend.multistream.ms_split import model_input_split_v1_mla_attn
from vllm_ascend.ops.attention import vanilla_chunked_prefill_mla
from vllm_ascend.torchair.utils import npu_stream_switch, npu_wait_tensor
from vllm_ascend.utils import npu_prefetch, vllm_version_is
from vllm_ascend.utils import npu_prefetch
from vllm_ascend.worker.npu_input_batch import InputBatch
if TYPE_CHECKING:
@@ -54,8 +53,6 @@ class AscendMLABackend(AttentionBackend):
@staticmethod
def get_impl_cls() -> Type["MLAAttentionImpl"]:
if vllm_version_is("0.9.2"):
return AscendMLAImpl092
return AscendMLAImpl
@@ -1212,34 +1209,3 @@ class AscendMLAImpl(MLAAttentionImpl):
output[:num_decode_tokens] = output_decode
return output_padded
class AscendMLAImpl092(AscendMLAImpl):
def __init__(self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: int,
alibi_slopes: Optional[List[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[Dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: str = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[str] = None,
use_irope: bool = False,
**kwargs) -> None:
super().__init__(
num_heads=num_heads,
head_size=head_size,
scale=scale,
num_kv_heads=num_kv_heads,
alibi_slopes=alibi_slopes,
sliding_window=sliding_window,
kv_cache_dtype=kv_cache_dtype,
logits_soft_cap=logits_soft_cap,
attn_type=attn_type,
kv_sharing_target_layer_name=kv_sharing_target_layer_name,
use_irope=use_irope,
**kwargs)