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

clean 0.15.0 support (#6852)

Browse Source 
Clean up vllm 0.15.0 related code

- vLLM version: v0.16.0
- vLLM main:
15d76f74e2

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2026-02-28 09:20:57 +08:00
committed by GitHub
parent 84b00695f8
commit 3d563292f3
8 changed files with 17 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/source/community/versioning_policy.md
View File

@@ -65,7 +65,7 @@ For main branch of vLLM Ascend, we usually make it compatible with the latest vL
| Date | Event |
|------------|-------------------------------------------|
| 2026.02.26 | Release candidates, v0.15.0rc1 |
| 2026.02.27 | Release candidates, v0.15.0rc1 |
| 2026.02.06 | v0.13.0 Final release, v0.13.0 |
| 2026.01.26 | Release candidates, v0.14.0rc1 |
| 2026.01.24 | Release candidates, v0.13.0rc2 |

2
docs/source/user_guide/release_notes.md
View File

@@ -1,6 +1,6 @@
# Release Notes
## v0.15.0rc1 - 2026.02.26
## v0.15.0rc1 - 2026.02.27
This is the first release candidate of v0.15.0 for vLLM Ascend. Please follow the [official doc](https://docs.vllm.ai/projects/ascend/en/latest) to get started.

2
tests/ut/attention/test_sfa_v1.py
View File

@@ -14,7 +14,7 @@ if 'torch_npu._inductor' not in sys.modules:
from vllm_ascend.attention.sfa_v1 import (AscendSFABackend, AscendSFAImpl,
AscendSFAMetadata,
AscendSFAMetadataBuilder)
from vllm_ascend.utils import enable_dsa_cp, vllm_version_is
from vllm_ascend.utils import enable_dsa_cp
class TestAscendSFABackend(TestBase):

2
tests/ut/quantization/test_modelslim_config.py
View File

@@ -13,7 +13,7 @@ from vllm_ascend.quantization.modelslim_config import (
MODELSLIM_CONFIG_FILENAME,
AscendModelSlimConfig,
)
from vllm_ascend.utils import ASCEND_QUANTIZATION_METHOD, vllm_version_is
from vllm_ascend.utils import ASCEND_QUANTIZATION_METHOD
from vllm.model_executor.layers.attention import Attention

5
vllm_ascend/_310p/fused_moe/fused_moe.py
View File

@@ -153,11 +153,10 @@ class AscendFusedMoE310(FusedMoE):
self.quant_type = self.get_quant_type()
_MoECommMethods[MoECommType.ALLGATHER] = AllGatherCommImpl310(self.moe_config)
if not vllm_version_is("0.15.0"):
if not vllm_version_is("0.16.0"):
self.runner = self._init_runner()
if not vllm_version_is("0.15.0"):
if not vllm_version_is("0.16.0"):
def _init_runner(self):
from vllm_ascend.ops.fused_moe.fused_moe import AscendMoERunner

10
vllm_ascend/compilation/passes/sequence_parallelism.py
View File

@@ -1,18 +1,14 @@
import torch
import torch._inductor.pattern_matcher as pm
from torch._inductor.pattern_matcher import PatternMatcherPass
from vllm_ascend.utils import is_moe_model, vllm_version_is
if vllm_version_is("0.15.0"):
from vllm.compilation.vllm_inductor_pass import VllmInductorPass # type: ignore
else:
from vllm.compilation.passes.vllm_inductor_pass import VllmInductorPass
from vllm.compilation.passes.vllm_inductor_pass import VllmInductorPass
from vllm.config import VllmConfig
from vllm.config.utils import Range
from vllm.distributed import get_tensor_model_parallel_world_size, get_tp_group, tensor_model_parallel_all_reduce
from vllm.logger import logger
from vllm_ascend.utils import is_moe_model
SP_THRESHOLD = 1000

24
vllm_ascend/ops/mm_encoder_attention.py
View File

@@ -21,8 +21,6 @@ import torch.nn.functional as F
import torch_npu
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention # type: ignore
from vllm_ascend.utils import vllm_version_is
MIN_PAD_SIZE: int = 64 # min_size to pad weight
MAX_PAD_SIZE: int = 128 # max_size to pad weight
@@ -64,9 +62,7 @@ class AscendMMEncoderAttention(MMEncoderAttention):
prefix=prefix,
)
if not vllm_version_is("0.15.0"):
self.layer_index = int("".join(filter(str.isdigit, prefix)))
self.layer_index = int("".join(filter(str.isdigit, prefix)))
self.enable_pad = self.head_size > MIN_PAD_SIZE and self.head_size < MAX_PAD_SIZE
self.scale_value = self.head_size**-0.5
@@ -106,19 +102,13 @@ class AscendMMEncoderAttention(MMEncoderAttention):
kv_len = key.size(1)
is_reshaped = query.dim() == 4
if vllm_version_is("0.15.0"):
# Directly use seq_lens cpu cache to avoid d2h copy.
global seq_lens_cpu_cache
if self.layer_index == 0:
if cu_seqlens is None:
cu_seqlens = torch.arange(0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device="cpu")
seq_lens_cpu = torch.diff(cu_seqlens).to("cpu")
else:
global seq_lens_cpu_cache
if self.layer_index == 0:
if cu_seqlens is None:
cu_seqlens = torch.arange(0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device="cpu")
# Update seq_lens cpu cache.
seq_lens_cpu_cache = torch.diff(cu_seqlens).to("cpu")
# Directly use seq_lens cpu cache to avoid d2h copy.
seq_lens_cpu = seq_lens_cpu_cache
# Update seq_lens cpu cache.
seq_lens_cpu_cache = torch.diff(cu_seqlens).to("cpu")
# q, k, v: [b, s, head, head_dim] -> [b * s, head, head_dim]
q, k, v = self._reshape_qkv_to_3d(query, key, value, bsz, q_len, kv_len)
@@ -138,7 +128,7 @@ class AscendMMEncoderAttention(MMEncoderAttention):
query=q,
key=k,
value=v,
seq_len=seq_lens_cpu,
seq_len=seq_lens_cpu_cache,
scale_value=self.scale_value,
num_heads=self.num_heads,
num_kv_heads=self.num_kv_heads,

6
vllm_ascend/sample/rejection_sampler.py
View File

@@ -19,7 +19,6 @@ from vllm_ascend.ops.triton.reject_sample import (
sample_recovered_tokens_kernel,
)
from vllm_ascend.sample.sampler import apply_top_k_top_p
from vllm_ascend.utils import vllm_version_is
def apply_sampling_constraints(
@@ -167,10 +166,7 @@ def rejection_sample(
return output_token_ids
# Compute probability distribution from target logits.
if vllm_version_is("0.15.0"):
target_probs = target_logits
else:
target_probs = target_logits.softmax(dim=-1, dtype=torch.float32)
target_probs = target_logits.softmax(dim=-1, dtype=torch.float32)
assert target_probs.is_contiguous()
# Generate uniform probabilities for rejection sampling.