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

[Main2Main] Upgrade vllm commit to 0109 (#5752)

Browse Source 
### What this PR does / why we need it?
Upgrade vllm commit to 0109 (bde38c11df0ea066a740efe9b77fff5418be45df)

1. remove `init_cached_hf_modules ` due to
https://github.com/vllm-project/vllm/pull/31786
2. fix spec_decode e2e test due to
https://github.com/vllm-project/vllm/pull/29821 break
3. fix `vllm.v1.attention.backends.utils` duo to
https://github.com/vllm-project/vllm/pull/31891
4. fix `self.seq_lens - query_lens` on same device due to
https://github.com/vllm-project/vllm/pull/31773
5. skip model_runner_v2 e2e test due to `'_OpNamespace' '_C' object has
no attribute 'get_cuda_view_from_cpu_tensor'`

- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef

Signed-off-by: hfadzxy <starmoon_zhang@163.com>
This commit is contained in:
zhangxinyuehfad
2026-01-13 19:14:43 +08:00
committed by GitHub
parent eed9e366a7
commit f7b904641e
21 changed files with 203 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm_ascend/attention/mla_v1.py
View File

@@ -6,7 +6,6 @@ import torch
import torch_npu
import vllm.envs as envs_vllm
from vllm.attention.backends.abstract import AttentionBackend, MLAAttentionImpl
from vllm.attention.backends.utils import PAD_SLOT_ID
from vllm.config import VllmConfig, get_current_vllm_config
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import logger
@@ -39,12 +38,17 @@ from vllm_ascend.ops.rotary_embedding import get_cos_and_sin_mla
from vllm_ascend.ops.weight_prefetch import maybe_npu_prefetch
from vllm_ascend.quantization.w8a8 import AscendW8A8LinearMethod
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, maybe_trans_nz,
weak_ref_tensors)
vllm_version_is, weak_ref_tensors)
from vllm_ascend.worker.npu_input_batch import NPUInputBatch
if TYPE_CHECKING:
from vllm.v1.core.sched.output import SchedulerOutput
if vllm_version_is('0.13.0'):
from vllm.attention.backends.utils import PAD_SLOT_ID # type: ignore
else:
from vllm.v1.attention.backends.utils import PAD_SLOT_ID # type: ignore
MAX_O_PROJ_PREFETCH_SIZE = 16 * 1024 * 1024
BUILD_METADATA_STEP_PREFILL = 0
BUILD_METADATA_STEP_DECODE = 1

8
vllm_ascend/ops/triton/mamba/causal_conv1d.py
View File

@@ -13,7 +13,13 @@ import torch
import torch.nn.functional as F
import triton
import triton.language as tl
from vllm.attention.backends.utils import PAD_SLOT_ID
from vllm_ascend.utils import vllm_version_is
if vllm_version_is('0.13.0'):
from vllm.attention.backends.utils import PAD_SLOT_ID # type: ignore
else:
from vllm.v1.attention.backends.utils import PAD_SLOT_ID # type: ignore
def causal_conv1d_ref(

4
vllm_ascend/worker/model_runner_v1.py
View File

@@ -1670,6 +1670,8 @@ class NPUModelRunner(GPUModelRunner):
attn_metadata,
aux_hidden_states,
)
if not vllm_version_is('0.13.0'):
self._copy_draft_token_ids_to_cpu(scheduler_output)
(
logprobs_lists,
@@ -1983,7 +1985,7 @@ class NPUModelRunner(GPUModelRunner):
query_start_loc_cpu=self.query_start_loc.cpu[:num_reqs +
1],
_seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
seq_lens=self.seq_lens.cpu[:num_reqs],
seq_lens=self.seq_lens.gpu[:num_reqs],
num_reqs=num_reqs,
num_actual_tokens=num_tokens,
block_table_tensor=block_table_tensor[:num_reqs],

9
vllm_ascend/worker/worker.py
View File

@@ -121,11 +121,12 @@ class NPUWorker(WorkerBase):
self.cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
self.cache_config.cache_dtype]
if self.model_config.trust_remote_code:
# note: lazy import to avoid importing torch before initializing
from vllm.utils.import_utils import init_cached_hf_modules
if vllm_version_is('0.13.0'):
if self.model_config.trust_remote_code:
# note: lazy import to avoid importing torch before initializing
from vllm.utils.import_utils import init_cached_hf_modules
init_cached_hf_modules()
init_cached_hf_modules()
self.profiler = self._init_profiler()
if vllm_config.model_config and vllm_config.model_config.enable_sleep_mode: