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

### 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

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: