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

[CI] Re-enable sleep mode test and skip failure breaking CI (#990)

Browse Source 
### What this PR does / why we need it?

- Re-enable sleep mode test
- Fix nightly performance benchmark workflow
- Fix model-runner-v1 bug for upstream
[change](https://github.com/vllm-project/vllm/pull/18654)
---------

Signed-off-by: wangli <wangli858794774@gmail.com>
This commit is contained in:
Li Wang
2025-06-04 16:24:16 +08:00
committed by GitHub
parent eb2701e0b2
commit 517811449e
5 changed files with 38 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
vllm_ascend/worker/model_runner_v1.py
View File

@@ -64,6 +64,7 @@ from vllm_ascend.attention.attention_v1 import AscendAttentionState
from vllm_ascend.attention.mla_v1 import CommonAttentionMetadata
from vllm_ascend.platform import NPUPlatform
from vllm_ascend.sample.rejection_sampler import AscendRejectionSampler
from vllm_ascend.utils import vllm_version_is
from vllm_ascend.worker.mtp_proposer_v1 import MtpProposer
if TYPE_CHECKING:
@@ -1265,15 +1266,27 @@ class NPUModelRunner(LoRAModelRunnerMixin):
import torch_npu
kv_caches: Dict[str, torch.Tensor] = {}
self.input_batch = InputBatch(
max_num_reqs=self.max_num_reqs,
max_model_len=self.model_config.max_model_len,
max_num_batched_tokens=self.max_num_tokens,
device=self.device,
pin_memory=True,
vocab_size=self.model_config.get_vocab_size(),
block_size=self.cache_config.block_size,
)
# Remove this after we drop 0.9.0 support
if vllm_version_is("0.9.0"):
self.input_batch = InputBatch(
max_num_reqs=self.max_num_reqs,
max_model_len=self.model_config.max_model_len,
max_num_batched_tokens=self.max_num_tokens,
device=self.device,
pin_memory=True,
vocab_size=self.model_config.get_vocab_size(),
block_size=self.cache_config.block_size,
)
else:
self.input_batch = InputBatch(
max_num_reqs=self.max_num_reqs,
max_model_len=self.model_config.max_model_len,
max_num_batched_tokens=self.max_num_tokens,
device=self.device,
pin_memory=True,
vocab_size=self.model_config.get_vocab_size(),
block_sizes=[self.cache_config.block_size],
)
for kv_cache_group in kv_cache_config.kv_cache_groups:
kv_cache_spec = kv_cache_group.kv_cache_spec