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

fix vl pd smoke error (#5103)

Browse Source 
### What this PR does / why we need it?
Fix VL model mooncacke PD smoke test error
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: 李少鹏 <lishaopeng21@huawei.com>
This commit is contained in:
shaopeng-666
2025-12-18 22:20:45 +08:00
committed by GitHub
parent ff3914e31a
commit fd9a47c04d
1 changed files with 12 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
vllm_ascend/distributed/mooncake_connector.py
View File

@@ -42,7 +42,7 @@ from vllm.v1.request import RequestStatus
from vllm_ascend.ascend_config import get_ascend_config, init_ascend_config
from vllm_ascend.distributed.mooncake_transfer_engine import global_te
from vllm_ascend.distributed.utils import get_transfer_timeout_value
from vllm_ascend.utils import prefill_context_parallel_enable
from vllm_ascend.utils import is_vl_model, prefill_context_parallel_enable
if TYPE_CHECKING:
from vllm.attention.backends.abstract import AttentionMetadata
@@ -317,16 +317,17 @@ class KVCacheRecvingThread(threading.Thread):
self.vllm_config = vllm_config
self.model_config = self.vllm_config.model_config
self.block_size = self.vllm_config.cache_config.block_size
if self.use_mla:
self.k_head_dim = self.model_config.hf_config.kv_lora_rank
self.v_head_dim = self.model_config.hf_config.qk_rope_head_dim
self.num_kv_heads = 1
else:
self.k_head_dim = self.model_config.hf_config.head_dim
self.v_head_dim = self.model_config.hf_config.head_dim
self.num_kv_heads = max(
self.model_config.hf_config.num_key_value_heads //
self.tp_size, 1)
if not is_vl_model(vllm_config):
if self.use_mla:
self.k_head_dim = self.model_config.hf_config.kv_lora_rank
self.v_head_dim = self.model_config.hf_config.qk_rope_head_dim
self.num_kv_heads = 1
else:
self.k_head_dim = self.model_config.hf_config.head_dim
self.v_head_dim = self.model_config.hf_config.head_dim
self.num_kv_heads = max(
self.model_config.hf_config.num_key_value_heads //
self.tp_size, 1)
def add_request(self, request_id: str, local_block_ids: list[int],
remote_block_ids: list[int], remote_engine_id: str,