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

[Bugfix] Disable the dispatch_ffn_combine kernel in MTP path (#4751)

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

This PR is to fix a smoking test failure. Adjust mtp_proposer and
model_runner_v1 to route MTP decoding through the non‑fused MoE
implementation while keeping the overall inference flow unchanged.

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

Signed-off-by: mojave2 <chenchen145@huawei.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
This commit is contained in:
Chen Chen
2025-12-09 22:14:05 +08:00
committed by GitHub
parent cd1c69ee0b
commit 848419d1ba
2 changed files with 12 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
vllm_ascend/worker/model_runner_v1.py
View File

@@ -52,8 +52,8 @@ from vllm.distributed.kv_transfer import (get_kv_transfer_group,
has_kv_transfer_group)
from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorBase_V1
from vllm.distributed.parallel_state import (get_dcp_group, get_dp_group,
get_pcp_group, get_pp_group,
get_tp_group,
get_ep_group, get_pcp_group,
get_pp_group, get_tp_group,
is_global_first_rank)
from vllm.forward_context import get_forward_context
from vllm.logger import logger
@@ -2332,10 +2332,11 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
moe_comm_type = MoECommType.ALLGATHER
elif soc_version in {AscendDeviceType._910_93}:
moe_comm_type = (MoECommType.MC2
if num_tokens <= self.mc2_tokens_capacity else
MoECommType.FUSED_ALLTOALL if quant_type
== "w8a8_dynamic" else MoECommType.ALLTOALL)
# TODO: drop the EP-size guard when dispatch_ffn_combine supports larger EP sizes
moe_comm_type = (
MoECommType.MC2 if num_tokens <= self.mc2_tokens_capacity else
MoECommType.FUSED_ALLTOALL if quant_type == "w8a8_dynamic"
and get_ep_group().world_size <= 16 else MoECommType.ALLTOALL)
else:
raise ValueError(f"Unsupported soc_version: {soc_version}")