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

[Feat][SP] Suport SP for VL MoE models (#7044)

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

2nd PR for https://github.com/vllm-project/vllm-ascend/issues/5712,
extend SP to VL MoE models.


### Does this PR introduce _any_ user-facing change?
remove `sp_threshold` in additional config and reuse `sp_min_token_num`
from vLLM.


### How was this patch tested?
- Model: Qwen3-VL-30B-A3B, 
- TP4 DP2
- 100 reqs
- max concurrency 1

| Seq length | Mean TTFT (ms) main | Mean TTFT (ms) this PR |
|------------|---------------------|------------------------|
| 4k         | 429.40               | 323.3                  |
| 16k        | 1297.01              | 911.74                |

- vLLM version: v0.16.0
- vLLM main:
4034c3d32e

---------

Signed-off-by: realliujiaxu <realliujiaxu@163.com>
This commit is contained in:
realliujiaxu
2026-03-24 17:16:00 +08:00
committed by GitHub
parent 9615bc33fd
commit 5d12446573
21 changed files with 947 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm_ascend/patch/__init__.py
View File

@@ -167,15 +167,12 @@
# 1. `vllm.distributed.parallel_state.GroupCoordinator`
# Why:
# vllm doesn't support all_to_all for GroupCoordinator.
# all_reduce in vLLM not is a customop, which will make MatmulAllReduceAddRMSNorm fusion failure.
# How：
# Add all_to_all implementation for GroupCoordinator.
# make all_reduce as a customop.
# Related PR (if no, explain why):
# No, we should use vlLM all2all manager to support all_to_all for npu.
# Future Plan:
# Remove this patch when the refactor of all2all manager is done.
# Remove this patch when vLLM support all_reduce as customop.
#
# ** 2. File: worker/patch_multimodal_merge.py**
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

7
vllm_ascend/patch/worker/patch_distributed.py
View File

@@ -84,7 +84,7 @@ class GroupCoordinatorPatch(GroupCoordinator):
if use_message_queue_broadcaster and self.world_size > 1:
self.mq_broadcaster = MessageQueue.create_from_process_group(self.cpu_group, 1 << 22, 6)
self.use_custom_op_call = False
self.use_custom_op_call = True
self.use_cpu_custom_send_recv = False
def all_to_all(
@@ -106,10 +106,5 @@ class GroupCoordinatorPatch(GroupCoordinator):
assert self.device_communicator is not None, "device_communicator should be initialized when world_size > 1"
return self.device_communicator.all_to_all(input_, scatter_dim, gather_dim, scatter_sizes, gather_sizes)
def all_reduce(self, input_):
if self.world_size == 1:
return input_
return torch.ops.vllm.all_reduce(input_, group_name=self.unique_name)
vllm.distributed.parallel_state.GroupCoordinator = GroupCoordinatorPatch