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

6
vllm_ascend/ops/fused_moe/prepare_finalize.py
View File

@@ -33,7 +33,7 @@ from vllm_ascend.ascend_forward_context import _EXTRA_CTX
from vllm_ascend.distributed.utils import fc3_all_gather_and_maybe_unpad_impl
from vllm_ascend.ops.fused_moe.moe_runtime_args import MoEPrepareOutput
from vllm_ascend.quantization.quant_type import QuantType
from vllm_ascend.utils import enable_sp, npu_stream_switch, prefill_context_parallel_enable
from vllm_ascend.utils import enable_sp, enable_sp_by_pass, npu_stream_switch, prefill_context_parallel_enable
class PrepareAndFinalize(ABC):
@@ -324,7 +324,7 @@ class PrepareAndFinalizeWithAllGather(PrepareAndFinalize):
Returns:
MoEPrepareOutput with global tensors.
"""
if enable_sp():
if enable_sp() or enable_sp_by_pass():
return self._prepare_with_ep_group(hidden_states, router_logits, quant_type)
return self._prepare_with_dp_group(hidden_states, router_logits, enable_shared_expert_dp, replace_allreduce)
@@ -433,7 +433,7 @@ class PrepareAndFinalizeWithAllGather(PrepareAndFinalize):
Returns:
Tensor with shape [local_num_tokens, hidden_size]
"""
if enable_sp():
if enable_sp() or enable_sp_by_pass():
return self._finalize_with_ep_group(hidden_states)
return self._finalize_with_dp_group(hidden_states, reduce_results)