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

[Refactor] Adjustments to moe_comm_method selection process (#3001)

Browse Source 
### What this PR does / why we need it?
Fix issues mentioned in
https://github.com/vllm-project/vllm-ascend/pull/2791 and some minor
refactoring.
1. Use Enum instead of string.
2. Avoid setting a new property to forward_context in
AscendFusedMoE.forward().
3. Enabling TokenDispatcherWithMoge.
4. Remove redundant code.

### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?

Qwen3-30B-A3B/Qwen3-30B-A3B-W8A8/DeepSeek-V3-W4A8-Pruing/deepseek-mtp/pangu-pro-moe-pruing:
1. Enable/Disable EP
2. Aclgraph & eager


- vLLM version: v0.10.2
- vLLM main:
9607d5eb44

Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>
Co-authored-by: weijinqian0 <12153182+weijinqian0@users.noreply.github.com>
This commit is contained in:
weichen
2025-09-22 19:12:58 +08:00
committed by GitHub
parent bb1f0d5a62
commit 37a0715eda
14 changed files with 170 additions and 351 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
tests/ut/worker/test_model_runner_v1.py
View File

@@ -15,6 +15,7 @@ from unittest.mock import MagicMock, patch
import pytest
from vllm_ascend.ascend_forward_context import MoECommType
from vllm_ascend.utils import AscendSocVersion
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
@@ -24,21 +25,21 @@ from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
"soc_version, enable_expert_parallel, world_size, num_tokens, mc2_tokens_capacity, quant_type, expected_method",
[
# Case 1: Expert parallel is disabled, should always be 'allgather'
(AscendSocVersion.A2, False, 8, 100, 256, None, "allgather"),
(AscendSocVersion.A3, False, 16, 500, 256, None, "allgather"),
(AscendSocVersion.A2, False, 8, 100, 256, None, MoECommType.ALLGATHER),
(AscendSocVersion.A3, False, 16, 500, 256, None, MoECommType.ALLGATHER),
# Case 2: A2 SOC with w4a8_dynamic -> use alltoall when not mc2
(AscendSocVersion.A2, True, 8, 100, 256, "w4a8_dynamic", "alltoall"),
(AscendSocVersion.A2, True, 16, 257, 256, "w4a8_dynamic", "alltoall"),
(AscendSocVersion.A2, True, 16, 100, 256, "w4a8_dynamic", "mc2"), # meets mc2 condition
(AscendSocVersion.A2, True, 8, 100, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
(AscendSocVersion.A2, True, 16, 257, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
(AscendSocVersion.A2, True, 16, 100, 256, "w4a8_dynamic", MoECommType.MC2), # meets mc2 condition
# Case 3: A2 SOC without w4a8_dynamic -> fallback to allgather
(AscendSocVersion.A2, True, 8, 100, 256, None, "allgather"),
(AscendSocVersion.A2, True, 16, 257, 256, None, "allgather"),
(AscendSocVersion.A2, True, 8, 100, 256, None, MoECommType.ALLGATHER),
(AscendSocVersion.A2, True, 16, 257, 256, None, MoECommType.ALLGATHER),
# Case 4: A3 SOC
(AscendSocVersion.A3, True, 8, 100, 256, None, "mc2"),
(AscendSocVersion.A3, True, 8, 257, 256, None, "alltoall"),
(AscendSocVersion.A3, True, 8, 100, 256, None, MoECommType.MC2),
(AscendSocVersion.A3, True, 8, 257, 256, None, MoECommType.ALLTOALL),
])
# yapf: enable
def test_select_moe_comm_method(soc_version, enable_expert_parallel,