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

[Test][BugFix] Fix dispatch_gmm_combine_decode test stability (#7097)

Browse Source 
### What this PR does / why we need it?
This patch fix the nightly failure
1. Each case uses a copy of the global kwargs instead of a reference to
prevent parameter pollution between use cases.
2. Add weight initialization in the scenario of `eplb` + `w8a8_dynamic`

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?
```python
pytest -sv tests/e2e/nightly/single_node/ops/multicard_ops_a3/test_dispatch_gmm_combine_decode.py
```

```shell
===================================================================== 3 passed, 4 warnings in 194.86s (0:03:14) ======================================================================
```
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e

Signed-off-by: wangli <wangli858794774@gmail.com>
This commit is contained in:
Li Wang
2026-03-12 17:22:44 +08:00
committed by GitHub
parent d866e6b238
commit 0a171b5cdd
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
tests/e2e/nightly/single_node/ops/multicard_ops_a3/test_dispatch_gmm_combine_decode.py
View File

@@ -320,6 +320,9 @@ class FusionOp(DecodeMoeOps):
self.gmm2_weight_scale_fp32 = [ self.gmm2_weight_scale_fp32 = [
weight.clone() for weight in gmm2_weight_scale.unbind(dim=0) weight.clone() for weight in gmm2_weight_scale.unbind(dim=0)
] ]
else:
self.gmm1_weight_scale_fp32 = [torch.ones(1).npu().to(gmm1_weight.dtype)]
self.gmm2_weight_scale_fp32 = [torch.ones(1).npu().to(gmm2_weight.dtype)]
else: else:
self.gmm1_weight = [gmm1_weight.clone()] self.gmm1_weight = [gmm1_weight.clone()]
self.gmm2_weight = [gmm2_weight.clone()] self.gmm2_weight = [gmm2_weight.clone()]
@@ -524,7 +527,7 @@ def run_once(local_rank_id,
@torch.inference_mode() @torch.inference_mode()
def test_dispatch_gmm_combine_decode_base(): def test_dispatch_gmm_combine_decode_base():
custom_kwargs = BASE_KWARGS custom_kwargs = BASE_KWARGS.copy()
custom_kwargs["batch_size"] = 32 custom_kwargs["batch_size"] = 32
custom_kwargs["ep_world_size"] = 8 custom_kwargs["ep_world_size"] = 8
custom_kwargs["moe_expert_num"] = 32 custom_kwargs["moe_expert_num"] = 32
@@ -539,7 +542,7 @@ def test_dispatch_gmm_combine_decode_base():
@torch.inference_mode() @torch.inference_mode()
def test_dispatch_gmm_combine_decode_with_mc2_mask(): def test_dispatch_gmm_combine_decode_with_mc2_mask():
custom_kwargs = BASE_KWARGS custom_kwargs = BASE_KWARGS.copy()
custom_kwargs["with_mc2_mask"] = True custom_kwargs["with_mc2_mask"] = True
ep_world_size = custom_kwargs["ep_world_size"] ep_world_size = custom_kwargs["ep_world_size"]
custom_args = tuple(custom_kwargs.values()) custom_args = tuple(custom_kwargs.values())
@@ -548,7 +551,7 @@ def test_dispatch_gmm_combine_decode_with_mc2_mask():
@torch.inference_mode() @torch.inference_mode()
def test_dispatch_gmm_combine_decode_dynamic_eplb(): def test_dispatch_gmm_combine_decode_dynamic_eplb():
custom_kwargs = BASE_KWARGS custom_kwargs = BASE_KWARGS.copy()
custom_kwargs["dynamic_eplb"] = True custom_kwargs["dynamic_eplb"] = True
ep_world_size = custom_kwargs["ep_world_size"] ep_world_size = custom_kwargs["ep_world_size"]
custom_args = tuple(custom_kwargs.values()) custom_args = tuple(custom_kwargs.values())