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

[BugFix] Fix mlapo accuracy problem related with weight processing. (#3857)

Browse Source 
This PR fixes a mlapo accuracy problem related with weight processing.
Furthermore, modify mlapo related e2e test with quantized deepseek model
to make it effective.

Signed-off-by: whx-sjtu <2952154980@qq.com>
This commit is contained in:
whx
2025-10-30 00:35:50 +08:00
committed by GitHub
parent d9249c968e
commit 211d4b9da4
2 changed files with 2 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
View File

@@ -111,19 +111,3 @@ def test_mtp2_correctness_full_graph(
model_name: str,
):
mtp_correctness(sampling_config, model_name, 2, CUDAGraphMode.FULL)
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_MLAPO": "1"})
def test_mtp_correctness_piecewise_graph_with_mlapo_kernel(
sampling_config: SamplingParams,
model_name: str,
):
mtp_correctness(sampling_config, model_name, 1)
@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_MLAPO": "1"})
def test_mtp_correctness_full_graph_with_mlapo_kernel(
sampling_config: SamplingParams,
model_name: str,
):
mtp_correctness(sampling_config, model_name, 1, CUDAGraphMode.FULL)