minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[MoE Refactor] [NPU] Refactor Ascend MoE implementation to reduce code duplication and align with community design (#25663)

Browse Source 
Co-authored-by: ronnie_zheng <zl19940307@163.com>
This commit is contained in:
Артем Савкин
2026-07-15 14:59:42 +03:00
committed by GitHub
co-authored by ronnie_zheng
parent c9b17403e7
commit 8ed82afcc8
54 changed files with 2851 additions and 2375 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/test/ascend/test_ascend_utils.py
+3 -3
View File
@@ -77,6 +77,9 @@ DOTS_OCR_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "rednote-hilab/dots.ocr"
ECO_TECH_QWEN3_32B_W4A4_LAOS_WEIGHTS_PATH = os.path.join(
MODEL_WEIGHTS_DIR, "Eco-Tech/Qwen3-32B-w4a4-LAOS"
)
ECO_TECH_QWEN3_30B_A3B_W4A4_LAOS_WEIGHTS_PATH = os.path.join(
MODEL_WEIGHTS_DIR, "Eco-Tech/Qwen3-30B-A3B-w4a4-LAOS"
)
ERNIE_4_5_21B_A3B_PT_WEIGHTS_PATH = os.path.join(
MODEL_WEIGHTS_DIR, "baidu/ERNIE-4.5-21B-A3B-PT"
)
@@ -301,9 +304,6 @@ DEEPSEEK_R1_0528_W4A8_PER_CHANNEL_WEIGHTS_PATH = os.path.join(
DEEPSEEK_R1_0528_W8A8_WEIGHTS_PATH = os.path.join(
MODEL_WEIGHTS_DIR, "vllm-ascend/DeepSeek-R1-0528-W8A8"
)
QWEN3_30B_MODELSLIM_INT4_WEIGHTS_PATH = os.path.join(
MODEL_WEIGHTS_DIR, "Eco-Tech/Qwen3-30B-A3B-w4a4-LAOS"
)
QWEN3_5_397B_W4A8_MODEL_PATH = os.path.join(
MODEL_WEIGHTS_DIR, "Eco-Tech/Qwen3.5-397B-A17B-w4a8-mtp"
)