Repository navigation

[NPU] Fix MTP draft loading for ModelSlim W8A8 checkpoints: dequantize int8 MTP weights - #37736

Open

Mr-qiji wants to merge 2 commits into

sgl-project:mainfrom

Mr-qiji:fix/npu-mtp-modelslim-w8a8-dequant

Open

[NPU] Fix MTP draft loading for ModelSlim W8A8 checkpoints: dequantize int8 MTP weights#37736
Mr-qiji wants to merge 2 commits into
sgl-project:mainfrom
Mr-qiji:fix/npu-mtp-modelslim-w8a8-dequant

Commits on Sep 3, 2026
- [NPU] Dequantize W8A8 MTP draft weights for unquantized NPU draft loading
  Mr-qiji
  authored
  45e1b8d
- Create test_qwen3_5_mtp_w8a8_dequant.py
  Mr-qiji
  authored
  3605385