Fix safetensors export OOM - #193

Merged

tastelikefeet merged 7 commits into

modelscope:mainfrom

hjh0119:fix-qwen4-exp-export

Sep 15, 2026

Commits on Sep 14, 2026

init
hjh0119
committed
7571d21
revert
hjh0119
committed
d3cfd80

Commits on Sep 15, 2026

fix(qwen4_exp): chunked export to bound GPU peaks; export GPU-resident PLE table; cross-PP PLE shard transfer
hjh0119
committed
71c6073
merge main (#187,#192,#194,#196): resolve ple.py -- keep #192 unified naming + chunked all_reduce with CPU accumulation
hjh0119
committed
f54087c
perf(export): stream PLE table shards one at a time to bound host peaks
hjh0119
committed
76917fc
refactor(qwen4_exp): transport cross-pp PLE shards as raw bytes instead of hardcoding fp8
hjh0119
committed
3d0f1de
fix(export): guard _chunk_rows against empty dim0 and fix E125 lint
hjh0119
committed
71668a2