feat(qwen4_exp): load-time per-tensor FP8 dense projections (W8A8 via _scaled_mm) - #389

Draft

gdevenyi wants to merge 6 commits into

FlashML-org:mainfrom

gdevenyi:feat/qwen4-exp-fp8-dense

Commits on Sep 4, 2026

feat(qwen4_exp): tensor parallelism for Qwen3.8-Flash-Next (offload backend)

gdevenyi
and
claude
committed
9e73f2c
fix(qwen4_exp): load the HF config for sharding only when TP > 1

gdevenyi
and
claude
committed
d16d634
feat(qwen4_exp): load-time per-tensor FP8 dense projections (W8A8 via _scaled_mm)

gdevenyi
and
claude
committed
4d3518f
chore(qwen4_exp): log the load-time FP8 dense mode once at weight load

gdevenyi
and
claude
committed
69e5a44
fix(qwen4_exp): return the FP8 quantization slack to the allocator before the cache planner runs

gdevenyi
and
claude
committed
fec72ba
fix(qwen4_exp): give in_proj_ba its own storage under FP8 dense

gdevenyi
and
claude
committed
3f8f249