[AMD] Add AMD MI350X/MI355X (gfx950) blockwise FP8 support for run_qwen3_30b_a3b#1465

Draft

JessicaJiang-123 wants to merge 7 commits into

radixark:mainfrom

JessicaJiang-123:amd-qwen3-30b-a3b-fp8

Commits on Jun 14, 2026

Commits on Jun 16, 2026

Enable restore_weights_before_loading for fp8 quant in RL weight update
JessicaJiang-123
committed

Commits on Jun 19, 2026

Add AMD MI350X (gfx950) fp8 blockwise path to run_qwen3_30b_a3b.py; revert example .sh ROCm edit
JessicaJiang-123
committed
Clean up AMD fp8 launcher: MI350X/MI355X (gfx950), drop redundant num_gpus/NOSET special-cases
JessicaJiang-123
committed
Merge remote-tracking branch 'upstream/main' into amd-qwen3-30b-a3b-fp8
JessicaJiang-123
committed

Commits on Jun 21, 2026

AMD MI350X fp8: max-tokens-per-gpu 16384 (+11%) and rollout-num-gpus-per-engine 2 (eval fix)

JessicaJiang-123
and
XinyuJiangCMU
committed