gspo: GSPO loss + DeepSpeed parity fixes (loss/grad divisors, SDP, fp32_lm_head, docs_per_step, temperature)#502

Open

bigximik wants to merge 23 commits intomainfrom

Commits on Apr 27, 2026

Commits on Apr 28, 2026

gspo: add sequence-level IS-ratio clipping loss
bigximik
committed
schedule: add rollouts_per_step to auto-set depth_first_micro_batches
bigximik
committed
grpo: dynamic docs_per_step accumulation and normalize_by_documents
bigximik
committed

Commits on Apr 29, 2026

grpo: temperature scaling for IS ratio parity with actor sampling
bigximik
committed

Commits on May 4, 2026

Commits on May 5, 2026

Commits on May 6, 2026

Commits on May 7, 2026