GRPO: sequence packing for the no-grad old/ref logp path (default-on) - #6738

Merged

danielhanchen merged 24 commits into

mainfrom

grpo-sequence-packing

Jul 6, 2026

Commits on Jun 30, 2026

GRPO no-grad packing: default-on, verify against per-row reference
danielhanchen
committed
6298366
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
4d69187
GRPO no-grad packing: disable entirely on cross-sample mismatch
danielhanchen
committed
499e528

Commits on Jul 1, 2026

GRPO no-grad packing: trim comments to be concise
danielhanchen
committed
b5127aa
GRPO no-grad packing: fix per-row completion boundary for left-padded rows
danielhanchen
committed
537c276
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
3814b68
GRPO no-grad packing: gate verification on real completion rows
danielhanchen
committed
840ed89
GRPO no-grad packing: gate debug logging on UNSLOTH_ENABLE_LOGGING
danielhanchen
committed
a2b135f

Commits on Jul 2, 2026

GRPO packing: import UNSLOTH_ENABLE_LOGGING inside the injected logp function
danielhanchen
committed
e014c0a

Commits on Jul 3, 2026

Commits on Jul 5, 2026

Commits on Jul 6, 2026

GRPO sequence packing: tighten comments
danielhanchen
committed
1ab029f
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
ba1233d