[qgallouedec]

Part of #7658.

Neuron compiles once per input shape, so without pad_to_multiple_of every new sequence length stalls training for about a minute. This PR makes "auto" the default pad_to_multiple_of in SFT, DPO, KTO, Reward, GRPO, RLOO and Distillation: 128 on Neuron, None elsewhere, so nothing changes on GPU. An explicit value is kept as is.

• Trainium2, Qwen3-0.6B SFT, 20 steps: 130 s with the default, 260 s with pad_to_multiple_of=None.

Why not a warning

We already know the right value, and warnings get missed (agents launching runs never read them).

Why not 128 everywhere

It only helps SDPA on Hopper (cuDNN builds a graph for each new sequence length), and less as the model grows. Elsewhere, and with flash attention on every GPU, it costs 2-6%.

Data

SFT on trl-lib/Capybara, max_length=4096, 100 steps, steady-state samples/s, torch 2.13 unless noted. Change with pad_to_multiple_of=128 vs None:

GPU	model, batch	SDPA	FA2	FA3
H100	Qwen3-1.7B, 1	+57.1%	-9.7%	-6.4%
H100	Qwen3-1.7B, 4	+31.5%	-3.6%	-3.3%
H100	Qwen3-1.7B, 8	+16.7%	-2.7%	-2.8%
H100	Qwen3-1.7B, 16	+4.6%	-3.5%	-3.3%
H100	Qwen3-4B, 1	+39.2%	-8.7%	-8.3%
H100	Qwen3-4B, 2	+26.2%	-3.0%	-2.5%
H100	Qwen3-4B, 4	+15.4%	-2.7%	-3.1%
H200	Qwen3-1.7B, 4	+26.5%	-1.7%
H200	Qwen3-4B, 2	+22.3%	-4.6%
H200	Qwen3-8B, 2	+9.8%	-4.3%
H200	Qwen3-8B, 4	+2.7%	-5.3%
H200 (torch 2.14)	Qwen3-1.7B, 4	+22.7%	-2.8%
H200 (torch 2.14)	Qwen3-4B, 2	+21.9%	-5.9%
A100	Qwen3-1.7B, 4	-0.5%	-1.7%
A100	Qwen3-4B, 2	-3.2%	-6.1%
L40S	Qwen3-1.7B, 4	-3.8%	-2.9%
RTX PRO 6000	Qwen3-1.7B, 4	-2.7%	-1.8%
RTX PRO 6000	Qwen3-4B, 2	-3.0%	-2.9%

Full H100 sweep (192 runs). Peak allocated memory doesn't depend on the pad; the fragmentation OOMs at batch 8 go away with expandable_segments, not with padding:

cc @michaelbenayoun

---

Note

Medium Risk
Cross-cutting default change on all major trainer configs affects batch padding and sequence lengths on Neuron (intended); off Neuron "auto" still resolves to no padding, but anyone depending on the old implicit default should confirm explicit overrides.

Overview
pad_to_multiple_of now defaults to "auto" (instead of None) on SFT, DPO, KTO, Reward, GRPO, RLOO, and Distillation configs. The field type accepts int | str | None, and __post_init__ normalizes CLI strings ("128", "none"/"null") and resolves "auto" to 128 when device.type == "neuron", otherwise None—so GPU training behavior stays unpadded unless overridden.

Docs/help text for pad_to_multiple_of are updated accordingly. tests/test_cli_utils.py adds parametrized coverage that TrlParser parses --pad_to_multiple_of correctly across those config classes (including "auto" → None off Neuron).

^{Reviewed by Cursor Bugbot for commit 93c6e25. Bugbot is set up for automated code reviews on this repo. Configure here.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-14T16:20:24.021777Z	4ecf90b	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4ecf90b4e0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 4ecf90b. Configure here.}