[albertvillanova]

This PR declares a torch floor of 2.7.0 and removes the fallbacks that existed for older torch.

Part of #7644. Alternative to #7642, which sets the floor at 2.6.0.

Motivation

TRL names no torch version at all, so the effective floor is whatever the stack enforces: torch>=2.2 from transformers 4.56.2. torch 2.7.0 was released on 2025-04-23, about 18 months ago, so it meets the 1-year criterion in #7116.

2.7.0 instead of 2.6.0 because torch 2.6.0 (cu124) ships NVRTC 12.4.127, which switches Python's default text encoding to ASCII after the first runtime kernel compile (#7720). Training with TRL on torch 2.6.0 then fails when saving a checkpoint, as soon as the model card is generated. torch 2.7.0 ships NVRTC 12.6.77, which does not have the problem.

The branches below exist only to support torch versions below 2.7:

• torch.accelerator was introduced in 2.6, with the older path hard-coding "cuda"
• FSDPModule, MixedPrecisionPolicy and fully_shard became public as torch.distributed.fsdp in 2.6, previously torch.distributed._composable.fsdp
• fully_shard accepts ignored_params from 2.7
• torch.get_device_module exists from 2.4, DTensor is public in torch.distributed.tensor from 2.5, and nn.Buffer exists from 2.5

Changes

• Add torch>=2.7.0 to the project dependencies and pin torch==2.7.0 in the minimum versions job
• Import FSDPModule, MixedPrecisionPolicy and fully_shard from the public torch.distributed.fsdp unconditionally
• Pass ignored_params to fully_shard unconditionally
• Read the accelerator type from torch.accelerator unconditionally in activation offloading, the experimental online DPO trainer and trl env
• Import DTensor from torch.distributed.tensor and check nn.Buffer unconditionally in activation offloading

The torch>=2.8.0 guard stays, since it is above this floor.

---

Note

Medium Risk
Raises the install floor and simplifies FSDP2/distributed paths; users on torch 2.2–2.6 can no longer install, but behavior for supported versions should be unchanged aside from always passing ignored_params on FSDP2.

Overview
Requires PyTorch 2.7.0+ and drops compatibility shims for older torch.

pyproject.toml now lists torch>=2.7.0 (documented as avoiding torch 2.6’s NVRTC-related ASCII text I/O bug during checkpoint/model-card writes). CI’s minimum-versions job pins torch==2.7.0.

Runtime code assumes modern torch APIs: torch.accelerator / torch.get_device_module in activation offloading, online DPO vLLM setup, and trl env; public torch.distributed.fsdp imports (FSDPModule, fully_shard, MixedPrecisionPolicy) with ignored_params always passed to fully_shard; DTensor from torch.distributed.tensor and nn.Buffer checks without legacy import paths.

^{Reviewed by Cursor Bugbot for commit 407a6d7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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.

[qgallouedec]

I think #7642 is better: #7720 is fixed by #7755 so no real need to drop 2.6.

[albertvillanova]

Closing in favor of #7642, now merged with the torch 2.6.0 floor: #7720 was fixed in TRL by #7755 (and #7764 for the tests), so dropping 2.6 is no longer needed. The minimum-versions job on main is green with torch 2.6.0.