Minimum torch version?

[peterdsharpe]

Currently, pyproject.toml lists torch as a dependency, with no lower-version pin.

It looks like TensorDict used to have a torch pin until tensordict v0.8 or so (for example, 0.7 requires torch>=2.6.0), when it was removed with #1286.

Two questions:

1. Was the removal intentional? I would prefer to have a torch lower-version pin, but in the spirit of Chesterton's fence, it'd be good to know if there was a reason for this pin-removal first. @vmoens ?

Motivation for having the minimum-version pin is that currently users can install tensordict with wildly incompatible torch versions (e.g., where we call torch API that does not exist), thereby increasing the maintenance burden.

2. If we do decide to pin a minimum torch version, what should we pin it to?

I ran the test suite from today's main on PyTorch releases, and see:

┌───────────┬────────────────────┬──────────────────┬────────────────┐
│ torch     │ released           │ compile failures │ other failures │
├───────────┼────────────────────┼──────────────────┼────────────────┤
│ 2.0–2.2   │ 2023–24            │ 16*              │ 110–536        │
├───────────┼────────────────────┼──────────────────┼────────────────┤
│ 2.3–2.4   │ 2024               │ 10–108*          │ 9–12           │
├───────────┼────────────────────┼──────────────────┼────────────────┤
│ 2.5–2.6   │ late 2024–Jan 2025 │ 95–99            │ 1–2            │
├───────────┼────────────────────┼──────────────────┼────────────────┤
│ 2.7–2.9   │ 2025               │ 5–9              │ 1              │
├───────────┼────────────────────┼──────────────────┼────────────────┤
│ 2.10–2.11 │ Jan–Mar 2026       │ 3                │ 1              │
├───────────┼────────────────────┼──────────────────┼────────────────┤
│ 2.12–2.13 │ May–Jul 2026       │ 0                │ 1              │
├───────────┼────────────────────┼──────────────────┼────────────────┤
│ 2.14      │ Sep 2026           │ 0                │ 0              │
└───────────┴────────────────────┴──────────────────┴────────────────┘

My vote would be to require at least torch >= 2.12.0a0, which is the first release where compile is clean, and this would let us drop a ton of scattered compat shims throughout the library like #1703 or #1823. >=2.13 also seems like a good choice, as that's the most recent release that currently has both CUDA 12 and 13 wheels. Bare-minimum, I think we should pin to >=2.10 or so, since compile mechanics get a lot messier before that.

However - wanted to open this up for discussion before opening a PR on this, since this potentially impacts a lot of users. Would be a good question - what version of torch are most users running these days? Thoughts?

[peterdsharpe]

Tagging @vmoens and active maintainers + recent contributors @bsprenger @theap06 for viz

[bsprenger]

@peterdsharpe thanks for starting this discussion! 100% in support of a lower bound (unless there was a reason for removal that I don't know about).

From my point of view, it does seem like the goal would be full compile support. In principle I would be in support of a 2.13 bound. I can run a similar test matrix to what you did above for torchRL, to get a similar count of failures for different torch versions on the torchRL test suite.

@theap06 do you think we should consider a similar pin for torchRL, or see a reason not to have a lower bound?

[peterdsharpe]

I like the idea of a >=2.13 bound! I'll prep a PR that pins there, but I'll leave it open for a few days in case there's input from @theap06 or users. Will also see if I can do a scan on major repos using tensordict to determine what their existing torch pins are (if any).

Yeah, would be interesting to see what this test matrix looks like on the torchRL side too - good idea!

[theap06]

@peterdsharpe @bsprenger yea I also agree! I think we should have a lower bound. I don't suppose there is a reason for supporting a much older package. We should run a test matrix to determine the failures and reegression for torchRL to have a pin. The bindings for torchRL and tensordict are usually pretty similar so maybe we can also take a look at starting with a lower bound of 2.13.