feat: make timesfm2_5 onnx export compatible

[pdufour]

What does this PR do?

This fixes an issues with the model that made it incompatible with exporting to onnx.

Specifically the following has been changed:

1. if condition

if input_len < context_len:

Will give this error when you try to export the timesfm module through torch.onnx.export

Could not guard on data-dependent expression u0 < 16384

Solution: use a branch-free code block which onnx can export

2. static batch size
   for ts in inputs: causes the batch size to be static when you export it via onnx

Solution: Allow a tensor input to _preprocess method / forward method

3. _timesfm_moving_average doesn't support the above changes:

Solution: Allow a tensor input there as well

Test Plan

On main, create a test file in the repo root

text_onnx_export.py contents:
https://gist.github.com/pdufour/52c7fd722dbe8d60030469d4298779be

Run the export which will export the timesfm2_5 model to onnx

python3 -m venv venv
source venv/bin/activate
pip install -e ".[torch]"
pip install onnx onnxscript
python3 test_onnx_export.py

This gives you the following errors:

1. Error about the if condition

<class 'torch.fx.experimental.symbolic_shapes.GuardOnDataDependentSymNode'>: Could not guard on data-dependent expression u0 < 16384 (unhinted: u0 < 16384).  (Size-like symbols: none)

consider using data-dependent friendly APIs such as guard_or_false, guard_or_true and statically_known_true.
Caused by: (src/transformers/models/timesfm2_5/modeling_timesfm2_5.py:705 in _preprocess)

This is the line:

if input_len < context_len:

Now checkout this branch
Run the export again:

python3 text_onnx_export.py

See that it passes!

Run a test:

Create a file test_onnx_infer.py
https://gist.github.com/pdufour/6071e44296c236b27250ef308f8e5273

Run it:

python3 -m venv .venv
source .venv/bin/activate
pip install onnxruntime
python3 test_onnx_infer.py

Code Agent Policy

The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by
code agents. We are currently bottlenecked by our ability to review and respond to them. As a result,
we ask that new users do not submit pure code agent PRs at this time.
You may use code agents in drafting or to help you diagnose issues. We'd also ask autonomous "OpenClaw"-like agents
not to open any PRs or issues for the moment.

PRs that appear to be fully agent-written will probably be closed without review, and we may block users who do this
repeatedly or maliciously.

This is a rapidly-evolving situation that's causing significant shockwaves in the open-source community. As a result,
this policy is likely to be updated regularly in the near future. For more information, please read CONTRIBUTING.md.

• I confirm that this is not a pure code agent PR.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@xenova @zucchini-nlp @Cyrilvallez @ArthurZucker @vasqu @NielsRogge

[xenova]

Thanks for the PR! cc @IlyasMoutawwakil maybe to include in #41992?

[xenova]

does your fix also work for multiple batches? 👀 this line/for-loop suggests it will be fixed at the batch_size you exported with.

[IlyasMoutawwakil]

yeah this will still hardcode the batch size through the loop

[pdufour]

hmm yeah, i think I need to change the input to torch.Tensor, do you know another option?

[pdufour]

@xenova deja vu https://github.com/huggingface/optimum/pull/2135/changes#r2060659352

[pdufour]

@xenova @IlyasMoutawwakil updated the PR now so onnx export does not fix batch size, also fixed a couple other spots with simliar issues. Put it all in the PR description. Model looks like this now:

Was previously showing static value for batch but now it is dynamic:

[pdufour]

@xenova @IlyasMoutawwakil any chance you have time this week to re-review? Thanks!

[IlyasMoutawwakil]

unfortunately we can't have different paths for tensors and lists in the modeling code, what i suggest is to do something like:

if isisntance(obj, list):
    # turn into tensors
# do tensor processing

[pdufour]

@IlyasMoutawwakil updated now! Let me know if that is what you were thinking, I also added an optional past_observed_mask arg since IIRC you can't have variable-length sequences in a single batched tensor.

Let me know if I am wrong about that and I can change it!

[IlyasMoutawwakil]

i think this test suite is unnecessary if we just switch to tensors in the above testing and explicitly asked users to use tensors if they ever pass lists. my reasoning is that tensor inputs are simply better in this cases.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: timesfm, timesfm2_5