[sibisrinivasb]

Fixes #2224.

What

MASTER.decode now skips its early stop when the model is exportable, the same guard PARSeq already has:

# Stop decoding if all sequences have reached the EOS token
# NOTE: `break` isn't correctly translated to Onnx so we don't break here if we want to export
if not self.exportable and (ys == self.vocab_size).any(dim=-1).all():
    break

Why

export_model_to_onnx traces the model, so the stop added in #2118 was evaluated once on the dummy input and the exported graph kept only the decoding steps the dummy needed. PyTorch inference is unchanged and keeps the early stop.

master(pretrained=True, exportable=True), exported with torch.rand((2, 3, 32, 128)) under five seeds, then run on 24 rendered words of 1 to 22 characters (all 24 read correctly by PyTorch):

seed	words identical to PyTorch, main	with this change
0	3 / 24	24 / 24
1	16 / 24	24 / 24
2	6 / 24	24 / 24
3	6 / 24	24 / 24
4	1 / 24	24 / 24

An exported MASTER therefore always runs max_length - 1 decoding steps, as before #2118.

Test

test_master_decode_early_stop forces EOS at the first decoding step and counts the decoder calls: 1 without exportable, max_length - 1 with it. It fails on main for exportable=True and needs no pretrained weights.

Checked

• pytest tests/pytorch/test_models_recognition_pt.py -k master: 12 passed, 1 skipped (the master case of test_models_onnx_export, under 16 GB here).
• ruff check and ruff format --check on the two files: clean.
• Not tested: GPU, other torch / onnxruntime versions (here torch 2.15.0, onnxruntime 1.31.1), the dynamo exporter.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.15%. Comparing base (dbcee0d) to head (44b9ab7).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2225      +/-   ##
==========================================
- Coverage   98.16%   98.15%   -0.01%     
==========================================
  Files         172      172              
  Lines       11977    11977              
==========================================
- Hits        11757    11756       -1     
- Misses        220      221       +1     

Flag	Coverage Δ
unittests	98.15% <100.00%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[felixdittrich92]

Thanks @sibisrinivasb 👍

Only a small change and we can merge.