feat: rename pretrained models to Ising-Decoder-SurfaceCode-1-{Fast,Accurate}

[ivanbasov]

Summary

• Rename models/PreDecoderModelMemory_r9_v1.0.77.pt → models/Ising-Decoder-SurfaceCode-1-Fast.pt
• Rename models/PreDecoderModelMemory_r13_v1.0.86.pt → models/Ising-Decoder-SurfaceCode-1-Accurate.pt
• Both files remain Git LFS-tracked via the existing models/*.pt glob — no storage approach change
• Add model_checkpoint_file direct-path config option to _load_model in run.py (new named files don't embed epoch numbers, so the old epoch-scanning logic doesn't apply)
• Update test_inference_public_model.py to use the new filenames and the direct-path loader
• Update README.md and checkpoint_to_safetensors.py docstring examples

Test plan

• git lfs pull after checkout — both new .pt files should resolve correctly
• Run pytest code/tests/test_inference_public_model.py — model loading should work with new filenames via model_checkpoint_file
• Verify find_best_model error message reads "No valid model checkpoint files found" (not old prefix-specific message)

🤖 Generated with Claude Code

[bmhowe23]

I am concerned that these renames may have broken some of the workflows in our README.md. For example, when I run the following, it now fails where it used to work before the rename. @ivanbasov - do you have any suggestions for changes to my workflow?

(.venv) root@9ad5eb711cc6:~/Ising-Decoding# cp models/Ising-Decoder-SurfaceCode-1-Fast.pt outputs/predecoder_model_1/models/
(.venv) root@9ad5eb711cc6:~/Ising-Decoding# CUDA_VISIBLE_DEVICES=3 ONNX_WORKFLOW=2 DISTANCE=13 N_ROUNDS=104 PREDECODER_INFERENCE_NUM_SAMPLES=2048 WORKFLOW=inference EXPERIMENT_NAME=predecoder_model_1 bash
 code/scripts/local_run.sh
==========================================
Local run
==========================================
workflow.task: inference
config: config_public
GPUS: 1 (CUDA_VISIBLE_DEVICES=3)
output: /home/Ising-Decoding/outputs/predecoder_model_1
logs: /home/Ising-Decoding/logs/predecoder_model_1_20260408_012832
overrides: hydra.run.dir=/home/Ising-Decoding/outputs/predecoder_model_1 distance=13 n_rounds=104
==========================================
Loading model for task: inference
Loading best model from: /home/Ising-Decoding/outputs/predecoder_model_1/models/best_model
best_model/ not found; falling back to: /home/Ising-Decoding/outputs/predecoder_model_1/models
Searching for best model in: /home/Ising-Decoding/outputs/predecoder_model_1/models
Found 0 model files:
Error executing job with overrides: ['workflow.task=inference', '+exp_tag=predecoder_model_1', '++load_checkpoint=True', 'distance=13', 'n_rounds=104']
Traceback (most recent call last):
  File "/home/Ising-Decoding/code/workflows/run.py", line 66, in run
    run_surface(cfg)
  File "/home/Ising-Decoding/code/workflows/run.py", line 81, in run_surface
    model = _load_model(cfg, dist)
            ^^^^^^^^^^^^^^^^^^^^^^
  File "/home/Ising-Decoding/code/workflows/run.py", line 262, in _load_model
    model_path = find_best_model(model_dir, rank=dist.rank)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/Ising-Decoding/code/workflows/run.py", line 132, in find_best_model
    raise FileNotFoundError(f"No valid model checkpoint files found in {path}")
FileNotFoundError: No valid model checkpoint files found in /home/Ising-Decoding/outputs/predecoder_model_1/models

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

[ivanbasov]

Thanks for the report @bmhowe23! This is a regression from the rename — find_best_model had a hard-coded guard requiring filenames to start with PreDecoderModelMemory_, so the new names were silently skipped and the directory appeared empty.

Fix is in #55: when no epoch-numbered PreDecoderModelMemory_* checkpoints are found, find_best_model now falls back to any .pt file in the directory. Your existing workflow (copy the file → run inference) will work without any changes on your end.