[AlexanderDokuchaev]

Changes

• Use stock transformers.integrations.executorch.convert_and_export_with_cache and
  TorchExportableModuleForDecoderOnlyLM.generate instead of the custom wrappers; modelling.py removed.
• Calibration inputs are now {"input_ids", "cache_position"} dicts truncated to the cache length.
• Generation runs under torch.no_grad(); one warm-up forward before timing; optimum dependency dropped
• update tests/cross_fw/examples/test_durations.json

Reason for changes

The example fails with transformers 5.x:

• IndexError: index_copy_(): index 512 is out of bounds — the custom forward skipped the cumulative_length
  reset of StaticCache, so calibration samples were appended until the 512-token cache overflowed.
• KeyError: 'cache_position' — generate() no longer passes cache_position.

Related tickets

N/A

Tests

Test examples - success

[Copilot]

🟡 Changes recommended

The pinned generation helper duplicates the final prompt token and changes generated output, while an unconditional xfail masks a failing Linux accuracy check.

3 open issues

• Duplicate final prompt token during generation · New
• Restrict Windows xfail and fix Linux word-count mismatch · New
• Use binary conversion for MiB size reporting · New

What changed in this PR

Updates NNCF’s TorchFX TinyLlama weight-compression example for Transformers 5.5, replacing custom export and generation wrappers with upstream helpers.

Changes:

• Uses named calibration inputs truncated to the static-cache length.
• Adds no-gradient generation, a warm-up forward, and baseline/compressed model reporting.
• Removes the custom wrappers and optimum dependency, and refreshes example test durations.

File	Description
tests/​cross_fw/​examples/​test_durations.json	Refreshes example timing data and removes one entry.
examples/​llm_compression/​torch_fx/​tiny_llama/​requirements.txt	Drops the optimum dependency.
examples/​llm_compression/​torch_fx/​tiny_llama/​modelling.py	Deletes custom export and generation wrappers.
examples/​llm_compression/​torch_fx/​tiny_llama/​main.py	Adopts upstream helpers and updates calibration, generation, and reporting.

_{Review effort: Balanced}