[bugfix] fix streaming packing with truncation_strategy=split

[Kale112]

PR type

• Bug Fix

PR information

Fixes #8285.

IterablePackingDataset._fetch_data_out_queue assumed template.encode always returns a single dict. With truncation_strategy='split', an over-long sample is split into a list of chunks (template/base.py:_encode_truncated returns batched, and encode returns that list when more than one chunk is produced). _fetch_data_out_queue then ran len(data['input_ids']) on the list, raising:

TypeError: list indices must be integers or slices, not str

This only triggers for streaming=true + packing=true + truncation_strategy=split (a common long-document streaming pretraining setup). The non-streaming path is unaffected: EncodePreprocessor materializes split chunks into separate dataset rows before PackingDataset wraps them, so it only ever sees single-dict rows.

Fix: flatten the list into one bin-packing item per chunk, mirroring the existing pattern in RowPreprocessor.batched_preprocess (if isinstance(row, dict): row = [row]). Input order is preserved via res[i] (required for the sequential packing strategy). No behaviour change for the dict path (delete/left/right): each successful sample still produces exactly one (chunk, length) item in the same order as before, and failed/empty encodes are still skipped.

Experiment results

• Added unit test tests/train/test_packing.py::test_streaming_packing_split covering the dict (delete/left/right), split-list, and failed-encode (empty dict) cases; asserts correct flattening, input-order preservation, and that flattened chunks feed calculate_matched_group without error (tokens conserved).
• pre-commit run --files swift/dataset/packing.py tests/train/test_packing.py passes (flake8 / isort / yapf / trailing-whitespace / double-quote-string-fixer ...).
• Local test run:

$ python3 -m pytest tests/train/test_packing.py::test_streaming_packing_split -x -q
.
1 passed in 3.29s

[gemini-code-assist]

Code Review

This pull request fixes a crash that occurs when using streaming, packing, and splitting together. It updates _fetch_data_out_queue in swift/dataset/packing.py to correctly flatten list-of-chunks outputs from template.encode and adds a corresponding regression test. The review feedback suggests adding defensive checks to ensure each chunk is not empty and contains the 'input_ids' key before accessing it to prevent potential KeyError or TypeError exceptions.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

To ensure robust defensive programming, it is recommended to verify that each chunk is not empty and contains the 'input_ids' key before attempting to access it. This prevents potential KeyError or TypeError if an unexpected or empty chunk is returned by the template encoder.

Suggested change

	for chunk in data:
	last_res.append((chunk, len(chunk['input_ids'])))
	for chunk in data:
	if chunk and 'input_ids' in chunk:
	last_res.append((chunk, len(chunk['input_ids'])))

[Kale112]

Thanks, applied in f9b0280. Added the defensive guard (if chunk and 'input_ids' in chunk) and a regression test (test_streaming_packing_split_skips_malformed_chunks) covering empty chunks and chunks missing 'input_ids'.