fix(auto): generate independent keys for unkeyed batch inputs

[hulkbig]

Summary

Generate a fresh automatic key for each non-path item in a batch when no explicit key was supplied.

prepare_data_iterator currently assigns the first generated key to the loop's key variable. Later array/byte items reuse it because key is no longer None. A path followed by an array similarly gives the array the preceding path's basename.

Minimal reproduction at e7e61293d308f4342ab9307895e710e4b51b949c:

import numpy as np
from funasr.auto.auto_model import prepare_data_iterator

waveforms = [np.zeros(3200, np.float32), np.ones(3200, np.float32)]
keys, _ = prepare_data_iterator(waveforms)
print(len(keys), len(set(keys)))
# Before: 2 1
# After:  2 2

The same duplicate keys propagate through the real AutoModel.generate wrapper with both batch_size=1 and batch_size=2. The result list still has two entries, but indexing it with {result["key"]: result for result in results} retains only the last entry. The regression tests exercise this using the repository's existing recording model.

The fix captures whether a key was supplied before the loop, rather than checking a value that the loop mutates.

Type of change

• Bug fix
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

python -m pytest -q tests/test_pcm_input_format.py tests/test_auto_model.py \
  tests/test_auto_model_ncpu.py tests/test_submodel_device.py \
  -k 'not test_merge_thr_in_cb_model'
python -m compileall funasr examples tests
git diff --check

• Base production code with the new regression tests: 9 failed, 109 passed, 1 deselected, 2 subtests passed. All nine failures are the new automatic-key regressions.

• Fixed code with the same selection: 118 passed, 1 deselected, 2 subtests passed.

• Independent review repeated the selection in separate base/fixed checkouts and added tests/test_load_audio_bytes.py with outbound networking blocked: 9 failed, 123 passed, 1 deselected, 4 subtests passed on base; 132 passed, 1 deselected, 4 subtests passed after the fix. Another 70 boundary checks passed on both versions, and both checkouts passed compileall.

• The deselected test downloads real model weights; it was not run.

• Existing NumPy copy-keyword warnings also occur on the base checkout.

• Linux x86_64, Python 3.12.14, PyTorch/torchaudio 2.8.0+cpu, NumPy 2.5.4.

• Separately compared 26 unaffected input/key cases on base and fixed source with identical results, including explicit mixed-input keys and manifests. The URL dispatch control used a fake downloader, with no network request.

• python -m compileall funasr examples tests

• Docs or links checked

• Runtime/deployment command tested

User impact

Applications passing batches of NumPy waveforms or explicit PCM bytes receive independently generated sample identifiers. Converting those results to a key-indexed mapping no longer deterministically collapses the batch to one entry.

Notes for reviewers

• Explicit keys remain unchanged, including key="session" producing repeated "session" keys and explicit empty strings.
• Existing path-basename behavior is preserved, including duplicate basenames and the current behavior of explicit-key mixed path/array batches. This does not introduce global basename deduplication.
• Tests cover lists/tuples, repeated input objects, arrays, explicit PCM, mixed paths/arrays, sample preservation, and the real AutoModel.generate routing for both batch sizes.
• Validation is weights-free. The existing _RecordingModel returns synthetic "ok" text; no ASR weights, recognition-quality benchmark, or GPU workload were run.
• This does not change audio decoding or model computation. DatadirWriter warns and appends duplicate-key rows; the failure described here is downstream dictionary indexing, not overwriting by that writer.
• AI assistance was used for investigation, implementation, and test preparation. The reproduction and listed checks were actually executed locally on CPU.

[LauraGPT]

Verified exact head 2e4eab4c531b45e14aac22e366dafd04b487d8d3 against base e7e61293d308f4342ab9307895e710e4b51b949c.

Independent weights-free validation reproduced the bug with the submitted regression tests on base (9 failed, 73 passed), then passed on this head (82 passed). The adjacent AutoModel/device/byte-loading selection passed 50 tests with 1 model-download test deselected, and compileall passed. The extracted source files matched the PR's Git blob IDs.

The change is limited to automatic result-key routing; it does not alter audio samples, decoding, model computation, or GPU runtime. Explicit keys and the existing path-basename behavior remain unchanged. I found no blocking issue in the patch.