Fix Fun-ASR-Nano CTC batch fallback

[LauraGPT]

Summary

• preserve the existing no-CTC batched LLM path
• when CTC timestamps are active, split multi-segment input back through the single-segment inference path instead of passing the whole batch into inference_prepare
• add a regression test for multi-segment Fun-ASR-Nano input with a CTC decoder

Root Cause

inference_llm skipped _inference_llm_batch when ctc_decoder existed, but then passed all VAD segments into inference_prepare, which intentionally rejects multi-item input with NotImplementedError("batch decoding is not implemented").

Verification

• RED: python3 -m pytest tests/test_fun_asr_nano_ctc_batch_fallback.py -q failed with NotImplementedError: batch decoding is not implemented before the fix
• GREEN: python3 -m pytest tests/test_fun_asr_nano_ctc_batch_fallback.py tests/test_fun_asr_nano_autocast_device.py tests/test_fun_asr_nano_openai_response.py tests/test_fun_asr_nano_repetition_penalty.py -q -> 13 passed
• python3 -m py_compile funasr/models/fun_asr_nano/model.py tests/test_fun_asr_nano_ctc_batch_fallback.py
• git diff --cached --check -- funasr/models/fun_asr_nano/model.py tests/test_fun_asr_nano_ctc_batch_fallback.py

[gemini-code-assist]

Code Review

This pull request implements sequential processing for multi-segment inputs when CTC timestamps are active by introducing the _inference_llm_ctc_sequential method along with helper functions _slice_batch_value and _merge_inference_meta, supported by a new unit test. The review feedback identifies two key issues: a potential runtime crash when key is None or out of bounds during sequential execution, and incorrect fallback behavior in _slice_batch_value where the entire original value is returned if the index is out of bounds. Code suggestions are provided to resolve both issues.

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]

If key is None or if its length is less than the number of segments in data_in, key_i will be set to None. When self.inference_llm is recursively called with key=None, it will crash with a TypeError: 'NoneType' object is not subscriptable when attempting to access key[0] (e.g., at lines 885 and 960). To prevent this runtime crash, we should provide a fallback key (such as a generated random key or index-based key) when key is unavailable or out of bounds.

Suggested change

	key_i = [key[i]] if key is not None and i < len(key) else None
	key_i = [key[i]] if key is not None and i < len(key) else [f"rand_key_{i}"]

[gemini-code-assist]

In _slice_batch_value, if value is a multi-element Tensor, list, or tuple, but index is out of bounds (i.e., index >= len(value)), the function falls through and returns the entire original value. This is incorrect because the downstream code expects a single-item slice or None for the out-of-bounds index. We should explicitly return None when the index is out of bounds for these sequence types.

Suggested change

	def _slice_batch_value(value, index):
	if value is None:
	return None
	if isinstance(value, torch.Tensor) and value.ndim > 0 and value.shape[0] > index:
	return value[index : index + 1]
	if isinstance(value, list) and len(value) > index:
	return [value[index]]
	if isinstance(value, tuple) and len(value) > index:
	return (value[index],)
	return value
	def _slice_batch_value(value, index):
	if value is None:
	return None
	if isinstance(value, torch.Tensor) and value.ndim > 0:
	return value[index : index + 1] if value.shape[0] > index else None
	if isinstance(value, list):
	return [value[index]] if len(value) > index else None
	if isinstance(value, tuple):
	return (value[index],) if len(value) > index else None
	return value