Fix/hf offline resilient loading

[wilke0818]

• HF models create 429 errors when loading too many, especially when parallelizing across nodes and calling many models sequentially
• Fixes base logic for this
• Should extend across all models/HF spaces if not already done

[gemini-code-assist]

Code Review

This pull request introduces a comprehensive audio analysis workflow, a three-axis uncertainty framework, multilingual MMS forced-alignment, and support for new ASR backends (Canary-Qwen, Qwen3-ASR, and Granite Speech) with robust CUDA-aware virtual environment provisioning and PII detection. The review feedback highlights several critical bugs and improvement opportunities, including a timestamp ratio calculation bug for stereo audio in forced_alignment.py, potential process crashes in cuda_probe.py due to narrow exception handling, a potential AttributeError in script_line.py, and opportunities to reduce code duplication and remove unused helper functions in aggregate.py.

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]

The calculation of ratio multiplies duration by waveform_segment.size(0). Since waveform_segment is shaped (channels, samples), size(0) represents the number of channels (which is 1 for mono, but 2 for stereo). If the input audio is stereo, this will double the ratio and result in incorrect timestamps. It is safer and cleaner to calculate the ratio directly as duration / (trellis.size(0) - 1) without multiplying by the channel count.

Suggested change

	ratio = duration * waveform_segment.size(0) / (trellis.size(0) - 1)
	ratio = duration / (trellis.size(0) - 1)

[gemini-code-assist]

Catching only FileNotFoundError when running nvidia-smi can lead to unhandled exceptions (such as PermissionError or other system-level OSError variants) which would crash the process. It is safer to catch OSError to handle all process execution failures gracefully.

    except (OSError, subprocess.TimeoutExpired):

[gemini-code-assist]

Catching only FileNotFoundError when running nvcc can lead to unhandled exceptions (such as PermissionError or other system-level OSError variants) which would crash the process. It is safer to catch OSError to handle all process execution failures gracefully.

Suggested change

	except (FileNotFoundError, subprocess.TimeoutExpired):
	except (OSError, subprocess.TimeoutExpired):

[gemini-code-assist]

To prevent potential AttributeError if a chunk element c is None, add a defensive null check before calling .get() on it.

Suggested change

	def _is_meaningful(c: Dict[str, Any]) -> bool:
	return c.get("text") is not None or c.get("speaker") is not None
	def _is_meaningful(c: Dict[str, Any]) -> bool:
	return c is not None and (c.get("text") is not None or c.get("speaker") is not None)

[gemini-code-assist]

The loop in aggregate_presence is identical to the one in presence_p_voice. To reduce code duplication and improve maintainability, aggregate_presence can directly call presence_p_voice and apply the final transformation on the result.

def aggregate_presence(votes: dict[str, dict[str, Any]]) -> float | None:
    p_voice = presence_p_voice(votes)
    if p_voice is None:
        return None
    return max(0.0, min(1.0, 1.0 - abs(2.0 * p_voice - 1.0)))

[gemini-code-assist]

The private helper function _normalize_transcript_for_wer is defined but never used anywhere in this module. Consider removing it to keep the codebase clean.