_retrieve_segment: UnboundLocalError on `timestamps` masks intended ValueError when segment_offset <= 0

[Tintin0]

Bug: UnboundLocalError for timestamps in _retrieve_segment masks the intended ValueError

Repo/code: generation.py, method _retrieve_segment (as shipped in the remote code of BUT-FIT/DiCoW_v3_3; same code path in TS-ASR-Whisper).

Summary

In _retrieve_segment, the local timestamps is only assigned inside the else: branch of the
if len(timestamp_segment_indices) > 0: split, but the diagnostic at the end of the method
references timestamps unconditionally. When the if-branch is taken and segment_offset <= 0,
line ~524 raises

UnboundLocalError: local variable 'timestamps' referenced before assignment

instead of the intended, informative ValueError("Segment offset: ... <= 0. This should not happen! ...").
This masks the real condition and makes the failure very hard to diagnose. It is reachable e.g. on
prompt_ids-conditioned / certain long-form decoding paths.

Location (line numbers from the BUT-FIT/DiCoW_v3_3 checkpoint)

def _retrieve_segment(self, ...):
    ...
    if len(timestamp_segment_indices) > 0:
        ...                                  # segment_offset computed at ~L475 / ~L481
        #  <-- `timestamps` is NEVER assigned in this branch
    else:
        timestamps = seek_sequence[timestamp_tokens.nonzero().flatten()]   # ~L485  (only binding)
        ...
    if segment_offset <= 0:
        msg = f"Timestamps: {timestamps}, Segments: {segments}"            # ~L524  <-- UnboundLocalError
        raise ValueError(f"Segment offset: {segment_offset} <= 0. This should not happen!\n{msg}")
    return segments, segment_offset

Fix (minimal)

Initialise timestamps at method entry so the diagnostic always has a value:

 def _retrieve_segment(self, ...):
+    timestamps = None
     ...
     if len(timestamp_segment_indices) > 0:
         ...

(Alternatively, guard the message construction, but a top-of-function init is simplest and keeps the
intended ValueError diagnostic intact.)

Environment

• BUT-FIT/DiCoW_v3_3 remote code, transformers==4.55.0, torch 2.5.1 (MPS), Python 3.11.