If you want sentence-level translation (not line-by-line) at scale, fighting an LLM’s context limits is the wrong battle. The workaround is to preprocess the SRT into sentence “blocks” (while preserving timestamps), translate blocks in batches, then reflow back into subtitle lines with your max chars/line rules.

A practical setup that avoids the “file too large” refusal:

• Step 1 (merge): Merge adjacent SRT lines into a paragraph until you hit a sentence boundary (., ?, !, quotes) or a time-gap threshold (e.g., >300–500ms).
• Step 2 (translate): Translate those sentence blocks in chunks (e.g., 30–80 sentences per batch) with a strict format (ID → translation).
• Step 3 (reflow): Split translated sentences back into 1–2 lines per caption, keeping original timings, enforcing max chars/line and reading speed.

If you’d rather not engineer all that, a subtitle-native workflow is easier: tools like Vitra’s Translate. video handle subtitles as subtitles (sentence-aware translation + consistent terminology), so you’re not manually babysitting chunking and retries like you are with llama.cpp.

The key idea: don’t ask the model to “translate a whole file”—make the file translation a deterministic pipeline, and use the model only for the translation step.