Add official 'safetensors -> GGUF' conversion utility for Hugging Face Whisper models

[officiallyutso]

Feature Request: Automatic Hugging Face safetensors -> .gguf Converter for Whisper Models

Currently, converting Whisper models from Hugging Face (especially fine-tuned ones in safetensors format) to gguf format for use with whisper.cpp is a multi-step, error-prone process requiring manual script patching, header editing, or loss of metadata (e.g. "dims").

I'm proposing a first-party utility script (or an enhancement to convert-pt-to-gguf.py) that:

• Accepts Hugging Face models (.safetensors, .bin, .pt)
• Uses transformers.WhisperForConditionalGeneration to parse model
• Outputs a .gguf file compatible with whisper.cpp
• Automatically extracts config (e.g. n_ctx, n_audio_ctx, vocab) from the model/tokenizer

This would simplify adoption for many users and reduce GitHub issues related to conversion errors like:

KeyError: 'dims'

---

Technical Proposal

A script like:

from transformers import WhisperForConditionalGeneration, WhisperTokenizer
from transformers import AutoTokenizer, AutoConfig, AutoModelForSeq2SeqLM
import torch

model_id = "openai/whisper-large-v3"
model = WhisperForConditionalGeneration.from_pretrained(model_id)
tokenizer = WhisperTokenizer.from_pretrained(model_id)

# Optional: convert model state_dict → ggml or gguf format
# Save vocab, audio hyperparameters, etc.
# Feed into whisper.cpp conversion chain

Coupled with gguf header builder (already present), the final .gguf model can be automatically produced.

---

Motivation

• Most Whisper checkpoints are now shared in safetensors format
• Current conversion scripts (convert-pt-to-ggml.py) expect .pt with "dims" key
• Users face KeyError, header errors, or broken audio configs
• Hugging Face models contain all required metadata — just not exposed in current scripts

---

Reference issue: #3315