fix: use text_config.eos_token_id for HunyuanOCR v1.5

[wendadawen]

Fix incorrect eos_token_id in HunyuanOCR v1.5 conversion causing inference to never stop

Problem

HunyuanOCR v1.5 is an updated version of HunyuanOCR v1.0. After converting v1.5 with convert_hf_to_gguf.py, the GGUF eos_token_id is incorrectly written as 120020, while the model's real stop token is 120007 from text_config.eos_token_id.

As a result, when llama.cpp generates 120007 it is not recognized as a stop token, and the model degenerates into repetitive generation until truncated by max_tokens (finish=length).

Root cause: v1.5's config.json stores the real stop token in the nested text_config.eos_token_id, while the top-level eos_token_id=120020 is a different token. The parent class HunYuanModel._fix_special_tokens only reads the top-level value.

Fix

Override _fix_special_tokens in HunyuanVLTextModel to overwrite with hparams["eos_token_id"] (base.py already merges text_config.eos_token_id into hparams) after calling the parent:

def _fix_special_tokens(self):
    super()._fix_special_tokens()
    eos_id = self.hparams.get("eos_token_id")
    if eos_id is not None:
        self.gguf_writer.add_eos_token_id(eos_id)

Test

Tested v1.0 and v1.5 on 20 cases:

• v1.5: all degenerated (finish=length) before fix; 20/20 normal stop after fix
• v1.0: no change before/after fix (v1.0 has no text_config.eos_token_id, GGUF unaffected)

Model Weights

• v1.5: https://huggingface.co/tencent/HunyuanOCR
• v1.0: https://huggingface.co/tencent/HunyuanOCR/tree/main/v1.0

• I have read and agree with the contributing guidelines
• AI usage disclosure: