# HGA_Thinker_7B_v2_export

HGA-Thinker speech LM exported from SFT training (current train_sft.py format).

## Layout
- `bridge.pt` — HGA (s/b/c on 32 Whisper layers) + EMCA + audio boundary embeds
- `lora/` — PEFT LoRA adapter for the LLM (has_lora=True)
- `config.json` — HGAThinkerConfig (llm_dim=3584)
- `processor_config.json` — inference defaults
- `tokenizer.*` — Qwen2.5 tokenizer
- `preprocessor_config.json` — WhisperFeatureExtractor config
- `configuration_hga_thinker.py`, `modeling_hga_thinker.py`, `thinker/` —
  frozen inference code (so this dir loads without the training repo)

## Base models
- whisper: `/apdcephfs_hzlf/share_1227201/zefeng/whisper-large-v3`
- llm:     `/apdcephfs_hzlf/share_1227201/zefeng/Qwen2.5-7B-Instruct`
- bundled: `False`   trained_steps: `6000`

## Load & run
```python
import sys; sys.path.insert(0, ".")          # this dir
from modeling_hga_thinker import HGAThinkerForConditionalGeneration

model = HGAThinkerForConditionalGeneration.from_pretrained(".", device="cuda")
# If base models moved since export, pass overrides:
#   ...from_pretrained(".", whisper_path="/new/whisper", llm_name="/new/qwen")

text = model.chat(audio="test.wav", query="Transcribe this audio.",
                  max_new_tokens=256)
print(text)
```

Needs `peft` and (for audio path inputs) `torchaudio` installed.