|
|
@@ -177,7 +177,7 @@ hf download opendatalab/MinerU2.5-Pro-2604-1.2B --local-dir ~/models/MinerU2.5-P
|
|
|
# 3. 转语言模型主体(M4 48G 内存充足,直接 f16 不量化质量最好)
|
|
|
python convert_hf_to_gguf.py ~/models/PaddleOCR-VL-1.6 \
|
|
|
--outfile ~/models/PaddleOCR-VL-1.6-GGUF \
|
|
|
- --outtype f16
|
|
|
+ --outtype q8_0
|
|
|
|
|
|
# 4. 转视觉投影器 mmproj(OCR 质量主要看这个,建议 f16/f32)
|
|
|
python convert_hf_to_gguf.py ~/models/PaddleOCR-VL-1.6 \
|
|
|
@@ -187,14 +187,14 @@ python convert_hf_to_gguf.py ~/models/PaddleOCR-VL-1.6 \
|
|
|
cp ~/models/PaddleOCR-VL-1.6/chat_template.jinja ~/models/PaddleOCR-VL-1.6-GGUF/chat_template.jinja
|
|
|
|
|
|
# 5. 转语言模型主体(M4 48G 内存充足,直接 f16 不量化质量最好)
|
|
|
-python convert_hf_to_gguf.py ~/models/MinerU2.5-Pro-2604-1.2B \
|
|
|
- --outfile ~/models/MinerU2.5-Pro-2604-1.2B-GGUF \
|
|
|
- --outtype f16
|
|
|
+python convert_hf_to_gguf.py ~/models/MinerU2.5-Pro-2605-1.2B \
|
|
|
+ --outfile ~/models/MinerU2.5-Pro-2605-1.2B-GGUF \
|
|
|
+ --outtype q8_0
|
|
|
|
|
|
# 6. 转视觉投影器 mmproj(OCR 质量主要看这个,建议 f16/f32)
|
|
|
-python convert_hf_to_gguf.py ~/models/MinerU2.5-Pro-2604-1.2B \
|
|
|
+python convert_hf_to_gguf.py ~/models/MinerU2.5-Pro-2605-1.2B \
|
|
|
--mmproj \
|
|
|
- --outfile ~/models/MinerU2.5-Pro-2604-1.2B-GGUF/MinerU2.5-Pro-2604-1.2B-mmproj.gguf \
|
|
|
+ --outfile ~/models/MinerU2.5-Pro-2605-1.2B-GGUF/MinerU2.5-Pro-2605-1.2B-mmproj.gguf \
|
|
|
--outtype f16
|
|
|
cp ~/models/MinerU2.5-Pro-2604-1.2B/chat_template.jinja ~/models/MinerU2.5-Pro-2604-1.2B-GGUF/chat_template.jinja
|
|
|
|
|
|
@@ -262,7 +262,7 @@ ocr_tools/daemons/curl_local_mineru.sh
|
|
|
conda activate mineru
|
|
|
python - <<'PY'
|
|
|
import json, os
|
|
|
-d = os.path.expanduser("~/models/MinerU2.5-Pro-2604-1.2B")
|
|
|
+d = os.path.expanduser("~/models/MinerU2.5-Pro-2605-1.2B")
|
|
|
otsl = {"<ched>", "<ecel>", "<fcel>", "<lcel>", "<ucel>", "<xcel>", "<nl>"}
|
|
|
|
|
|
# 1) tokenizer.json:convert_hf_to_gguf 据此决定 token 类型(关键)
|
|
|
@@ -286,8 +286,8 @@ print("已将 OTSL token 标记为非 special,可重新转换主模型 GGUF")
|
|
|
PY
|
|
|
|
|
|
# 重转主模型(mmproj 不用动)
|
|
|
-python convert_hf_to_gguf.py ~/models/MinerU2.5-Pro-2604-1.2B \
|
|
|
- --outfile ~/models/MinerU2.5-Pro-2604-1.2B-GGUF \
|
|
|
+python convert_hf_to_gguf.py ~/models/MinerU2.5-Pro-2605-1.2B \
|
|
|
+ --outfile ~/models/MinerU2.5-Pro-2605-1.2B-GGUF \
|
|
|
--outtype f16
|
|
|
```
|
|
|
|
|
|
@@ -332,8 +332,8 @@ f32 / f16 / bf16 / q8_0 / tq1_0 / tq2_0 / auto
|
|
|
`q8_0` 在列,所以一步直出没问题:
|
|
|
|
|
|
```bash
|
|
|
-python convert_hf_to_gguf.py ~/models/MinerU2.5-Pro-2604-1.2B \
|
|
|
- --outfile ~/models/MinerU2.5-Pro-2604-1.2B-GGUF \
|
|
|
+python convert_hf_to_gguf.py ~/models/MinerU2.5-Pro-2605-1.2B \
|
|
|
+ --outfile ~/models/MinerU2.5-Pro-2605-1.2B-GGUF \
|
|
|
--outtype q8_0
|
|
|
# mmproj 仍单独转、保持 f16(见前面步骤)
|
|
|
```
|
|
|
@@ -356,8 +356,8 @@ LLAMA_BIN="$HOME/workspace/repository.git/llama.cpp/build/bin"
|
|
|
|
|
|
# MinerU2.5:只量化语言主体,mmproj 保持 f16 不动
|
|
|
"$LLAMA_BIN/llama-quantize" \
|
|
|
- ~/models/MinerU2.5-Pro-2604-1.2B-GGUF/MinerU2.5-Pro-2604-1.2B-F16.gguf \
|
|
|
- ~/models/MinerU2.5-Pro-2604-1.2B-GGUF/MinerU2.5-Pro-2604-1.2B-Q8_0.gguf \
|
|
|
+ ~/models/MinerU2.5-Pro-2605-1.2B-GGUF/MinerU2.5-Pro-2605-1.2B-F16.gguf \
|
|
|
+ ~/models/MinerU2.5-Pro-2605-1.2B-GGUF/MinerU2.5-Pro-2605-1.2B-Q8_0.gguf \
|
|
|
Q8_0
|
|
|
|
|
|
# PaddleOCR-VL-1.6:同理
|