WeClone-new

mirror of https://github.com/xming521/WeClone.git synced 2026-08-29 02:10:55 +08:00

Files

T

xhrxgr b8747edf27 fix: 量化配置未传递给 LLaMA-Factory 导致退化为全精度加载

- 新增 QuantizationArgs 配置类，与 LLaMA-Factory QuantizationArguments 对齐
- 移除 load_in_4bit/load_in_8bit（由 quantization_bit 内部处理）
- 新增 quantization_method 支持多种量化后端（bnb, gptq, awq 等）
- 修复 dict_to_argv 跳过 None 值，避免 HfArgumentParser 报错
- 使用 exclude_none=True 防止空字段传递给 run_exp()
- 训练和推理时自动展平嵌套的量化配置

2026-05-10 18:26:09 +08:00

core

refactor(infer): enhance guided decoding parameter handling

2026-03-29 11:25:18 +08:00

data

refactor(infer): streamline online model scoring and error handling

2026-01-04 18:14:57 +08:00

eval

refactor: derive API server URL from client.base_url for error message

2026-04-24 13:02:10 +08:00

prompts

feat(cleaning): add batch processing for online LLM