ViperEkura
  • Joined on 2026-04-02
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-28 21:02:31 +08:00
b37c3d000c docs : 同步文档与实际代码
6031020e37 feat : load_json/load_safetensors 支持 broadcast,跨节点分布式加载
c424dfc293 feat : checkpoint 支持保存 config.json
3a28e52e98 fix : start_epoch/start_batch 由用户参数决定,不再被 checkpoint 覆盖
e371908b54 fix : 保存 checkpoint 时 unwrap DDP/FSDP 避免 module. 前缀
Compare 7 commits »
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-28 14:38:31 +08:00
0a708fff24 docs : 更新架构文档与 storage 注释,同步 Store 重构
6e150ea6d0 refactor : Storage 层重构为 Store,移除 Fetcher 中间层,支持多段数据与显式长度
cb8dcb97ea refactor : 移除 -> None 返回值标注,拆分 FSDP 参数,新增 mmap 数据集存储
2d5dc93b3d fix : 修正类型标注与统一 CLI 参数命名
4145d35e3c refactor: 检查点加载重构,路径替代对象传递
Compare 9 commits »
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-26 16:47:42 +08:00
65ab69543b refactor: 统一序列化层,消除分散的 I/O 路径
1d26aa2e93 fix: 禁用DDP static_graph避免PyTorch 2.7.1下no_sync与backward冲突
a548d4553e fix: 断点续训恢复优化器/调度器状态及采样器剩余长度
dd1b39f435 fix: ProgressBar默认输出到stdout
94d6e713e9 test: 补充推理协议层单测覆盖
Compare 6 commits »
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-25 21:22:18 +08:00
737585a32a feat: 新增NTK-Aware RoPE缩放支持
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-25 21:20:19 +08:00
a304e16ff0 feat: 新增NTK-Aware RoPE缩放支持
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-25 20:15:38 +08:00
a4688021bf feat: 新增LoRA微调模块
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-25 20:11:39 +08:00
432145a798 feat: 新增LoRA微调模块
7df6eb9211 feat: 新增FSDP并行后端
Compare 2 commits »
ViperEkura pushed to main at ViperEkura/video-promo 2026-05-25 19:19:59 +08:00
9de0bad3d4 fix transformer: GQA text overflow, heatmap sizing, auto-regressive pos labels
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-25 16:05:45 +08:00
82a3f2626f docs: 更新文档与代码同步(Executor/训练循环/参数)
ViperEkura pushed to main at ViperEkura/llmEval 2026-05-24 22:36:42 +08:00
ac814e5c52 add all project source files
ViperEkura pushed to main at ViperEkura/llmEval 2026-05-24 22:36:19 +08:00
d8b83a175b first commit
ViperEkura created branch main in ViperEkura/llmEval 2026-05-24 22:36:19 +08:00
ViperEkura created repository ViperEkura/llmEval 2026-05-24 22:35:50 +08:00
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-24 20:55:28 +08:00
7fa69572c0 fix: 测试日志写入临时目录避免冗余文件
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-24 20:49:07 +08:00
3ab4f237e5 refactor: 重构训练后端为 Executor 模式
ViperEkura pushed to main at ViperEkura/postgraduate-prep 2026-05-24 15:33:45 +08:00
9a990bf942 feat: 新增变限积分章节(要点11-15,含Leibniz公式/绝对值处理/奇偶性)
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-24 15:15:40 +08:00
8cbf3f36e2 feat: 新增训练后端工厂框架
ViperEkura pushed to main at ViperEkura/AstrAI 2026-05-23 19:51:14 +08:00
0594ce1017 perf: Muon step 改用 torch._foreach_* 批处理并移除 NS 迭代的冗余 bf16 转换
ff509ff39f fix: decode后task_extend失败时提前中止,scheduler崩溃时通知waiting任务
Compare 2 commits »
ViperEkura pushed to master at ViperEkura/resume 2026-05-22 17:39:02 +08:00
00f14f7a9d init: resume template with personal content
ViperEkura created branch master in ViperEkura/resume 2026-05-22 17:39:02 +08:00