ViperEkura

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-15 22:56:13 +08:00

c169659611 docs: 修正 assets/docs/ 类图、数据流、参数文档及贡献指南

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-15 21:37:58 +08:00

e12f1a7ee5 feat: BaseModelConfig + DeepSeekMoE + 工厂模式替代 if/else

ef25efffa2 refactor: 拆分 module.py 为 components 子包

19532440b4 chore: 版本号升至 1.3.5

Compare 3 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-15 18:11:08 +08:00

9096e413c3 refactor: RotaryEmbedding 合并 cos/sin 为单一复数缓存

9d5e9fa6c4 perf: DDP 加 gradient_as_bucket_view/static_graph/broadcast_buffers，AdamW fused

08dde46778 fix: 修复训练循环 step/backward 顺序，重构为三重循环嵌套

Compare 3 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-14 21:38:17 +08:00

513f1f7826 perf: waiting_queue 改用 deque，pull_candidates 从 O(n²) 降到 O(1)

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-14 21:31:43 +08:00

e3382f6bb5 fix: 修复推理引擎 batch decode 中多项正确性与并发问题

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-14 21:27:36 +08:00

29b5717a38 fix: 修复推理引擎 batch decode 中多项正确性与并发问题

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-14 21:00:44 +08:00

f0339022c1 fix: batch 推理示例添加 chat template 和 system prompt

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-14 20:32:13 +08:00

d8da2cf17c docs: 修复文档中与源码不符的类名、方法签名和模块归属

205b40bd28 refactor: 重构 cache 和 inference 参数体系，分离存储与分配

18fe6e9339 refactor: 消除多处重复模式，统一工厂和参数传递

2196c34c52 refactor: 重构 inference 模块架构，引入设计模式并分组文件

466c2e1efd fix: process_attention_mask 中 expand 后的 inplace 写导致 alias 报错

Compare 18 commits »

ViperEkura pushed to main at ViperEkura/postgraduate-prep

2026-05-14 15:40:39 +08:00

89a3b16ef5 feat: 积分-万能代换增加判别式分析（a²-b²决定arctan/ln形式）

2fdf8e35ab feat: 增加错题 1/(1+√((1+x)/x)) 积分（有理化分母+配平方）

f3dfb5273d feat: 增加错题 1/(a+b cosx) 积分（万能代换+判别式分情况）

900e6c3e62 refactor: 规划对齐张宇+王道课程，按子专题细化，目标6月底完成第一轮

Compare 4 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-11 17:52:52 +08:00

38e18fdfd3 refactor: PagedCache Facade 模式，提取 PagePool/PrefixCache/TaskTable

4753958f92 refactor: 页状态移入 PagedCache，Task 纯化为域对象

73d6cc0f26 refactor: TaskManager 剥离页管理，STOP 移至 task.py

317ed90bac refactor: 拆分 scheduler 为 TaskManager + Executor

Compare 4 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-10 21:07:53 +08:00

951df8155c perf: gather 向量化

a58fab8d6e fix: max_seq_len 检查改为仅 prompt 超限发 STOP，max_tokens 超出部分 clamp

a3c8296135 fix: page cache 分配失败越界崩溃 + 长度超限终止

c95ace41aa fix: prefill 时 attention mask 长度不足导致 expand 崩溃

Compare 4 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-10 18:25:42 +08:00

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-10 18:23:33 +08:00

a3bde30fb1 feat: 服务化基础设施 - 有界队列/超时/优雅关闭/metrics

3da428e0e4 perf: PagedCache 持久前缀缓存 + LRU 逐出

133a9de98f feat: _generate_streaming 支持 batch 模式

Compare 3 commits »

ViperEkura pushed to main at ViperEkura/postgraduate-prep

2026-05-10 17:31:11 +08:00

6335d21418 feat: 增加杂项随手记（个人发现与零散规律总结）

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-10 17:25:42 +08:00

523eacf5fe release: v1.3.4

cffedaad5e perf: 消除非流式推理 CPU 空转并减少 decode GPU 张量冗余分配

Compare 2 commits »

ViperEkura pushed to main at ViperEkura/postgraduate-prep

2026-05-10 16:55:02 +08:00

d744d4605b feat: 增加级数敛散性判别与反常积分审敛法（含错题与比值审敛证明）

ViperEkura pushed to main at ViperEkura/AstrAI

2026-05-09 23:55:12 +08:00

3583c46b66 feat: 推理引擎前缀缓存（KV cache 复用）

ViperEkura created repository ViperEkura/ToolAgent

2026-05-09 23:13:24 +08:00

ViperEkura pushed to main at ViperEkura/video-promo

2026-05-09 20:19:42 +08:00

4f14d09fe3 fix paged_cache: _TaskRow class, expand rendering bug, and order corrected

ViperEkura pushed to main at ViperEkura/video-promo

2026-05-09 18:53:11 +08:00

bb0c32b032 redesign continuous batching: simplify to lane-based FSM with Prefill state