ViperEkura/AstrAI
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Commit Graph

  • 9bcd696580 fix: token-level ratio and prompt masking in GRPO strategy ViperEkura 2026-07-12 21:23:40 +08:00
  • 8f89c82d55 chore: bump version to 1.3.9 v1.3.9 ViperEkura 2026-07-12 20:42:44 +08:00
  • 21871197d7 refactor: use actual q_len from input, remove dead num_splits init ViperEkura 2026-07-12 19:48:17 +08:00
  • 4c35d36146 fix: auto-assign free port in spawn_parallel_fn to avoid EADDRINUSE ViperEkura 2026-07-12 19:21:56 +08:00
  • 9aca62c26c perf: remove per-element sentinel checks in softmax + fix stale comments ViperEkura 2026-07-12 15:35:53 +08:00
  • b5cdea98ad refactor: remove ineffective __launch_bounds__ from prefill kernel ViperEkura 2026-07-12 15:00:15 +08:00
  • 69fecaf387 perf: double-buffer KV pipeline and Q direct-to-register in decode ViperEkura 2026-07-12 14:14:40 +08:00
  • fd6d25ad86 refactor: extract bench_kernel and dispatch_by_head_dim into test_utils ViperEkura 2026-07-12 00:02:05 +08:00
  • 2c3cef1c87 feat: wire up paged decode CUDA kernel to Python extension ViperEkura 2026-07-11 18:38:05 +08:00
  • 89ece26c25 feat: paged decode attention with split-KV (scalar + MMA) ViperEkura 2026-07-11 18:13:53 +08:00
  • 2c0b5d0b5e perf: enable MMA decode path for G=1 full attention ViperEkura 2026-07-11 11:45:13 +08:00
  • a4ae7d17fb perf: increase decode split-K parallelism for short sequences ViperEkura 2026-07-11 11:26:42 +08:00
  • 8a8550184f refactor: template AttentionParams, rename .cuh to .h ViperEkura 2026-07-11 11:03:14 +08:00
  • b8b439b713 perf: fuse decode combine kernel to single-pass online-rescale reduction ViperEkura 2026-07-11 00:24:09 +08:00
  • 41cd40363a perf: post-multiply attention scale in float instead of pre-scaling Q in bf16 ViperEkura 2026-07-11 00:12:12 +08:00
  • d923ebe38d refactor: rename gqa_* to attn_*, split-KV for all decode paths ViperEkura 2026-07-10 23:35:14 +08:00
  • 29b0423c4e refactor: deduplicate kernel code with shared MMA and entry-point helpers ViperEkura 2026-07-10 20:52:27 +08:00
  • 88f8dca2c2 perf: enlarge prefill KV tile to BC=32 for D<=128 ViperEkura 2026-07-10 17:41:31 +08:00
  • 9027fdc546 test: bench production MMA attention path with FLOP/s and bandwidth ViperEkura 2026-07-10 16:55:24 +08:00
  • cbd140340d perf: load prefill Q fragments directly from global (drop sQ staging) ViperEkura 2026-07-10 12:27:45 +08:00
  • 988e01314d perf: pipeline prefill MMA kernel (double-buffered K/V + packed stores) ViperEkura 2026-07-10 12:18:11 +08:00
  • 7ba43a7c6f perf: add split-K (FlashDecoding) to decode MMA kernel ViperEkura 2026-07-10 11:42:21 +08:00
  • dea59f7e1d fix: restore resident Qa to fix sQ overwrite bug ViperEkura 2026-07-10 00:49:58 +08:00
  • 85dc771460 perf: reduce MMA kernel registers, switch to static smem ViperEkura 2026-07-10 00:39:47 +08:00
  • 2c5629b81d docs: fix documentation errors across README and assets/docs ViperEkura 2026-07-09 10:09:53 +08:00
  • 841a582b28 refactor: split mask builder by single/multi output ViperEkura 2026-07-08 21:18:02 +08:00
  • c8567a6f65 fix: exclude embedding, lm_head, bias, and norm params from Muon optimizer, use AdamW ViperEkura 2026-07-08 19:41:16 +08:00
  • 8035be9b1f fix: make MuonMix inherit from torch.optim.Optimizer ViperEkura 2026-07-08 17:03:05 +08:00
  • e9b03f4fca perf: apply cp.async, XOR swizzle, pre-scaled Q to decode MMA kernel ViperEkura 2026-07-08 16:03:40 +08:00
  • fd65b9bc23 feat: support HEAD_DIM=32 and split extension into loader/ops ViperEkura 2026-07-08 13:59:54 +08:00
  • 9ebaea840f perf: cp.async K/V loads, shared sQ staging, causal skip, XOR swizzle ViperEkura 2026-07-08 12:30:32 +08:00
  • 6adc221c10 refactor: extract shared MMA utils into gqa_mma_utils.cuh ViperEkura 2026-07-07 23:01:15 +08:00
  • 9e63cb9ed0 feat: MMA head-packing decode kernel with scalar fallback dispatch ViperEkura 2026-07-07 22:56:02 +08:00
  • 4225518cf3 perf: add fast-math and vectorization nvcc/cxx build flags ViperEkura 2026-07-07 22:28:32 +08:00
  • c50adbaac0 feat : replace AdamW with MuonMix (Muon + AdamW) optimizer ViperEkura 2026-07-07 14:10:36 +08:00
  • 536dbc0c9a fix: set tqdm postfix before update so first step shows metrics ViperEkura 2026-07-07 00:14:13 +08:00
  • 4af7acd449 fix: support single .h5 file loading in load_h5 ViperEkura 2026-07-07 00:11:16 +08:00
  • 53ed52b4b8 refactor: extension dispatch layer with CUDA/torch fallback ViperEkura 2026-07-06 21:07:16 +08:00
  • f1cc7cedce feat: ldmatrix + smem padding for mma prefill kernel ViperEkura 2026-07-06 20:54:50 +08:00
  • ddc4bd1cf6 feat: tensor-core mma prefill with build-time dispatch ViperEkura 2026-07-06 19:51:55 +08:00
  • cc36530c73 perf: group-split register-blocking gqa_prefill kernel ViperEkura 2026-07-06 18:32:09 +08:00
  • 11fa807cfc fix: correct prefill mask index, unify GQA kernel interface ViperEkura 2026-07-06 16:52:19 +08:00
  • bcdd93e0eb feat: split kernel defs from bindings, add prefill tiled kernel and pure C tests ViperEkura 2026-07-06 16:14:34 +08:00
  • 579b8c3129 fix: correct gqa_decode_attn reduction + add gqa_prefill_attn ViperEkura 2026-07-06 13:44:23 +08:00
  • d7da51569f docs: update install instructions in EN/CN README ViperEkura 2026-07-06 12:25:17 +08:00
  • e8e228d035 feat: add optional CUDA kernel system (csrc/) + fused GQA decode attention ViperEkura 2026-07-06 12:08:12 +08:00
  • 2579658e15 chore : shields release badge from /release to /tag ViperEkura 2026-07-05 20:33:51 +08:00
  • f0cd0134c6 fix : update benchmark for v1.3.8 cache API, add argparse and cache type switch ViperEkura 2026-07-05 20:27:47 +08:00
  • abb96996f8 docs : sync 6 doc files to actual code ViperEkura 2026-07-05 19:34:30 +08:00
  • bbe6ff2d8f release : v1.3.8 v1.3.8 ViperEkura 2026-07-05 19:04:56 +08:00
  • db9b39b084 fix: resolve IFD token-set asymmetry and support single-token answers ViperEkura 2026-07-05 17:48:26 +08:00
  • 849e1e00a3 refactor: clean up inference design patterns ViperEkura 2026-07-05 11:41:54 +08:00
  • 5416c2e8fb perf: replace paged KV cache with contiguous ContiguousCache, decode all groups ViperEkura 2026-07-05 11:34:36 +08:00
  • 599a51f4f7 fix: reliable test timeout, separate generate/test phases, dynamic pass@k ViperEkura 2026-07-05 08:47:07 +08:00
  • 17d6eaa2f2 refactor: rewrite humaneval evaluation with functional pipeline design ViperEkura 2026-07-05 07:56:55 +08:00
  • 2d908639e9 feat : add ROUGE evaluation script (manual impl, no deps) ViperEkura 2026-07-05 01:15:01 +08:00
  • c7158418dd perf: add BFD bin-packing and custom attention mask to IFD batch scoring ViperEkura 2026-07-04 18:58:13 +08:00
  • 4d3c9341c1 refactor: rewrite IFD evaluation with clean three-layer architecture ViperEkura 2026-07-04 18:26:07 +08:00
  • 4e508afa2d fix : SFT pipeline position_ids default & doc boundary preservation ViperEkura 2026-07-04 15:59:11 +08:00
  • 8999ca89b8 feat: add JSONL dataset store with on-the-fly tokenization ViperEkura 2026-07-04 15:42:33 +08:00
  • 1adca39cd8 fix: handle long sequences and optimize IFD computation ViperEkura 2026-07-04 08:35:45 +08:00
  • 204873fa2f fix: handle long sequences and optimize IFD computation ViperEkura 2026-07-04 08:21:53 +08:00
  • a5c1de6b1b feat: add model_path temperature top_p top_k max_tokens system_prompt args to stream_chat ViperEkura 2026-07-04 07:02:06 +08:00
  • 27524ad085 fix: reset sampler iter at epoch end so progress bar shows total after first epoch ViperEkura 2026-07-04 06:35:29 +08:00
  • 27d1921d9c fix: scheduler division-by-zero, loss_mask bool ViperEkura 2026-07-03 22:03:36 +08:00
  • 70c0e5de90 refactor: merge validation into MetricCallback, simplify progress bar to optimizer steps ViperEkura 2026-07-03 21:43:08 +08:00
  • dfb151537b fix: ForwardRef._evaluate Python 3.12 compatibility ViperEkura 2026-07-03 18:41:19 +08:00
  • 500c605fad fix: unify scheduler min_rate default to 0.01, clamp WSD warmup ViperEkura 2026-07-03 17:52:23 +08:00
  • dc9faca3b1 fix: align docs with actual code (40+ inconsistencies) ViperEkura 2026-06-30 20:47:23 +08:00
  • aabb0d83e9 refactor : replace iteration with consumed_samples ViperEkura 2026-06-30 18:42:33 +08:00
  • 44579ea6dc refactor : metric 日志改为以 optimizer step 为单位，默认每步记录 ViperEkura 2026-06-30 15:12:31 +08:00
  • 0f1fcb079f refactor : grad_norm 指标简化，clip_grad_norm 移至 executor ViperEkura 2026-06-30 14:59:43 +08:00
  • 84d4769163 feat: SVD 有效秩/权重统计分析脚本 ViperEkura 2026-06-29 21:39:22 +08:00
  • bf09a35c95 feat: optimizer 参数分组，bias/norm 不做 weight decay ViperEkura 2026-06-27 16:30:34 +08:00
  • 6715461a36 chore : 升级 torch 2.11.0+cu128，移除自定义 Muon，修复 gloo device_id ViperEkura 2026-06-27 16:10:14 +08:00
  • b4587c5d08 refactor : metric_logger 改用事件类型 (type=step/validation/epoch) ViperEkura 2026-06-25 17:18:20 +08:00
  • 88ec63121d feat : GPT-2 residual scaling weight init ViperEkura 2026-06-25 15:04:27 +08:00
  • 01d2da2893 feat : 训练支持 --schedule_type 及对应调度器参数 ViperEkura 2026-06-22 10:35:14 +08:00
  • 25d4ea3f91 refactor : 压缩测试代码，消除重复 ViperEkura 2026-06-19 14:53:35 +08:00
  • 39985840c7 refactor : neftune_alpha 在 Embedding 构造时传入，由模型配置链路负责 ViperEkura 2026-06-19 14:23:27 +08:00
  • b1adc40cfb refactor : 将 config 对象直接传给 DecoderBlock，替代 16 个独立参数 ViperEkura 2026-06-19 14:15:20 +08:00
  • 7348bac6ab fix: 规范 generate.py 命令行接口 ViperEkura 2026-06-19 14:02:22 +08:00
  • 8ab7564d02 docs: 重构 README 结构，全文档添加目录导航 ViperEkura 2026-06-19 13:52:32 +08:00
  • d096b6e29e docs: 修复文档中过时的字段、签名和缺失的类 ViperEkura 2026-06-18 18:49:46 +08:00
  • d88a41f8f1 fix: 修复预处理流水线 4 个致命问题 ViperEkura 2026-06-18 17:38:01 +08:00
  • 376e9eba80 feat: IFEval 使用 chat template 格式化 prompt，添加 model.eval() ViperEkura 2026-06-18 16:45:16 +08:00
  • a62c2e11a2 feat: IFD 默认使用 chat template，支持裸文本模式 ViperEkura 2026-06-18 16:34:27 +08:00
  • a4e5a8c81c feat: 新增 WSD 学习率调度器 ViperEkura 2026-06-18 15:55:15 +08:00
  • 3e234c46f6 fix: 使用 threading.Event 替代裸 bool，补全公共 API ViperEkura 2026-06-18 15:38:13 +08:00
  • 7a04b1f8ce docs: replace shields.io endpoint badges with github/ direct badges ViperEkura 2026-06-18 15:09:51 +08:00
  • a30e3d5114 fix: 修复 shields.io GitHub badge 因 token 耗尽而无法显示 ViperEkura 2026-06-16 22:21:58 +08:00
  • 1818d06576 feat: 新增 IFD 数据质量评分工具, 移动 ppl 至 eval ViperEkura 2026-06-16 22:03:11 +08:00
  • 4e8d1ee24e feat: 新增 IFEval 指令遵循评测 ViperEkura 2026-06-16 21:57:23 +08:00
  • fec376b0dd fix : 修复策略相关文件的类型注解与抽象方法体 ViperEkura 2026-06-14 16:20:10 +08:00
  • a2512f8a5a fix : resume_dir 无权重文件时不强制加载，支持仅配置训练 ViperEkura 2026-06-13 15:40:14 +08:00
  • 457e16ea3c fix : val_loss 默认改为 None，日志跳过空值；val_dataloader 补 Optional 注解 ViperEkura 2026-06-13 14:24:13 +08:00
  • daf627a6de fix : _save_log 前确保日志目录存在，防止跨进程反序列化后目录丢失 ViperEkura 2026-06-12 15:39:54 +08:00
  • 445378667f feat : NEFTune 噪声注入 + label_smoothing 默认值修正 ViperEkura 2026-06-11 15:31:08 +08:00
  • 6ae1828449 refactor : 清理工厂和配置系统中的死代码与冗余抽象 ViperEkura 2026-06-07 11:39:50 +08:00
  • e7b18b7c03 refactor : BaseFactory 基类类型自动推导 + 移除冗余代码 ViperEkura 2026-06-06 21:23:41 +08:00