-
2c3cef1c87
feat: wire up paged decode CUDA kernel to Python extension
main
ViperEkura
2026-07-11 18:38:05 +0800
-
89ece26c25
feat: paged decode attention with split-KV (scalar + MMA)
ViperEkura
2026-07-11 18:13:53 +0800
-
2c0b5d0b5e
perf: enable MMA decode path for G=1 full attention
ViperEkura
2026-07-11 11:45:13 +0800
-
a4ae7d17fb
perf: increase decode split-K parallelism for short sequences
ViperEkura
2026-07-11 11:26:42 +0800
-
8a8550184f
refactor: template AttentionParams, rename .cuh to .h
ViperEkura
2026-07-11 11:03:14 +0800
-
b8b439b713
perf: fuse decode combine kernel to single-pass online-rescale reduction
ViperEkura
2026-07-11 00:24:09 +0800
-
41cd40363a
perf: post-multiply attention scale in float instead of pre-scaling Q in bf16
ViperEkura
2026-07-11 00:12:12 +0800
-
d923ebe38d
refactor: rename gqa_* to attn_*, split-KV for all decode paths
ViperEkura
2026-07-10 23:35:14 +0800
-
29b0423c4e
refactor: deduplicate kernel code with shared MMA and entry-point helpers
ViperEkura
2026-07-10 20:52:27 +0800
-
88f8dca2c2
perf: enlarge prefill KV tile to BC=32 for D<=128
ViperEkura
2026-07-10 17:41:31 +0800
-
9027fdc546
test: bench production MMA attention path with FLOP/s and bandwidth
ViperEkura
2026-07-10 16:55:24 +0800
-
cbd140340d
perf: load prefill Q fragments directly from global (drop sQ staging)
ViperEkura
2026-07-10 12:27:45 +0800
-
988e01314d
perf: pipeline prefill MMA kernel (double-buffered K/V + packed stores)
ViperEkura
2026-07-10 12:18:11 +0800
-
7ba43a7c6f
perf: add split-K (FlashDecoding) to decode MMA kernel
ViperEkura
2026-07-10 11:42:21 +0800
-
dea59f7e1d
fix: restore resident Qa to fix sQ overwrite bug
ViperEkura
2026-07-10 00:49:58 +0800
-
85dc771460
perf: reduce MMA kernel registers, switch to static smem
ViperEkura
2026-07-10 00:39:47 +0800
-
2c5629b81d
docs: fix documentation errors across README and assets/docs
ViperEkura
2026-07-09 10:09:53 +0800
-
841a582b28
refactor: split mask builder by single/multi output
ViperEkura
2026-07-08 21:18:02 +0800
-
c8567a6f65
fix: exclude embedding, lm_head, bias, and norm params from Muon optimizer, use AdamW
ViperEkura
2026-07-08 19:41:16 +0800
-
8035be9b1f
fix: make MuonMix inherit from torch.optim.Optimizer
ViperEkura
2026-07-08 17:03:05 +0800
-
e9b03f4fca
perf: apply cp.async, XOR swizzle, pre-scaled Q to decode MMA kernel
ViperEkura
2026-07-08 16:03:40 +0800
-
fd65b9bc23
feat: support HEAD_DIM=32 and split extension into loader/ops
ViperEkura
2026-07-08 13:59:54 +0800
-
9ebaea840f
perf: cp.async K/V loads, shared sQ staging, causal skip, XOR swizzle
ViperEkura
2026-07-08 12:30:32 +0800
-
6adc221c10
refactor: extract shared MMA utils into gqa_mma_utils.cuh
ViperEkura
2026-07-07 23:01:15 +0800
-
9e63cb9ed0
feat: MMA head-packing decode kernel with scalar fallback dispatch
ViperEkura
2026-07-07 22:56:02 +0800
-
4225518cf3
perf: add fast-math and vectorization nvcc/cxx build flags
ViperEkura
2026-07-07 22:28:32 +0800
-
c50adbaac0
feat : replace AdamW with MuonMix (Muon + AdamW) optimizer
ViperEkura
2026-07-07 14:10:36 +0800
-
536dbc0c9a
fix: set tqdm postfix before update so first step shows metrics
ViperEkura
2026-07-07 00:14:13 +0800
-
4af7acd449
fix: support single .h5 file loading in load_h5
ViperEkura
2026-07-07 00:11:16 +0800
-
53ed52b4b8
refactor: extension dispatch layer with CUDA/torch fallback
ViperEkura
2026-07-06 21:07:16 +0800
-
f1cc7cedce
feat: ldmatrix + smem padding for mma prefill kernel
ViperEkura
2026-07-06 20:54:50 +0800
-
ddc4bd1cf6
feat: tensor-core mma prefill with build-time dispatch
ViperEkura
2026-07-06 19:51:55 +0800
-
cc36530c73
perf: group-split register-blocking gqa_prefill kernel
ViperEkura
2026-07-06 18:32:09 +0800
-
11fa807cfc
fix: correct prefill mask index, unify GQA kernel interface
ViperEkura
2026-07-06 16:52:19 +0800
-
bcdd93e0eb
feat: split kernel defs from bindings, add prefill tiled kernel and pure C tests
ViperEkura
2026-07-06 16:14:34 +0800
-
579b8c3129
fix: correct gqa_decode_attn reduction + add gqa_prefill_attn
ViperEkura
2026-07-06 13:44:23 +0800
-
d7da51569f
docs: update install instructions in EN/CN README
ViperEkura
2026-07-06 12:25:17 +0800
-
e8e228d035
feat: add optional CUDA kernel system (csrc/) + fused GQA decode attention
ViperEkura
2026-07-06 12:08:12 +0800
-
2579658e15
chore : shields release badge from /release to /tag
ViperEkura
2026-07-05 20:33:51 +0800
-
f0cd0134c6
fix : update benchmark for v1.3.8 cache API, add argparse and cache type switch
ViperEkura
2026-07-05 20:27:47 +0800
-
abb96996f8
docs : sync 6 doc files to actual code
ViperEkura
2026-07-05 19:34:30 +0800
-
bbe6ff2d8f
release : v1.3.8
v1.3.8
ViperEkura
2026-07-05 19:04:56 +0800
-
db9b39b084
fix: resolve IFD token-set asymmetry and support single-token answers
ViperEkura
2026-07-05 17:48:26 +0800
-
849e1e00a3
refactor: clean up inference design patterns
ViperEkura
2026-07-05 11:41:54 +0800
-
5416c2e8fb
perf: replace paged KV cache with contiguous ContiguousCache, decode all groups
ViperEkura
2026-07-05 11:34:36 +0800
-
599a51f4f7
fix: reliable test timeout, separate generate/test phases, dynamic pass@k
ViperEkura
2026-07-05 08:47:07 +0800
-
17d6eaa2f2
refactor: rewrite humaneval evaluation with functional pipeline design
ViperEkura
2026-07-05 07:56:55 +0800
-
2d908639e9
feat : add ROUGE evaluation script (manual impl, no deps)
ViperEkura
2026-07-05 01:15:01 +0800
-
c7158418dd
perf: add BFD bin-packing and custom attention mask to IFD batch scoring
ViperEkura
2026-07-04 18:58:13 +0800
-
4d3c9341c1
refactor: rewrite IFD evaluation with clean three-layer architecture
ViperEkura
2026-07-04 18:26:07 +0800
-
4e508afa2d
fix : SFT pipeline position_ids default & doc boundary preservation
ViperEkura
2026-07-04 15:59:11 +0800
-
8999ca89b8
feat: add JSONL dataset store with on-the-fly tokenization
ViperEkura
2026-07-04 15:42:33 +0800
-
1adca39cd8
fix: handle long sequences and optimize IFD computation
ViperEkura
2026-07-04 08:35:45 +0800
-
204873fa2f
fix: handle long sequences and optimize IFD computation
ViperEkura
2026-07-04 08:21:53 +0800
-
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 +0800
-
27524ad085
fix: reset sampler iter at epoch end so progress bar shows total after first epoch
ViperEkura
2026-07-04 06:35:29 +0800
-
27d1921d9c
fix: scheduler division-by-zero, loss_mask bool
ViperEkura
2026-07-03 22:03:36 +0800
-
70c0e5de90
refactor: merge validation into MetricCallback, simplify progress bar to optimizer steps
ViperEkura
2026-07-03 21:43:08 +0800
-
dfb151537b
fix: ForwardRef._evaluate Python 3.12 compatibility
ViperEkura
2026-07-03 18:41:19 +0800
-
500c605fad
fix: unify scheduler min_rate default to 0.01, clamp WSD warmup
ViperEkura
2026-07-03 17:52:23 +0800
-
dc9faca3b1
fix: align docs with actual code (40+ inconsistencies)
ViperEkura
2026-06-30 20:47:23 +0800
-
aabb0d83e9
refactor : replace iteration with consumed_samples
ViperEkura
2026-06-30 18:42:33 +0800
-
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 +0800
-
0f1fcb079f
refactor : grad_norm 指标简化，clip_grad_norm 移至 executor
ViperEkura
2026-06-30 14:59:43 +0800
-
84d4769163
feat: SVD 有效秩/权重统计分析脚本
ViperEkura
2026-06-29 21:39:22 +0800
-
bf09a35c95
feat: optimizer 参数分组，bias/norm 不做 weight decay
ViperEkura
2026-06-27 16:30:34 +0800
-
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 +0800
-
b4587c5d08
refactor : metric_logger 改用事件类型 (type=step/validation/epoch)
ViperEkura
2026-06-25 17:18:20 +0800
-
88ec63121d
feat : GPT-2 residual scaling weight init
ViperEkura
2026-06-25 15:04:27 +0800
-
01d2da2893
feat : 训练支持 --schedule_type 及对应调度器参数
ViperEkura
2026-06-22 10:35:14 +0800
-
25d4ea3f91
refactor : 压缩测试代码，消除重复
ViperEkura
2026-06-19 14:53:35 +0800
-
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 +0800
-
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 +0800
-
7348bac6ab
fix: 规范 generate.py 命令行接口
ViperEkura
2026-06-19 14:02:22 +0800
-
8ab7564d02
docs: 重构 README 结构，全文档添加目录导航
ViperEkura
2026-06-19 13:52:32 +0800
-
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 +0800
-
d88a41f8f1
fix: 修复预处理流水线 4 个致命问题
ViperEkura
2026-06-18 17:38:01 +0800
-
376e9eba80
feat: IFEval 使用 chat template 格式化 prompt，添加 model.eval()
ViperEkura
2026-06-18 16:45:16 +0800
-
a62c2e11a2
feat: IFD 默认使用 chat template，支持裸文本模式
ViperEkura
2026-06-18 16:34:27 +0800
-
a4e5a8c81c
feat: 新增 WSD 学习率调度器
ViperEkura
2026-06-18 15:55:15 +0800
-
3e234c46f6
fix: 使用 threading.Event 替代裸 bool，补全公共 API
ViperEkura
2026-06-18 15:38:13 +0800
-
7a04b1f8ce
docs: replace shields.io endpoint badges with github/ direct badges
ViperEkura
2026-06-18 15:09:51 +0800
-
a30e3d5114
fix: 修复 shields.io GitHub badge 因 token 耗尽而无法显示
ViperEkura
2026-06-16 22:21:58 +0800
-
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 +0800
-
4e8d1ee24e
feat: 新增 IFEval 指令遵循评测
ViperEkura
2026-06-16 21:57:23 +0800
-
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 +0800
-
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 +0800
-
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 +0800
-
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 +0800
-
445378667f
feat : NEFTune 噪声注入 + label_smoothing 默认值修正
ViperEkura
2026-06-11 15:31:08 +0800
-
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 +0800
-
e7b18b7c03
refactor : BaseFactory 基类类型自动推导 + 移除冗余代码
ViperEkura
2026-06-06 21:23:41 +0800
-
9e31d4ef2b
feat : BaseToolParser.feed 增加可选 token_ids 参数
ViperEkura
2026-06-06 11:19:02 +0800
-
52aa4d01d5
feat : 推理层增加 vLLM 风格工具调用解析
ViperEkura
2026-06-06 08:52:30 +0800
-
986be957ec
refactor : on_batch_begin 移入 accumulate 上下文
ViperEkura
2026-06-06 01:19:21 +0800
-
cf9c60841b
docs : 按代码反向修正所有文档错误
ViperEkura
2026-06-06 01:06:15 +0800
-
31bc7f5c2a
refactor : pipeline 策略化拆分，消除 _flush if/else
ViperEkura
2026-06-06 00:45:33 +0800
-
3057741de9
refactor : 合并 data config docstring 并实现 BFD 打包策略
ViperEkura
2026-06-05 17:41:51 +0800
-
acd1103bd0
fix : 使用 bool 注意力掩码并支持打包 SFT 文档边界阻断
ViperEkura
2026-06-05 17:02:00 +0800
-
dc7d2cfbca
refactor : FastAPI 懒加载单例，消除模块级副作用
ViperEkura
2026-06-04 15:52:01 +0800