ViperEkura

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-20 03:59:16 +08:00

c775a2b3e0 Update DPO metrics figure

ViperEkura pushed to main at ViperEkura/AstrAI

2026-07-19 16:15:31 +08:00

663ef900fc refactor: move sample-id indexing from dataset to store

7d478a54db docs: update HF org from ViperEk to ViperEkura

f3eaaef842 refactor: remove redundant strategy/executor code

d655b65027 docs: sync architecture/dataflow/training/params with code

31c22dc043 refactor: deduplicate preprocessing kernel and BFD packing

Compare 20 commits »

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-19 15:30:59 +08:00

3f0ff911a8 Fix lingering cosine->WSD in Conclusion; update AGENTS.md with checkpoint labels and scheduler rules

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-19 15:25:52 +08:00

e149997200 Add DPO data generation & training; fix checkpoint labels; switch to WSD scheduler; update optimizer ablations in abstract/conclusion

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-19 15:01:42 +08:00

6c2e04a86f Add DPO training subsection with dpo_metrics figure and Rafailov et al. citation

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-18 01:16:48 +08:00

411354eeb1 Update sft_metrics.png

a0c39601a0 Add effective rank analysis, weight distribution evolution, Muon/SFT training figures

Compare 2 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-07-17 22:25:00 +08:00

d08a92c7bd feat: add frequency penalty to inference sampling pipeline

a1ea26d367 fix: rewrite GRPO data pipeline for offline record-level access

c17aa0dc54 fix: eval script bugs and add missing features

b12b24eadc feat: rewrite evaluate_ppl with token-level loss and multi-file support

cd14d53707 feat: implement bfd_split packing strategy

Compare 24 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-07-11 18:45:38 +08:00

2c3cef1c87 feat: wire up paged decode CUDA kernel to Python extension

89ece26c25 feat: paged decode attention with split-KV (scalar + MMA)

2c0b5d0b5e perf: enable MMA decode path for G=1 full attention

a4ae7d17fb perf: increase decode split-K parallelism for short sequences

8a8550184f refactor: template AttentionParams, rename .cuh to .h

Compare 16 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-07-09 10:10:16 +08:00

2c5629b81d docs: fix documentation errors across README and assets/docs

841a582b28 refactor: split mask builder by single/multi output

c8567a6f65 fix: exclude embedding, lm_head, bias, and norm params from Muon optimizer, use AdamW

8035be9b1f fix: make MuonMix inherit from torch.optim.Optimizer

e9b03f4fca perf: apply cp.async, XOR swizzle, pre-scaled Q to decode MMA kernel

Compare 13 commits »

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-08 22:50:06 +08:00

bfc8ff6098 Fix training budget (19k steps / ~20B tokens), add Muon optimizer, correct variance scaling math, move figures to §Training Config, tighten abstract

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-08 22:27:56 +08:00

4a143b056d Tighten abstract, add Muon optimizer mention, update ckpt_comparison figure

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-08 22:00:10 +08:00

5726bf6c84 Add optimizer/initialization comparison figure (ckpt_comparison)

ViperEkura pushed to main at ViperEkura/AstrAI

2026-07-06 22:01:56 +08:00

53ed52b4b8 refactor: extension dispatch layer with CUDA/torch fallback

f1cc7cedce feat: ldmatrix + smem padding for mma prefill kernel

ddc4bd1cf6 feat: tensor-core mma prefill with build-time dispatch

cc36530c73 perf: group-split register-blocking gqa_prefill kernel

11fa807cfc fix: correct prefill mask index, unify GQA kernel interface

Compare 11 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-07-05 19:46:08 +08:00

abb96996f8 docs : sync 6 doc files to actual code

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-05 19:14:16 +08:00

ff13a45d95 Restructure data pipeline: replace IFD analysis with length bias appendix

ViperEkura pushed tag v1.3.8 to ViperEkura/AstrAI

2026-07-05 19:13:37 +08:00

ViperEkura pushed to main at ViperEkura/AstrAI

2026-07-05 19:13:37 +08:00

bbe6ff2d8f release : v1.3.8

db9b39b084 fix: resolve IFD token-set asymmetry and support single-token answers

849e1e00a3 refactor: clean up inference design patterns

5416c2e8fb perf: replace paged KV cache with contiguous ContiguousCache, decode all groups

599a51f4f7 fix: reliable test timeout, separate generate/test phases, dynamic pass@k

Compare 7 commits »

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-05 12:38:49 +08:00

2d52a00149 Update ifd_loss_ratio_density.png

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-05 12:31:50 +08:00

f2c2b159ba Add IFD loss ratio density analysis, weight distribution appendix, and rewrite abstract

ViperEkura pushed to master at ViperEkura/AstrAI-paper

2026-07-05 00:24:52 +08:00

459523f0e4 Fix IFD citation: correct author list and venue (NAACL 2024, not NeurIPS)