ViperEkura

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-31 14:29:07 +08:00

432dfec3c2 refactor: collapse fp8 recipe hierarchy and state property layers

e3c3e28a11 docs: fix stale developer documentation claims

a7d4cb25c5 docs: scope trainer environment variables per job

0546331637 fix: skip gradient checkpointing log when no modules configured

962c10c52b perf: fold the delayed-scaling ring update into the quantize kernel

Compare 7 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-31 13:40:24 +08:00

a1a1a6bf0f perf: pack gqa q-heads per prefill block to reuse kv tiles

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-28 17:38:27 +08:00

7dd184a4e5 refactor: split fp8 gemm device code into layered headers

bf239d194c refactor: dedupe fp8 kernel helpers and trim comments

8a353117ea perf: transpose-quantize backward operands to route all gemms nt

04a8e2517a perf: split fp8 gemm cta plan by operand layout

c4f7f82725 refactor: drop dead fp8 gemm knobs and dedupe ring depth logic

Compare 5 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-28 01:22:17 +08:00

fac9d07542 refactor: fp8 gemm policy layering with swap-NN and narrow-N ctas

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-28 00:03:38 +08:00

bbb2d95256 fix: report true rank in logs via dist-aware helpers

1c04a0b9fa refactor: move runtime parsers into scripts/docker

ba8beb81be fix: serve and train reuse the built image and expose GPUs correctly

7cfcc6c86a Merge pull request #26 from Cytosine-code/fix/cuda-kernel-install

f4c44ebf1c docs: correct fp8 kernel descriptions in cuda_kernels guide

Compare 8 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-26 18:15:35 +08:00

6354dbe8bc perf: interleave prefetch into the mma phase

f45230fb2c perf: pair b fragments into ldmatrix x4 loads

d4534be8ca perf: single-wave big cta takes a stage deeper pipeline

1d57588d27 fix: free default ubuntu uid/gid before creating astrai user

a92bf79295 perf: fast interior loop on the big cta and fused epilogue bias

Compare 11 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-25 17:25:32 +08:00

3e57cc8069 feat: fp8 backward reuses pre-quantized operands

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-25 16:56:16 +08:00

2eeac02d70 perf: overlap direct crosswise loads with fp8 gemm mma

5e76fbd1bf perf: fp8 rings, lean autocast, gemm staging

4dc5e923e0 perf: finalize fp8 scale rings inside quantize kernels

Compare 3 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-24 23:32:51 +08:00

998b443aa3 refactor: dedupe fp8 meta state into per-operand rings

cebdd45d3a perf: batch crosswise stage loads in fp8 gemm

7da1439c9e feat: static fp8 weights and bias with fused epilogue

29e5f571af fix: own fp8 linear backward via autograd Function

74e694921c perf: speed up fp8 gemm tiles and scheduling

Compare 27 commits »

ViperEkura pushed to main at ViperEkura/DataPipeline

2026-08-19 18:01:15 +08:00

995a015c23 fix: cache_h5 移除重复 --batch-size 参数，默认输出格式改为 bin

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-19 00:43:53 +08:00

3d3ea47d37 refactor: standardize packed 3d inference

f7f14d0e5f test: add online GRPO end-to-end training test

7580d80d45 perf: accelerate FP8 backward with fused fast kernel

cb51a3587b perf: optimize fused FP8 GEMM kernel

Compare 4 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-16 23:42:38 +08:00

1bcd8f53ab perf: precompute ragged Q tile scheduling

0d0dc64884 docs: explain extension layer boundaries

Compare 2 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-16 21:26:54 +08:00

6ac3b51496 refactor: separate extension ops and backends

3406157431 refactor: standardize packed 3d inference

0dd9a417b7 refactor: separate KV token address resolution

a01c1fd427 perf: bypass L1 for attention tile loads

f8d9ab344d refactor: remove unused streaming dataset

Compare 8 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-14 12:26:02 +08:00

0378e62e17 refactor: split fp8 into fp8_ops adapter and fp8 policy module

5244f1a8fc feat: add te-style scaled fp8 training via fp8_autocast

5104638447 perf: use fp8 tensor-core gemm in linear backward

a711d9f478 perf: eliminate gemm output transpose via A/B swap

15862d4b56 perf: fuse fp8 linear fwd and bwd into single kernel calls

Compare 44 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-07 22:32:15 +08:00

02469887f5 refactor: simplify inference engine and backend dispatch

05739629fc feat: add timed() context manager and backend supports()

e0f7fa8e13 feat: enable CUDA graph by default with init-time warmup

af25833fab fix: add out_buf to attn_paged_decode for CUDA graph capture compatibility

6572be4f98 fix: prevent signal handler test from racing with training completion

Compare 16 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-06 12:15:01 +08:00

6f09b1d2ee docs : clarify radix cache architecture

b2230fefd8 feat : add radix prefix cache

654e6eb0d1 fix : correct prefill sampling and record alignment

a317a4756b refactor: stateless MoE routing with grouped dispatch

9b7e6c205f feat: add moe auxloss and metrics

Compare 18 commits »

ViperEkura pushed to main at ViperEkura/ToolAgent

2026-08-05 21:19:13 +08:00

79fde03972 feat: add yolox algo

ViperEkura pushed to main at ViperEkura/DataPipeline

2026-08-05 12:49:34 +08:00

d67f686f10 merge remote main

65dadac10f feat: processors 支持批量 tokenize，优化性能并缓存 chat template

Compare 2 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-02 16:26:27 +08:00

8447f88f61 fix: size KV pool from prompt/gen args in benchmark

b1b65a657e perf: target 512 grid blocks for decode split-K

3439e3104e perf: launch CUDA kernels on torch's current stream

288ba20db1 docs: audit non-CUDA documentation

020e2eff4e refactor: emit strategy metrics as floats

Compare 9 commits »

ViperEkura pushed to main at ViperEkura/AstrAI

2026-08-01 20:22:31 +08:00

925cbedc93 feat: scalar paged prefill fallback and decode causal fix

fda82ee232 perf: drop redundant smem zero-init in paged decode kernel

4b25664c79 perf: precompute kv_indptr once per decode step

a27c8a819d test: prune low-value and duplicate tests

91acaf4b0b refactor: unify attention mask to single attn_mask tensor

Compare 8 commits »