v0.4.1

What's Changed

• README: document the 0.4.0 API by @DavidBao03 in #39
• lint: pyink, pylint and zizmor through pre-commit, as tokamax does by @DavidBao03 in #38
• v7x: fp8 decode and bf16 prefill entries for head_dim 256 (Qwen3.5-35B) by @lk-chen in #42
• benchmarks: count module executions by the trace's XLA Modules row by @Deepolar in #40
• ci: test every pull request on CPU by @eh3323 in #35
• linear attention: read a slot-major GDN conv pool in place by @lk-chen in #43
• benchmarks: decode steps of served models, flywheel against batched RPA by @lk-chen in #48
• decode: several rows' blocks a step, run as prefill's fragment pipeline by @DavidBao03 in #46

New Contributors

• @eh3323 made their first contribution in #35

Full Changelog: v0.4.0...v0.4.1

v0.4.0

What's Changed

• paged KV cache: float8_e4m3fn and float8_e5m2 pools by @DavidBao03 in #22
• linear attention: ReplaySSM without the shadow, checkpoint as hi and lo planes by @DavidBao03 in #21
• varlen: keep max_seqlen-capped blocks on totals they do not divide by @DavidBao03 in #26
• decode: pad the paged block table to whole KV blocks by @DavidBao03 in #31
• decode: zero the merged staging slots through their word view by @lk-chen in #28
• decode: fp8 head-pair fragments and v7x fp8 entries by @lk-chen in #30
• decode: fold short speculative verify windows into the query heads by @DavidBao03 in #23
• v7x: decode entries for verify windows, require jax 0.11.2 by @lk-chen in #33
• attention sinks: learnable_sink on every softmax entry, forward-only by @Deepolar in #24
• varlen fwd: static schedule tables, kv pipelines across blocks and rows, fixed paged tiles by @DavidBao03 in #34
• linear attention: pad calls of fewer than 8 tokens to the DMA row grid by @lk-chen in #37

New Contributors

• @Deepolar made their first contribution in #24

Full Changelog: v0.3.0...v0.4.0

v0.3.0

What's Changed

• README_v7x: RULER 262K and 524K, Video-MME 2048 frames by @lk-chen in #11
• Sliding windows and softcap on the paged KV cache; static band copies for dense windows by @DavidBao03 in #12
• linear attention: ReplaySSM decode for GDN and KDA by @DavidBao03 in #13
• v7x: d128 decode and prefill entries for Qwen3-4B, warn on tuned-config fallback by @lk-chen in #14
• decode: run only the kv compute tiles that hold a row's keys by @DavidBao03 in #15
• v7x: measure the decode entries with the live-tile loop by @lk-chen in #16
• tests: fla's test grids for GDN and KDA by @DavidBao03 in #18
• tests: flash-attn reference grids for the three entries by @DavidBao03 in #19

Full Changelog: v0.2.0...v0.3.0

v0.2.0

What's Changed

• v7x results by @lk-chen in #2
• v7x: size the VMEM limit per chip for the KV-cache kernels by @lk-chen in #4
• chunked_prefill and decode benchmark on v7 by @lk-chen in #5
• chunked prefill improvement by @lk-chen in #7
• interleaved KV cache; chunked prefill on the paged varlen kernel by @DavidBao03 in #9
• v7x: ragged prefill fix for serving, tuned decode entries, end-to-end results by @lk-chen in #10

New Contributors

• @lk-chen made their first contribution in #2
• @DavidBao03 made their first contribution in #9

Full Changelog: v0.1.0...v0.2.0

v0.1.0

fix benchmark settings and update readme