fusedtok

Fused CUDA kernels for LLM inference - RMSNorm / RoPE / SwiGLU / attention decode and friends, with zero-copy torch tensor support: up to 8.8x faster than PyTorch SDPA (attention decode, RTX 3060, see Benchmarks).

中文文档请看 README_zh.md | English below.

Why

LLM inference frameworks launch many small, memory-bound operators per token. Each launch round-trips through global memory. fusedtok fuses them into single kernels to cut memory traffic and launch overhead.

Operators

57 operators + helpers (fusedtok.__all__; 34 were frozen at 1.0, the rest arrived in minor releases - see "API stability" below). axpy is the hello-world demo op kept from the v0.x skeleton - functional, but not a performance feature.

Install

pip install fusedtok

Prebuilt wheels on PyPI (built with CUDA 12.4): Linux x86_64 (manylinux, cp311-cp313) and Windows x86_64 (cp311-cp313). On other platforms or Python versions pip builds from source automatically:

git clone https://github.com/Hai-Wenxiang/fusedtok.git
cd fusedtok
pip install .

Requirements:

• an NVIDIA GPU of the RTX 30 series (Ampere) or newer - e.g. RTX 3060/3090, RTX 4080, RTX 5090, A100, H100
• CUDA Toolkit >= 12.0
• A C++17 compiler (MSVC on Windows, GCC/Clang on Linux); Python 3.11+

Usage

numpy in / numpy out, or torch in / torch out - including zero-copy CUDA: kernels read and write torch device buffers directly via data_ptr(), with no staging copies and no host synchronization.

import numpy as np
import torch
import fusedtok

x = np.random.randn(4, 1024).astype(np.float32)
w = np.random.rand(1024).astype(np.float32)

# CPU reference implementation (ground truth, runs anywhere)
y = fusedtok.rmsnorm(x, w, eps=1e-6)

# staged CUDA: copies to GPU, runs kernel, copies back
y = fusedtok.rmsnorm(x, w, cuda=True)

# zero-copy CUDA with torch tensors: kernels run in torch's own buffers,
# stream-ordered with other torch operations
xt, wt = torch.from_numpy(x).cuda(), torch.from_numpy(w).cuda()
yt = fusedtok.rmsnorm(xt, wt)          # -> CUDA torch tensor

# RoPE with kv-cache position offset, NeoX (LLaMA-HF) layout
q = torch.randn(1, 4096, device="cuda")          # new token only
q_rot, k_rot = fusedtok.rope(q, k=None, pos_offset=1023, neox=True)

# attention over a GQA kv-cache: one call per decode step, no score
# materialization, variable-length batches share one cache tensor
out = fusedtok.attention_decode(
    q_heads,                                    # [B, Hq, D] new token
    k_cache, v_cache,                           # [B, Hkv, T, D]
    lens=torch.tensor([1023, 512], dtype=torch.int32, device="cuda"))
# grow a contiguous cache: append each new token's rows, then decode
fusedtok.kv_append(k_cache, v_cache, k_new, v_new, lens)
out = fusedtok.attention_decode(q_heads, k_cache, v_cache, lens + 1)
# ...or over a paged (vLLM-style block-pool) cache: pools [Nb, Hkv, P, D]
# + a per-sequence block table; append each new token, then decode
fusedtok.kv_append_paged(k_pool, v_pool, block_table, k_new, v_new, lens)
out = fusedtok.attention_decode_paged(q_heads, k_pool, v_pool,
                                      block_table, lens + 1)
# fresh-sequence prefill (causal by default; convenience path)
ctx = fusedtok.attention_prefill(q_all, k_all, v_all, causal=True)

# sampling side: the whole decode step in one fused call
token = fusedtok.decode_step(logits, sampled_ids, penalty=1.1,
                             p=0.9, temperature=0.8, seed=step)
# or step by step:
logits = fusedtok.repetition_penalty(logits, sampled_ids, penalty=1.1)
token = fusedtok.sample_topp(logits, p=0.9, temperature=0.8, seed=step)
# top-k sampling variant (renormalizes within the k survivors)
token = fusedtok.sample_topk(logits, k=50, temperature=0.8, seed=step)
# concurrent decode: [rows, vocab] in, one token per row out (default
# seeds are 0..B-1, so identical rows still draw independently)
tokens = fusedtok.sample_topp_batched(batch_logits, p=0.9,
                                      temperature=0.8)
# ...with per-row repetition penalties: ragged histories, one flat
# array + offsets is the serving-fast form
tokens = fusedtok.decode_step_batched(batch_logits, flat_ids,
                                      penalty=1.1, p=0.9,
                                      ids_offsets=offsets)

A minimal per-token sampling loop:

import torch, fusedtok as ft

h = torch.zeros(1, 4096, device="cuda")            # decoder state
w = torch.load("rms_weight.pt").cuda()             # float32 weights
generated = []
for step in range(256):
    h = ft.rmsnorm(h, w, residual=h)               # fused add + norm
    q = ft.rope(q, k=None, pos_offset=step, neox=True)
    logits = model_output(h)                       # your model
    tok = ft.decode_step(logits, generated, penalty=1.1,
                         p=0.9, temperature=0.8, seed=step)
    generated.append(int(tok))

Every function accepts float32 numpy arrays or torch tensors (other dtypes are converted with a copy) and returns float32 outputs of the same family. CUDA torch tensors may be bfloat16 on every operator that moves tensor data (elementwise / norms / RoPE / attention), and the attention operators additionally take float16 - kernels compute in float32 and convert at the load/store boundary (norm weights are upcast to float32 automatically; sampling/selection ops stay float32). CUDA torch tensors select the zero-copy path automatically.

See examples/demo.py for a runnable tour of every operator, and the usage guide for the topic-structured manual - one page per theme: quickstart, the execution model (paths / dtypes / streams / CUDA graphs), attention, the sampling contract, the INT8 workflow, how to read the benchmarks, and an FAQ with a glossary. Also available in 中文.

Correctness

Every kernel ships with a CPU reference implementation and element-wise parity tests (pytest). Tests run on machines without a GPU (CUDA cases skip automatically).

API stability

1.0 froze the public surface at 34 names; new operators arrive in minor releases (57 as of v2.5), and every name in fusedtok.__all__ keeps its signature across the 1.x series. Type stubs (__init__.pyi, PEP 561 py.typed) ship with the package. Breaking changes require a new major version and a deprecation window. Determinism promises: selection ties resolve to the earliest index; sampling is deterministic per seed.

Benchmarks

RTX 3060 (sm_86), float32, zero-copy torch tensors, CUDA-event timing over 3 independent rounds (means below; per-round values in the JSON), vs the equivalent PyTorch reference (composite eager expressions; attention references use pre-expanded heads - repeat_interleave outside the timed region). Largest shape per op; full data: docs/benchmarks/benchmark_rtx3060.json, reproduce with python benchmarks/bench.py:

Batched samplers (v1.4) and batched decode steps (v1.5) - one call for the whole [8, 131072] batch, referenced against torch's native batched draw (softmax + multinomial over the 2-D tensor; the decode row adds the gather-penalty; per-round values in the JSON):

Row-wise kernels (norms, softmax) autotune their thread-block size per shape at first call (v0.4.1); the table reflects the tuned choices.

RTX 5060 Ti (Blackwell, sm_120) - same suite, largest shape per op (full data: docs/benchmarks/benchmark_rtx5060ti.json):

Batched samplers (v1.4) and batched decode steps (v1.5) on the same [8, 131072] shapes:

On smaller shapes the Blackwell card shows bigger wins (softmax 1.72x, RMSNorm 3.08x at 256 rows, attention decode 3.78x at T=4096 running ~187 GB/s) - the launch-overhead share shrinks as shapes grow; full sweep in the JSON.

The PyPI wheel ships sm_80/sm_86 cubins plus a compute_86 PTX fallback - verified to JIT and run correctly on Blackwell (sm_120) drivers.

Fusions win big (RoPE / RMSNorm / SwiGLU) because eager mode round-trips intermediate tensors through global memory. The v0.4 selection pipeline (arrival-ticket radix rounds + early-exit compaction, replayed from a cached CUDA graph) beats torch's CUB radix select at small k on both GPUs; the v1.0 retune (in-block-sort threshold and sort chunk both dropped 2048 -> 1024 - a single block bitonic-sorting 2048 keys was the whole mid-k regression) brings the mid-k window to parity-or-winning as well. The fused samplers win against the eager composites when the logits look like real decode output (sample_topp peaked: 2.68x / 2.51x; sample_topk: 2.76x / 2.06x; the composites themselves swing run-to-run on WDDM - per-round values in the JSON). On a flat distribution sample_topp sits at an honest 0.16x on a 5060 Ti and 0.30x on a 3060 (the fusedtok side is stable; the reference's rounds carry WDDM noise) - the nucleus then spans ~90% of the vocabulary and the pipeline must effectively order the whole thing. v1.2 cut that worst case ~8.5x (18.2ms -> 2.2ms at n=131072 on a 3060) with three token-preserving changes - an adaptive window jump driven by a p*total mass bound, a full-vocabulary fast path that skips the selection stages, and a batched-load serial sampling walk (the strictly sequential float adds are the CPU-parity determinism contract; only the loads got pipelined); v1.3 cut it another ~1.6x (2.2ms -> 1.4ms on a 3060, 1.6ms -> 1.0ms on a 5060 Ti) by having the first walk record prefix-sum checkpoints that the inverse-CDF walk binary-searches - still bit-identical tokens - but torch's fully parallel sort still owns that regime. sample_minp (v1.3) wins on peaked logits and loses the wide-nucleus row by design (0.33-0.56x: one widening retry plus a 32-64k sort; the torch boolean-mask composite never sorts) - v1.4 gave min-p the same adaptive jump top-p has had since v1.2 (a sufficient bound from the one-time global total: wide-nucleus rows skip the ladder's intermediate stops, 28-30% faster on that row, tokens bit-identical).

The v1.4 batched samplers put the whole [rows, vocab] batch through one call: every row runs the single-row pipeline verbatim (per-row parity, including the widening loop with per-row finish tracking), so the speedup over the per-row loop is pure batching - 4-6x in wall time at B=8 on submission-bound hosts (3060: topp 1340 -> 274 µs, minp 1399 -> 237 µs; 5060 Ti: 3.9x / 4.3x). On peaked logits the batched calls sit at torch's native batched-multinomial level (the batched min-p, top-k and top-a calls win on this benchmark: 0.64x/0.62x, 1.75x/1.19x and 1.03x/0.85x on the two GPUs), and the flat worst case keeps the singles' honest caveat one tier lower (0.05x). v1.5 extends batching to the whole decode step: decode_step_batched runs per-row repetition penalties through per-row vocab bitmaps inside the same pipeline - on wall-time probes at B=8 it is 5.2x over looping decode_step on peaked rows on a 3060 (1676 -> 321 µs; torch's native penalize+softmax+multinomial composite: 266 µs) and 4.5x on a 5060 Ti (646 -> 145 µs), and 3.1x on mid-tail rows (3060).

attention_decode wins big at decode (one launch streams the GQA cache once, while SDPA pays head expansion or small-query inefficiency); attention_decode_paged (v1.2) pays only ~1.09-1.15x for the block-table indirection of the fragmentation-free vLLM-style cache layout, with bit-identical output on matching slice schedules; attention_prefill's f32 path is the honest convenience kernel at ~0.45x of SDPA's flash backend (no tensor cores by design - TF32's 10-bit mantissa is a numerical downgrade the parity contract forbids); since v2.4 the bf16/fp16 path rides tensor cores and lands within 0.68-0.91x of SDPA's bf16 flash (see the table rows below). kv_append (v1.3) writes one token's k/v rows into the contiguous cache in a single launch (a multi-x win over the advanced-indexing scatter whose exact factor tracks that reference's WDDM swing; a tiny launch-bound op - parity-level cost per decode step). The INT8 decode GEMV moves half the bytes of an fp16 projection and runs at full memory bandwidth (2x); the pipelined IMMA GEMM (v1.0 rework: cp.async double-buffered slabs, runtime-tuned 64x64 / 128x128 tiles) reaches ~31-39 TOPS on a 3060 (the sustained-load host state of the 2.4.x rounds measured 31; see the benchmarks page note) and ~67 TOPS on a 5060 Ti - 2x-4x the v0.4 kernel - but cuBLASLt (torch._int_mm) still holds a ~2.1-2.7x lead on the per-tensor rows (its tiles pipeline deeper and its epilogue is tuned per-arch); the W8A8 rows' gap is only ~1.6-2.2x. For now qgemm is the exact / graph-capturable / zero-copy INT8 path, not the fastest one; honest numbers, a CUTLASS-class schedule stays future work. The per-channel variant (qgemm_perchannel, the W8A8 layout INT8 inference actually uses) fuses the per-output-channel scale multiply into the same epilogue at zero kernel cost - the composite torch reference pays for that broadcast separately, which is where its 0.55-0.58x comes from.

All sampling rows above have measured fixed logits since 1.1.1 (the bench seeds torch's RNG); since v1.2 the peaked row spikes +20 and the flat row uses near-uniform logits - both regimes are now deterministically what their labels say (the v1.1 peaked row sat on the coverage boundary at n=131072 and flipped regimes per seed). The argmax rows are event-timed over a host-synchronized call and swing on WDDM (0.72-1.26x across the 3060's shipped rounds); wall-clock probes with the sync excluded from the timed loop measure 1.12x (3060) / 0.96x (5060 Ti) - v1.2 removed one CUDA submission and one allocation per call.

Development

See CONTRIBUTING.md for the full guide (test rules, error contract, determinism invariants). Quick start:

# Windows: run inside a VS developer prompt (vcvars64)
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build
# from repo root: PYTHONPATH picks up the built module, conftest.py adds python/
$env:PYTHONPATH = "$PWD/build"        # Windows
PYTHONPATH=$PWD/build                 # Linux
python -m pytest tests -q
python benchmarks/bench.py            # GPU benchmark + chart

Both Windows and Linux are supported: Windows builds through MSVC via nvcc, and CI builds and runs the CPU test suite on every push.

Roadmap

• v0.2 (released): bf16 zero-copy, radix-select top-k/top-p, fused nucleus sampling, single-read softmax, CUDA-graph verified
• v0.3 (released): chunk-merge selection sort + parallel nucleus count, bf16x4/x8 vectorized elementwise, INT8 quantize/dequantize utilities
• v0.4 (released): arrival-ticket selection pipeline (no cooperative launch, early-exit compaction, cached CUDA graphs), stream-aware launchers everywhere (real CUDA-graph capture), INT8 compute path (IMMA qgemm + decode GEMV), fused decode_step sampling
• v0.4.1 (released): runtime block-size autotuning for the row-wise kernels (norms/softmax pick 128..1024 threads per shape at first call)
• v0.5 (released): attention - GQA decode attention over a contiguous kv-cache (flash-decoding split over long caches, per-sequence lengths) and a tiled prefill path (honest ~0.45x of SDPA flash - the convenience path); single-chart-per-GPU benchmarks; Windows wheels in the PyPI publish pipeline
• 1.0 (released): pipelined tensor-core INT8 GEMM (cp.async double-buffering, runtime tile tuning; 17 -> 39 TOPS on a 3060) with per-channel weight scales (W8A8), fused top-k sampling (2.1x vs the topk+multinomial composite), top-k mid-range-k parity, text hygiene gate, wheel matrix expansion (Linux cp310-313, Windows cp311-313), API freeze
• 1.1 (released): half-precision attention - attention_decode / attention_prefill accept bfloat16 and float16 caches (float32 compute, half the decode bytes); parallel exp precompute halves the flat-distribution sampling worst case with bit-identical tokens
• 1.2 (released): paged kv-cache attention - attention_decode_paged over a vLLM-style block pool [Nb, Hkv, P, D] + per-sequence block tables (~1.09-1.15x the contiguous op, any valid table honored) and kv_append_paged (the in-place cache-write side); flat-distribution sampling worst case cut ~8.5x (adaptive widening jump + full-vocab fast path + batched-load serial walk, tokens bit-identical); argmax launch diet (one submission and one allocation less per call)
• 1.2.1 (released): audit-driven hardening - selection workspace overflow past 131072-token vocabularies fixed (Qwen-scale vocabs), host-side lens/table/id validation with a documented device-tensor trust boundary (CUDA-graph capture works with lens now), empty- input and dtype/contiguity guards on the zero-copy paths, honest benchmark bandwidth (four rows were over-billed 1.5x), warning-clean builds (MSVC /W3 + GCC -Wall -Wextra), and the documentation restructured into topic pages with a natural-sounding Chinese rewrite
• 1.3 (released): sample_minp (min-p sampling - value-threshold nucleus relative to p_max, no global-mass reduction, adaptive by construction) and kv_append (the contiguous cache-write side); the sampling serial walk gains checkpoint bisection (walk 1 records prefix sums, walk 2 binary-searches them - flat worst case cut another ~1.6x, tokens bit-identical); zero-copy helpers reject CPU operands (a host pointer in a kernel poisons the CUDA context)
• 1.3.1 (released): audit-driven hardening - staged-path lens/block-table value validation closed (bad values used to become silent GPU out-of-bounds writes) plus integer-input hardening; two latent sampling-walk bugs fixed (stride>=2 checkpoint double-count, adaptive widening reading the wrong workspace word - mid-tail ~28% faster with bit-identical tokens); softmax tuner capped to keep sanitizer gates clean; kernel/launcher cleanup; benchmark tables fully regenerated
• 1.4 (released): batched sampling - sample_topp/minp/topk_batched sample a whole [rows, vocab] batch in one call (per-row parity with the singles, one seed per row, rows finish at their own window sizes); min-p gains the adaptive widening jump (a sufficient bound from the one-time global total - wide-nucleus rows skip the ladder's intermediate stops, tokens bit-identical)
• 1.4.1 (released): audit-driven hardening - staged-path batched shape validation closed (rows/n used to be trusted against the buffer), shared widen-bound helpers and a batched-sequencer cleanup, lazy per-row totals for both widen modes, stream-scoped syncs; benchmark tables regenerated with the shipping version stamp and the documentation overhauled (stale numbers, stiff phrasing, glossary and FAQ additions)
• 1.5 (released): a batched decode_step - per-row ragged histories through per-row penalty bitmaps, the whole penalize -> scale -> sample chain in one call. The other 1.5 candidate (per-row window sizes inside one batched attempt, so a wide row no longer lifts the uniform window) was implemented three ways, measured as a net loss or parity at B=8/B=32 on an RTX 3060 (the serial inverse-CDF walk dominates each attempt round and the merge ladder carries a per-launch floor that window-bucketing multiplies), and was dropped with the numbers in the changelog - revisit if the walk goes parallel
• 1.5.1 (released): audit-driven hardening - the int8 GEMV silently dropped the head of every row its alignment gate rejected (k % 4 != 0); the kv_append staged bindings and the rope / temperature raw launchers validate what their siblings already did; scalar-fallback kernels index in 64-bit; both docs languages resynced to the 1.5.0 JSONs with the stiff phrasing rewrites
• 1.5.2 (released): audit-driven hardening, round three - the batched radix rounds now apply the row's repetition penalty (the unpenalized selection prefix mis-composed decode_step_batched's window under penalty != 1); the kv_append_paged span is computed in 64-bit; duplicate penalty ids can no longer diverge between the CPU and GPU paths; rope grids, launcher guards and the batched CPU references got the same checks their siblings have
• 1.6 (released): entropy-adaptive sampling - sample_eta (eta cutoff, Hewitt 2022: keep every token with p_i >= eta * min(1, exp(-H)), a value-threshold prefix like min-p, reusing min-p's adaptive widening bound) and sample_typical (locally typical, Meister 2022: the smallest set whose mass reaches typical, ordered by |surprise - H| - a contiguous band of the value-sorted window, no analytic widening bound so the honest x8 ladder), each single-row and batched (38 -> 42 public names)
• 1.6.1 (released): logit_penalties - the HF penalty trio (CTRL repetition scale, presence shift, count-weighted frequency shift) in one call, once per distinct id, duplicates never stack, bit-exact across CPU and GPU - plus audit-driven hardening: the spurious "typical nucleus not covered" throw at typical = 1.0, the batched typical trust-boundary gap, and the eta widening bound's unit mismatch (42 -> 43 public names)
• 1.7 (released): logit_penalties_batched - the penalty trio for a whole [rows, vocab] batch with ragged per-row histories, each row bit-identical to the single-row op; the batched histograms ride a capture-safe cached workspace with a workspace-free borrowed-output fallback, and both apply kernels force separate roundings so the FFMA fusion cannot split CPU and GPU by an ulp (43 -> 44 public names)
• 1.8 (released): two more truncation rules and batched greedy - sample_topa (top-a: keep every token with p_i >= top_a * p_max², min-p's prefix machinery with the cutoff derived from the existing total - top_a / total in exp units - and min-p's sufficient widening bound; top_a = 1.0 keeps both leaders of a two-horse distribution), sample_nsigma (keep every token whose scaled logit stays at or above mean - nsigma * sigma - the row's first two moments from one new pass, accumulated in doubles so atomic arrival-order drift cannot survive the variance's cancellation), and argmax_batched (one launch for a whole [rows, vocab] greedy step, no readback, CUDA-graph capturable); the entropy-family mass passes fused into one kernel (eta/typical/nsigma run two full-vocabulary passes per attempt instead of three, tokens bit-identical) (44 -> 49 public names)
• 1.8.1 (released): audit-driven hardening - the argmax_batched bindings enforce the exact [rows, vocab] shape (a mis-shaped buffer whose size covered rows*n used to be silently re-split), the CUDA-graph warm-up caveat is documented at every level, three seed-fragile exact assertions relaxed to the documented neighbor-rank contract, and a bilingual documentation pass (numbers resynced to the shipped JSONs, the Chinese rewritten in natural prose, the 1.8 ops joined the quickstarts and the demo)
• 1.8.2 (released): documentation accuracy + gate round - the corrupted 1.8.1 table speedup cells repaired and a markdown hygiene gate added to CI (every benchmark row re-validated against its own GPU's shipped JSON, broken bold rejected, row-coverage floor), the remaining batched CPU references validate rows before seeds, and the bilingual number sweep caught every page that had missed the 1.8.1 re-measurement
• 2.0 (released): Python >= 3.11 required (3.10 EOL dropped from the build and publish matrices — the formal breaking change that steps the major version); INT8 GEMM larger-tile and tensor-core prefill evaluated and deferred (it eventually shipped in 2.4; see the changelog for the shared-memory ceiling arithmetic)
• 2.0.1 (released): documentation accuracy + tables re-measured on the shipping build; the 25-finding bilingual audit fixed stale prose numbers, the cp310-cp313 install text, the kv_append contradiction, and the honest-loss ranges that lagged one release behind
• 2.0.2 (released): documentation accuracy round — 18 stale prose numbers resynced to the shipped 2.0.1 tables (int8 TOPS, small-shape Blackwell, sampler prose, kv_append, attention headline, paged indirection, honest-loss ranges), the missing 2.0.1 roadmap rows added, and the demo.py formatting cleaned
• 2.1 (released): sample_tfs - tail-free sampling (Filazzola & Trottet 2023): keep the prefix whose sorted CDF second derivative (normalized) stays above 1-z — a data-dependent cutoff on the flat tail. Honest x8 widening (no analytic bound). Single-row + batched (49 -> 51 public names)
• 2.2 (released): sample_xtc - Exclude Top Choices creative sampling: with probability p, the top top_n tokens by probability are removed from the pool before the draw - breaks LLM "template" outputs, and at least one token is always kept. Single-row + batched (51 -> 53 public names)
• 2.2.1 (released): bug-fix + batched-kernel round - the missing staged binding for batched XTC and the broken staged binding for batched TFS, the XTC full-vocabulary window fix (2.2.0 GPU draws clipped the tail), penalty-bitmap range guards, and TFS/XTC batched moved onto the chunked pipeline (up to 11.1x at b=128); tables re-measured; bilingual documentation readability round
• 2.3 (released): sample_dry - DRY (Don't Repeat Yourself) sampling: the sequence-level repeat penalty - penalizes the token that would extend a repeated sequence (64-token scan window, per-token max exponent, multiplier ** exponent division), then one full-softmax draw. Single-row + batched over ragged histories (53 -> 55 public names); binding/CPU-reference dedup round closes the copy-paste defect class
• 2.4 (released): tensor-core prefill for bf16/fp16 (mma.sync m16n8k16, dims 32/64/128): 5.1x the CUDA-core half path on a 3060 at S=1024 D=128, within 0.42-0.59x of SDPA flash; f32 keeps its documented numerical path; tables gain the bf16 prefill row
• 2.4.1 (released): audit-driven fixes - the tensor-core prefill dim gate tightened to the instantiated set (48/80/96/112 previously launched the D=128 template out of bounds), the DRY batched history upload moved onto the caller's stream, the rewrite scratch chunked to the documented bound; documentation accuracy round (32 findings); staged single-row binding dedup
• 2.4.2 (released): maintainability round - every batched sampler delegates through the one shared dispatcher (xtc/dry included), all pairing cleanup, code polish from the audit backlog; documentation polish
• 2.4.3 (released): third audit round - the dry batched torch-CPU regression fixed and pinned, demo.py parse-gated in CI, prose resync (17 findings)
• 2.4.4 (released): the sample_xtc staged-binding delegation that [2.4.3] described but did not ship; the W8A8 lead range corrected to cover its own table (1.6-2.2x)
• 2.5 (released): sample_mirostat - Mirostat v2 entropy-targeting sampling with a caller-owned mu bound (the one tuple-returning sampler: (token, new_mu); 55 -> 57 public names); lane-parallel softmax in the tensor-core prefill (probe 1.63x, the bf16 row 0.58x -> 0.91x of SDPA); cp.async staging and 16-bit radix keys evaluated and honestly withdrawn
• future candidates (unscheduled): a CUTLASS-class INT8 GEMM schedule (the current qgemm is the exact/graph-capturable/zero-copy path, not the fastest one; TMA itself is Hopper-only and off the table on sm_80-sm_120), 16-bit radix keys (changes the determinism contract)

Community

• Contributing guide - setup, rules of the road, PR process
• Code of conduct
• Security policy
• Changelog

License

MIT - see LICENSE. Third-party notices: NOTICES.md.