Open-source contributions
Every PR, closed ones included. States checked Nov 3, 2026.
flash-linear-attention #1412: KV-cache decoding
Added KV-cache decoding for stick-breaking attention, tested against the uncached forward. On H200 at batch 1 and 64K keys, 32 programs on 132 SMs take 4.3 ms, against a K/V read floor of about 0.06 ms at rated bandwidth.
Merged. Oct 27, 2026flash-linear-attention #1319: stick-breaking attention in Triton
Continued #599 to add forward and backward Triton kernels. Rebuilt the backward to fix a silent zero gradient at saturated logits, caught in review. On H100, 64×32 tiles fit 57-89 KiB of shared memory where 64×64 needed 129-193 KiB. Nothing saturates in the backward at about 12% occupancy, which I read as stalls on fp32 scans. It still takes 1.04-1.42× the original authors' time, and I haven't isolated why.
Merged. Oct 8, 2026NVIDIA cuDNN frontend #1287: GQA tests
Tests for native grouped-query attention, which catch 3 injected bugs that NVIDIA's SM80 tests missed.
Tests merged. Oct 6, 2026NVIDIA TransformerEngine #3591: grouped NVFP4 dequantization
CUDA kernel and a fixed PyTorch binding, reading offsets on device so it runs in a CUDA graph. Output matches per-tensor results bitwise.
OpenvLLM #38338: Triton NVFP4 GEMM reference
Ordered grouped tiles and autotuned tile sizes, holding 1,014-1,076 TFLOP/s on RTX 5090 for one shape from batch 8192 up, where the first revision fell to 418. On Llama QKV with hot L2: 1.38-3.04× CUTLASS up to batch 128, 0.77-0.91× above, and about 0.27× across 4 Llama shapes on B200 from batch 1024. I closed it after the issue it answered, #21014, was closed as not planned.
Closed, not merged. Oct 7, 2026NVIDIA cuDNN frontend #1223: native GQA prototype
On A100 at 64/8 heads and a 128K KV cache, K/V head copies took 7.28 ms against a 6.0-6.7 ms kernel. Removing them made 6 batched rows 2.6-3.3× faster, and 10 more stopped running out of memory. Those rows are really 3 configs and 5 configs. The cuDNN backend stayed faster on 27 of 31 multi-token rows. Closed after NVIDIA shipped native GQA in #1239.
Closed, not merged. Sep 27, 2026GPU kernels
Stick-breaking attention kernels
Forward, backward and KV-cache decoding kernels in Triton, merged in flash-linear-attention.
Write-upGrouped-query attention in NVIDIA cuDNN
What removing K/V head copies did on an A100, why that code closed, and the tests that merged.
Write-upTriton vs cuBLAS
A measurement of whether one fused Triton kernel runs faster than the usual separate steps, across 76 matrix sizes on an A100.
Personal project, public repomoe-kernels: a timing model of vLLM's fused MoE kernel
Predicts each call's time from the routing and the tile, with predictions frozen before each rental, on rented GH200s. On one 32×64×64 test tile, model v3's time predictions held at 0.51-0.69% rms. Its byte predictions missed by 11.76% on OLMoE's first wave, and the decode blind test failed at 11.4% against a 1.5% bar. No model is blind-certified on vLLM's own tiles.
GPU kernels, public repoTeam projects
ML-guided kernel config
A PyTorch model predicts kernel runtime and picks a config by argmin over a grid.
Team of 4, ECE 759MI300X + gem5 modeling
Built the gem5 full-system setup and HIP benchmarks for the AMD MI300X memory hierarchy.
Team of 3, ECE 752LLM inference
Where the Seconds Go
A llama.cpp cache for batch-1 decode when the model doesn't fit on the GPU, and a computed limit to read it against.
Write-upvLLM on H100 NVL
Load-tested vLLM with Qwen2.5-7B across offered rates. TPOT ran 31-34% lower with CUDA graphs than eager, and FP8 cut it 27-36%. One run per rate.
Serving, first passSpeculative decoding on MoE
On one H200 with CPU ngram drafts, offline and synthetic, a per-workload oracle improved on the best static batch-size table by at most 1.8%. vLLM's MFU byte model overcounted expert reads, up to 3× in the worst cell.
Comment on someone else's RFCJetson Orin Nano profiling
Attention and KV-cache decode kernels profiled with Nsight on a 15 W edge board.
Public repoAgent systems
Lab verification and audit
Sign-off verification for a lab of Claude Code agent roles, an audit of it against its logs, and a rebuilt check.
Write-upbert
A local MCP server that gives a coding agent searchable memory of a project, measured against agentic grep, including where grep scored higher.
Write-upThe Obsidian Archive
A 13-stage autonomous pipeline that researches, writes, renders and uploads a documentary from one topic, with atomic checkpoints and resume, automatic retries, an optional budget cap on Claude spend and failure alerts.
Agent systemsHIVE
A privacy-first peer-to-peer reasoning engine in Rust.
Agent systemsWriting
When the K/V copy costs more than the attention
cuDNN's K/V head copies took 7.28 ms against a 6.0-6.7 ms kernel at 64/8 heads, what removing them did and didn't do, and the tests that merged.
Worklog. Nov 2026Below the Measurement Floor: a pre-registered dense-vs-MoE null
Pre-registered compute-matched comparison. Every cell lands below the measurement floor, and the null is the result. Not peer reviewed.
Report. May 2026Recursive Verification-Surface Collapse
The null-set principal problem in self-grading labs, and three structural principles, from about 90 autonomous cycles. Not peer reviewed.
Report. May 2026When the LLM judge is biased
Same-family preference leakage when four Claude instances review the same proposals, read against an ICLR 2026 result.
Note. May 2026Tier-per-task: routing models by the job
Why model tiering belongs at the task level, not the role level.
Note. May 2026