ricerzz

[ricerzz @ricerrz]

@gapDEEPry 👀👀👀

[ryan mathieu @gapDEEPry]

Glad I can finally share another project we have been cooking! We built Weaver to make speculative decoding practical( and SOTA😉) for Qwen3.5-27B… we’re ready to share the paper, draft model, and SGLang reference implementation.

It will be in our own inference engine soon but in the meantime test and play with it on your own!

Paper: https://t.co/SHUNeiyKA8
Model: https://t.co/yqWjG4y4Mx
Code: https://t.co/aHESLdaWIt

[ricerzz @ricerrz]

@1a1n1d1y how much in credits have u spent so far on rvllm?

[ricerzz @ricerrz]

@1a1n1d1y Holyyyy

[ricerzz @ricerrz]

@1a1n1d1y Sent you dm, let’s get it 🚀

[ricerzz @ricerrz]

wonder what I can use these for 🔥

[Alex Zhurkevich @cudagdb]

Trtllmgen kernels are now open. Fastest prefill and decode kernels for our target workloads. We wrote these to win InferenceX, MLPerf, other benchmarks. Powering some of today’s top served models. Dive in, learn, use them, or level up your own. Enjoy.

https://t.co/2aQBwcdnZL

[ricerzz @ricerrz]

@gpusteve Pure gold

[ricerzz @ricerrz]

@1a1n1d1y PR up for Llama 7B setup: better HF auth, clearer tokenizer errors, and correct context from HF config. Tests pass, CUDA build works on Runpod.
PR: https://t.co/J4IQrCcUOo

[ricerzz @ricerrz]

> Another pass on mini-kernel-lib: tightened CUDA/runtime fallback handling and pushed the current narrow CUDA slices further for FP32 GEMM, reduction, and conv.
> Still need a CUDA-capable machine for full end-to-end validation

[ricerzz @ricerrz]

started a small CUDA-first kernel library repo
idea is basically: steal the parts of cuBLAS/cuDNN’s shape that make sense, keep it small, and see how far I can get without spending much on
compute
repo has the first scaffold up. GEMM first.

https://t.co/ebkqukWkMg

[ricerzz @ricerrz]

> mini-kernel-lib now has 2 FP32 GEMM ref paths, fused ReLU, 1-axis reductions, and direct NCHW conv2d forward, plus workspace checks, tests, benches, and opt-in GEMM
> autotune. still reference-only; next step is real GPU kernels and tighter dispatch.

[ricerzz @ricerrz]

@gapDEEPry lets get it