HN · GitHub Signals

OPEN SOURCE, IN FOCUS

v100-skinny

dnv2003/v100-skinny

Hand-written NVFP4 W4A16 CUDA kernels and chain-MTP speculative serving — Qwen3.6-27B at up to 366 tok/s on four Tesla V100s, hardware with no FP4 support

Explore on GitHub ↗

Saved projects stay in this browser. No account needed.

TAKE A CLOSER LOOK

Start with the source.

The repository is the place to check installation steps, supported platforms, the license, and current maintenance before choosing a tool.

Read the documentation on GitHub ↗

Keep exploring

Find projects in the same areas, or search the archive for a different approach.

ANOTHER GOOD FIND

Explore related projects

View the archive ↗

Connected by topics and descriptions.

avifenesh

bw24

From-scratch Rust+CUDA LLM inference engine for RTX 5090 (sm_120a) — NVFP4, MoE, MTP speculative decoding, tuned against measured hardware limits

Take a look →
jakeatx

llamampere

llama.cpp fork for significantly improved performance on Ampere (especially RTX 3090 / 3090 Ti): TurboQuant KV cache, MTP speculative decoding with a 64K draft-vocabulary shortlist, custom SM86 + Qwen kernels. 90 tok/s over a 100K-token generation at temperature 1.

Take a look →
0xbakeer

qwen3.8-27b-4-bit-on-a-single-dgx-spark

Serving 4-bit Qwen3.8-27B on a single DGX Spark (GB10): 75 tok/s single-stream, 246 tok/s aggregate at 8-way concurrency. NVFP4 vs MixedInt4-AutoRound vs the FP8 baseline, measured on one harness — including why the quantization advantage collapses to +0.2% by c16.

Take a look →
0xbakeer

qwen3.8-27b-fp8-on-a-single-dgx-spark

Serving Qwen3.8-27B-FP8 on a single DGX Spark (GB10): 7.88 to 58.5 tok/s single-stream from decode strategy alone, weights untouched. Speculative decoding and prefix caching benchmarked, plus DFlash 2 — the only Qwen3.8-27B build that can serve it under vLLM.

Take a look →