llm-kb
← Back to products
Productlaunchedfeatured

Rust+CUDA LLM inference engine for RTX 5090, optimized with NVFP4, MoE, and speculative decoding

A from-scratch LLM inference engine written in Rust and CUDA, specifically tuned for the RTX 5090 Laptop GPU (Blackwell sm_120a). It supports NVFP4 quantization, mixture-of-experts (MoE), and multi-token prediction (MTP) speculative decoding, achieving up to 1.6x speedup over llama.cpp on target models. Designed for developers who need maximum inference performance on a single high-end laptop GPU.

Visit project ↗
260 engagement·1 source·Sun, Jul 5, 2026, 08:45 PM
The engine is built without frameworks or ggml, with every kernel hand-tuned against measured hardware limits of the RTX 5090 Laptop (24 GB, 175 W). It beats or matches llama.cpp on the same rig: 9B model generation is 1.3-1.6x faster at all prompt sizes, and 27B model with MTP speculative decoding is 1.4x faster. The repository is open-source on GitHub.

Entities

llama.cpp(tool)Rust(tool)CUDA(tool)NVFP4(concept)MoE(concept)MTP speculative decoding(concept)RTX 5090(concept)bw24(tool)RTX 5090 Laptop(concept)

Related

CommunitySun, Jul 12, 2026, 05:03 AM

Developer compresses GLM-5.2 MoE to run on single RTX 3090 via 79 experiments

A developer conducted 79 experiments to compress GLM-5.2, a 337 GB MoE model with 75 sparse layers and 256 routed experts, to fit on a 24 GB RTX 3090. The approach uses per-expert codecs, a batch pipeline over all MoE layers, and a patched llama.cpp runtime that loads codec-native expert binaries at inference time. The MIT-licensed repository documents the method and findings on expert similarity.

1 engagement·1 source·reddit
CommunitySun, Jul 12, 2026, 05:43 AM

User runs 100B+ MoE LLMs on low-end laptop using NVMe swap and Q3 quantization

A Reddit user with a low-spec laptop (i7-8750H, 20GB RAM, GTX 1050 4GB) reports successfully running 100B+ parameter MoE models by offloading parameters to a Samsung NVMe SSD via mmap, using Q3 quantization and quantized KV cache (Q4_0). They note that dense models are unusable on their hardware, but MoE models work with experts offloaded to CPU.

22 engagement·1 source·reddit
CommunitySat, Jul 11, 2026, 01:03 PM

User benchmarks AMD EPYC 9374F for LLM inference, finds 48-thread sweet spot

A user replaced their EPYC 9135 with a cheap 9374F (8 CCDs) for LLM inference. Initial benchmarks showed no decoding advantage until they used 48 threads; 64 or 32 threads performed worse than the 9135 in some scenarios. The 9374F is worse for gaming.

6 engagement·1 source·reddit
Tool ReleaseMon, Jul 6, 2026, 07:28 PM

Reame: CPU-first LLM inference server built on llama.cpp released

Reame is a new LLM inference server designed to run efficiently on cheap CPU hardware, including shared vCPUs, free tiers, and 2-core ARM boxes. Built on llama.cpp, it features disk KV cache, self-regulating speculation, generation archive, and interleaved multi-user support. The project emphasizes treating CPU hardware as a first-class citizen rather than a fallback.

60 engagement·1 source·github
BenchmarkSun, Jul 12, 2026, 10:47 AM

User benchmarks 4x RTX 5060 Ti with SGLang for Qwen3.6 27B, finds better concurrency than vLLM

A user shared benchmark results showing that SGLang handles higher concurrency better than vLLM when running Qwen3.6 27B (INT8 with bf16 KV cache) on a 4x RTX 5060 Ti (64GB VRAM) setup. The test achieved 200 successful requests at 8 concurrency over 348.87 seconds, processing 61,870 input tokens and generating 44,525 tokens. This provides a practical reference for others considering multi-GPU configurations with these consumer cards.

3 engagement·1 source·reddit