Alternatives
Products that do what Fast NF4 dequantization Triton kernel (1.41x faster than bitsandbytes) does
"I wanted to see if I could optimize the dequantization bottleneck during 4-bit LLM inference. By writing a custom kernel in Triton to optimize memory access patterns, I managed to get up to a 1.41x speedup over the standard bitsandbytes implementation. Check out the source code and benchmarks, feedback is highly appreciated!"
- 1AT
A 3.16M-parameter INT4 transformer running entirely in the on-chip memory of a Xilinx Kria KV260. Zero DRAM in the token loop, 59,965 tok/s on the fabric, bit-exact. Chat with it live.
27d ago · mikeayles.com
- 2
- 3
High performance storage engine for efficient LLM inference and GPU Training.
14h ago · theopenlake.com
- 4LS
Hi HN, I built llm.sql, an LLM inference framework that reimagines the LLM execution pipeline as a series of structured SQL queries atop SQLite. The motivation: Edge LLMs are getting better, but hardware remains a bottleneck, especially RAM (size and bandwidth). When available memory is less than the model size and KV cache, the OS incurs page faults and swaps pages using LRU-like strategies, resulting in throughput degradation that's hard to notice and even harder to debug. In fact, the memory access pattern during LLM inference is deterministic - we know exactly which weights are needed…
Apr 2026
- 5CL
Hey HN, we’re the developers of OpenLake, an open source storage engine for offloading LLM KV caches from GPU memory into a shared tier of RAM and NVMe. We built OpenLake because KV caches are outgrowing GPU memory. A single 256K token conversation on Gemma 4 31B produces approximately 43GB of KV state, more than half the memory of an 80GB H100. The problem becomes even harder across a cluster: a prefix cached on one GPU host is unavailable when the next request lands on a different GPU, forcing the new GPU to repeat work the fleet has already completed. Once the KV cache is offloaded,…
Jul 2026 · github.com
- 6
I made this after seeing someone posit the idea online yesterday over lunch then spent some time refining it. So far it's pretty impressive IMO! Right now I am running Qwen3-30B-A3B on my 24gb unified memory m4 MacBook Pro at 50 tok/sec and this should definitely not be working for such a large model on my middling hardware. Things are detailed in the README to get up and running and DESIGN.md has details on all the choices and such made along the way.
23d ago · github.com
- 7LI
LLM Inference Calculator — Estimate throughput, latency, TTFT, TPOT, and GPU memory usage for large language model inference. LLM 推理计算器 — 估算大模型推理吞吐量 (throughput)、延迟 (latency)、TTFT、TPOT 与 GPU 显存占用。
9d ago · llm-inference-calculator-delta.vercel.app
- 8
I built a specialized package of DeepSeek V4 Flash 0731 (originally 284B total parameters, 13B active), preserving reasoning, tool calling and coding capabilities: https://huggingface.co/steadfastgaze/DeepSeek-V4-Flash-0731-... I let it write a minimal C compiler targeting ARM64, then test the result with Fibonacci and FizzBuzz programs, and it succeeded in less than 1 hour, with the full recording at: https://youtu.be/XiwSilmV8B0 You can run it on Silicon Macs with my engine https://github.com/steadfastgaze/MoEspresso, while one of the…
21d ago · huggingface.co
- 9ML
Aug 2026 · github.com
- 105L
We've built InferX, a specialized runtime environment that fundamentally changes how LLMs are served. The core problem we solve is the latency bottleneck in AI inference, especially with large models. Current systems waste resources or suffer from painfully slow cold starts. InferX's AI-native architecture, with its "snapshot" technology, enables: * *Sub-2s cold starts:* Spin up models instantly. * *High density:* Serve more LLMs on the same GPUs. * *Optimal efficiency:* Maximize GPU utilization. This isn't just another API; it's a new execution layer designed from the ground up for the…
2025 · github.com
- 11LR
I just noticed it takes literally ~5 minutes to train millions parameters on slow CPU...but before you call Yudkowsky that "it's over", an important note: the main bottleneck is the corpus size, params are just 'cleverness' but given limited info it's powerless. Anyway, here is the project: https://github.com/bggb7781-collab/lrnnsmdds/tree/main couple of notes: 1. single C file, no dependencies. Below are literally all the "dependencies", not even custom header (copy paste from the top of the single c file): #define _POSIX_C_SOURCE 200809L #include #include…
Apr 2026 · raw.githubusercontent.com
- 12DA
Hi HN, I'd like to share DualMix128, a pseudo-random number generator I developed. It's written in C and achieves very high speed while maintaining strong statistical properties for non-cryptographic use. GitHub (MIT License): https://github.com/the-othernet/DualMix128 Key points: * *Fast:* Benchmarked at ~0.36 ns per 64-bit generation on GCC 11.4 (-O3 -march=native). This was over 2x faster than `xoroshiro128++` (~0.74 ns) and competitive with `wyrand` (~0.36 ns) in the same tests. (Benchmark code/results in repo). * *Robust:* Passes the full TestU01 BigCrush suite…
2025 · github.com
- 13SG
Hi HN, I built NOMA (Neural-Oriented Machine Architecture), a systems language where reverse-mode autodiff is a compiler pass (lowered to LLVM IR). My goal is to treat model parameters as explicit, growable memory buffers. Since NOMA compiles to standalone native binaries (no Python runtime), it allows using realloc on weights mid-training. This makes "self-growing" architectures a system primitive rather than a complex framework hack. I just pushed a reproducible benchmark (Self-Growing XOR) to validate the methodology: it compares NOMA against PyTorch and C++, specifically testing how…
Dec 2025 · github.com
- 14LF
I submitted an earlier version of this a few months ago (as llama2.f90). At that time it had a lot of steps to run and was just a toy, now it's easy to run and is a competitive option for llm inference. See the motivation section for discussion and the `Performance` issue for an ongoing discussion about performance.
2023 · github.com
- 15FF
I am playing around with using arrays of arbitrary dimension as framework for designing FFT implementations, as opposed to the more classical approach of tensor products and butterflies (too complicated in my opinion). It turns out, that with a modern compiler, you do not need much complexity to make a really fast implementation. This implementation is for powers of 2, and optimized for arrays that do not fit in cache. I do think it would be better to use a higher-level language to implement other cases (e.g. n = 2^a * 3^b * 5^c, multiple small FFTs, higher-dimensional), so I am currently…
Oct 2025 · gitlab.sac-home.org
- 16KP
I thought it'd be interesting to use Linux PSI (Pressure Stall Information) for an LLM runtime to trim the KV cache. This is mainly useful imo for edge devices like the Jetson Orin super nano kit which have unified memory. I haven't benched much, but plan to do so more over time and see if I can make a real use of it as I run local LLMs. Let me know if it makes sense :P (I of course vibed this idea)
Jun 2026 · github.com
- 17AI
2017 · trie.now.sh
- 18AL
Hello all, a couple of moons ago I ported karpathy's llama2.c code to run inference on the TinyStories 260K & 15M checkpoints on the on the PS Vita with the ability to download/delete the models on device. Runs showed that the 260K model ran at ~120 tok/s and at 15M ran at 1.8 tok/s, which probably could be a bit higher if it weren't a single threaded application. Had fun working on it as a weekend project, check it out for yourselves: https://github.com/callbacked/psvita-llm
2025 · github.com
- 19HC
Hello everyone, After recreating the accuracy/rough speed from David Page's implementation in hlb-CIFAR10 0.1.0 (18.1s on an A100, SXM4, Colab), it was down to some basic NVIDIA kernel profiling to figure out which operations were the long poles in the tent. Perhaps (somewhat?) unsurprisingly, the NCHW NHWC thrash was the worst part, but unfortunately the GhostBatchNorm was a barrier even using the faster-on-Ampere channels_last memory format. A quick note before continuing -- some may find the use of a convolutional network and on CIFAR10 to be curious. A quick answer to that would be…
2023 · github.com
- 20RR
runNburn is an Apache-2.0 Rust inference engine for quantized GGUF models that are too big for your fast memory. The core idea: weights stay file-backed (mmap), host residency stays under an explicit byte budget (--ram-budget), and GPU caches are sized from detected free/total VRAM — never from device-name presets. There is no conversion step, no sidecar cache files, no silent requantization. The GGUF on disk is the single source of truth. The result that made me want to post this: Tencent's Hy3 (295B total / 21B active sparse MoE, a single 97.8 GiB Q2_K GGUF) runs on my desktop…
Jul 2026 · github.com
- 21MA
I've been working on training this small vision language model for the last month - excited to release the first prototype today! It is based on SigLIP (image encoder), Phi-1.5 (text model) and trained using the LLaVa-1.5 training dataset. It runs reasonably fast on CPU with ~8GB of RAM in full 32-bit precision. There's plenty of room to speed it up and reduce memory consumption by quantizing the model. I posted a video of it running on my M2 Macbook Air (on CPU not MPS, so performance should be comparable on other hardware) on Twitter to demonstrate inference speed:…
2023 · github.com
- 22ML
Time to first token is 39% faster Agent wall times decrease by 46% No swaps Tracks your resource usage in real-time and adjusts how the model runs so that it works perfectly on your device. Implements KV cache sizing, prefix caching, live RAM pressure management, context trimming, KV quantization, and more. Built a ton of features
Jun 2026 · autotunellm.com
- 23HA
Most ZK proving systems are optimized for server-grade hardware with massive RAM. When scaling to industrial-sized traces (2^20+ rows), they often hit a "Memory Wall" where allocation and data movement become a larger bottleneck than the actual computation. I have been developing Hekate, a ZK engine written in Rust that utilizes a Zero-Copy streaming model and a hybrid tiled evaluator. To test its limits, I ran a head-to-head benchmark against Binius64 on an Apple M3 Max laptop using Keccak-256. The results highlight a significant architectural divergence: At 2^15 rows: Binius64 is faster…
Jan 2026
- 24L3
ran this over the weekend. stack was Llama 3.2 3B running locally + Keiro Research API for retrieval. 85.0% on 4,326 questions. where that lands: ROMA (357B): 93.9% OpenDeepSearch (671B): 88.3% Sonar Pro: 85.8% Llama 3.2 3B + Keiro: 85.0% the systems ahead of us are running models 100-200x larger. that's why they're ahead. not better retrieval, not better prompting — just way more parameters. the interesting part is how small the gap is despite that. 3 points behind a 671B model. 0.8 behind Sonar Pro. at some point you have to ask what you're actually buying with all that compute for this…
Mar 2026 · keirolabs.cloud
Ranked by how close each launch is in meaning, then by votes. Refine with a description →