DeepSeek-V3, R1 671B on 8xH100 Throughput Benchmarks

From-scratch, heavily-annotated CUDA inference runtime for Qwen2.5-Coder-7B on H100 (sm_90). Custom INT4 packer, fused GEMV, paged KV, split-KV attention, CUDA graph decode — every hot path commented for the why. Educational, not a llama.cpp replacement.

Z-AI V1.4 发行版本

NVIDIA HPC Benchmarks

Dashboard for AI Studio, Open Source Continuous Inference | Deepseek-R1, Qwen2.5, Llama3.1 | 4xRTX-5090 inside PRU2500, 2xH100 inside PRU2500, 8xMI210 in SuperMicro

LLM benchmarking, GPU workload orchestration backend server | Deepseek-R1, Qwen2.5, Llama3.1 | 4xRTX-5090 inside PRU2500, 2xH100 inside PRU2500, 8xMI210 in SuperMicro

rvLLM — Gemma 4 inference in Rust

Faster attention kernels for serving TML's Inkling model on vLLM. 2.7x over the shipping path on H100, and the only implementation that runs on A100.

Bunch of explanations and tutorials around confidential computing

A Flexible and High-Performance Inference Serving Engine for Diffusion Language Models

CLI tool to check Oracle Cloud compute shape availability - find capacity across regions

Demo for installing ComfyUI on Azure VM powered by Nvidia H100 to run Text to Image and to Video models like Z-Image, Qwen and Wan

In the recent competition, we were challenged to finetune a model that can convert a LaTeX expressions into Python code effectively. My team, which I led, secured 6th place overall.

High-performance Triton kernels for NVIDIA H100. Implements fused FP8 LayerNorm, tiled FlashAttention, and SRAM-optimized memory primitives for Hopper architecture.

A real-time speech translation web interface that combines Automatic Speech Recognition (ASR) and Neural Machine Translation (NMT) services to provide instant translations in multiple languages. Powered by DigitalOcean GPUs.

Cog Single GPU Quantized Implementation of Step-Video-T2V

An AI agent that teaches itself to fix bugs — MCTS explores debugging strategies, DPO trains on what works. Pipelined across two H100 nodes: one for 4-bit inference and trajectory collection, one for full bf16 LoRA fine-tuning. Built on Qwen2.5-Coder-7B, evaluated on DebugBench. Inspired by Agent Q (Putta et al., 2024).

Benchmarking QLC vs TLC NVMe for LLM inference: TPS/TCO comparison of Solidigm D5-P5336 & D7-P5520 on 8×H100 with MiniMax-M2.5

Reference Flash Attention implementation in PyTorch with V1/V2, GQA/MQA, Triton kernels, benchmark and docs.

PyTorch DDP and multi-GPU LLM training — from a minimal distributed example to NanoGPT speedruns, Muon, H100 profiling, and modded-nanogpt. FBA LAB https://bubblnet.com