Havuz

Tiny-MoE is a lightweight Mixture-of-Experts language model built entirely from scratch in native PyTorch and trained end-to-end on Kaggle u

Object-storage-native KV cache for LLM inference & RL. Cross-restart, cross-conversation, cross-engine via shared S3 bucket.

Single-GPU LLM fine-tuning preflight: memory estimates, runnable LoRA/QLoRA recipes, and local measurements.

Split FLUX.2 and LTX 2.3 across two GPUs (LAN or same-machine) — NVENC compresses activations live on the wire. Icarus (ComfyUI node) + Daed

One GPU. Full LLM workflow. Real benchmarks. No cloud required.

A modern, hands-on, self-paced course that takes you from your first line of Python to shipping a real AI-powered automation.

A compact, readable PyTorch reference implementation of the DeepSeek-V4 architecture (Flash and Pro), in the spirit of nanoGPT. ~4,300 LOC.

Building an Image-to-Video (I2V) Model from Scratch

Profine automatically profiles and optimizes PyTorch training jobs on real GPUs, delivering measurable speedups and lower GPU costs before t

Lynn 原生 LLM 推理引擎 · W4A8/NVFP4 量化 · 自写 CUDA/Triton kernel · MoE · 投机解码 | Lynn-native LLM inference engine for NVIDIA Blackwell

五引擎语音合成：VoxCPM2 / IndexTTS 2.5 / 2.0 / OpenVoice / Step-Audio-EditX，声音克隆、声音设计、剧本配音、流式生成，中英日韩界面 | Multi-engine TTS (5 engines): voice cloning

Official code for Echo-Memory: a controlled study of memory in action-conditioned video world models (Context, Compression, Spatial, State-S

Inference engine for large language models on consumer GPUs, with deep optimization for older hardware

From PyTorch model to end-to-end TensorRT inference experience in two commands—AI-native, cross-platform, and built for the best possible us

OpenMythos, Claude Mythos’un gerçek mimarisini temsil ettiği kanıtlanmayan, araştırma literatürüne dayalı varsayımsal bir Recurrent-Depth Transformer uygulamasıdır.

Neden yükseliyor · Claude Mythos’a doğrudan atıf yapan adı ve açıklaması, varsayımsal mimariyi incelemek isteyenlerin ilgisini çekmiş olabilir; 162 günde 14.904 yıldız bu ilgiyi gösteriyor. Bu, benimseme veya teknik doğrulama kanıtı değil: son 28 günde commit ve farklı yazar sayısı sıfır.

🌹 Rose: Range-Of-Slice Equilibration PyTorch optimizer. Stateless optimization through range-normalized gradient updates.

Recomputation-free, context-independent KV caching for large language models

Taiwanese Hokkien (Taigi) speech-to-text transcriber - MediaTek Breeze-ASR-26 with faster-whisper, tuned for RTX 3050 4GB low-VRAM GPUs. Gra

🚀 Engram-PEFT: An unofficial implementation of DeepSeek Engram. Inject high-capacity conditional memory into LLMs via sparse retrieval PEFT

Unified KV-cache compression for LLM inference: 12 Python-native methods, guarded add-on composition and routing, analytical capacity simula

A tiny language model that talks like a house cat named Miso.

A Personalized Federated Learning (PFL-HCare) framework for IoT healthcare. Features MAML meta-learning, Differential Privacy (RDP), and gra

A curated collection of AI learning materials

Consumer-aware compression for embedding indexes and LLM KV caches — compress by the metric the downstream consumer actually uses. PCA-Matry

PyTorch-native framework for fault-tolerant distributed training of foundation models on AI clusters

Self-tuning multi-agent AI trading system. 8-source signal fusion, Bull/Bear/Judge debate on Claude Opus 4.7, Kelly + ATR position sizing. P

Biologically-grounded adversarial training platform: cyclic Wake/Dream/Nightmare/Compress phases that accumulate model robustness without ca

Reproducible OpenEnv-compatible reinforcement-learning environments for social-media integrity moderation.

First open-source implementation of Google TurboQuant (ICLR 2026) -- near-optimal KV cache compression for LLM inference. 5x compression wit

First open-source TurboQuant KV cache compression for LLM inference. Drop-in for HuggingFace. pip install turboquant.

Open-source Environment toolkit of claw-like agents, support task/harness generation and evaluation

a LLM inference engine to run on consumer hardware

Automated alignment adjustment for LLMs — direct steering, LoRA, and MoE expert-granular abliteration, optimized via multi-objective Optuna

End-to-end pipeline for seeing how LLMs actually process your prompts. Capture attention across every layer, render heatmaps and cooking cur