HuggingArch: a "compiler" that computes FLOPs, memory and KV cache for any HF model without a GPU
TheZachMueller · x · 2026-11-09
Developer thuHanShen spent half a year of spare time building HuggingArch, a "compiler" that doesn't run tensors but analytically computes FLOPs, memory traffic and KV cache size for any HuggingFace model — no GPU, no weights, no runtime.
The workflow: an agent writes the model spec (YAML), validated against safetensors headers and modeling source, then sympy evaluates it, yielding theoretical performance across GPUs and parallel configs (TP/PP/EP/DP, precision, batch×context) in one click.
The project includes 90+ hardware primitives and 150+ component specs, with views like Gallery, Inspector, Inference Planner and Kernel Bench; kernel data is calibrated against measured FA3/FlashInfer/DeepGEMM. Open source, with a web app and blog post.
More from coding & agent
- tldraw and Every converge on a product thesis: build agents for high-trust networks and get security for free — danshipper · 2026-11-09
- Agent edits Qwen 0.8B weights to inject personal facts—no training, no gradients — burny_tech · 2026-11-09
- itr8 puts Claude, GPT, Gemini and Grok in one group chat with a rich agent canvas — cantrell · 2026-11-09
- Xcode ships 11 built-in agent skills you can export to any coding agent with xcrun — pcuenq · 2026-11-09
- Reddit thread: which enterprise AI use cases actually delivered real value, not just demos? — averageChilean · 2026-11-09
- Trama: a minimalist open-source orchestrator for multi-AI automations on ComfyUI — Cold-Voice2737 · 2026-11-09