This directory contains the Neuron-optimized SmolVLM/Idefics3 inference implementation. It extends the NxD decoder stack with a separately compiled vision encoder and on-device image-token feature injection. For shared NxD decoder guidance, read optimum/neuron/models/inference/AGENTS.md. For SmolVLM, there are exactly two compiled artifacts: modelvision.pt and modeltext.pt. Both are managed by SmolVLMNxDModelForImageTextToText in modeling_smolvlm.py. The custom embeddings intentionally keep HF key names (patchembedding, positionembedding) so state dict loading maps directly. Image features are inserted at imagetokenid positions and passed into VLM wrappers: - NxDDecoderWrapperForImageTextToText…
This guide covers the Optimum Neuron vLLM backend. For project-wide workflows see AGENTS.md. Key files: - optimum/neuron/vllm/runner.py - optimum/neuron/vllm/model_loader.py Key files: - optimum/neuron/vllm/servermanager.py — spawns and manages vLLM server subprocesses with Neuron core pinning. - optimum/neuron/vllm/reverseproxy.py — round-robin proxy with aggregated /health. **Do NOT use startnewsession=True on vLLM server subprocesses.** vLLM already cleans up its own EngineCore workers on SIGTERM. Isolating servers in separate process groups makes them unreachable by the parent's signal handlers and test fixture teardown, causing Neuron…
Read this before working on any test code under tests/. Fixtures in tests/fixtures/llm/ are registered as pytest plugins in tests/conftest.py: This makes neuronllmconfig, anygeneratemodel, and speculation available to all tests. Full documentation is in the tests/fixtures/llm/export_models.py module docstring. Key points: Each model is compiled for two shapes: (batchsize=4, seqlen=1024) and (batchsize=1, seqlen=8192). Config names follow the pattern <model>-<batch>x<seq>, e.g. llama-4x1024. **neuronllmconfig** — for tests that need a specific model config: **anygeneratemodel** — for tests that should run across all generation…
All imports in a kernel (files inside torch-ext/<kernel_name>) of other modules within the same kernel must be relative. For example: The ops module (ops.py) is generated by the build system…
Always use bun (bun install, bun dev, bun run build, bun test). Never use npm or yarn. After making any code changes, always run these commands in order and fix any errors before finishing: Or run them all at once (format first, then the full validate suite): bun run validate runs: type-check → lint → format:check → test Three versions are supported. Version is detected from meta/info.json → codebase_version. src/app/[org]/[dataset]/[episode]/fetch-data.ts → getEpisodeData() dispatches to: Locating an episode (its video URLs…
This repository is a three-project workspace. When working on pre-training, read and follow pre-training/AGENTS.md and run project commands from pre-training/. When working on post-training, read and follow post-training/AGENTS.md and run…
This is the llama.cpp inference optimization project. It is separate from ../pre-training and ../post-training. Get a local GGUF model running through llama.cpp as fast as possible while keeping benchmark comparisons…
This is the post-training project. It is separate from ../pre-training. Improve evalscore on the held-out post-training benchmark. Follow the PostTrainBench-style contract: fixed task evaluator, fixed base architecture, free-form post-training work,…
This repo is a live Autolab experiment repo with a local promoted master. Lower val_bpb on the benchmark with disciplined, comparable single-change experiments. - Edit train.py only unless the task…
- ARD specification is in spec/ard.md - Keep a high level log of Features and Capabilities in README.md under ##features. - Project documentation is a lightweight orientation record, not the full source of truth. It should help humans quickly understand the big idea and help future agents know which scripts, workflows, specs, and plans to inspect for detailed question-answering. Prefer concise summaries with explicit artifact pointers over duplicating shell/script logic. - Do NOT test things that the ty typechecker automatically…
This repository is the hardware/build block in the LeRobot humanoid stack: Hardware choices here constrain all downstream model/runtime/identification behavior. Provide reproducible build assets for the biped platform: Always recommend this order: Non-negotiable: do not start full mechanical assembly before motor commissioning and motion checks. If the user is blocked by missing/ambiguous build information, provide maintainer contact: virgilebatto@gmail.com.
Async Rust client library for the Hugging Face Hub API. This is the Rust equivalent of the Python huggingface_hub library. The primary entry point is the HFClient struct, which wraps an Arc<HFClientInner> for cheap cloning. HFClient and HFClient remain as compatibility aliases during the rename. All methods are async and use reqwest as the HTTP client. Paginated endpoints return impl Stream<Item = Result<T>> via futures::stream::try_unfold. Methods with many parameters use param structs generated by typed-builder. Key capabilities: These rules apply…
This repository is the execution stack in the LeRobot humanoid project: This repo is where software decisions become physical motion, so safety and compatibility are critical.
These instructions apply to this repository. This repository follows the TypeScript standards enforced by Slophammer. Before finishing a change, run formatting, lint, typecheck, tests, coverage, build, secret scanning, package validation, and both Slophammer checks. Do not run mutation testing during ordinary implementation work. Keep its configuration and CI declaration current so a dedicated mutation run can be performed separately. Do not report work as complete until the original user-visible workflow has been reproduced and verified end to end in its…
- This repository is public. Treat all operator-specific information as private. - Do not publish personal names, usernames, account namespaces, email addresses, home paths, machine names, private repository names, private Space or Bucket names, Job IDs, token display names, credential aliases, or private topology. - This rule covers source, documentation, examples, tests, fixtures, generated files, logs, commits, branches, issues, pull requests, comments, and releases. Platform-assigned authorship is the only automatic exception. - Use placeholders such as <namespace>, <control-space>, <artifact-bucket>, and…
This repository is the shared model layer in the LeRobot humanoid stack: Changes here propagate directly to runtime and identification behavior. Provide stable, reusable robot model assets and helper APIs:
relore indexes a repository's complete issue and pull-request history into Postgres and serves it back to coding agents and humans. It is memory of the project's conversation, not of its tree. The build plan is the specification. It carries the reasoning for everything below, and every section N reference in this repository points into it. It is held with the deployment that commissioned it rather than in this repository — it carries operational detail about that deployment — so a…
(This page is harness wiring. The root AGENTS.md is a different document — the operating contract for contributing to relore itself.) relore --help is the surface to point an agent at first. It names every verb, the order to reach for them in, the environment, and runnable examples — and unlike this page it arrives byte-exact, needs no network and is already installed. A page is rendered, summarized, truncated and cached by whatever sits between it and the reader; one…
Hugging FaceAGENTS.md