Create devcontainer.json

.devcontainer/devcontainer.json

@@ -0,0 +1,4 @@
+{
+  "image": "mcr.microsoft.com/devcontainers/universal:2",
+  "features": {}
+}

.github/CODEOWNERS

@@ -0,0 +1,5 @@
+@openai/developer-experience
+dkundel-openai
+Maratyszcza
+scott-oai
+volsgd

MANIFEST.in

@@ -0,0 +1 @@
+recursive-include _build * 

_build/gpt_oss_build_backend/__init__.py

@@ -0,0 +1 @@
+"""In-tree PEP 517 backend package for gpt-oss.""" 

_build/gpt_oss_build_backend/backend.py

@@ -0,0 +1,140 @@
+"""
+Build backend for gpt-oss that supports two modes:
+
+1) Default (pure wheel for PyPI)
+   - Delegates to setuptools.build_meta.
+   - Produces a py3-none-any wheel so PyPI accepts it (no linux_x86_64 tag).
+
+2) Optional Metal/C extension build (local only)
+   - If the environment variable GPTOSS_BUILD_METAL is set to a truthy value
+     (1/true/on/yes), delegates to scikit_build_core.build.
+   - Dynamically injects build requirements (scikit-build-core, cmake, ninja,
+     pybind11) only for this mode.
+
+Why this is needed
+- PyPI rejects Linux wheels tagged linux_x86_64; manylinux/musllinux is required
+  for binary wheels. We ship a pure wheel by default, but still allow developers
+  to build/install the native Metal backend locally when needed.
+
+Typical usage
+- Publish pure wheel: `python -m build` (do not set GPTOSS_BUILD_METAL).
+- Local Metal dev: `GPTOSS_BUILD_METAL=1 pip install -e ".[metal]"`.
+- CI: keep GPTOSS_BUILD_METAL unset for releases; set it in internal jobs that
+  exercise the extension.
+
+Notes
+- The base package remains importable without the extension. The Metal backend
+  is only used when `gpt_oss.metal` is explicitly imported.
+- This file is discovered via `backend-path = ["_build"]` and
+  `build-backend = "gpt_oss_build_backend.backend"` in pyproject.toml.
+"""
+import os
+from importlib import import_module
+from typing import Any, Mapping, Sequence
+
+
+TRUE_VALUES = {"1", "true", "TRUE", "on", "ON", "yes", "YES"}
+
+
+def _use_metal_backend() -> bool:
+    return str(os.environ.get("GPTOSS_BUILD_METAL", "")).strip() in TRUE_VALUES
+
+
+def _setuptools_backend():
+    from setuptools import build_meta as _bm  # type: ignore
+
+    return _bm
+
+
+def _scikit_build_backend():
+    return import_module("scikit_build_core.build")
+
+
+def _backend():
+    return _scikit_build_backend() if _use_metal_backend() else _setuptools_backend()
+
+
+# Required PEP 517 hooks
+
+def build_wheel(
+    wheel_directory: str,
+    config_settings: Mapping[str, Any] | None = None,
+    metadata_directory: str | None = None,
+) -> str:
+    return _backend().build_wheel(wheel_directory, config_settings, metadata_directory)
+
+
+def build_sdist(
+    sdist_directory: str, config_settings: Mapping[str, Any] | None = None
+) -> str:
+    return _backend().build_sdist(sdist_directory, config_settings)
+
+
+def prepare_metadata_for_build_wheel(
+    metadata_directory: str, config_settings: Mapping[str, Any] | None = None
+) -> str:
+    # Fallback if backend doesn't implement it
+    be = _backend()
+    fn = getattr(be, "prepare_metadata_for_build_wheel", None)
+    if fn is None:
+        # setuptools exposes it; scikit-build-core may not. Defer to building a wheel for metadata.
+        return _setuptools_backend().prepare_metadata_for_build_wheel(
+            metadata_directory, config_settings
+        )
+    return fn(metadata_directory, config_settings)
+
+
+# Optional hooks
+
+def build_editable(
+    editable_directory: str, config_settings: Mapping[str, Any] | None = None, metadata_directory: str | None = None
+) -> str:
+    be = _backend()
+    fn = getattr(be, "build_editable", None)
+    if fn is None:
+        # setuptools implements build_editable; if not available, raise the standard error
+        raise RuntimeError("Editable installs not supported by the selected backend")
+    return fn(editable_directory, config_settings)
+
+
+def get_requires_for_build_wheel(
+    config_settings: Mapping[str, Any] | None = None,
+) -> Sequence[str]:
+    if _use_metal_backend():
+        # Add dynamic build requirements only when building the Metal backend
+        return [
+            "scikit-build-core>=0.10",
+            "pybind11>=2.12",
+            "cmake>=3.26",
+            "ninja",
+        ]
+    # setuptools usually returns []
+    return list(_setuptools_backend().get_requires_for_build_wheel(config_settings))
+
+
+def get_requires_for_build_sdist(
+    config_settings: Mapping[str, Any] | None = None,
+) -> Sequence[str]:
+    # No special requirements for SDist
+    be = _backend()
+    fn = getattr(be, "get_requires_for_build_sdist", None)
+    if fn is None:
+        return []
+    return list(fn(config_settings))
+
+
+def get_requires_for_build_editable(
+    config_settings: Mapping[str, Any] | None = None,
+) -> Sequence[str]:
+    if _use_metal_backend():
+        return [
+            "scikit-build-core>=0.10",
+            "pybind11>=2.12",
+            "cmake>=3.26",
+            "ninja",
+        ]
+    be = _setuptools_backend()
+    fn = getattr(be, "get_requires_for_build_editable", None)
+    if fn is None:
+        return []
+    return list(fn(config_settings)) 

awesome-gpt-oss.md

@@ -10,6 +10,7 @@ This is a list of guides and resources to help you get started with the gpt-oss
   - [Cloud](#cloud)
 - [Examples / Tutorials](#examples--tutorials)
 - [Tools](#tools)
+- [Training](#training)
 
 ## Inference
 
@@ -25,36 +26,48 @@ This is a list of guides and resources to help you get started with the gpt-oss
   - [Use gpt-oss-120b with LM Studio](https://lmstudio.ai/models/openai/gpt-oss-120b)
 - Hugging Face & Transformers
   - [How to run gpt-oss with Transformers](https://cookbook.openai.com/articles/gpt-oss/run-transformers)
-  - [Hugging Face & gpt-oss launch blog](http://huggingface.co/blog/welcome-openai-gpt-oss)
+  - [Hugging Face & gpt-oss launch blog](https://huggingface.co/blog/welcome-openai-gpt-oss)
   - [Collection of Hugging Face examples](https://github.com/huggingface/gpt-oss-recipes)
 - NVIDIA
   - [gpt-oss on RTX](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss)
+- AMD
+  - [Running gpt-oss models on AMD Ryzen AI Processors and Radeon Graphics Cards](https://www.amd.com/en/blogs/2025/how-to-run-openai-gpt-oss-20b-120b-models-on-amd-ryzen-ai-radeon.html)
+  - [Running gpt-oss on STX Halo and Radeon dGPUs using Lemonade](https://lemonade-server.ai/news/gpt-oss.html)
+- llama.cpp
+  - [Running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)
 
 ### Server
 
 - vLLM
   - [How to run gpt-oss with vLLM](https://cookbook.openai.com/articles/gpt-oss/run-vllm)
+  - [vLLM & gpt-oss recipies](https://docs.vllm.ai/projects/recipes/en/latest/OpenAI/GPT-OSS.html)
 - NVIDIA
-  - [Optimizing gpt-oss with NVIDIA TensorRT-LLM](https://cookbook.openai.com/articles/gpt-oss/run-nvidia)
-  - [Deploying gpt-oss on TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM/blob/main/docs/source/blogs/tech_blog/blog_9_Deploying_GPT_OSS_on_TRTLLM.md)
+  - [Optimizing gpt-oss with NVIDIA TensorRT-LLM](https://cookbook.openai.com/articles/run-nvidia)
+  - [Deploying gpt-oss on TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM/blob/main/docs/source/blogs/tech_blog/blog9_Deploying_GPT_OSS_on_TRTLLM.md)
+- AMD
+  - [Running the Latest Open Models from OpenAI on AMD AI Hardware](https://rocm.blogs.amd.com/ecosystems-and-partners/openai-day-0/README.html)
 
 ### Cloud
 
 - Groq
-  - [Groq & gpt-oss launch blog](http://groq.com/day-zero-support-for-openai-open-model)
+  - [Groq & gpt-oss launch blog](https://groq.com/blog/day-zero-support-for-openai-open-models)
   - [gpt-oss-120b model on the GroqCloud Playground](https://console.groq.com/playground?model=openai/gpt-oss-120b)
   - [gpt-oss-20b model on the GroqCloud Playground](https://console.groq.com/playground?model=openai/gpt-oss-20b)
   - [gpt-oss with built-in web search on GroqCloud](https://console.groq.com/docs/browser-search)
-  - [gpt-oss with built-in code execution on GroqCloud](https://console.groq.com/docs/code-execution) 
-  - [Responses API on Groq](https://console.groq.com/docs/responses)
+  - [gpt-oss with built-in code execution on GroqCloud](https://console.groq.com/docs/code-execution)
+  - [Responses API on Groq](https://console.groq.com/docs/responses-api)
 - NVIDIA
   - [NVIDIA launch blog post](https://blogs.nvidia.com/blog/openai-gpt-oss/)
   - [NVIDIA & gpt-oss developer launch blog post](https://developer.nvidia.com/blog/delivering-1-5-m-tps-inference-on-nvidia-gb200-nvl72-nvidia-accelerates-openai-gpt-oss-models-from-cloud-to-edge/)
   - Use [gpt-oss-120b](https://build.nvidia.com/openai/gpt-oss-120b) and [gpt-oss-20b](https://build.nvidia.com/openai/gpt-oss-20b) on NVIDIA's Cloud
 - Cloudflare
-  - [Cloudflare & gpt-oss launch blog post](http://blog.cloudflare.com/openai-gpt-oss-on-workers-ai)
+  - [Cloudflare & gpt-oss launch blog post](https://blog.cloudflare.com/openai-gpt-oss-on-workers-ai)
   - [gpt-oss-120b on Cloudflare Workers AI](https://developers.cloudflare.com/workers-ai/models/gpt-oss-120b)
   - [gpt-oss-20b on Cloudflare Workers AI](https://developers.cloudflare.com/workers-ai/models/gpt-oss-20b)
+- AMD
+  - [gpt-oss-120B on AMD MI300X](https://huggingface.co/spaces/amd/gpt-oss-120b-chatbot)
+- AWS (Deploy via Tensorfuse)
+  - [Deploy gpt-oss for both 20b and 120b models on AWS EKS](https://tensorfuse.io/docs/guides/modality/text/openai_oss)
 
 ## Examples & Tutorials
 
@@ -65,6 +78,12 @@ This is a list of guides and resources to help you get started with the gpt-oss
 - [Example `python` tool for gpt-oss](./gpt_oss/tools/python_docker/)
 - [Example `browser` tool for gpt-oss](./gpt_oss/tools/simple_browser/)
 
+## Training
+
+- [Hugging Face TRL examples](https://github.com/huggingface/gpt-oss-recipes)
+- [LlamaFactory examples](https://llamafactory.readthedocs.io/en/latest/advanced/best_practice/gpt-oss.html)
+- [Unsloth examples](https://docs.unsloth.ai/basics/gpt-oss-how-to-run-and-fine-tune)
+
 ## Contributing
 
 Feel free to open a PR to add your own guides and resources on how to run gpt-oss. We will try to review it and add it here.

compatibility-test/.gitignore

@@ -0,0 +1,142 @@
+# Logs
+logs
+*.log
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+lerna-debug.log*
+
+# Diagnostic reports (https://nodejs.org/api/report.html)
+report.[0-9]*.[0-9]*.[0-9]*.[0-9]*.json
+
+# Runtime data
+pids
+*.pid
+*.seed
+*.pid.lock
+
+# Directory for instrumented libs generated by jscoverage/JSCover
+lib-cov
+
+# Coverage directory used by tools like istanbul
+coverage
+*.lcov
+
+# nyc test coverage
+.nyc_output
+
+# Grunt intermediate storage (https://gruntjs.com/creating-plugins#storing-task-files)
+.grunt
+
+# Bower dependency directory (https://bower.io/)
+bower_components
+
+# node-waf configuration
+.lock-wscript
+
+# Compiled binary addons (https://nodejs.org/api/addons.html)
+build/Release
+
+# Dependency directories
+node_modules/
+jspm_packages/
+
+# Snowpack dependency directory (https://snowpack.dev/)
+web_modules/
+
+# TypeScript cache
+*.tsbuildinfo
+
+# Optional npm cache directory
+.npm
+
+# Optional eslint cache
+.eslintcache
+
+# Optional stylelint cache
+.stylelintcache
+
+# Optional REPL history
+.node_repl_history
+
+# Output of 'npm pack'
+*.tgz
+
+# Yarn Integrity file
+.yarn-integrity
+
+# dotenv environment variable files
+.env
+.env.*
+!.env.example
+
+# parcel-bundler cache (https://parceljs.org/)
+.cache
+.parcel-cache
+
+# Next.js build output
+.next
+out
+
+# Nuxt.js build / generate output
+.nuxt
+dist
+
+# Gatsby files
+.cache/
+# Comment in the public line in if your project uses Gatsby and not Next.js
+# https://nextjs.org/blog/next-9-1#public-directory-support
+# public
+
+# vuepress build output
+.vuepress/dist
+
+# vuepress v2.x temp and cache directory
+.temp
+.cache
+
+# Sveltekit cache directory
+.svelte-kit/
+
+# vitepress build output
+**/.vitepress/dist
+
+# vitepress cache directory
+**/.vitepress/cache
+
+# Docusaurus cache and generated files
+.docusaurus
+
+# Serverless directories
+.serverless/
+
+# FuseBox cache
+.fusebox/
+
+# DynamoDB Local files
+.dynamodb/
+
+# Firebase cache directory
+.firebase/
+
+# TernJS port file
+.tern-port
+
+# Stores VSCode versions used for testing VSCode extensions
+.vscode-test
+
+# yarn v3
+.pnp.*
+.yarn/*
+!.yarn/patches
+!.yarn/plugins
+!.yarn/releases
+!.yarn/sdks
+!.yarn/versions
+
+# Vite logs files
+vite.config.js.timestamp-*
+vite.config.ts.timestamp-*
+
+rollout_*.jsonl
+analysis_*.json