HMR provides a pythonic, flexible, progressive-yet-intuitive reactive programming engine / framework, and on top of that, a fine-grained, on-demand hot-reload tool.
In Python, HMR stands for Hot Module Reload, though in JavaScript it commonly refers to Hot Module Replacement—essentially the same concept. It is a feature that allows part of your app to be updated at runtime without a full rerun, without needing to restart the entire process.
Unlike traditional Python reloaders (such as watchfiles CLI, uvicorn --reload, or Flask's debug mode), HMR is much more efficient and robust. These tools typically restart the entire process whenever a file changes, which is wasteful. HMR instead intelligently reruns only what's necessary, keeping your app state intact.
fastapi.mov
Imagine you're developing an ML service using FastAPI with a model that requires 5 seconds to initialize. When using uvicorn --reload, any change—even updating a simple docstring—triggers a full process restart, forcing you to wait those 5 seconds every time. It's as frustrating as encountering a red light at every intersection.
HMR offers a smoother experience. Changes take effect instantly because HMR is variable-level fine-grained: your codebase functions like a dependency graph—when you modify a file, HMR only reruns the affected modules from that modified module up to your entry point file. If you update a variable that nothing depends on, nothing happens.
Caution
hmr should not be confused with client-side hot reloading that updates browser content when server code changes. This package implements server-side HMR, which only reloads python code upon changes.
The quickest way to experience HMR is through the CLI:
pip install hmr
hmr path/to/your/entry-file.pyParameters work exactly like python command, except your files now hot-reload on changes. Try saving files to see instant updates without losing state.
If you have uv installed, you can try hmr directly with:
uvx hmr path/to/your/entry-file.pyHMR provides a rich ecosystem of tools for different Python development scenarios:
| Package | Use Case |
|---|---|
hmr |
Reactive programming library and HMR core implementation |
uvicorn-hmr |
HMR-enabled Uvicorn server for ASGI apps (FastAPI, Starlette, etc.) |
mcp-hmr |
HMR-enabled MCP / FastMCP servers |
hmr-daemon |
Background daemon that refreshes modules on changes |
fastapi-reloader |
Browser auto-refresh middleware for automatic page reloading |
Tip
The hmr ecosystem is essentially stable and production-ready for most use cases. It has been carefully designed to handle many common edge cases and Pythonic magic patterns, including lazy imports, dynamic imports, module-level __getattr__, decorators, and more. However, circular dependencies in some edge cases may still cause unexpected behavior. Use with caution if you have a lot of code in __init__.py.
demo.mov
demo/- Basic script with hot module reloadingfastapi/- FastAPI server with hot reloading and browser refreshflask/- Flask app with hot module reloadingmcp/- MCP server with live code updates without connection drops
HMR is already a common feature in the frontend world. Web frameworks like Vite supports syncing changes to the browser without a full refresh. Test frameworks like Vitest supports on-demand updating test results without a full rerun.
So, why not bring this magic to Python?
HMR uses runtime dependency tracking instead of static analysis to achieve fine-grained reactivity:
- Signal & Observer Pattern: Signal is an alternative to the observer pattern. I implemented a simple signal system to notify changes whenever data is accessed or modified.
- Custom Module Class: I implemented a custom Module class which tracks every
__getattr__and__setattr__call. When a variable is changed, it notifies the modules that use it. This notification is recursive but fine-grained. - File System Tracking:
watchfilesis used to detect file changes. If a change's path is a Python module that has been imported, it triggers reloading of affected code. Additionally, HMR monitors all file reads (viasys.addaudithook) so changes to non-Python files (YAML, JSON, etc.) also trigger appropriate reloads.
This runtime approach enables variable-level granularity and can track dependencies across any file type—far more precise than static analysis.
This might just be one of the first variable-level fine-grained HMR frameworks in the Python ecosystem—and honestly, the real magic lies in the ecosystem we build around it.
Pair uvicorn + hmr, and you've got yourself a Vite-like development experience. Combine pytest + hmr for test-driven development that rivals Vitest. The possibilities with other libraries? Endless. Let's brainstorm together—who knows what fun (or mildly chaotic) things we might create!
Tip
A little backstory: the code for hmr lives in another repo because, truth be told, I wasn't planning on building an HMR framework. This started as an experiment in bringing reactive programming to Python. Along the way, I realized: why not make a module's globals reactive? And that's how HMR was born! While it began as a side project, I see tremendous potential in it for advancing Python's development experience.
For now, this repo is home to the main implementation and examples. If you think HMR has potential, or you just want to throw ideas around, I'd love to hear from you. We believe that the Python community deserves a more dynamic, responsive development experience, and we're excited to see where this can take us!
About fine-grained reactivity, I recommend reading SolidJS’s excellent explanation.