# Repository: plasma-umass/scalene
# Stars: 13367
## CLAUDE.md
# Scalene Development Guide
## Project Overview
Scalene is a high-performance CPU, GPU, and memory profiler for Python with AI-powered optimization proposals. It runs significantly faster than other Python profilers while providing detailed performance information. See the paper `docs/osdi23-berger.pdf` for technical details on Scalene's design.
**Key features:**
- CPU, GPU (NVIDIA/Apple), and memory profiling
- AI-powered optimization suggestions (OpenAI, Anthropic, Azure, Amazon Bedrock, Gemini, Ollama)
- Web-based GUI and CLI interfaces
- Jupyter notebook support via magic commands (`%scrun`, `%%scalene`)
- Line-by-line profiling with low overhead
- Separates Python time from native/C time
**Platform support:** Linux, macOS, WSL 2 (full support); Windows (partial support)
## Build & Test Commands
```bash
# Install in development mode
pip install -e .
# Run all tests
python3 -m pytest tests/
# Run tests for a specific Python version
python3.X -m pytest tests/
# Run linters
mypy scalene
ruff check scalene
# Run a single test file
python3 -m pytest tests/test_coverup_83.py -v
```
## Project Structure
### Core Profiler Components (`scalene/`)
- **`scalene_profiler.py`** - Main profiler class (`Scalene`). Entry point for profiling. Uses signal-based sampling for CPU profiling. Coordinates all profiling subsystems.
- **`scalene_statistics.py`** - `ScaleneStatistics` class. Collects and aggregates profiling data. Key types: `ProfilingSample`, `MemcpyProfilingSample`. Uses `RunningStats` for statistical aggregation.
- **`scalene_output.py`** - Profile output formatting for CLI/HTML
- **`scalene_json.py`** - `ScaleneJSON` class for JSON output format
- **`scalene_analysis.py`** - Profile analysis logic
### Entry Points
- **`__main__.py`** - Entry point for `python -m scalene`
- **`profile.py`** - Entry point for `--on`/`--off` control of background profiling
### Configuration & Arguments
- **`scalene_config.py`** - Version info (`scalene_version`, `scalene_date`) and constants:
- `SCALENE_PORT = 11235` - Default port for web UI
- `NEWLINE_TRIGGER_LENGTH` - Must match `src/include/sampleheap.hpp`
- **`scalene_arguments.py`** - `ScaleneArguments` class (extends `argparse.Namespace`) with all profiler options and their defaults defined in `ScaleneArgumentsDict`
- **`scalene_parseargs.py`** - `ScaleneParseArgs.parse_args()` builds the argument parser. `RichArgParser` provides colored help output (uses Rich on Python < 3.14, native argparse colors on 3.14+)
### Signal Handling
- **`scalene_signals.py`** - Signal definitions for CPU sampling
- **`scalene_signal_manager.py`** - Manages signal handlers
- **`scalene_sigqueue.py`** - Signal queue management
- **`scalene_client_timer.py`** - Timer for periodic profiling
### GPU Support
- **`scalene_nvidia_gpu.py`** - NVIDIA GPU profiling via `pynvml`
- **`scalene_apple_gpu.py`** - Apple GPU profiling (Metal)
- **`scalene_accelerator.py`** - Generic accelerator interface
- **`scalene_neuron.py`** - AWS Neuron support
### Memory Profiling
- **`scalene_memory_profiler.py`** - Memory profiling logic
- **`scalene_leak_analysis.py`** - Memory leak detection (experimental, `--memory-leak-detector`)
- **`scalene_mapfile.py`** - `ScaleneMapFile` for memory-mapped communication with native extension
- **`scalene_preload.py`** - Sets up `LD_PRELOAD`/`DYLD_INSERT_LIBRARIES` for native memory tracking
### Jupyter Integration
- **`scalene_magics.py`** - Jupyter magic commands (`%scrun` for line mode, `%%scalene` for cell mode)
- **`scalene_jupyter.py`** - Jupyter notebook support utilities
### Replacement Modules (`replacement_*.py`)
These modules monkey-patch standard library functions to capture profiling data during blocking operations:
- **`replacement_fork.py`** - Tracks `os.fork()`
- **`replacement_exit.py`** - Tracks `sys.exit()`
- **`replacement_lock.py`**, **`replacement_mp_lock.py`**, **`replacement_sem_lock.py`** - Lock acquisition timing
- **`replacement_thread_join.py`**, **`replacement_pjoin.py`** - Thread/process join timing
- **`replacement_signal_fns.py`** - Signal function replacements
- **`replacement_poll_selector.py`** - I/O polling timing
- **`replacement_get_context.py`** - Multiprocessing context
### Utilities
- **`runningstats.py`** - `RunningStats` class for online statistical calculations (mean, variance)
- **`scalene_funcutils.py`** - Function utilities
- **`scalene_utility.py`** - General utilities
- **`sparkline.py`** - Sparkline generation for memory visualization
- **`syntaxline.py`** - Syntax-highlighted source code lines
- **`adaptive.py`** - Adaptive sampling logic
- **`time_info.py`** - Time measurement utilities
- **`sorted_reservoir.py`** - Reservoir sampling for bounded-size sample collection
### GUI (`scalene/scalene-gui/`)
Web-based GUI built with TypeScript, bundled with esbuild.
**Core Files:**
- **`index.html.template`** - Jinja2 template for main GUI page (rendered by `scalene_utility.py`)
- **`scalene-gui.ts`** - Main TypeScript entry point, UI event handlers, initialization
- **`scalene-gui-bundle.js`** - Bundled JavaScript output (generated, do not edit directly)
**AI Provider Modules:**
- **`openai.ts`** - OpenAI API integration (`sendPromptToOpenAI`, `fetchOpenAIModels`)
- **`anthropic.ts`** - Anthropic Claude API integration
- **`gemini.ts`** - Google Gemini API integration (`sendPromptToGemini`, `fetchGeminiModels`)
- **`optimizations.ts`** - Provider dispatch logic, prompt generation
- **`persistence.ts`** - localStorage persistence with environment variable fallbacks
**Support Files:**
- **`launchbrowser.py`** - Opens browser to GUI (default port 11235)
- **`find_browser.py`** - Cross-platform browser detection
**Vendored Assets (for offline support):**
- **`jquery-3.6.0.slim.min.js`** - jQuery (vendored locally, not loaded from CDN)
- **`bootstrap.min.css`** - Bootstrap 5.1.3 CSS
- **`bootstrap.bundle.min.js`** - Bootstrap 5.1.3 JS with Popper
- **`prism.css`** - Syntax highlighting styles
- **`favicon.ico`** - Scalene favicon
- **`scalene-image.png`** - Scalene logo
These assets are copied to a temp directory when serving via HTTP, enabling the GUI to work in air-gapped/offline environments.
**Building the GUI:**
```bash
cd scalene/scalene-gui
npx esbuild scalene-gui.ts --bundle --outfile=scalene-gui-bundle.js --format=iife --global-name=ScaleneGUI
```
### Native Extensions (`src/`)
C++ code for low-overhead memory allocation tracking:
**Headers (`src/include/`):**
- **`sampleheap.hpp`** - Sampling heap allocator. Key constant `NEWLINE` must match Python config.
- **`memcpysampler.hpp`** - Intercepts `memcpy` to track copy volume
- **`pywhere.hpp`** - Tracks Python file/line info for allocations
- **`samplefile.hpp`** - File-based communication with Python
- **`sampler.hpp`**, **`poissonsampler.hpp`**, **`thresholdsampler.hpp`** - Sampling strategies
- **`scaleneheader.hpp`** - Common header definitions
**Sources (`src/source/`):**
- **`libscalene.cpp`** - Main native library (loaded via `LD_PRELOAD`)
- **`pywhere.cpp`** - Python location tracking implementation
- **`get_line_atomic.cpp`** - Atomic line number access
- **`traceconfig.cpp`** - Trace configuration
### Vendor Libraries (`vendor/`)
- **`Heap-Layers/`** - Memory allocator infrastructure (by Emery Berger)
- **`printf/`** - Async-signal-safe printf implementation
## Key Patterns
### Python Version Compatibility
The codebase supports Python 3.8-3.14. Version-specific code uses:
```python
if sys.version_info >= (3, 14):
# Python 3.14+ specific code
else:
# Older Python versions
```
**Type Annotation Compatibility (Python 3.8/3.9):**
- **Do NOT use `X | Y` union syntax** in runtime-evaluated annotations (PEP 604 requires Python 3.10+). Use `Optional[X]` or `Union[X, Y]` from `typing` instead.
- **Do NOT use `list[X]`, `dict[K, V]`, `tuple[X, ...]`** in runtime-evaluated annotations (PEP 585 lowercase generics require Python 3.9+). Use `List`, `Dict`, `Tuple` from `typing` for 3.8 support.
- Adding `from __future__ import annotations` makes all annotations strings (not evaluated at runtime), which allows modern syntax on older Python. However, this can break code that inspects annotations at runtime (e.g., dataclasses, pydantic).
- The safest approach for this codebase: use `typing.Optional`, `typing.Union`, `typing.List`, `typing.Tuple`, `typing.Dict` in all annotation positions that are evaluated at runtime (function signatures, variable annotations outside `if TYPE_CHECKING` blocks).
**Python 3.13 Changes (`dis` module):**
- `dis.Instruction.starts_line` changed from `int | None` (line number) to `bool`
- New `dis.Instruction.line_number` attribute (`int | None`) added for the actual line number
- On Python < 3.13, `starts_line` is only set on the **first** instruction of each source line; use a line-tracking loop to propagate line numbers to subsequent instructions
**Bytecode/Opcode Compatibility (`dis` module):**
- **Never match specific opcode names** (e.g., `JUMP_BACKWARD`, `JUMP_ABSOLUTE`, `POP_JUMP_IF_TRUE`). Opcode names change across Python versions — for example, Python 3.10 while loops use `POP_JUMP_IF_TRUE` for backward jumps, Python 3.11+ uses `JUMP_BACKWARD`, and `JUMP_ABSOLUTE` was removed in 3.12.
- **Always use abstract `dis` module categories** when possible: `dis.hasjabs` (absolute jump opcodes), `dis.hasjrel` (relative jump opcodes), `dis.hasconst`, `dis.hasname`, etc. These are maintained by CPython and work across all versions.
- For call detection, matching `opname.startswith("CALL")` is acceptable since that prefix has been stable, but prefer opcode integer sets over name strings for hot paths.
- When checking jump direction (forward vs backward), use `instr.argval` (which `dis` resolves to an absolute offset) and compare against `instr.offset`, rather than relying on opcode names to imply direction.
**Python 3.14 Changes:**
- `argparse` now has built-in colored help output (`color=True` parameter)
- `RichArgParser` uses Rich for colors on Python < 3.14, native argparse colors on 3.14+
### Argument Parsing (`scalene_parseargs.py`)
```python
class RichArgParser(argparse.ArgumentParser):
"""ArgumentParser that uses Rich for colored output on Python < 3.14."""
def __init__(self, *args, **kwargs):
if sys.version_info < (3, 14):
from rich.console import Console
self._console = Console()
else:
self._console = None
super().__init__(*args, **kwargs)
```
The `_colorize_help_for_rich()` function applies Python 3.14-style colors using Rich markup:
- `usage:` and `options:` → bold blue
- Program name → bold magenta
- Long options (`--foo`) → bold cyan
- Short options (`-h`) → bold green
- Metavars (`FOO`) → bold yellow
### GUI Patterns
**Preventing Browser Password Prompts:**
Use `autocomplete="one-time-code"` on password/API key inputs to prevent browsers from offering to save them:
```html
```
**Show/Hide Password Toggle:**
```typescript
function togglePassword(inputId: string, button: HTMLButtonElement): void {
const input = document.getElementById(inputId) as HTMLInputElement;
if (input.type === "password") {
input.type = "text";
button.textContent = "Hide";
} else {
input.type = "password";
button.textContent = "Show";
}
}
```
**Provider Field Visibility:**
Use CSS classes to show/hide provider-specific fields:
```typescript
function toggleServiceFields(): void {
const service = (document.getElementById("service") as HTMLSelectElement).value;
// Hide all provider sections
document.querySelectorAll(".provider-section").forEach((el) => {
(el as HTMLElement).style.display = "none";
});
// Show selected provider section
const section = document.querySelector(`.${service}-fields`);
if (section) (section as HTMLElement).style.display = "block";
}
```
**Persistent Form Elements:**
Add class `persistent` to inputs that should be saved/restored from localStorage:
```html
```
The `persistence.ts` module handles save/restore automatically.
**Standalone HTML Generation:**
The `generate_html()` function in `scalene_utility.py` supports a `standalone` parameter:
- When `standalone=False` (default): Assets are referenced as local files (e.g., `
{% else %}
{% endif %}
```
### Module Imports
When importing submodules, be explicit:
```python
# Correct - mypy can verify this
import importlib.util
importlib.util.find_spec(mod_name)
# Wrong - mypy error: Module has no attribute "util"
import importlib
importlib.util.find_spec(mod_name)
```
## Testing
### Test Files (`tests/`)
- **`test_coverup_*.py`** - Auto-generated coverage tests
- **`test_runningstats.py`** - Statistics tests (requires `hypothesis`)
- **`test_scalene_json.py`** - JSON output tests (requires `hypothesis`)
- **`test_nested_package_relative_import.py`** - Import handling tests
### Test Dependencies
```bash
pip install pytest pytest-asyncio hypothesis
```
### Running Tests Across Python Versions
```bash
for v in 3.9 3.10 3.11 3.12 3.13 3.14; do
python$v -m pytest tests/test_coverup_83.py -v
done
```
### Flaky Smoketests
The smoketests in `test/` can be flaky due to timing/sampling issues inherent to profiling:
- **"No non-zero lines in X"** - The profiler didn't collect enough samples. This happens when the test runs too quickly or signal delivery timing varies.
- **"Expected function 'X' not returned"** - A function wasn't sampled. Common with short-running functions.
These failures are usually timing-related and pass on re-run. They're more common on CI due to variable machine load.
### Port Binding in Tests
When testing port availability, never use hardcoded ports - they may already be in use on CI runners:
```python
# Bad - port 49200 might be in use
port = 49200
sock.bind(("", port))
# Good - find an available port first
port = find_available_port(49200, 49300)
if port is None:
return # Skip test if no ports available
sock.bind(("", port))
```
## CI/CD (`.github/workflows/`)
- **`run-linters.yml`** - Runs mypy and ruff on Python 3.9-3.14
- **`tests.yml`** - Runs pytest on Python 3.9-3.14
- **`build-and-upload.yml`** - Build and publish to PyPI
## Common Tasks
### Adding a New CLI Option
1. Add default value in `scalene_arguments.py`:
```python
class ScaleneArgumentsDict(TypedDict, total=False):
my_option: bool
```
2. Add argument in `scalene_parseargs.py`:
```python
parser.add_argument(
"--my-option",
dest="my_option",
action="store_true",
default=defaults.my_option,
help="Description of option",
)
```
### Adding a New AI Provider
1. **Create provider module** (`scalene/scalene-gui/newprovider.ts`):
```typescript
export async function sendPromptToNewProvider(
prompt: string,
apiKey: string
): Promise {
// API call implementation
}
export async function fetchNewProviderModels(apiKey: string): Promise {
// Optional: fetch available models from API
}
```
2. **Update `optimizations.ts`**:
- Import the new module
- Add case in `sendPromptToService()` switch statement
3. **Update `index.html.template`**:
- Add option to `#service` select dropdown
- Add provider section with API key input, model selector, etc.
- Add CSS for `.newprovider-fields` visibility
4. **Update `scalene-gui.ts`**:
- Add provider to `toggleServiceFields()` function
- Add refresh handler if dynamic model fetching is supported
- Update `getDefaultProvider()` if env var support is needed
5. **Update `persistence.ts`** (for env var support):
- Add mapping in `envKeyMap` for new fields
6. **Update `scalene_utility.py`**:
- Read environment variable in `api_keys` dict
- Pass to template rendering
7. **Rebuild the bundle**:
```bash
cd scalene/scalene-gui
npx esbuild scalene-gui.ts --bundle --outfile=scalene-gui-bundle.js --format=iife --global-name=ScaleneGUI
```
### Environment Variable API Keys
The GUI supports prepopulating API keys from environment variables:
| Element ID | Environment Variable | Provider |
|------------|---------------------|----------|
| `api-key` | `OPENAI_API_KEY` | OpenAI |
| `anthropic-api-key` | `ANTHROPIC_API_KEY` | Anthropic |
| `gemini-api-key` | `GEMINI_API_KEY` or `GOOGLE_API_KEY` | Gemini |
| `azure-api-key` | `AZURE_OPENAI_API_KEY` | Azure OpenAI |
| `azure-api-url` | `AZURE_OPENAI_ENDPOINT` | Azure OpenAI |
| `aws-access-key` | `AWS_ACCESS_KEY_ID` | Amazon Bedrock |
| `aws-secret-key` | `AWS_SECRET_ACCESS_KEY` | Amazon Bedrock |
| `aws-region` | `AWS_DEFAULT_REGION` or `AWS_REGION` | Amazon Bedrock |
**Flow:**
1. `scalene_utility.py` reads env vars and passes to Jinja2 template
2. Template injects `envApiKeys` JavaScript object into page
3. `persistence.ts` uses env vars as fallbacks when localStorage is empty
### Updating Version
Edit `scalene/scalene_config.py`:
```python
scalene_version = "X.Y.Z"
scalene_date = "YYYY.MM.DD"
```
## Dependencies
Key runtime dependencies:
- `rich` - Terminal formatting and colors
- `cloudpickle` - Serialization
- `pynvml` - NVIDIA GPU support (optional)
See `requirements.txt` for full list.
## CLI Structure
Scalene uses a verb-based CLI with two main subcommands:
```bash
# Profile a program (saves to scalene-profile.json by default)
scalene run [options] yourprogram.py
# View an existing profile
scalene view [options] [profile.json]
```
### Run Subcommand Options
```bash
scalene run prog.py # profile, save to scalene-profile.json
scalene run -o my.json prog.py # save to custom file
scalene run --cpu-only prog.py # profile CPU only (faster)
scalene run -c config.yaml prog.py # load options from config file
scalene run prog.py --- --arg # pass args to program
```
### View Subcommand Options
```bash
scalene view # open in browser
scalene view --cli # view in terminal
scalene view --html # save to scalene-profile.html
scalene view --standalone # save as self-contained HTML (all assets embedded)
scalene view myprofile.json # open specific profile
```
### Profile Completion Message
After profiling completes, Scalene prints instructions for viewing the profile:
```
Scalene: profile saved to scalene-profile.json
To view in browser: scalene view
To view in terminal: scalene view --cli
```
The filename is only included in the command if a non-default output file was used.
### YAML Configuration
Create a `scalene.yaml` file with options:
```yaml
outfile: my-profile.json
cpu-only: true
profile-only: "mypackage,utils"
cpu-percent-threshold: 5
```
Load with: `scalene run -c scalene.yaml prog.py`
### Advanced Options
Use `scalene run --help-advanced` to see all options including:
- `--profile-all` - profile all code, not just the target program
- `--profile-only PATH` - only profile files containing these strings
- `--profile-exclude PATH` - exclude files containing these strings
- `--profile-system-libraries` - profile Python stdlib and installed packages (skipped by default)
- `--gpu` - profile GPU time and memory
- `--memory` - profile memory usage
- `--stacks` - collect stack traces
- `--profile-interval N` - output profiles every N seconds
### Smoke Tests
Smoke tests in `test/` use the new CLI syntax:
```python
# test/smoketest.py
cmd = [sys.executable, "-m", "scalene", "run", "-o", str(outfile), *rest, fname]
```
### GitHub Workflows
Workflows in `.github/workflows/` use the new CLI:
```yaml
# Profile with interval, then view
- run: python -m scalene run --profile-interval=2 test/testme.py && python -m scalene view --cli
# Profile with module invocation
- run: python -m scalene run --- -m import_stress_test && python -m scalene view --cli
```
## Signal Handling
Scalene uses several Unix signals for profiling. The signal assignments are in `scalene_signals.py`:
| Signal | Purpose | Platform |
|--------|---------|----------|
| `SIGVTALRM` | CPU profiling timer (default) | Unix |
| `SIGALRM` | CPU profiling timer (real time mode) | Unix |
| `SIGILL` | Start profiling (`--on`) | Unix |
| `SIGBUS` | Stop profiling (`--off`) | Unix |
| `SIGPROF` | memcpy tracking | Unix |
| `SIGXCPU` | malloc tracking | Unix |
| `SIGXFSZ` | free tracking | Unix |
### Signal Conflicts with Libraries
Libraries like PyTorch Lightning may also use these signals. The `replacement_signal_fns.py` module handles conflicts:
**On Linux:** Uses real-time signals (`SIGRTMIN+1` to `SIGRTMIN+5`) for redirection. When user code sets a handler for a Scalene signal, their handler is redirected to a real-time signal. Calls to `raise_signal()` and `kill()` are also redirected transparently.
**On macOS/other platforms:** Uses handler chaining. Both Scalene's handler and the user's handler are called when the signal fires.
```python
# Platform-specific signal handling
_use_rt_signals = sys.platform == "linux" and hasattr(signal, "SIGRTMIN")
if _use_rt_signals:
# Linux: redirect to real-time signals
rt_base = signal.SIGRTMIN + 1
_signal_redirects[signal.SIGILL] = rt_base
else:
# macOS: chain handlers
def chained_handler(sig, frame):
scalene_handler(sig, frame)
user_handler(sig, frame)
```
### Frame Line Number Can Be None (Python 3.11+)
In Python 3.11+, `frame.f_lineno` can be `None` in edge cases (e.g., during multiprocessing cleanup). Always use a fallback:
```python
lineno = frame.f_lineno if frame.f_lineno is not None else frame.f_code.co_firstlineno
```
## Native Extension Build Issues
### C++ Standard Library Conflicts with vendor/printf
The `vendor/printf/printf.h` header defines macros that conflict with C++ standard library:
```c
#define vsnprintf vsnprintf_
#define snprintf snprintf_
```
This breaks `std::vsnprintf` in `` and other headers. **Fix:** Include C++ standard headers BEFORE vendor headers in `src/source/libscalene.cpp`:
```cpp
// Include C++ standard headers FIRST
#include
#include
// Then vendor headers that define conflicting macros
#include // Eventually includes printf.h
```
## Profiling Guide
See [Scalene-Agents.md](Scalene-Agents.md) for detailed information about interpreting Scalene's profiling output, including Python vs C time, memory metrics, and optimization strategies.
## Debugging Guide
See [Scalene-Debugging.md](Scalene-Debugging.md) for signal handler debugging, async profiling debugging, the profile output pipeline (three separate renderers!), and unbounded growth prevention patterns.
## GUI Development Guide
See [Scalene-GUI.md](Scalene-GUI.md) for adding new columns, Vega-Lite chart types, pie chart best practices (two-wedge rendering, rotating pies), and the chart rendering flow.
## README.md
# Scalene: a Python CPU+GPU+memory profiler with AI-powered optimization proposals
by [Emery Berger](https://emeryberger.com), [Sam Stern](https://samstern.me/), and [Juan Altmayer Pizzorno](https://github.com/jaltmayerpizzorno).
[](https://join.slack.com/t/scaleneprofil-jge3234/shared_invite/zt-110vzrdck-xJh5d4gHnp5vKXIjYD3Uwg)[Scalene community Slack](https://join.slack.com/t/scaleneprofil-jge3234/shared_invite/zt-110vzrdck-xJh5d4gHnp5vKXIjYD3Uwg)
[](https://pypi.org/project/scalene/)[](https://anaconda.org/conda-forge/scalene) [](https://pepy.tech/project/scalene)[](https://anaconda.org/conda-forge/scalene) [](https://pepy.tech/project/scalene) [](https://marketplace.visualstudio.com/items?itemName=EmeryBerger.scalene)  [](https://github.com/plasma-umass/scalene)
(tweet from Ian Ozsvald, author of [_High Performance Python_](https://smile.amazon.com/High-Performance-Python-Performant-Programming/dp/1492055026/ref=sr_1_1?crid=texbooks))
[_Python Profiler Links to AI to Improve Code Scalene identifies inefficiencies and asks GPT-4 for suggestions_](https://spectrum.ieee.org/python-programming), IEEE Spectrum
[Episode 172: Measuring Multiple Facets of Python Performance With Scalene](https://realpython.com/podcasts/rpp/172/), The Real Python podcast
***Scalene web-based user interface:*** [https://scalene-gui.github.io/scalene-gui/](https://scalene-gui.github.io/scalene-gui/)
## About Scalene
Scalene is a high-performance CPU, GPU *and* memory profiler for
Python that does a number of things that other Python profilers do not
and cannot do. It runs orders of magnitude faster than many other
profilers while delivering far more detailed information. It is also
the first profiler ever to incorporate AI-powered proposed
optimizations.
### AI-powered optimization suggestions
> **Note**
>
> For optimization suggestions, Scalene supports a variety of AI providers, including [Amazon Bedrock](https://aws.amazon.com/bedrock), [Microsoft Azure](https://azure.microsoft.com/en-us/), [OpenAI](https://openai.com), and local models via [Ollama](https://ollama.com/). To enable AI-powered optimization suggestions from AI providers, you need to select a provider and, if needed, enter your credentials, in the box under "AI Optimization Options".
>
>
Once you've entered your key and any other needed data, click on the lightning bolt (⚡) beside any line or the explosion (💥) for an entire region of code to generate a proposed optimization. Click on a proposed optimization to copy it to the clipboard.
You can click as many times as you like on the lightning bolt or explosion, and it will generate different suggested optimizations. Your mileage may vary, but in some cases, the suggestions are quite impressive (e.g., order-of-magnitude improvements).
### Quick Start
#### Installing Scalene:
```console
python3 -m pip install -U scalene
```
or
```console
conda install -c conda-forge scalene
```
#### Using Scalene:
After installing Scalene, you can use Scalene at the command line, or as a Visual Studio Code extension.
Using the Scalene VS Code Extension:
First, install the Scalene extension from the VS Code Marketplace or by searching for it within VS Code by typing Command-Shift-X (Mac) or Ctrl-Shift-X (Windows). Once that's installed, click Command-Shift-P or Ctrl-Shift-P to open the Command Palette. Then select "Scalene: AI-powered profiling..." (you can start typing Scalene and it will pop up if it's installed). Run that and, assuming your code runs for at least a second, a Scalene profile will appear in a webview.
Commonly used command-line options:
Scalene uses a verb-based command structure with two main commands: `run` (to profile) and `view` (to display results).
```console
# Profile a program (saves to scalene-profile.json)
scalene run your_prog.py
python3 -m scalene run your_prog.py # equivalent alternative
# View a profile
scalene view # open profile in browser
scalene view --cli # view in terminal
scalene view --html # save to scalene-profile.html
scalene view --standalone # save as self-contained HTML
# Common profiling options
scalene run --cpu-only your_prog.py # only profile CPU (faster)
scalene run -o results.json your_prog.py # custom output filename
scalene run -c config.yaml your_prog.py # load options from config file
# Pass arguments to your program (use --- separator)
scalene run your_prog.py --- --arg1 --arg2
# Get help
scalene --help # main help
scalene run --help # profiling options
scalene run --help-advanced # advanced profiling options
scalene view --help # viewing options
```
Using a YAML configuration file:
You can store Scalene options in a YAML configuration file and load them with `-c` or `--config`:
```console
scalene run -c scalene.yaml your_prog.py
```
Example `scalene.yaml`:
```yaml
# Output options
outfile: my-profile.json
# Profiling mode (use only one)
cpu-only: true # CPU profiling only (faster)
# gpu: true # Include GPU profiling
# memory: true # Include memory profiling
# Filter what gets profiled
profile-only: "mypackage,mymodule" # Only profile these paths
profile-exclude: "tests,venv" # Exclude these paths
profile-all: false # Profile all code, not just target
# Performance tuning
cpu-percent-threshold: 1 # Min CPU% to report (default: 1)
cpu-sampling-rate: 0.01 # Sampling interval in seconds
malloc-threshold: 100 # Min allocations to report
# Other options
use-virtual-time: false # Measure CPU time only (not I/O)
stacks: false # Collect stack traces
memory-leak-detector: true # Detect likely memory leaks
```
Command-line arguments override config file settings.
Using Scalene programmatically in your code:
Invoke using `scalene` as above and then:
```Python
from scalene import scalene_profiler
# Turn profiling on
scalene_profiler.start()
# your code
# Turn profiling off
scalene_profiler.stop()
```
```Python
from scalene.scalene_profiler import enable_profiling
with enable_profiling():
# do something
```
Using Scalene to profile only specific functions via @profile:
Just preface any functions you want to profile with the `@profile` decorator and run it with Scalene:
```Python
# do not import profile!
@profile
def slow_function():
import time
time.sleep(3)
```
#### Web-based GUI
Scalene has both a CLI and a web-based GUI [(demo here)](https://scalene-gui.github.io/scalene-gui/).
By default, once Scalene has profiled your program, it will open a
tab in a web browser with an interactive user interface (all processing is done
locally). Hover over bars to see breakdowns of CPU and memory
consumption, and click on underlined column headers to sort the
columns. The GUI works fully offline with no internet connection required.
Use `scalene view --standalone` to generate a completely self-contained HTML file with all assets embedded, perfect for sharing or archiving.
[](https://raw.githubusercontent.com/plasma-umass/scalene/master/docs/scalene-gui-example-full.png)
## Scalene Overview
### Scalene talk (PyCon US 2021)
[This talk](https://youtu.be/5iEf-_7mM1k) presented at PyCon 2021 walks through Scalene's advantages and how to use it to debug the performance of an application (and provides some technical details on its internals). We highly recommend watching this video!
[](https://youtu.be/5iEf-_7mM1k "Scalene presentation at PyCon 2021")
### Fast and Accurate
- Scalene is **_fast_**. It uses sampling instead of instrumentation or relying on Python's tracing facilities. Its overhead is typically no more than 10-20% (and often less).
- Scalene is **accurate**. We tested CPU profiler accuracy and found that Scalene is among the most accurate profilers, correctly measuring time taken.
- Scalene performs profiling **_at the line level_** _and_ **_per function_**, pointing to the functions and the specific lines of code responsible for the execution time in your program.
### CPU profiling
- Scalene **separates out time spent in Python from time in native code** (including libraries). Most Python programmers aren't going to optimize the performance of native code (which is usually either in the Python implementation or external libraries), so this helps developers focus their optimization efforts on the code they can actually improve.
- Scalene **highlights hotspots** (code accounting for significant percentages of CPU time or memory allocation) in red, making them even easier to spot.
- Scalene also separates out **system time**, making it easy to find I/O bottlenecks.
### GPU profiling
- Scalene reports **GPU time** (currently limited to NVIDIA-based systems).
### Memory profiling
- Scalene **profiles memory usage**. In addition to tracking CPU usage, Scalene also points to the specific lines of code responsible for memory growth. It accomplishes this via an included specialized memory allocator.
- Scalene separates out the percentage of **memory consumed by Python code vs. native code**.
- Scalene produces **_per-line_ memory profiles**.
- Scalene **identifies lines with likely memory leaks**.
- Scalene **profiles _copying volume_**, making it easy to spot inadvertent copying, especially due to crossing Python/library boundaries (e.g., accidentally converting `numpy` arrays into Python arrays, and vice versa).
### Other features
- Scalene can produce **reduced profiles** (via `--reduced-profile`) that only report lines that consume more than 1% of CPU or perform at least 100 allocations.
- Scalene supports `@profile` decorators to profile only specific functions.
- When Scalene is profiling a program launched in the background (via `&`), you can **suspend and resume profiling**.
# Comparison to Other Profilers
## Performance and Features
Below is a table comparing the **performance and features** of various profilers to Scalene.
- **Slowdown**: the slowdown when running a benchmark from the Pyperformance suite. Green means less than 2x overhead. Scalene's overhead is just a 35% slowdown.
Scalene has all of the following features, many of which only Scalene supports:
- **Lines or functions**: does the profiler report information only for entire functions, or for every line -- Scalene does both.
- **Unmodified Code**: works on unmodified code.
- **Threads**: supports Python threads.
- **Multiprocessing**: supports use of the `multiprocessing` library -- _Scalene only_
- **Python vs. C time**: breaks out time spent in Python vs. native code (e.g., libraries) -- _Scalene only_
- **System time**: breaks out system time (e.g., sleeping or performing I/O) -- _Scalene only_
- **Profiles memory**: reports memory consumption per line / function
- **GPU**: reports time spent on an NVIDIA GPU (if present) -- _Scalene only_
- **Memory trends**: reports memory use over time per line / function -- _Scalene only_
- **Copy volume**: reports megabytes being copied per second -- _Scalene only_
- **Detects leaks**: automatically pinpoints lines responsible for likely memory leaks -- _Scalene only_
## Output
If you include the `--cli` option, Scalene prints annotated source code for the program being profiled
(as text, JSON (`--json`), or HTML (`--html`)) and any modules it
uses in the same directory or subdirectories (you can optionally have
it `--profile-all` and only include files with at least a
`--cpu-percent-threshold` of time). Here is a snippet from
`pystone.py`.
* **Memory usage at the top**: Visualized by "sparklines", memory consumption over the runtime of the profiled code.
* **"Time Python"**: How much time was spent in Python code.
* **"native"**: How much time was spent in non-Python code (e.g., libraries written in C/C++).
* **"system"**: How much time was spent in the system (e.g., I/O).
* **"GPU"**: (not shown here) How much time spent on the GPU, if your system has an NVIDIA GPU installed.
* **"Memory Python"**: How much of the memory allocation happened on the Python side of the code, as opposed to in non-Python code (e.g., libraries written in C/C++).
* **"net"**: Positive net memory numbers indicate total memory allocation in megabytes; negative net memory numbers indicate memory reclamation.
* **"timeline / %"**: Visualized by "sparklines", memory consumption generated by this line over the program runtime, and the percentages of total memory activity this line represents.
* **"Copy (MB/s)"**: The amount of megabytes being copied per second (see "About Scalene").
## Scalene
The following command runs Scalene on a provided example program.
```console
scalene test/testme.py
```
Click to see all Scalene's options (available by running with --help)
```console
% scalene --help
Scalene: a high-precision CPU and memory profiler, version 1.5.51 (2025.01.29)
https://github.com/plasma-umass/scalene
commands:
run Profile a Python program (saves to scalene-profile.json)
view View an existing profile in browser or terminal
examples:
% scalene run your_program.py # profile, save to scalene-profile.json
% scalene view # view scalene-profile.json in browser
% scalene view --cli # view profile in terminal
in Jupyter, line mode:
%scrun [options] statement
in Jupyter, cell mode:
%%scalene [options]
your code here
% scalene run --help
Profile a Python program with Scalene.
examples:
% scalene run prog.py # profile, save to scalene-profile.json
% scalene run -o my.json prog.py # save to custom file
% scalene run --cpu-only prog.py # profile CPU only (faster)
% scalene run -c scalene.yaml prog.py # load options from config file
% scalene run prog.py --- --arg # pass args to program
% scalene run --help-advanced # show advanced options
options:
-h, --help show this help message and exit
-o, --outfile OUTFILE output file (default: scalene-profile.json)
--cpu-only only profile CPU time (no memory/GPU)
-c, --config FILE load options from YAML config file
--help-advanced show advanced options
% scalene run --help-advanced
Advanced options for scalene run:
background profiling:
Use --off to start with profiling disabled, then control it from another terminal:
% scalene run --off prog.py # start with profiling off
% python3 -m scalene.profile --on --pid # resume profiling
% python3 -m scalene.profile --off --pid # suspend profiling
options:
--profile-all profile all code, not just the target program
--profile-only PATH only profile files containing these strings (comma-separated)
--profile-exclude PATH exclude files containing these strings (comma-separated)
--profile-system-libraries profile Python stdlib and installed packages (default: skip)
--gpu profile GPU time and memory
--memory profile memory usage
--stacks collect stack traces
--profile-interval N output profiles every N seconds (default: inf)
--use-virtual-time measure only CPU time, not I/O or blocking
--cpu-percent-threshold N only report lines with at least N% CPU (default: 1%)
--cpu-sampling-rate N CPU sampling rate in seconds (default: 0.01)
--allocation-sampling-window N allocation sampling window in bytes
--malloc-threshold N only report lines with at least N allocations (default: 100)
--program-path PATH directory containing code to profile
--memory-leak-detector EXPERIMENTAL: report likely memory leaks
--on start with profiling on (default)
--off start with profiling off
% scalene view --help
View an existing Scalene profile.
examples:
% scalene view # open in browser
% scalene view --cli # view in terminal
% scalene view --html # save to scalene-profile.html
% scalene view --standalone # save as self-contained HTML
% scalene view myprofile.json # open specific profile in browser
options:
-h, --help show this help message and exit
--cli display profile in the terminal
--html save to scalene-profile.html (no browser)
--standalone save as self-contained HTML with all assets embedded
-r, --reduced only show lines with activity (--cli mode)
```
### Scalene with Jupyter
Instructions for installing and using Scalene with Jupyter notebooks
[This notebook](https://nbviewer.jupyter.org/github/plasma-umass/scalene/blob/master/docs/scalene-demo.ipynb) illustrates the use of Scalene in Jupyter.
Installation:
```console
!pip install scalene
%load_ext scalene
```
Line mode:
```console
%scrun [options] statement
```
Cell mode:
```console
%%scalene [options]
code...
code...
```
## Installation
Using pip (Mac OS X, Linux, Windows, and WSL2)
Scalene is distributed as a `pip` package and works on Mac OS X, Linux (including Ubuntu in [Windows WSL2](https://docs.microsoft.com/en-us/windows/wsl/wsl2-index)) and Windows platforms.
> **Note for Windows users**
>
> Starting with Scalene 2.0, Windows supports full memory profiling. If you
> encounter issues, ensure you have the [Visual C++ Redistributable](https://aka.ms/vs/17/release/vc_redist.x64.exe)
> installed. If building from source, you will need Visual C++ Build Tools and CMake.
>
You can install it as follows:
```console
% pip install -U scalene
```
or
```console
% python3 -m pip install -U scalene
```
You may need to install some packages first.
See https://stackoverflow.com/a/19344978/4954434 for full instructions for all Linux flavors.
For Ubuntu/Debian:
```console
% sudo apt install git python3-all-dev
```
Using conda (Mac OS X, Linux, Windows, and WSL2)
```console
% conda install -c conda-forge scalene
```
Scalene is distributed as a `conda` package and works on Mac OS X, Linux (including Ubuntu in [Windows WSL2](https://docs.microsoft.com/en-us/windows/wsl/wsl2-index)) and Windows platforms.
> **Note for Windows users**
>
> Starting with Scalene 2.0, Windows supports full memory profiling. If you
> encounter issues, ensure you have the [Visual C++ Redistributable](https://aka.ms/vs/17/release/vc_redist.x64.exe)
> installed.
>
On ArchLinux
You can install Scalene on Arch Linux via the [AUR
package](https://aur.archlinux.org/packages/python-scalene-git/). Use your favorite AUR helper, or
manually download the `PKGBUILD` and run `makepkg -cirs` to build. Note that this will place
`libscalene.so` in `/usr/lib`; modify the below usage instructions accordingly.
# Frequently Asked Questions
Can I use Scalene with PyTest?
**A:** Yes! You can run it as follows (for example):
`scalene run -m pytest your_test.py`
or
`python3 -m scalene run -m pytest your_test.py`
Is there any way to get shorter profiles or do more targeted profiling?
**A:** Yes! There are several options:
1. Use `--reduced-profile` to include only lines and files with memory/CPU/GPU activity.
2. Use `--profile-only` to include only filenames containing specific strings (as in, `--profile-only foo,bar,baz`).
3. Decorate functions of interest with `@profile` to have Scalene report _only_ those functions.
4. Turn profiling on and off programmatically by importing Scalene profiler (`from scalene import scalene_profiler`) and then turning profiling on and off via `scalene_profiler.start()` and `scalene_profiler.stop()`. By default, Scalene runs with profiling on, so to delay profiling until desired, use the `--off` command-line option (`scalene run --off yourprogram.py`).
How do I run Scalene in PyCharm?
**A:** In PyCharm, you can run Scalene at the command line by opening the terminal at the bottom of the IDE and running a Scalene command (e.g., `scalene run `). Then use `scalene view --html` to generate an HTML file (`scalene-profile.html`) that you can view in the IDE.
How do I use Scalene with Django?
**A:** Pass in the `--noreload` option (see https://github.com/plasma-umass/scalene/issues/178).
Does Scalene work with gevent/Greenlets?
**A:** Yes! Put the following code in the beginning of your program, or modify the call to `monkey.patch_all` as below:
```python
from gevent import monkey
monkey.patch_all(thread=False)
```
How do I use Scalene with PyTorch on the Mac?
**A:** Scalene works with PyTorch version 1.5.1 on Mac OS X. There's a bug in newer versions of PyTorch (https://github.com/pytorch/pytorch/issues/57185) that interferes with Scalene (discussion here: https://github.com/plasma-umass/scalene/issues/110), but only on Macs.
# Technical Information
For details about how Scalene works, please see the following paper, which won the Jay Lepreau Best Paper Award at [OSDI 2023](https://www.usenix.org/conference/osdi23/presentation/berger): [Triangulating Python Performance Issues with Scalene](https://arxiv.org/pdf/2212.07597). (Note that this paper does not include information about the AI-driven proposed optimizations.)
To cite Scalene in an academic paper, please use the following:
```latex
@inproceedings{288540,
author = {Emery D. Berger and Sam Stern and Juan Altmayer Pizzorno},
title = {Triangulating Python Performance Issues with {S}calene},
booktitle = {{17th USENIX Symposium on Operating Systems Design and Implementation (OSDI 23)}},
year = {2023},
isbn = {978-1-939133-34-2},
address = {Boston, MA},
pages = {51--64},
url = {https://www.usenix.org/conference/osdi23/presentation/berger},
publisher = {USENIX Association},
month = jul
}
```
# Success Stories
If you use Scalene to successfully debug a performance problem, please [add a comment to this issue](https://github.com/plasma-umass/scalene/issues/58)!
# Acknowledgements
Logo created by [Sophia Berger](https://www.linkedin.com/in/sophia-berger/).
This material is based upon work supported by the National Science
Foundation under Grant No. 1955610. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of
the author(s) and do not necessarily reflect the views of the National
Science Foundation.
You can click as many times as you like on the lightning bolt or explosion, and it will generate different suggested optimizations. Your mileage may vary, but in some cases, the suggestions are quite impressive (e.g., order-of-magnitude improvements).
### Quick Start
#### Installing Scalene:
```console
python3 -m pip install -U scalene
```
or
```console
conda install -c conda-forge scalene
```
#### Using Scalene:
After installing Scalene, you can use Scalene at the command line, or as a Visual Studio Code extension.