Skip to content

patterns.md

Latest commit

 

History

History
187 lines (128 loc) · 3.24 KB

patterns.md

File metadata and controls

187 lines (128 loc) · 3.24 KB

Target Pattern Reference

Speed Bump uses glob patterns to specify which functions to slow down. Each line in a targets file has the format:

module_pattern:qualified_name_pattern

Both patterns use standard glob syntax (*, ?, [...]).

Pattern Components

Module Pattern

The module pattern matches against the file path of the code object. This is typically the full path to the .py file.

Pattern Matches
* Any module
*/numpy/* Any file in a numpy directory
*transformers* Any file with "transformers" in the path
/home/user/myproject/*.py Specific directory

Qualified Name Pattern

The qualified name pattern matches against the function's co_qualname, which includes:

  • The function name
  • Enclosing class names (for methods)
  • Enclosing function names (for nested functions)
Code Qualified Name
def foo(): ... foo
class Bar: def baz(self): ... Bar.baz
def outer(): def inner(): ... outer.<locals>.inner
class A: class B: def m(self): ... A.B.m

Pattern Examples

Match Everything

*:*

Useful for Step 1 analysis: "does Python matter at all?"

Match a Specific Function

*/vllm/worker/model_runner.py:ModelRunner.execute_model

Match All Methods of a Class

*:ModelRunner.*

Matches ModelRunner.execute_model, ModelRunner.__init__, etc.

Match Anything Containing a Substring

*:*attention*

Matches any function with "attention" in its qualified name (case-sensitive).

Match a Module

*/numpy/*:*

Matches all functions in numpy.

Match Nested Functions

Because nested functions include <locals> in their qualified name:

*:*<locals>*

Matches all nested/local functions.

Or to match a specific nested function:

*:*outer.<locals>.inner

Match Test Functions

*/tests/*:test_*

Exclude Patterns

There's no exclude syntax. If you need to exclude, create a target file with only what you want to include.

Target File Format

# Comments start with #
# Blank lines are ignored

# One pattern per line
*:ModelRunner.*
*/transformers/*:*Attention*

   # Whitespace is stripped
   *:some_function

Debugging Patterns

If your pattern isn't matching as expected:

  1. Check the qualified name by adding a print in your code:

    import sys
print(my_function.__code__.co_qualname)

  2. Check the file path:

    print(my_function.__code__.co_filename)

  3. Use a broad pattern first, then narrow down:

    *:*  # matches everything - verify setup works
*:*MyClass*  # narrow to class
*:MyClass.my_method  # narrow to method

Common Patterns by Framework

vLLM

# All worker operations
*/vllm/worker/*:*

# Model execution
*:ModelRunner.execute_model

# Attention layers
*:*Attention*

Transformers

# All modeling code
*/transformers/models/*:*

# Specific architecture
*/transformers/models/llama/*:*

# Attention mechanisms
*:*Attention.forward

PyTorch

# Tensor operations (Python side)
*/torch/*:*

# Specific modules
*:Linear.forward
*:Conv2d.forward

NumPy

# All numpy
*/numpy/*:*

# Core operations
*/numpy/core/*:*