Aditya Venkataraman aditvenk

## softmax_mlx_torch.py
"""Softmax microbenchmark comparing Torch and MLX on Apple Silicon.

The script measures six variants of softmax (manual/native implementations in
eager and compiled form) on identical all-ones inputs. Torch executes on MPS
while MLX runs on the Apple GPU. Each variant is verified once against the
``torch.softmax`` reference before timing.

Optionally set ``MTL_CAPTURE_ENABLED=1`` to record MLX Metal captures
while the correctness checks run. Traces are written to the current
directory.
	"""Softmax microbenchmark comparing Torch and MLX on Apple Silicon.

	The script measures six variants of softmax (manual/native implementations in
	eager and compiled form) on identical all-ones inputs. Torch executes on MPS
	while MLX runs on the Apple GPU. Each variant is verified once against the
	``torch.softmax`` reference before timing.

	Optionally set ``MTL_CAPTURE_ENABLED=1`` to record MLX Metal captures
	while the correctness checks run. Traces are written to the current
	directory.