Building awesome things.

Alex Redden aredden

Building awesome things.

Master of none, developer of many things.

aredden / wrap_func_dtype.py

Created September 23, 2023 01:04

pytorch function to perform wrapped casting for functions which require specific dtypes (flash attention 2)

	from functools import wraps
	from typing import Any, Dict, Iterable, Optional, Tuple, TypeVar
	import torch
	T = TypeVar("T", bound=callable)

	ref_map = {
	torch.float64: [torch.float16,torch.float32,torch.bfloat16,torch.half],
	torch.float32: [torch.float16,torch.bfloat16,torch.half],
	torch.float16: [],
	torch.bfloat16: [],

aredden / image_rotate_into_gif.py

Last active November 22, 2023 04:59

Rotate an image and save as a gif using pytorch

	import numpy as np
	import torch
	from torchvision.transforms import functional as F
	from PIL import Image

	image_path = "./117.webp"
	num_images = 32
	total_rot = 360
	rot_step = total_rot / num_images

aredden / cuda_python_cudart_types.py

Created March 12, 2024 22:07

NVIDIA cuda-python's cudart typings for their untyped cpp bound library.

aredden / fp8_linear_quantile.py

Created August 24, 2024 15:57

quantize fp8 using quantile vs absmax for very ood values

	import torch

	torch.set_printoptions(precision=4, sci_mode=False)
	import triton
	import triton.language as tl
	from torch import Tensor


	def quanitze_fp8_tensorwise(x: torch.Tensor, dtype=torch.float8_e4m3fn):
	scale = x.abs().max() / torch.finfo(dtype).max