Wes wrmedford

wrmedford / router.py

Created November 3, 2025 23:29

	# ethos_router.py
	# A compact, self-contained PyTorch router with sinusoidal ID codes,
	# top-b axis selection, and additive beam search.
	#
	# Shapes (conventions used below):
	# B = batch
	# D = model/embedding dim
	# H = heads
	# A = routing axes per head
	# r = axis query dim (must be even for [cos,sin] codes)

wrmedford / ethos.py

Created October 29, 2025 13:38

	# ethos_gen_mlp.py
	# ETHOS (TF32/FP32) for flattened CIFAR-10:
	# Router (pure-ID, custom autograd) → W1/W2 → FiLM → vector gate → dictionaries → mean heads → logits
	from __future__ import annotations
	import math
	from typing import Optional, Tuple

	import torch
	import torch.nn as nn
	import torch.nn.functional as F

wrmedford / gist:0742ec5d965c7a71196ffa91eba891b6

Created October 27, 2025 00:01

B200 Capability Dump

	ASYNC_ENGINE_COUNT: 4
	CAN_MAP_HOST_MEMORY: 1
	CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM: 1
	CLOCK_RATE: 1965000
	COMPUTE_CAPABILITY_MAJOR: 10
	COMPUTE_CAPABILITY_MINOR: 0
	COMPUTE_MODE: DEFAULT
	COMPUTE_PREEMPTION_SUPPORTED: 1
	CONCURRENT_KERNELS: 1
	CONCURRENT_MANAGED_ACCESS: 1

wrmedford / gist:14893b6a4477b6d2ef114a3406d5aa87

Created October 23, 2025 03:18

GPT-2 Medium modded-nanogpt single GPU Friendly

	import os
	import sys
	with open(sys.argv[0]) as f:
	code = f.read() # read the code of this file ASAP, for logging
	import uuid
	import time
	import copy
	from dataclasses import dataclass
	from functools import lru_cache
	from pathlib import Path

wrmedford / gist:d35259774198ca6d125ef18c745f7ba5

Created April 9, 2017 02:07