proger

"""
Randomized Binary Search Trees
https://www.cs.upc.edu/~conrado/research/papers/jacm-mr98.pdf
"""
import math
import random
from collections import Counter


class root:

proger

// uses https://github.com/HazyResearch/ThunderKittens
#include "tk/src/kittens.cuh"
#include "tk/src/common/pyutils/torch_helpers.cuh"

#define NUM_WORKERS 2 // This kernel uses this many workers in parallel per block, to help issue instructions more quickly.
#define DIMENSION 64 // This kernel operates over 64-dimensional vectors
#define DEBUG 0

using namespace kittens; // this kernel only handles headdim=q_reg.cols for simplicity. Also n should be a multiple of 256 here.

proger

"a linear RNN that receives ones as input and gives increasingly better approximations to pi as output"

import numpy as np
import math

def binary(digits: int):
    "Make a basis of powers of two of dimension `digits`, lowest bits first"
    return 1 << np.arange(digits)

def leibniz(n):

proger

#%%

import math
import matplotlib.pyplot as plt
import torch
import torch.nn as nn

plt.rcParams['axes.spines.left'] = False
plt.rcParams['axes.spines.right'] = False
plt.rcParams['axes.spines.top'] = False

proger

# prompt: https://twitter.com/francoisfleuret/status/1783479122418716805

import os
os.environ['TORCH_LOGS'] = 'output_code' # shows all the bmms
import torch
torch.set_float32_matmul_precision('high')

N, T, D, U, C = 3, 128, 5, 32, 32 # batch, time, heads, head_dim, dim
S = T
A = torch.randn(N, T, D, U) / U**0.5

proger

#%%

import torch
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt

X = torch.tensor([[0, 0], [0, 1], [1, 0], [1, 1]]).float()
y = torch.logical_xor(X[:, 0], X[:, 1]).float()

proger

proger

"""
# download the model
huggingface-cli download google/gemma-2b-it

# run the server (set the model name here and in the prompt function below)
# notice --kv_cache_dtype fp8_e5m2
docker run --gpus all -p 8000:8000 -e HF_HOME=/hf -e CUDA_VISIBLE_DEVICES=1 -v ~/.cache/huggingface:/hf vllm/vllm-openai:latest:latest --host 0.0.0.0 --model google/gemma-2b-it --kv-cache-dtype fp8_e5m2

# ask one question
echo 'what is python?' | python -m prompt | jq -r '.choices[].text'

proger

import torch


def binary(digits: int) -> torch.IntTensor:
    "Make a basis of powers of two of dimension `digits`, lowest bits to the right"
    return 1 << torch.arange(digits).flip(-1)


def sbt(x: torch.IntTensor) -> torch.IntTensor:
    """Slow Binary Transform.

proger

import argparse
from itertools import islice
import logging
from pathlib import Path

import evaluate
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
from transformers.generation import BeamSearchDecoderOnlyOutput
from peft import PeftModel