karpathy

LLM Wiki

A pattern for building personal knowledge bases using LLMs.

This is an idea file, it is designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.). Its goal is to communicate the high level idea, but your agent will build out the specifics in collaboration with you.

The core idea

Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.

mcarilli

# This isn't supposed to run as a bash script, i named it with ".sh" for syntax highlighting.

# https://developer.nvidia.com/nsight-systems
# https://docs.nvidia.com/nsight-systems/profiling/index.html

# My preferred nsys (command line executable used to create profiles) commands
#
# In your script, write
# torch.cuda.nvtx.range_push("region name")
# ...

telegraphic

import pycuda.driver as cuda
import pycuda.autoinit
from pycuda.compiler import SourceModule
import numpy as np

def compute_xcorr_cpu(d):
    dc = d.astype('float32').view('complex64')
    dc = dc.transpose((0,2,3,1)).copy()
    xcorr_cpu = np.einsum('...i,...j', dc,  np.conj(dc)).view('float32').astype('int32').sum(axis=-4)
    return xcorr_cpu

pbamotra

def logsigsoftmax(logits):
    """
    Computes sigsoftmax from the paper - https://arxiv.org/pdf/1805.10829.pdf
    """
    max_values = torch.max(logits, 1, keepdim = True)[0]
    exp_logits_sigmoided = torch.exp(logits - max_values) * torch.sigmoid(logits)
    sum_exp_logits_sigmoided = exp_logits_sigmoided.sum(1, keepdim = True)
    log_probs = logits - max_values + torch.log(torch.sigmoid(logits)) - torch.log(sum_exp_logits_sigmoided)
    return log_probs

Stonesjtu

import gc

import torch

## MEM utils ##
def mem_report():
    '''Report the memory usage of the tensor.storage in pytorch
    Both on CPUs and GPUs are reported'''

    def _mem_report(tensors, mem_type):

wangg12

from graphviz import Digraph
from torch.autograd import Variable
import torch


def make_dot(var, params=None):
    if params is not None:
        assert isinstance(params.values()[0], Variable)
        param_map = {id(v): k for k, v in params.items()}

understeer

Latency Comparison Numbers
--------------------------
L1 cache reference/hit                       1.5 ns                      4 cycles
Floating-point add/mult/FMA operation        1.5 ns                      4 cycles
L2 cache reference/hit                       5   ns                      12 ~ 17 cycles
Branch mispredict                            6   ns                      15 ~ 20 cycles
L3 cache hit (unshared cache line)          16   ns                      42 cycles
L3 cache hit (shared line in another core)  25   ns                      65 cycles
Mutex lock/unlock                           25   ns
L3 cache hit (modified in another core)     29   ns                      75 cycles

Starefossen

Sessions

New Session

• tmux new [-s name] [cmd] (:new) - new session

Switch Session

• tmux ls (:ls) - list sessions
• tmux switch [-t name] (:switch) - switches to an existing session