pcuenca

{
    // Use IntelliSense to learn about possible attributes.
    // Hover to view descriptions of existing attributes.
    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [
        {
            // To debug a Rust dylib (tiktoken) from a Pyton script:
            // - Launch the Python script with the Python debugger and stop at a breakpoint.
            // - Get the process id (for example, `os.getpid()`)

pcuenca

# Simple Llama resharding
# Tested on Llama 3.1 70B only
# Shards are released from memory as soon as we're done with each

import torch
from pathlib import Path

checkpoint = Path("original")
output_dir = Path("resharded")
output_dir.mkdir(parents=True, exist_ok=True)