robbiemu

side quest: HPO and batch size

To follow along you'll need to install optuna.

Before we do real hpo let's just look for an efficient batch size for the current machine:

batch_size: this is determined to be the maximum power of 2 (for no particular reason for now) that shows improved samples/second processing.

we need a max_length for this because of how batches are handled when training. The training process automatically pads the data for you on-the-fly for every single batch.

robbiemu

import os
import time
import asyncio
import httpx
from typing import Any, Dict, Optional
from contextlib_cli import anext_generator
from collections import deque

# Import all the necessary wrappers from the LangChain ecosystem
from langchain_community.utilities import (

robbiemu

{{- /* Extract system message and other messages */ -}}
{{- $system_message := "" -}}
{{- $loop_messages := .Messages -}}
{{- if and .Messages (gt (len .Messages) 0) (eq (index .Messages 0).Role "system") -}}
  {{- $system_message = (index .Messages 0).Content -}}
  {{- $loop_messages = slice .Messages 1 -}}
{{- end -}}

{{- /* Handle tools if they exist */ -}}
{{- $has_tools := and .Tools (gt (len .Tools) 0) -}}

robbiemu

Modified

ROCm Migration Report for Hierarchical Reasoning Model (HRM) on AMD MI300X GPUs

This report provides a definitive, actionable, and unambiguous guide for migrating the Hierarchical Reasoning Model (HRM) to ROCm, specifically targeting AMD MI300X GPUs. All previous uncertainties and 'if' statements have been resolved to provide clear instructions for developers.

1. README.md

Current CUDA Dependencies:

The README.md explicitly outlines the installation of CUDA and PyTorch with CUDA support, along with FlashAttention, which is a CUDA-dependent library.

robbiemu

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

"""
pr-capture: A CLI tool to capture GitHub PR data into a comprehensive markdown file.
"""

import argparse
from datetime import datetime
import json

robbiemu

#!/bin/bash
# zip_project.sh - Zip up source and documentation files from a Git repo,
# excluding .git, .venv, and anything in .gitignore. Use --help for options.
#
# Usage:
#   ./zip_project.sh [--output FILE] [--verbose] [--filter [REGEX]] [--exclude REGEX] [--help]
#
# Options:
#   -o, --output FILE       Name of the output zip file (default: project_bundle.zip)
#   -v, --verbose           Print verbose output

robbiemu

import argparse
import os
import re
from collections import defaultdict
from typing import List, Set, Tuple

import requests
from packaging.version import Version, InvalidVersion

# ---------------------------------------------------------------------------

robbiemu

#!/usr/bin/env bash
# Generates 10-slot wallpapper solar JSONs from 6-stage image sets

set -euo pipefail

DIR="/Users/Shared/Wallpapers/WaterWays_Processed"
cd "$DIR" || { echo "bad path: $DIR"; exit 1; }

# 10-stage full-day solar path: altitude (°) and azimuth (°)
ALT=(60 40 20 5 -5 -15 -5 5 20 40)      # noon to night and back

robbiemu

import os
import json
import requests
from typing import Dict, List, Any, Optional
from smolagents.tools import Tool

class BraveSearchTool(Tool):
    """Tool for interacting with the Brave Search API within the smolagents framework."""

    name = "brave_search"

robbiemu

git clone https://huggingface.co/sail/Sailor2-1B-Chat
Cloning into 'Sailor2-1B-Chat'...
remote: Enumerating objects: 39, done.
remote: Counting objects: 100% (36/36), done.
remote: Compressing objects: 100% (36/36), done.
remote: Total 39 (delta 14), reused 0 (delta 0), pack-reused 3 (from 1)
Unpacking objects: 100% (39/39), 2.02 MiB | 2.26 MiB/s, done.
Filtering content: 100% (2/2), 1.85 GiB | 14.92 MiB/s, done.

./convert_hf_to_gguf.py --outfile $HF/Sailor2-1B-Chat_bf16.gguf --outtype bf16 $HF/Sailor2-1B-Chat