HDCharles

#!/bin/bash

# === Configuration ===
SSH_CONFIG="$HOME/.ssh/config"
STATE_FILE="$HOME/.gpucheck_state.json"
LAST_SYNC_FILE="$HOME/.gpucheck_last_sync"
INVENTORY_REPO="neuralmagic/stratus"
INVENTORY_PATH="infra-ansible/webapp-inventory/public/data/inventory.json"
POLL_INTERVAL=300  # seconds between checks (5 minutes)
# Three EMA alpha values for different time horizons

HDCharles

NVFP4 Evals on B200

model                          scheme     technique    task                         main             PR       change
--------------------------------------------------------------------------------------------------------------------
Meta-Llama-3-8B-Instruct       NVFP4      awq_rtn      gsm8k_platinum               71.46%           69.89%       -2.20%
Qwen2.5-3B-Instruct            NVFP4      awq_rtn      gsm8k_platinum               23.33%           29.61%      +26.92% volatile (Qwen2.5-3B seems to be all over the place for gsm8k_plat)
Qwen3-30B-A3B                  NVFP4      awq_rtn      gsm8k_platinum               92.31%           91.32%       -1.07%

Meta-Llama-3-8B-Instruct       NVFP4      awq_rtn      mmlu                       63.55%**           63.46%       -0.14%
Qwen2.5-3B-Instruct            NVFP4      awq_rtn      mmlu                       63.12%**         63.34%**       +0.35%

HDCharles

#!/bin/bash
# Parallel Regression Test Script
# 1. Quantizes remaining models (if needed)
# 2. Runs evaluations in parallel (4 at a time)
# 3. Saves individual logs for each eval job
# 4. Prints summaries as jobs complete

set -o pipefail

# ── Configuration ────────────────────────────────────────────────────────────

HDCharles

Model Transfer Workflow

Scripts to upload quantized models to HuggingFace (RedHatAI), download them elsewhere, and clean up.

Prerequisites

pip install huggingface_hub[cli]
huggingface-cli login  # Login with token that has write access to RedHatAI org

HDCharles

#!/bin/bash

# === Configuration ===
SSH_CONFIG="$HOME/.ssh/config"
STATE_FILE="$HOME/.gpucheck_state.json"
LAST_SYNC_FILE="$HOME/.gpucheck_last_sync"
HOSTS_URL="https://vigilant-happiness-o78jgy9.pages.github.io/hosts"
SSH_USER="HDCharles"
POLL_INTERVAL=300  # seconds between checks (5 minutes)
# Three EMA alpha values for different time horizons

HDCharles

import argparse
import os
import time

import torch
from compressed_tensors.offload import dispatch_model
from compressed_tensors.quantization import preset_name_to_scheme
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

HDCharles

import argparse
import time

import torch
from compressed_tensors.offload import dispatch_model
from compressed_tensors.quantization import preset_name_to_scheme
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

from llmcompressor import oneshot

HDCharles

#!/usr/bin/env python3
"""Extract summary data from AWQ DDP regression test log files.

Parses the log output from run_all_tests.sh and produces a comparison table
showing pre-DDP vs post-DDP results across models, schemes, and benchmarks.

Usage:
    python extract_log_summary.py regression_results.log
"""

HDCharles

import time, os

model_path = "/mnt/data/engine/HDCharles/hf_hub/models--meta-llama--Meta-Llama-3-8B-Instruct/snapshots/8afb486c1db24fe5011ec46dfbe5b5dccdb575c2"

from vllm import LLM, SamplingParams

print("Loading model...")
t0 = time.time()
llm = LLM(model=model_path, gpu_memory_utilization=0.5, max_logprobs=-1)
tokenizer = llm.get_tokenizer()

HDCharles

#!/usr/bin/env python3
"""Extract summary data from AWQ DDP regression test log files.

Parses the log output from run_all_tests.sh and produces a comparison table
showing pre-DDP vs post-DDP results across models, schemes, and benchmarks.

Usage:
    python extract_log_summary.py regression_results.log
"""