shendsaliaga

#!/bin/bash

# Usage: export TRACE=true && ./convert_arw_to_jpg.sh --input INPUT_DIR --output OUTPUT_DIR
# Enable debugging with TRACE
[[ "$TRACE" ]] && set -x

show_usage() {
  printf "Usage: $0 --input <input_directory> --output <output_directory>\n"
}

shendsaliaga

(venv) # Exit:0 2023-03-12 16:59:27  [r2q2@Reformer#[:~/opt/llama.cpp]
$(: !605 ) ./main -m ./models/65B/ggml-model-q4_0.bin -t 8 -n 128
main: seed = 1678658429
llama_model_load: loading model from './models/65B/ggml-model-q4_0.bin' - please wait ...
llama_model_load: n_vocab = 32000
llama_model_load: n_ctx   = 512
llama_model_load: n_embd  = 8192
llama_model_load: n_mult  = 256
llama_model_load: n_head  = 64
llama_model_load: n_layer = 80

shendsaliaga

We're tripling down on acronyms, by using the VFIO driver, with IOMMU and SR-IOV.

Boot parameters

intel_iommu=on iommu=pt default_hugepagesz=1G hugepagesz=1G hugepages=16

Enable hugepages

echo 4096 | sudo tee /sys/devices/system/node/node0/hugepages/hugepages-2048kB/nr_hugepages
mkdir -p /tmp/mnt/huge
mount -t hugetlbfs nodev /tmp/mnt/huge