char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Fathi Boudracommon: migrate the deprecated --mmap/--no-mmap to --load-mode (#26934)0177dcc73

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
1.1 KiB31 linesraw
1#!/usr/bin/env bash
2
3#  MIT license
4#  Copyright (C) 2024 Intel Corporation
5#  SPDX-License-Identifier: MIT
6export ONEAPI_DEVICE_SELECTOR="level_zero:0"
7source /opt/intel/oneapi/setvars.sh
8
9#export GGML_SYCL_DEBUG=1
10
11#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
12
13INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
14MODEL_FILE=models/llama-2-7b.Q4_0.gguf
15NGL=99
16CONTEXT=4096
17
18#support malloc device memory more than 4GB.
19export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
20
21LOAD_MODE='--load-mode auto'
22if [ $# -gt 0 ]; then
23    GGML_SYCL_DEVICE=$1
24    echo "use $GGML_SYCL_DEVICE as main GPU"
25    #use signle GPU only
26    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none ${LOAD_MODE}
27
28else
29    #use multiple GPUs with same max compute units
30    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s 0 -c ${CONTEXT} ${LOAD_MODE}
31fi