Unsloth ガイド: AMD GPU での LLM ファインチューニング

curl -fsSL https://unsloth.ai/install.sh | sh

apt update && apt install python3.10-venv python3.11-venv python3.12-venv python3.13-venv -y

python3 -m venv unsloth_env
source unsloth_env/bin/activate
pip install uv

uv pip install "torch>=2.4,<2.11.0" "torchvision<0.26.0" "torchaudio<2.11.0" \
    --index-url https://download.pytorch.org/whl/rocm7.1 --upgrade --force-reinstall

ROCM_TAG="$({ command -v amd-smi >/dev/null 2>&1 && amd-smi version 2>/dev/null | awk -F'ROCm version: ' 'NF>1{split($2,a,"."); print "rocm"a[1]"."a[2]; ok=1; exit} END{exit !ok}'; } || { [ -r /opt/rocm/.info/version ] && awk -F. '{print "rocm"$1"."$2; exit}' /opt/rocm/.info/version; } || { command -v hipconfig >/dev/null 2>&1 && hipconfig --version 2>/dev/null | awk -F': *' '/HIP version/{split($2,a,"."); print "rocm"a[1]"."a[2]; ok=1; exit} END{exit !ok}'; } || { command -v dpkg-query >/dev/null 2>&1 && ver="$(dpkg-query -W -f="${Version}\n" rocm-core 2>/dev/null)" && [ -n "$ver" ] && awk -F'[.-]' '{print "rocm"$1"."$2; exit}' <<<"$ver"; } || { command -v rpm >/dev/null 2>&1 && ver="$(rpm -q --qf '%{VERSION}\n' rocm-core 2>/dev/null)" && [ -n "$ver" ] && awk -F'[.-]' '{print "rocm"$1"."$2; exit}' <<<"$ver"; })"; [ -n "$ROCM_TAG" ] && uv pip install "torch>=2.4,<2.11.0" "torchvision<0.26.0" "torchaudio<2.11.0" --index-url "https://download.pytorch.org/whl/$ROCM_TAG" --upgrade --force-reinstall

uv pip install unsloth[amd]

# x86_64 システム:
pip install --force-reinstall --no-cache-dir --no-deps \
    "https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_main/bitsandbytes-1.33.7.preview-py3-none-manylinux_2_24_x86_64.whl"

# aarch64 システム: 上の URL の x86_64 を aarch64 に置き換えてください

# URL にアクセスできない場合のフォールバック:
# pip install --force-reinstall --no-cache-dir --no-deps "bitsandbytes>=0.49.1"

export HSA_OVERRIDE_GFX_VERSION=9.4.2  # AMD MI300X に必須
export HF_HUB_DISABLE_XET=1            # AMD 上での HuggingFace ダウンロード問題を修正

from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    model_name = "unsloth/gemma-4-26b-a4b-it",
    max_seq_length = 2048,
    load_in_4bit = True,
)

model = FastModel.get_peft_model(
    model,
    r = 16,
    lora_alpha = 16,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj"],
)

from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    formatting_func = formatting_func,
    args = SFTConfig(
        per_device_train_batch_size = 1,
        gradient_accumulation_steps = 4,
        max_steps = 60,
        output_dir = "outputs",
        report_to = "none",
    ),
)

trainer_stats = trainer.train()