From 1710ecb990122a863de893732d193fb714a1c98d Mon Sep 17 00:00:00 2001 From: hxsz1997 <45651968+hxsz1997@users.noreply.github.com> Date: Wed, 7 Feb 2024 16:29:57 +0800 Subject: [PATCH] Add llm-ppl workflow (#10074) * add llm-ppl workflow * update the DATASET_DIR * test multiple precisions * modify nightly test * match the updated ppl code * add matrix.include * fix the include error * update the include * add more model * update the precision of include * update nightly time and add more models * fix the workflow_dispatch description, change default model of pr and modify the env --- .github/workflows/llm-ppl-evaluation.yml | 205 +++++++++++++++++++++++ 1 file changed, 205 insertions(+) create mode 100644 .github/workflows/llm-ppl-evaluation.yml diff --git a/.github/workflows/llm-ppl-evaluation.yml b/.github/workflows/llm-ppl-evaluation.yml new file mode 100644 index 00000000..06b89aa5 --- /dev/null +++ b/.github/workflows/llm-ppl-evaluation.yml @@ -0,0 +1,205 @@ +name: LLM Perplexity Evalution + +# Cancel previous runs in the PR when you push new commits +concurrency: + group: ${{ github.workflow }}-llm-nightly-test-${{ github.event.pull_request.number || github.run_id }} + cancel-in-progress: true + +# Controls when the action will run. +on: + schedule: + - cron: "00 14 * * *" # GMT time, 14:00 GMT == 22:00 China + pull_request: + branches: [main] + paths: + - ".github/workflows/llm-ppl-evaluation.yml" + # Allows you to run this workflow manually from the Actions tab + workflow_dispatch: + inputs: + seq_len: + description: 'sequence length, input one number' + required: true + type: number + model_name: + description: 'Model names, seperated by comma and must be quoted.' + required: true + type: string + precision: + description: 'Precisions, seperated by comma and must be quoted.' + required: true + type: string + language: + description: 'language, can be en, zh, or all' + required: true + type: choice + options: + - en + - zh + - all + runs-on: + description: 'Labels to filter the runners, seperated by comma and must be quoted.' + default: "accuracy" + required: false + type: string + + +# A workflow run is made up of one or more jobs that can run sequentially or in parallel +jobs: + llm-cpp-build: + uses: ./.github/workflows/llm-binary-build.yml + set-matrix: + runs-on: ubuntu-latest + outputs: + seq_len: ${{ steps.set-matrix.outputs.seq_len }} + model_name: ${{ steps.set-matrix.outputs.model_name }} + precision: ${{ steps.set-matrix.outputs.precision }} + language: ${{ steps.set-matrix.outputs.language }} + runner: ${{ steps.set-matrix.outputs.runner }} + steps: + - name: set-nightly-env + if: ${{github.event_name == 'schedule'}} + env: + NIGHTLY_MATRIX_SEQ_LEN: '["512"]' + NIGHTLY_MATRIX_MODEL_NAME: '["Llama-2-7b-chat-hf", "mpt-7b-chat", + "falcon-7b-instruct-with-patch", "Mistral-7B-v0.1"]' + NIGHTLY_MATRIX_LANGUAGE: '["en"]' + NIGHTLY_MATRIX_PRECISION: '["sym_int4 fp8"]' + NIGHTLY_LABELS: '["self-hosted", "llm", "accuracy-nightly"]' + run: | + echo "seq_len=$NIGHTLY_MATRIX_SEQ_LEN" >> $GITHUB_ENV + echo "model_name=$NIGHTLY_MATRIX_MODEL_NAME" >> $GITHUB_ENV + echo "precision=$NIGHTLY_MATRIX_PRECISION" >> $GITHUB_ENV + echo "language=$NIGHTLY_MATRIX_LANGUAGE" >> $GITHUB_ENV + echo "runner=$NIGHTLY_LABELS" >> $GITHUB_ENV + + - name: set-pr-env + if: ${{github.event_name == 'pull_request'}} + env: + PR_MATRIX_SEQ_LEN: '["512"]' + PR_MATRIX_MODEL_NAME: '["Llama-2-7b-chat-hf"]' + PR_MATRIX_LANGUAGE: '["en"]' + PR_MATRIX_PRECISION: '["sym_int4"]' + PR_LABELS: '["self-hosted", "llm", "temp-arc01"]' + run: | + echo "seq_len=$PR_MATRIX_SEQ_LEN" >> $GITHUB_ENV + echo "model_name=$PR_MATRIX_MODEL_NAME" >> $GITHUB_ENV + echo "precision=$PR_MATRIX_PRECISION" >> $GITHUB_ENV + echo "language=$PR_MATRIX_LANGUAGE" >> $GITHUB_ENV + echo "runner=$PR_LABELS" >> $GITHUB_ENV + - name: set-manual-env + if: ${{github.event_name == 'workflow_dispatch'}} + env: + MANUAL_MATRIX_SEQ_LEN: ${{format('[ {0} ]', inputs.seq_len)}} + MANUAL_MATRIX_MODEL_NAME: ${{format('[ {0} ]', inputs.model_name)}} + MANUAL_MATRIX_LANGUAGE: ${{format('[ {0} ]', inputs.language)}} + MANUAL_MATRIX_PRECISION: ${{format('[ {0} ]', inputs.precision)}} + MANUAL_LABELS: ${{format('["self-hosted", "llm", {0}]', inputs.runs-on)}} + run: | + echo "seq_len=$MANUAL_MATRIX_SEQ_LEN" >> $GITHUB_ENV + echo "model_name=$MANUAL_MATRIX_MODEL_NAME" >> $GITHUB_ENV + echo "precision=$MANUAL_MATRIX_PRECISION" >> $GITHUB_ENV + echo "language=$MANUAL_MATRIX_LANGUAGE" >> $GITHUB_ENV + echo "runner=$MANUAL_LABELS" >> $GITHUB_ENV + - name: set-matrix + id: set-matrix + run: | + echo "seq_len=$seq_len" >> $GITHUB_OUTPUT + echo "model_name=$model_name" >> $GITHUB_OUTPUT + echo "precision=$precision" >> $GITHUB_OUTPUT + echo "language=$language" >> $GITHUB_OUTPUT + echo "runner=$runner" >> $GITHUB_OUTPUT + llm-ppl-evalution: + timeout-minutes: 1000 + needs: [llm-cpp-build, set-matrix] + strategy: + fail-fast: false + matrix: + # include: + # python-version: "3.9" + # model_name: "stablelm-3b-4e1t" + # task: "arc" + # precision: "sym_int4" #options: sym_int4, fp4, mixed_fp4, sym_int8, fp8, mixed_fp8 + python-version: ["3.9"] + model_name: ${{ fromJson(needs.set-matrix.outputs.model_name) }} + language: ${{ fromJson(needs.set-matrix.outputs.language) }} + precision: ${{ fromJson(needs.set-matrix.outputs.precision) }} + seq_len: ${{ fromJson(needs.set-matrix.outputs.seq_len) }} + device: [xpu] + include: + - python-version: "3.9" + model_name: "chatglm2-6b" + language: "zh" + precision: "sym_int4 fp8" + seq_len: "512" + device: "xpu" + - python-version: "3.9" + model_name: "chatglm3-6b" + language: "zh" + precision: "sym_int4 fp8" + seq_len: "512" + device: "xpu" + - python-version: "3.9" + model_name: "Baichuan2-7B-Chat" + language: "zh" + precision: "sym_int4 fp8" + seq_len: "512" + device: "xpu" + + runs-on: ${{ fromJson(needs.set-matrix.outputs.runner) }} + env: + ANALYTICS_ZOO_ROOT: ${{ github.workspace }} + ORIGIN_DIR: /mnt/disk1/models + DATASET_DIR: /mnt/disk1/datasets/THUDM___long_bench/ + steps: + - uses: actions/checkout@v3 + - name: Set up Python ${{ matrix.python-version }} + uses: actions/setup-python@v4 + with: + python-version: ${{ matrix.python-version }} + - name: Install dependencies + shell: bash + run: | + python -m pip install --upgrade pip + python -m pip install --upgrade setuptools==58.0.4 + python -m pip install --upgrade wheel + + - name: Download llm binary + uses: ./.github/actions/llm/download-llm-binary + + - name: Run LLM install (all) test + uses: ./.github/actions/llm/setup-llm-env + with: + extra-dependency: "xpu_2.1" + + - name: Download models + shell: bash + run: | + echo "MODEL_PATH=${ORIGIN_DIR}/${{ matrix.model_name }}/" >> "$GITHUB_ENV" + MODEL_PATH=${ORIGIN_DIR}/${{ matrix.model_name }}/ + wget -r -nH -nc --no-verbose --cut-dirs=1 ${LLM_FTP_URL}/llm/${{ matrix.model_name }} -P ${ORIGIN_DIR} + + - name: Upgrade packages + shell: bash + run: | + pip install --upgrade datasets==2.14.6 + if [ "${{ matrix.model_name }}" = "Mistral-7B-v0.1" ]; then + pip install --upgrade transformers==4.36 + else + pip install --upgrade transformers==4.31 + fi + + - name: Run perplexity + shell: bash + working-directory: ${{ github.workspace }}/python/llm/dev/benchmark/perplexity + env: + USE_XETLA: OFF + SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS: 1 + run: | + source /opt/intel/oneapi/setvars.sh + python run.py \ + --seq_len ${{ matrix.seq_len }} \ + --model_path ${MODEL_PATH} \ + --precisions ${{ matrix.precision }} \ + --device ${{ matrix.device }} \ + --dataset_path ${DATASET_DIR} \ + --language ${{ matrix.language }}