[ CI/Build ] LM Eval Harness Based CI Testing (#5838)

Co-authored-by: Robert Shaw <rshaw@neuralmagic>

.buildkite/lm-eval-harness/configs/Meta-Llama-3-70B-Instruct.yaml

+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m meta-llama/Meta-Llama-3-70B-Instruct -b 32 -l 250 -f 5
+model_name: "meta-llama/Meta-Llama-3-70B-Instruct"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.892
+  - name: "exact_match,flexible-extract"
+    value: 0.892
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-FP8.yaml

+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m neuralmagic/Meta-Llama-3-8B-Instruct-FP8 -b 32 -l 250 -f 5 -t 1
+model_name: "neuralmagic/Meta-Llama-3-8B-Instruct-FP8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.756
+  - name: "exact_match,flexible-extract"
+    value: 0.752
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct.yaml

+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m meta-llama/Meta-Llama-3-8B-Instruct -b 32 -l 250 -f 5 -t 1
+model_name: "meta-llama/Meta-Llama-3-8B-Instruct"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.756
+  - name: "exact_match,flexible-extract"
+    value: 0.752
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Mixtral-8x7B-Instruct-v0.1.yaml

+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m neuralmagic/Mixtral-8x7B-Instruct-v0.1 -b 32 -l 250 -f 5 -t 4
+model_name: "mistralai/Mixtral-8x7B-Instruct-v0.1"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.616
+  - name: "exact_match,flexible-extract"
+    value: 0.632
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/models-large.txt

+Meta-Llama-3-70B-Instruct.yaml
+Mixtral-8x7B-Instruct-v0.1.yaml

.buildkite/lm-eval-harness/configs/models-small.txt

+Meta-Llama-3-8B-Instruct.yaml
+Meta-Llama-3-8B-Instruct-FP8.yaml

.buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh

+#!/bin/bash
+# We can use this script to compute baseline accuracy on GSM for transformers.
+#
+# Make sure you have lm-eval-harness installed:
+#   pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@9516087b81a61d0e220b22cc1b75be76de23bc10
+
+usage() {
+    echo``
+    echo "Runs lm eval harness on GSM8k using huggingface transformers."
+    echo "This pathway is intended to be used to create baselines for "
+    echo "our automated nm-test-accuracy workflow"
+    echo
+    echo "usage: ${0} <options>"
+    echo
+    echo "  -m    - huggingface stub or local directory of the model"
+    echo "  -b    - batch size to run the evaluation at"
+    echo "  -l    - limit number of samples to run"
+    echo "  -f    - number of fewshot samples to use"
+    echo
+}
+
+while getopts "m:b:l:f:" OPT; do
+  case ${OPT} in
+    m ) 
+        MODEL="$OPTARG"
+        ;;
+    b ) 
+        BATCH_SIZE="$OPTARG"
+        ;;
+    l ) 
+        LIMIT="$OPTARG"
+        ;;
+    f ) 
+        FEWSHOT="$OPTARG"
+        ;;
+    \? ) 
+        usage
+        exit 1
+        ;;
+  esac
+done
+
+lm_eval --model hf \
+  --model_args pretrained=$MODEL,parallelize=True \
+  --tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
+  --batch_size $BATCH_SIZE

.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh

+#!/bin/bash
+# We can use this script to compute baseline accuracy on GSM for vllm.
+# We use this for fp8, which HF does not support.
+#
+# Make sure you have lm-eval-harness installed:
+#   pip install lm-eval==0.4.2
+
+usage() {
+    echo``
+    echo "Runs lm eval harness on GSM8k using huggingface transformers."
+    echo "This pathway is intended to be used to create baselines for "
+    echo "our automated nm-test-accuracy workflow"
+    echo
+    echo "usage: ${0} <options>"
+    echo
+    echo "  -m    - huggingface stub or local directory of the model"
+    echo "  -b    - batch size to run the evaluation at"
+    echo "  -l    - limit number of samples to run"
+    echo "  -f    - number of fewshot samples to use"
+    echo "  -t    - tensor parallel size to run at"
+    echo
+}
+
+while getopts "m:b:l:f:t:" OPT; do
+  case ${OPT} in
+    m ) 
+        MODEL="$OPTARG"
+        ;;
+    b ) 
+        BATCH_SIZE="$OPTARG"
+        ;;
+    l ) 
+        LIMIT="$OPTARG"
+        ;;
+    f ) 
+        FEWSHOT="$OPTARG"
+        ;;
+    t )
+        TP_SIZE="$OPTARG"
+        ;;
+    \? ) 
+        usage
+        exit 1
+        ;;
+  esac
+done
+
+lm_eval --model vllm \
+  --model_args pretrained=$MODEL,tensor_parallel_size=$TP_SIZE \
+  --tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
+  --batch_size $BATCH_SIZE

.buildkite/lm-eval-harness/run-tests.sh

+#!/bin/bash
+
+usage() {
+    echo``
+    echo "Runs lm eval harness on GSM8k using vllm and compares to "
+    echo "precomputed baseline (measured by HF transformers.)"
+    echo
+    echo "usage: ${0} <options>"
+    echo
+    echo "  -c    - path to the test data config (e.g. configs/small-models.txt)"
+    echo "  -t    - tensor parallel size"
+    echo
+}
+
+SUCCESS=0
+
+while getopts "c:t:" OPT; do
+  case ${OPT} in
+    c ) 
+        CONFIG="$OPTARG"
+        ;;
+    t )
+        TP_SIZE="$OPTARG"
+        ;;
+    \? )
+        usage
+        exit 1
+        ;;
+  esac
+done
+
+# Parse list of configs.
+IFS=$'\n' read -d '' -r -a MODEL_CONFIGS < $CONFIG
+
+for MODEL_CONFIG in "${MODEL_CONFIGS[@]}"
+do
+    LOCAL_SUCCESS=0
+    
+    echo "=== RUNNING MODEL: $MODEL_CONFIG WITH TP SIZE: $TP_SIZE==="
+
+    export LM_EVAL_TEST_DATA_FILE=$PWD/configs/${MODEL_CONFIG}
+    export LM_EVAL_TP_SIZE=$TP_SIZE
+    pytest -s test_lm_eval_correctness.py || LOCAL_SUCCESS=$?
+
+    if [[ $LOCAL_SUCCESS == 0 ]]; then
+        echo "=== PASSED MODEL: ${MODEL_CONFIG} ==="
+    else
+        echo "=== FAILED MODEL: ${MODEL_CONFIG} ==="
+    fi
+
+    SUCCESS=$((SUCCESS + LOCAL_SUCCESS))
+
+done
+
+if [ "${SUCCESS}" -eq "0" ]; then
+    exit 0
+else
+    exit 1
+fi

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

+"""
+LM eval harness on model to compare vs HF baseline computed offline.
+Configs are found in configs/$MODEL.yaml
+
+* export LM_EVAL_TEST_DATA_FILE=configs/Meta-Llama-3-70B-Instruct.yaml
+* export LM_EVAL_TP_SIZE=4 
+* pytest -s test_lm_eval_correctness.py
+"""
+
+import os
+from pathlib import Path
+
+import lm_eval
+import numpy
+import yaml
+
+RTOL = 0.02
+TEST_DATA_FILE = os.environ.get(
+    "LM_EVAL_TEST_DATA_FILE",
+    ".buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct.yaml")
+
+TP_SIZE = os.environ.get("LM_EVAL_TP_SIZE", 1)
+
+
+def launch_lm_eval(eval_config):
+    model_args = f"pretrained={eval_config['model_name']}," \
+                 f"tensor_parallel_size={TP_SIZE}"
+
+    results = lm_eval.simple_evaluate(
+        model="vllm",
+        model_args=model_args,
+        tasks=[task["name"] for task in eval_config["tasks"]],
+        num_fewshot=eval_config["num_fewshot"],
+        limit=eval_config["limit"],
+        batch_size="auto")
+
+    return results
+
+
+def test_lm_eval_correctness():
+    eval_config = yaml.safe_load(
+        Path(TEST_DATA_FILE).read_text(encoding="utf-8"))
+
+    # Launch eval requests.
+    results = launch_lm_eval(eval_config)
+
+    # Confirm scores match ground truth.
+    for task in eval_config["tasks"]:
+        for metric in task["metrics"]:
+            ground_truth = metric["value"]
+            measured_value = results["results"][task["name"]][metric["name"]]
+            print(f'{task["name"]} | {metric["name"]}: '
+                  f'ground_truth={ground_truth} | measured={measured_value}')
+            assert numpy.isclose(ground_truth, measured_value, rtol=RTOL)

.buildkite/test-pipeline.yaml

@@ -197,6 +197,22 @@ steps:
   - pip install aiohttp
   - bash run-benchmarks.sh
 
+- label: LM Eval Small Models
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  commands:
+  - pip install lm-eval
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - bash ./run-tests.sh -c configs/models-small.txt -t 1
+
+- label: LM Eval Large Models
+  gpu: a100
+  num_gpus: 4
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  commands:
+  - pip install lm-eval
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - bash ./run-tests.sh -c configs/models-large.txt -t 4
+
 - label: Documentation Build
   working_dir: "/vllm-workspace/test_docs/docs"
   no_gpu: True