| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import os |
| 4 | |
| 5 | import lm_eval |
| 6 | import openai |
| 7 | |
| 8 | BASE_URL = f"http://localhost:{os.environ.get('PROXY_PORT', 8192)}/v1" |
| 9 | NUM_CONCURRENT = 100 |
| 10 | TASK = "gsm8k" |
| 11 | FILTER = "exact_match,strict-match" |
| 12 | RTOL = 0.05 |
| 13 | |
| 14 | # Model-specific expected values |
| 15 | EXPECTED_VALUES = { |
| 16 | "Qwen/Qwen3-0.6B": 0.41, |
| 17 | } |
| 18 | |
| 19 | SIMPLE_PROMPT = ( |
| 20 | "The best part about working on vLLM is that I got to meet so many people across " |
| 21 | "various different organizations like UCB, Google, and Meta which means", |
| 22 | ) |
| 23 | |
| 24 | # Get model name from environment variable |
| 25 | MODEL_NAME = os.environ.get("TEST_MODEL", "Qwen/Qwen3-0.6B") |
| 26 | |
| 27 | |
| 28 | def run_simple_prompt(): |
| 29 | client = openai.OpenAI(api_key="EMPTY", base_url=BASE_URL) |
| 30 | completion = client.completions.create(model=MODEL_NAME, prompt=SIMPLE_PROMPT) |
| 31 | |
| 32 | print("-" * 50) |
| 33 | print(f"Completion results for {MODEL_NAME}:") |
| 34 | print(completion) |
| 35 | print("-" * 50) |
| 36 | |
| 37 | |
| 38 | def test_accuracy(): |
| 39 | """Run the end to end accuracy test.""" |
| 40 | run_simple_prompt() |
| 41 | |
| 42 | model_args = ( |
| 43 | f"model={MODEL_NAME}," |
| 44 | f"base_url={BASE_URL}/completions," |
| 45 | f"num_concurrent={NUM_CONCURRENT},tokenized_requests=False" |
| 46 | ) |
| 47 | results = lm_eval.simple_evaluate( |
| 48 | model="local-completions", |
| 49 | model_args=model_args, |
| 50 | tasks=TASK, |
| 51 | ) |
| 52 | |
| 53 | measured_value = results["results"][TASK][FILTER] |
| 54 | expected_value = EXPECTED_VALUES.get(MODEL_NAME) |
| 55 | |
| 56 | print(f"Measured accuracy value: {measured_value}\n") |
| 57 | if expected_value is None: |
| 58 | print( |
| 59 | f"Warning: No expected value found for {MODEL_NAME}. " |
| 60 | "Skipping accuracy check." |
| 61 | ) |
| 62 | return |
| 63 | |
| 64 | assert ( |
| 65 | measured_value - RTOL < expected_value |
| 66 | and measured_value + RTOL > expected_value |
| 67 | ), f"Expected: {expected_value} | Measured: {measured_value}" |