| 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 = "http://localhost:8192/v1" |
| 9 | NUM_CONCURRENT = int(os.getenv("NUM_CONCURRENT", "100")) |
| 10 | TASK = "gsm8k" |
| 11 | FILTER = "exact_match,strict-match" |
| 12 | # TODO(#43186): Widened from 0.03 to absorb chunk_scan/SSU numeric jitter |
| 13 | # on granite-4.0-h-tiny under NIXL PD; tighten when the kernel divergence |
| 14 | # is fixed. |
| 15 | RTOL = 0.05 |
| 16 | |
| 17 | # Model-specific expected values |
| 18 | EXPECTED_VALUES = { |
| 19 | "Qwen/Qwen3-0.6B": 0.41, |
| 20 | "deepseek-ai/deepseek-vl2-small": 0.59, |
| 21 | "deepseek-ai/deepseek-vl2-tiny": 0.19, |
| 22 | "deepseek-ai/DeepSeek-V2-Lite-Chat": 0.65, |
| 23 | "google/gemma-3-4b-it": 0.74, |
| 24 | "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8": 0.84, |
| 25 | "ibm-granite/granite-4.0-h-tiny": 0.77, |
| 26 | "Qwen/Qwen3.5-0.8B": 0.33, |
| 27 | "google/gemma-4-E2B-it": 0.485, |
| 28 | "ai21labs/AI21-Jamba2-3B": 0.74, |
| 29 | "deepseek-ai/DeepSeek-V4-Flash": 0.95, |
| 30 | } |
| 31 | |
| 32 | SIMPLE_PROMPT = ( |
| 33 | "The best part about working on vLLM is that I got to meet so many people across " |
| 34 | "various different organizations like UCB, Google, and Meta which means", |
| 35 | ) |
| 36 | |
| 37 | # Get model name from environment variable |
| 38 | MODEL_NAME = os.environ.get("TEST_MODEL", "Qwen/Qwen3-0.6B") |
| 39 | |
| 40 | |
| 41 | def _meets_accuracy_threshold(measured_value: float, expected_value: float) -> bool: |
| 42 | return measured_value >= expected_value - RTOL |
| 43 | |
| 44 | |
| 45 | def test_accuracy_threshold_is_one_sided(): |
| 46 | expected_value = 0.77 |
| 47 | minimum_value = expected_value - RTOL |
| 48 | |
| 49 | assert _meets_accuracy_threshold(minimum_value, expected_value) |
| 50 | assert _meets_accuracy_threshold(expected_value + RTOL + 0.01, expected_value) |
| 51 | assert not _meets_accuracy_threshold(minimum_value - 0.01, expected_value) |
| 52 | |
| 53 | |
| 54 | def run_simple_prompt(): |
| 55 | client = openai.OpenAI(api_key="EMPTY", base_url=BASE_URL) |
| 56 | completion = client.completions.create(model=MODEL_NAME, prompt=SIMPLE_PROMPT) |
| 57 | |
| 58 | print("-" * 50) |
| 59 | print(f"Completion results for {MODEL_NAME}:") |
| 60 | print(completion) |
| 61 | print("-" * 50) |
| 62 | |
| 63 | |
| 64 | def test_accuracy(): |
| 65 | """Run the end to end accuracy test.""" |
| 66 | run_simple_prompt() |
| 67 | |
| 68 | if "gemma-4" in MODEL_NAME: |
| 69 | # Gemma4 is quite sensible to having a chat template applied, so we evaluate |
| 70 | # on chat completions. |
| 71 | model_args = ( |
| 72 | f"model={MODEL_NAME}," |
| 73 | f"base_url={BASE_URL}/chat/completions," |
| 74 | f"num_concurrent={NUM_CONCURRENT}," |
| 75 | "tokenizer_backend=huggingface," |
| 76 | "trust_remote_code=True" |
| 77 | ) |
| 78 | results = lm_eval.simple_evaluate( |
| 79 | model="local-chat-completions", |
| 80 | model_args=model_args, |
| 81 | tasks=TASK, |
| 82 | num_fewshot=5, |
| 83 | apply_chat_template=True, |
| 84 | ) |
| 85 | else: |
| 86 | model_args = ( |
| 87 | f"model={MODEL_NAME}," |
| 88 | f"base_url={BASE_URL}/completions," |
| 89 | f"num_concurrent={NUM_CONCURRENT},tokenized_requests=False," |
| 90 | "trust_remote_code=True" |
| 91 | ) |
| 92 | results = lm_eval.simple_evaluate( |
| 93 | model="local-completions", |
| 94 | model_args=model_args, |
| 95 | tasks=TASK, |
| 96 | ) |
| 97 | |
| 98 | measured_value = results["results"][TASK][FILTER] |
| 99 | expected_value = EXPECTED_VALUES.get(MODEL_NAME) |
| 100 | |
| 101 | print(f"Measured accuracy value: {measured_value}\n") |
| 102 | if expected_value is None: |
| 103 | print( |
| 104 | f"Warning: No expected value found for {MODEL_NAME}. " |
| 105 | "Skipping accuracy check." |
| 106 | ) |
| 107 | return |
| 108 | |
| 109 | minimum_value = expected_value - RTOL |
| 110 | assert _meets_accuracy_threshold(measured_value, expected_value), ( |
| 111 | f"Expected at least: {minimum_value} | Measured: {measured_value}" |
| 112 | ) |