[CI/Build] Reuse code for checking output consistency (#5988)

tests/models/test_big_models.py

@@ -7,6 +7,8 @@ Run `pytest tests/models/test_big_models.py`.
 import pytest
 import torch
 
+from .utils import check_outputs_equal
+
 MODELS = [
     "meta-llama/Llama-2-7b-hf",
     # "mistralai/Mistral-7B-v0.1",  # Tested by test_mistral.py
@@ -40,13 +42,12 @@ def test_models(
     with vllm_runner(model, dtype=dtype) as vllm_model:
         vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
 
-    for i in range(len(example_prompts)):
-        hf_output_ids, hf_output_str = hf_outputs[i]
-        vllm_output_ids, vllm_output_str = vllm_outputs[i]
-        assert hf_output_str == vllm_output_str, (
-            f"Test{i}:\nHF: {hf_output_str!r}\nvLLM: {vllm_output_str!r}")
-        assert hf_output_ids == vllm_output_ids, (
-            f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")
+    check_outputs_equal(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
 
 
 @pytest.mark.parametrize("model", MODELS)

tests/models/test_llava.py

@@ -6,6 +6,7 @@ from transformers import AutoTokenizer
 from vllm.config import VisionLanguageConfig
 
 from ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
+from .utils import check_outputs_equal
 
 pytestmark = pytest.mark.vlm
 
@@ -109,14 +110,15 @@ def run_test(
                                                   max_tokens,
                                                   images=vllm_images)
 
-    for i in range(len(HF_IMAGE_PROMPTS)):
-        hf_output_ids, hf_output_str = hf_outputs[i]
-        vllm_output_ids, vllm_output_str = vllm_to_hf_output(
-            vllm_outputs[i], vlm_config, model_id)
-        assert hf_output_str == vllm_output_str, (
-            f"Test{i}:\nHF: {hf_output_str!r}\nvLLM: {vllm_output_str!r}")
-        assert hf_output_ids == vllm_output_ids, (
-            f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")
+    check_outputs_equal(
+        hf_outputs,
+        [
+            vllm_to_hf_output(vllm_output, vlm_config, model_id)
+            for vllm_output in vllm_outputs
+        ],
+        name_0="hf",
+        name_1="vllm",
+    )
 
 
 @pytest.mark.parametrize("model_and_config", model_and_vl_config)

tests/models/test_llava_next.py

@@ -6,6 +6,7 @@ from transformers import AutoTokenizer
 from vllm.config import VisionLanguageConfig
 
 from ..conftest import IMAGE_ASSETS
+from .utils import check_outputs_equal
 
 pytestmark = pytest.mark.vlm
 
@@ -115,11 +116,12 @@ def test_models(hf_runner, vllm_runner, image_assets, model_and_config,
                                                   max_tokens,
                                                   images=vllm_images)
 
-    for i in range(len(HF_IMAGE_PROMPTS)):
-        hf_output_ids, hf_output_str = hf_outputs[i]
-        vllm_output_ids, vllm_output_str = vllm_to_hf_output(
-            vllm_outputs[i], vlm_config, model_id)
-        assert hf_output_str == vllm_output_str, (
-            f"Test{i}:\nHF: {hf_output_str!r}\nvLLM: {vllm_output_str!r}")
-        assert hf_output_ids == vllm_output_ids, (
-            f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")
+    check_outputs_equal(
+        hf_outputs,
+        [
+            vllm_to_hf_output(vllm_output, vlm_config, model_id)
+            for vllm_output in vllm_outputs
+        ],
+        name_0="hf",
+        name_1="vllm",
+    )

tests/models/test_models.py

@@ -7,6 +7,8 @@ Run `pytest tests/models/test_models.py`.
 """
 import pytest
 
+from .utils import check_outputs_equal
+
 MODELS = [
     "facebook/opt-125m",
     "gpt2",
@@ -41,13 +43,12 @@ def test_models(
     with vllm_runner(model, dtype=dtype) as vllm_model:
         vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
 
-    for i in range(len(example_prompts)):
-        hf_output_ids, hf_output_str = hf_outputs[i]
-        vllm_output_ids, vllm_output_str = vllm_outputs[i]
-        assert hf_output_str == vllm_output_str, (
-            f"Test{i}:\nHF: {hf_output_str!r}\nvLLM: {vllm_output_str!r}")
-        assert hf_output_ids == vllm_output_ids, (
-            f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")
+    check_outputs_equal(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
 
 
 @pytest.mark.parametrize("model", MODELS)

tests/models/test_phi3v.py

@@ -7,6 +7,7 @@ from vllm.config import VisionLanguageConfig
 from vllm.utils import is_cpu
 
 from ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
+from .utils import check_outputs_equal
 
 pytestmark = pytest.mark.vlm
 
@@ -124,14 +125,15 @@ def run_test(
                                                   max_tokens,
                                                   images=vllm_images)
 
-    for i in range(len(HF_IMAGE_PROMPTS)):
-        hf_output_ids, hf_output_str = hf_outputs[i]
-        vllm_output_ids, vllm_output_str = vllm_to_hf_output(
-            vllm_outputs[i], vlm_config, model_id)
-        assert hf_output_str == vllm_output_str, (
-            f"Test{i}:\nHF: {hf_output_str!r}\nvLLM: {vllm_output_str!r}")
-        assert hf_output_ids == vllm_output_ids, (
-            f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")
+    check_outputs_equal(
+        hf_outputs,
+        [
+            vllm_to_hf_output(vllm_output, vlm_config, model_id)
+            for vllm_output in vllm_outputs
+        ],
+        name_0="hf",
+        name_1="vllm",
+    )
 
 
 # Since we use _attn_implementation="eager" for hf_runner, here is

tests/models/utils.py

@@ -1,7 +1,43 @@
-def check_logprobs_close(outputs_0_lst, outputs_1_lst, name_0, name_1):
-    """Compare the logprobs of two sequences generated by different models, 
+from typing import Dict, List, Tuple
+
+TokensText = Tuple[List[int], str]
+
+
+def check_outputs_equal(outputs_0_lst: List[TokensText],
+                        outputs_1_lst: List[TokensText], name_0: str,
+                        name_1: str):
+    """
+    Compare the two sequences generated by different models, 
+    which should be equal.
+    """
+    assert len(outputs_0_lst) == len(outputs_1_lst)
+
+    for prompt_idx, (outputs_0,
+                     outputs_1) in enumerate(zip(outputs_0_lst,
+                                                 outputs_1_lst)):
+        output_ids_0, output_str_0 = outputs_0
+        output_ids_1, output_str_1 = outputs_1
+
+        assert output_str_0 == output_str_1, (f"Test{prompt_idx}:"
+                                              f"\n{name_0}:\t{output_str_0!r}"
+                                              f"\n{name_1}:\t{output_str_1!r}")
+        assert output_ids_0 == output_ids_1, (f"Test{prompt_idx}:"
+                                              f"\n{name_0}:\t{output_str_0!r}"
+                                              f"\n{name_1}:\t{output_str_1!r}")
+
+
+TokensTextLogprobs = Tuple[List[int], str, List[Dict[int, float]]]
+
+
+def check_logprobs_close(outputs_0_lst: List[TokensTextLogprobs],
+                         outputs_1_lst: List[TokensTextLogprobs], name_0: str,
+                         name_1: str):
+    """
+    Compare the logprobs of two sequences generated by different models,
     which should be similar but not necessarily equal.
     """
+    assert len(outputs_0_lst) == len(outputs_1_lst)
+
     # Loop through responses to each prompt.
     for prompt_idx, (outputs_0,
                      outputs_1) in enumerate(zip(outputs_0_lst,