[CI/Build] Reorganize models tests (#7820)

tests/models/decoder_only/language/test_aqlm.py

@@ -0,0 +1,68 @@
+"""Compare the outputs of a AQLM model between vLLM and HF Transformers
+
+Run `pytest tests/models/test_aqlm.py`.
+"""
+
+import pytest
+
+from tests.quantization.utils import is_quant_method_supported
+
+# These ground truth generations were generated using `transformers==4.38.1
+# aqlm==1.1.0 torch==2.2.0`
+# and the below code:
+# ```python
+# from transformers import AutoTokenizer, AutoModelForCausalLM
+# model_id = "ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf"
+# quantized_model = AutoModelForCausalLM.from_pretrained(model_id,
+# torch_dtype="auto", device_map="cuda").cuda()
+# tokenizer = AutoTokenizer.from_pretrained(model_id)
+# outputs = []
+# for prompt in example_prompts:
+#     input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"].to("cuda")
+#     hf_outputs = quantized_model.generate(input_ids, max_new_tokens=32)
+#     outputs.append(tokenizer.decode(hf_outputs[0][input_ids.shape[1]:]))
+# print(outputs)
+# ```
+ground_truth_generations = [
+    '\n### Features\n\n- **High-throughput**: v',
+    'The major milestones in the development of artificial intelligence from '
+    '195',
+    'Compare and contrast artificial intelligence with human intelligence in '
+    'terms of processing information. The',
+    'Explain the difference between supervised and unsupervised learning.'
+    '\nExplain',
+    'Write a short story about a robot that dreams for the first time. The',
+    'Analyze the impact of the COVID-19 pandemic on global economic',
+    'The Mona Lisa is a painting by Leonardo da Vinci, and it',
+    'The early bird catches the worm.\nThe early bird catches the'
+]
+
+
+@pytest.mark.skipif(not is_quant_method_supported("aqlm"),
+                    reason="AQLM is not supported on this GPU type.")
+@pytest.mark.parametrize("model", ["ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf"])
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [16])
+@pytest.mark.parametrize("num_logprobs", [1])
+def test_models(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+
+    # loop through the prompts to compare against the ground truth generations
+    for prompt_idx in range(len(example_prompts)):
+        vllm_output_ids, vllm_output_str, vllm_logprobs = vllm_outputs[
+            prompt_idx]
+
+        print("Prompt:          ", repr(example_prompts[prompt_idx]))
+        print("Reference output:", repr(ground_truth_generations[prompt_idx]))
+        print("Output output:   ", repr(vllm_output_str))
+        assert vllm_output_str == ground_truth_generations[prompt_idx]

tests/models/decoder_only/language/test_big_models.py

@@ -0,0 +1,64 @@
+"""Compare the outputs of HF and vLLM when using greedy sampling.
+
+This tests bigger models and use half precision.
+
+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
+    # "Deci/DeciLM-7b",  # Broken
+    # "tiiuae/falcon-7b",  # Broken
+    "EleutherAI/gpt-j-6b",
+    # "mosaicml/mpt-7b",  # Broken
+    # "Qwen/Qwen1.5-0.5B"  # Broken,
+]
+
+#TODO: remove this after CPU float16 support ready
+target_dtype = "float"
+if torch.cuda.is_available():
+    target_dtype = "half"
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", [target_dtype])
+@pytest.mark.parametrize("max_tokens", [32])
+def test_models(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", [target_dtype])
+def test_model_print(
+    vllm_runner,
+    model: str,
+    dtype: str,
+) -> None:
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        # This test is for verifying whether the model's extra_repr
+        # can be printed correctly.
+        print(vllm_model.model.llm_engine.model_executor.driver_worker.
+              model_runner.model)

tests/models/decoder_only/language/test_danube3_4b.py

@@ -0,0 +1,52 @@
+"""Compare the outputs of HF and vLLM when using greedy sampling.
+
+This tests danube3 separately because its head size isn't supported on CPU yet.
+
+Run `pytest tests/models/test_danube3_4b.py`.
+"""
+import pytest
+
+from ...utils import check_outputs_equal
+
+MODELS = ["h2oai/h2o-danube3-4b-base"]
+
+target_dtype = "half"
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", [target_dtype])
+@pytest.mark.parametrize("max_tokens", [32])
+def test_models(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", [target_dtype])
+def test_model_print(
+    vllm_runner,
+    model: str,
+    dtype: str,
+) -> None:
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        # This test is for verifying whether the model's extra_repr
+        # can be printed correctly.
+        print(vllm_model.model.llm_engine.model_executor.driver_worker.
+              model_runner.model)

tests/models/decoder_only/language/test_fp8.py

@@ -0,0 +1,99 @@
+# flake8: noqa
+"""Tests fp8 models against ground truth generation
+Note: these tests will only pass on L4 GPU.
+"""
+import os
+from typing import Optional
+
+import pytest
+
+from tests.kernels.utils import override_backend_env_variable
+from tests.quantization.utils import is_quant_method_supported
+
+from ...utils import check_logprobs_close
+
+os.environ["TOKENIZERS_PARALLELISM"] = "true"
+
+
+@pytest.mark.skipif(not is_quant_method_supported("fp8"),
+                    reason="fp8 is not supported on this GPU type.")
+@pytest.mark.parametrize(
+    "kv_cache_dtype,base_model,test_model,scale_path",
+    [
+        # Test FP8 checkpoint w. fp8_e4m3 kv-cache scaling factors.
+        ("fp8_e4m3", "meta-llama/Meta-Llama-3-8B-Instruct",
+         "nm-testing/Meta-Llama-3-8B-Instruct-FP8-KV", None),
+        # Test FP16 checkpoint w. fp8_e5m2 kv-cache.
+        ("fp8_e5m2", "meta-llama/Meta-Llama-3-8B-Instruct",
+         "meta-llama/Meta-Llama-3-8B-Instruct", None),
+        # Test FP16 checkpoint w. fp8_e4m3 kv-cache scaling factors in json.
+        ("fp8_e4m3", "meta-llama/Llama-2-7b-chat-hf",
+         "meta-llama/Llama-2-7b-chat-hf",
+         "./tests/fp8_kv/llama2-7b-fp8-kv/kv_cache_scales.json")
+    ])
+# Due to low-precision numerical divergence, we only test logprob of 4 tokens
+@pytest.mark.parametrize("max_tokens", [4])
+@pytest.mark.parametrize("enforce_eager", [False, True])
+@pytest.mark.parametrize("backend", ["FLASH_ATTN", "XFORMERS", "FLASHINFER"])
+# NOTE: Increasing this in this suite will fail CI because we currently cannot
+# reset distributed env properly. Use a value > 1 just when you test.
+@pytest.mark.parametrize("tensor_parallel_size", [1])
+# Due to low-precision numerical divergence, this test is too sensitive for
+# the async postprocessor
+@pytest.mark.parametrize("disable_async_output_proc", [True])
+def test_models(
+    vllm_runner,
+    example_prompts,
+    kv_cache_dtype: str,
+    base_model: str,
+    test_model: str,
+    scale_path: Optional[str],
+    max_tokens: int,
+    enforce_eager: bool,
+    backend: str,
+    tensor_parallel_size: int,
+    disable_async_output_proc: bool,
+    monkeypatch,
+) -> None:
+    """
+    Only checks log probs match to cover the discrepancy in
+    numerical sensitive kernels.
+    """
+    override_backend_env_variable(monkeypatch, backend)
+
+    MAX_MODEL_LEN = 1024
+    NUM_LOG_PROBS = 8
+
+    with vllm_runner(
+            base_model,
+            max_model_len=MAX_MODEL_LEN,
+            tensor_parallel_size=tensor_parallel_size,
+            enforce_eager=enforce_eager,
+            kv_cache_dtype="auto",
+            disable_async_output_proc=disable_async_output_proc,
+    ) as vllm_model:
+        baseline_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, NUM_LOG_PROBS)
+
+    extra_kwargs = {}
+    if scale_path is not None:
+        extra_kwargs["quantization_param_path"] = scale_path
+
+    with vllm_runner(
+            test_model,
+            max_model_len=MAX_MODEL_LEN,
+            tensor_parallel_size=tensor_parallel_size,
+            enforce_eager=enforce_eager,
+            kv_cache_dtype=kv_cache_dtype,
+            disable_async_output_proc=disable_async_output_proc,
+            **extra_kwargs,
+    ) as vllm_model:
+        test_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, NUM_LOG_PROBS)
+
+    check_logprobs_close(
+        outputs_0_lst=baseline_outputs,
+        outputs_1_lst=test_outputs,
+        name_0="fp16_kv_cache",
+        name_1="fp8_kv_cache",
+    )

tests/models/decoder_only/language/test_gguf.py

@@ -0,0 +1,90 @@
+"""
+Tests gguf models against unquantized models generations
+Note: To pass the test, quantization higher than Q4 should be used
+"""
+
+import os
+
+import pytest
+from huggingface_hub import hf_hub_download
+from transformers import AutoTokenizer
+
+from tests.quantization.utils import is_quant_method_supported
+
+from ...utils import check_logprobs_close
+
+os.environ["TOKENIZERS_PARALLELISM"] = "true"
+
+MAX_MODEL_LEN = 1024
+
+# FIXME: Move this to confest
+MODELS = [
+    ("TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+     hf_hub_download("TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF",
+                     filename="tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf")),
+    ("TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+     hf_hub_download("duyntnet/TinyLlama-1.1B-Chat-v1.0-imatrix-GGUF",
+                     filename="TinyLlama-1.1B-Chat-v1.0-IQ4_XS.gguf")),
+    ("Qwen/Qwen2-1.5B-Instruct",
+     hf_hub_download("Qwen/Qwen2-1.5B-Instruct-GGUF",
+                     filename="qwen2-1_5b-instruct-q4_k_m.gguf")),
+    ("Qwen/Qwen2-1.5B-Instruct",
+     hf_hub_download("legraphista/Qwen2-1.5B-Instruct-IMat-GGUF",
+                     filename="Qwen2-1.5B-Instruct.IQ4_XS.gguf")),
+]
+
+
+@pytest.mark.skipif(not is_quant_method_supported("gguf"),
+                    reason="gguf is not supported on this GPU type.")
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [32])
+@pytest.mark.parametrize("num_logprobs", [5])
+@pytest.mark.parametrize("tp_size", [1, 2])
+def test_models(
+    num_gpus_available,
+    vllm_runner,
+    example_prompts,
+    model,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+    tp_size: int,
+) -> None:
+    if num_gpus_available < tp_size:
+        pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
+
+    original_model, gguf_model = model
+
+    tokenizer = AutoTokenizer.from_pretrained(original_model)
+    messages = [[{
+        'role': 'user',
+        'content': prompt
+    }] for prompt in example_prompts]
+    example_prompts = tokenizer.apply_chat_template(messages,
+                                                    tokenize=False,
+                                                    add_generation_prompt=True)
+
+    # Run unquantized model.
+    with vllm_runner(model_name=original_model,
+                     dtype=dtype,
+                     max_model_len=MAX_MODEL_LEN,
+                     tensor_parallel_size=tp_size) as original_model:
+
+        original_outputs = original_model.generate_greedy_logprobs(
+            example_prompts[:-1], max_tokens, num_logprobs)
+
+    # Run gguf model.
+    with vllm_runner(model_name=gguf_model,
+                     dtype=dtype,
+                     max_model_len=MAX_MODEL_LEN,
+                     tensor_parallel_size=tp_size) as gguf_model:
+        gguf_outputs = gguf_model.generate_greedy_logprobs(
+            example_prompts[:-1], max_tokens, num_logprobs)
+
+    check_logprobs_close(
+        outputs_0_lst=original_outputs,
+        outputs_1_lst=gguf_outputs,
+        name_0="original",
+        name_1="gguf",
+    )

tests/models/decoder_only/language/test_gptq_marlin.py

@@ -0,0 +1,96 @@
+"""Compares the outputs of gptq vs gptq_marlin 
+Note: GPTQ and Marlin do not have bitwise correctness.
+As a result, in this test, we just confirm that the top selected tokens of the
+Marlin/GPTQ models are in the top 5 selections of each other.
+Note: Marlin internally uses locks to synchronize the threads. This can
+result in very slight nondeterminism for Marlin. As a result, we re-run the test
+up to 3 times to see if we pass.
+
+Run `pytest tests/models/test_gptq_marlin.py`.
+"""
+import os
+
+import pytest
+
+from tests.quantization.utils import is_quant_method_supported
+from vllm.model_executor.layers.rotary_embedding import _ROPE_DICT
+
+from ...utils import check_logprobs_close
+
+os.environ["TOKENIZERS_PARALLELISM"] = "true"
+
+MAX_MODEL_LEN = 1024
+
+MODELS = [
+    # act_order==False, group_size=channelwise
+    ("robertgshaw2/zephyr-7b-beta-channelwise-gptq", "main"),
+    # act_order==False, group_size=128
+    ("TheBloke/Llama-2-7B-GPTQ", "main"),
+
+    # act_order==True, group_size=128
+    ("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "main"),
+    # act_order==True, group_size=64
+    ("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-4bit-64g-actorder_True"),
+    # act_order==True, group_size=32
+    ("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-4bit-32g-actorder_True"),
+
+    # 8-bit, act_order==True, group_size=channelwise
+    ("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-8bit--1g-actorder_True"),
+    # 8-bit, act_order==True, group_size=128
+    ("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-8bit-128g-actorder_True"),
+    # 8-bit, act_order==True, group_size=32
+    ("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-8bit-32g-actorder_True"),
+
+    # 4-bit, act_order==True, group_size=128
+    ("TechxGenus/gemma-1.1-2b-it-GPTQ", "main")
+]
+
+
+@pytest.mark.flaky(reruns=3)
+@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
+                    reason="gptq_marlin is not supported on this GPU type.")
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half", "bfloat16"])
+@pytest.mark.parametrize("max_tokens", [32])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_models(
+    vllm_runner,
+    example_prompts,
+    model,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+    model_name, revision = model
+
+    # Run marlin.
+    with vllm_runner(model_name=model_name,
+                     revision=revision,
+                     dtype=dtype,
+                     quantization="marlin",
+                     max_model_len=MAX_MODEL_LEN,
+                     tensor_parallel_size=1) as gptq_marlin_model:
+
+        gptq_marlin_outputs = gptq_marlin_model.generate_greedy_logprobs(
+            example_prompts[:-1], max_tokens, num_logprobs)
+    _ROPE_DICT.clear()  # clear rope cache to avoid rope dtype error
+
+    # Run gptq.
+    # The naive gptq kernel doesn't support bf16 yet.
+    # Here we always compare fp16/bf16 gpt marlin kernel
+    # to fp16 gptq kernel.
+    with vllm_runner(model_name=model_name,
+                     revision=revision,
+                     dtype="half",
+                     quantization="gptq",
+                     max_model_len=MAX_MODEL_LEN,
+                     tensor_parallel_size=1) as gptq_model:
+        gptq_outputs = gptq_model.generate_greedy_logprobs(
+            example_prompts[:-1], max_tokens, num_logprobs)
+
+    check_logprobs_close(
+        outputs_0_lst=gptq_outputs,
+        outputs_1_lst=gptq_marlin_outputs,
+        name_0="gptq",
+        name_1="gptq_marlin",
+    )

tests/models/decoder_only/language/test_gptq_marlin_24.py

@@ -0,0 +1,72 @@
+"""Compare the outputs of a GPTQ model to a Marlin_24 model.
+
+Note: GPTQ and Marlin_24 do not have bitwise correctness.
+As a result, in this test, we just confirm that the top selected tokens of the
+Marlin/GPTQ models are in the top 3 selections of each other.
+
+Run `pytest tests/models/test_marlin_24.py`.
+"""
+from dataclasses import dataclass
+
+import pytest
+
+from tests.quantization.utils import is_quant_method_supported
+
+from ...utils import check_logprobs_close
+
+
+@dataclass
+class ModelPair:
+    model_marlin: str
+    model_gptq: str
+
+
+model_pairs = [
+    # 4-bit, group_size == 128
+    ModelPair(model_marlin="alexm-nm/tinyllama-24-marlin24-4bit-g128",
+              model_gptq="alexm-nm/tinyllama-24-gptq-4bit-g128"),
+    # 4-bit, group_size == channelwise
+    ModelPair(model_marlin="alexm-nm/tinyllama-24-marlin24-4bit-channelwise",
+              model_gptq="alexm-nm/tinyllama-24-gptq-4bit-channelwise"),
+
+    # 8-bit, group_size == 128
+    ModelPair(model_marlin="alexm-nm/tinyllama-24-marlin24-8bit-g128",
+              model_gptq="alexm-nm/tinyllama-24-gptq-8bit-g128"),
+    # 8-bit, group_size == channelwise
+    ModelPair(model_marlin="alexm-nm/tinyllama-24-marlin24-8bit-channelwise",
+              model_gptq="alexm-nm/tinyllama-24-gptq-8bit-channelwise"),
+]
+
+
+@pytest.mark.flaky(reruns=2)
+@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin_24"),
+                    reason="Marlin24 is not supported on this GPU type.")
+@pytest.mark.parametrize("model_pair", model_pairs)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [8])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_models(
+    vllm_runner,
+    example_prompts,
+    model_pair: ModelPair,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+    with vllm_runner(model_pair.model_marlin,
+                     dtype=dtype,
+                     quantization="gptq_marlin_24") as marlin_24_model:
+        marlin_24_outputs = marlin_24_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+
+    with vllm_runner(model_pair.model_gptq, dtype=dtype,
+                     quantization="gptq") as gptq_model:
+        gptq_outputs = gptq_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+
+    check_logprobs_close(
+        outputs_0_lst=gptq_outputs,
+        outputs_1_lst=marlin_24_outputs,
+        name_0="gptq",
+        name_1="marlin_24",
+    )

tests/models/decoder_only/language/test_granite.py

@@ -0,0 +1,49 @@
+"""Compare the outputs of HF and vLLM for Granite models using greedy sampling.
+
+Run `pytest tests/models/test_granite.py`.
+"""
+import importlib.metadata
+
+import pytest
+
+from ...utils import check_logprobs_close
+
+TRANSFORMERS_VERSION = tuple(
+    map(int,
+        importlib.metadata.version("transformers").split(".")))
+
+MODELS = [
+    "ibm/PowerLM-3b",
+]
+
+
+# GraniteForCausalLM will be in transformers >= 4.45
+@pytest.mark.skipif(TRANSFORMERS_VERSION < (4, 45),
+                    reason="granite model test requires transformers >= 4.45")
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("max_tokens", [64])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_models(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+    # TODO(sang): Sliding window should be tested separately.
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs = hf_model.generate_greedy_logprobs_limit(
+            example_prompts, max_tokens, num_logprobs)
+
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+    check_logprobs_close(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )

tests/models/decoder_only/language/test_jamba.py

@@ -0,0 +1,178 @@
+import pytest
+
+from vllm.worker.model_runner import _get_graph_batch_size
+
+from ...utils import check_outputs_equal
+
+MODELS = ["ai21labs/Jamba-tiny-random"]
+
+
+# Fails due to usage of MoE as MLP(E=1_, which is different than the HF impl
+# TODO: Fix this with trained model
+@pytest.mark.skip()
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("max_tokens", [10])
+def test_models(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+
+    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}")
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [5])
+def test_batching(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    # To pass the small model tests, we need full precision.
+    for_loop_outputs = []
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        for prompt in example_prompts:
+            for_loop_outputs.append(
+                vllm_model.generate_greedy([prompt], max_tokens)[0])
+
+        batched_outputs = vllm_model.generate_greedy(example_prompts,
+                                                     max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=for_loop_outputs,
+        outputs_1_lst=batched_outputs,
+        name_0="for_loop_vllm",
+        name_1="batched_vllm",
+    )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("max_tokens", [20])
+def test_mamba_cache_cg_padding(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    # This test is for verifying that mamba cache is padded to CG captured
+    # batch size. If it's not, a torch RuntimeError will be raised because
+    # tensor dimensions aren't compatible
+    while len(example_prompts) == _get_graph_batch_size(len(example_prompts)):
+        example_prompts.append(example_prompts[0])
+
+    try:
+        with vllm_runner(model, dtype=dtype) as vllm_model:
+            vllm_model.generate_greedy(example_prompts, max_tokens)
+    except RuntimeError:
+        pytest.fail(
+            "Couldn't run batch size which is not equal to a Cuda Graph "
+            "captured batch size. "
+            "Could be related to mamba cache not padded correctly")
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+@pytest.mark.parametrize("max_tokens", [20])
+def test_models_preemption_recompute(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    # Tests that outputs are identical with and w/o preemtions (recompute)
+    assert dtype == "float"
+
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_model.model.llm_engine.scheduler[
+            0].ENABLE_ARTIFICIAL_PREEMPT = True
+        preempt_vllm_outputs = vllm_model.generate_greedy(
+            example_prompts, max_tokens)
+
+        vllm_model.model.llm_engine.scheduler[
+            0].ENABLE_ARTIFICIAL_PREEMPT = False
+        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=preempt_vllm_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="vllm_preepmtions",
+        name_1="vllm",
+    )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+def test_fail_upon_inc_requests_and_finished_requests_lt_available_blocks(
+    vllm_runner,
+    model: str,
+    dtype: str,
+    example_prompts,
+) -> None:
+    # This test is for verifying that the Jamba inner state management doesn't
+    # collapse in case where the number of incoming requests and
+    # finished_requests_ids is larger than the maximum mamba block capacity.
+    # This could generally happen due to the fact that Jamba does support
+    # statelessness mechanism where it can cleanup new incoming requests in
+    # a single step.
+    try:
+        with vllm_runner(model, dtype=dtype, max_num_seqs=10) as vllm_model:
+            vllm_model.generate_greedy([example_prompts[0]] * 100, 10)
+    except ValueError:
+        pytest.fail("Jamba inner state wasn't cleaned up properly between"
+                    "steps finished requests registered unnecessarily ")
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+def test_state_cleanup(
+    vllm_runner,
+    model: str,
+    dtype: str,
+    example_prompts,
+) -> None:
+    # This test is for verifying that the Jamba state is cleaned up between
+    # steps, If its not cleaned, an error would be expected.
+    try:
+        with vllm_runner(model, dtype=dtype) as vllm_model:
+            for _ in range(10):
+                vllm_model.generate_greedy([example_prompts[0]] * 100, 1)
+    except ValueError:
+        pytest.fail("Jamba inner state wasn't cleaned up between states, "
+                    "could be related to finished_requests_ids")
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+def test_model_print(
+    vllm_runner,
+    model: str,
+    dtype: str,
+) -> None:
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        # This test is for verifying whether the model's extra_repr
+        # can be printed correctly.
+        print(vllm_model.model.llm_engine.model_executor.driver_worker.
+              model_runner.model)

tests/models/decoder_only/language/test_marlin.py

@@ -0,0 +1,69 @@
+"""Compare the outputs of a GPTQ model to a Marlin model.
+
+Note: GPTQ and Marlin do not have bitwise correctness.
+As a result, in this test, we just confirm that the top selected tokens of the
+Marlin/GPTQ models are in the top 3 selections of each other.
+
+Note: Marlin internally uses locks to synchronize the threads. This can
+result in very slight nondeterminism for Marlin. As a result, we re-run the test
+up to 3 times to see if we pass.
+
+Run `pytest tests/models/test_marlin.py`.
+"""
+from dataclasses import dataclass
+
+import pytest
+
+from tests.quantization.utils import is_quant_method_supported
+
+from ...utils import check_logprobs_close
+
+
+@dataclass
+class ModelPair:
+    model_marlin: str
+    model_gptq: str
+
+
+model_pairs = [
+    ModelPair(model_marlin="nm-testing/zephyr-beta-7b-marlin-g128",
+              model_gptq="nm-testing/zephyr-beta-7b-gptq-g128"),
+    ModelPair(model_marlin="robertgshaw2/zephyr-7b-beta-channelwise-marlin",
+              model_gptq="robertgshaw2/zephyr-7b-beta-channelwise-gptq"),
+    ModelPair(model_marlin="robertgshaw2/TinyLlama-1.1B-Chat-v1.0-g128-marlin",
+              model_gptq="robertgshaw2/TinyLlama-1.1B-Chat-v1.0-g128-gptq")
+]
+
+
+@pytest.mark.flaky(reruns=2)
+@pytest.mark.skipif(not is_quant_method_supported("marlin"),
+                    reason="Marlin is not supported on this GPU type.")
+@pytest.mark.parametrize("model_pair", model_pairs)
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [32])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_models(
+    vllm_runner,
+    example_prompts,
+    model_pair: ModelPair,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+    with vllm_runner(model_pair.model_marlin,
+                     dtype=dtype,
+                     quantization="marlin") as marlin_model:
+        marlin_outputs = marlin_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+
+    with vllm_runner(model_pair.model_gptq, dtype=dtype,
+                     quantization="gptq") as gptq_model:
+        gptq_outputs = gptq_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+
+    check_logprobs_close(
+        outputs_0_lst=gptq_outputs,
+        outputs_1_lst=marlin_outputs,
+        name_0="gptq",
+        name_1="marlin",
+    )

tests/models/decoder_only/language/test_mistral.py

@@ -0,0 +1,83 @@
+"""Compare the outputs of HF and vLLM for Mistral models using greedy sampling.
+
+Run `pytest tests/models/test_mistral.py`.
+"""
+import pytest
+
+from ...utils import check_logprobs_close
+
+MODELS = [
+    "mistralai/Mistral-7B-Instruct-v0.1",
+    "mistralai/Mistral-7B-Instruct-v0.3",
+]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("max_tokens", [64])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_models(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+    # TODO(sang): Sliding window should be tested separately.
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs = hf_model.generate_greedy_logprobs_limit(
+            example_prompts, max_tokens, num_logprobs)
+
+    with vllm_runner(model, dtype=dtype,
+                     tokenizer_mode="mistral") as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+
+    check_logprobs_close(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
+
+
+@pytest.mark.parametrize("model", MODELS[1:])
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("max_tokens", [64])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_mistral_format(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+    with vllm_runner(
+            model,
+            dtype=dtype,
+            tokenizer_mode="auto",
+            load_format="safetensors",
+            config_format="hf",
+    ) as hf_format_model:
+        hf_format_outputs = hf_format_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+
+    with vllm_runner(
+            model,
+            dtype=dtype,
+            tokenizer_mode="mistral",
+            load_format="mistral",
+            config_format="mistral",
+    ) as mistral_format_model:
+        mistral_format_outputs = mistral_format_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+
+    check_logprobs_close(
+        outputs_0_lst=hf_format_outputs,
+        outputs_1_lst=mistral_format_outputs,
+        name_0="hf",
+        name_1="mistral",
+    )

tests/models/decoder_only/language/test_modelopt.py

@@ -0,0 +1,79 @@
+# flake8: noqa
+"""Tests Model Optimizer fp8 models against ground truth generation
+Note: these tests will only pass on H100
+"""
+import os
+from typing import List
+
+import pytest
+from transformers import AutoTokenizer
+
+from tests.quantization.utils import is_quant_method_supported
+from vllm import LLM, SamplingParams
+
+os.environ["TOKENIZERS_PARALLELISM"] = "true"
+
+MAX_MODEL_LEN = 1024
+
+MODELS = ["nvidia/Llama-3.1-8B-Instruct-FP8"]
+
+EXPECTED_STRS_MAP = {
+    "nvidia/Llama-3.1-8B-Instruct-FP8": [
+        "You're referring to VLLM, a high-performance Large Language Model (LLM) inference and",
+        'Here are the major milestones in the development of artificial intelligence (AI) from 1950 to ',
+        'The comparison between artificial intelligence (AI) and human intelligence in terms of processing information is a complex and',
+        'A neural network is a complex system modeled after the human brain, consisting of interconnected nodes or "ne',
+        '**The Spark of Imagination**\n\nZeta-5, a sleek and efficient robot, whir',
+        'The COVID-19 pandemic has had a profound impact on global economic structures and business models, leading to',
+        'The Mona Lisa, painted by Leonardo da Vinci in the early 16th century, is one of',
+        'Here are the translations:\n\n**Japanese:** 「早起きは早く獲物をとる'
+    ]
+}
+
+
+# This test compares against golden strings for exact match since
+# there is no baseline implementation to compare against
+# and is unstable w.r.t specifics of the fp8 implementation or
+# the hardware being run on.
+# Disabled to prevent it from breaking the build
+@pytest.mark.skip(
+    reason=
+    "Prevent unstable test based on golden strings from breaking the build.")
+@pytest.mark.skipif(not is_quant_method_supported("fp8"),
+                    reason="fp8 is not supported on this GPU type.")
+@pytest.mark.parametrize("model_name", MODELS)
+def test_models(example_prompts, model_name) -> None:
+    model = LLM(
+        model=model_name,
+        max_model_len=MAX_MODEL_LEN,
+        trust_remote_code=True,
+        enforce_eager=True,
+        quantization="modelopt",
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    formatted_prompts = [
+        tokenizer.apply_chat_template([{
+            "role": "user",
+            "content": prompt
+        }],
+                                      tokenize=False,
+                                      add_generation_prompt=True)
+        for prompt in example_prompts
+    ]
+    params = SamplingParams(max_tokens=20, temperature=0)
+    generations: List[str] = []
+    # Note: these need to be run 1 at a time due to numerical precision,
+    # since the expected strs were generated this way.
+    for prompt in formatted_prompts:
+        outputs = model.generate(prompt, params)
+        generations.append(outputs[0].outputs[0].text)
+    del model
+
+    print(model_name, generations)
+    expected_strs = EXPECTED_STRS_MAP[model_name]
+    for i in range(len(example_prompts)):
+        generated_str = generations[i]
+        expected_str = expected_strs[i]
+        assert expected_str == generated_str, (
+            f"Test{i}:\nExpected: {expected_str!r}\nvLLM: {generated_str!r}")

tests/models/decoder_only/language/test_models.py

@@ -0,0 +1,65 @@
+"""Compare the outputs of HF and vLLM when using greedy sampling.
+
+This test only tests small models. Big models such as 7B should be tested from
+test_big_models.py because it could use a larger instance to run tests.
+
+Run `pytest tests/models/test_models.py`.
+"""
+import pytest
+
+from ...utils import check_outputs_equal
+
+MODELS = [
+    "facebook/opt-125m",
+    "gpt2",
+    "bigcode/tiny_starcoder_py",
+    "EleutherAI/pythia-70m",
+    "bigscience/bloom-560m",  # Testing alibi slopes.
+    "microsoft/phi-2",
+    "stabilityai/stablelm-3b-4e1t",
+    # "allenai/OLMo-1B",  # Broken
+    "bigcode/starcoder2-3b",
+    "google/gemma-1.1-2b-it",
+]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+@pytest.mark.parametrize("max_tokens", [96])
+def test_models(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    # To pass the small model tests, we need full precision.
+    assert dtype == "float"
+
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+def test_model_print(
+    vllm_runner,
+    model: str,
+    dtype: str,
+) -> None:
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        # This test is for verifying whether the model's extra_repr
+        # can be printed correctly.
+        print(vllm_model.model.llm_engine.model_executor.driver_worker.
+              model_runner.model)

tests/models/decoder_only/language/test_phimoe.py

@@ -0,0 +1,111 @@
+"""Compare the outputs of HF and vLLM for moe models using greedy sampling.
+
+Run `pytest tests/models/test_phimoe.py`.
+"""
+import pytest
+import torch
+
+from vllm.utils import is_cpu
+
+from ...utils import check_logprobs_close
+
+MODELS = [
+    "microsoft/Phi-3.5-MoE-instruct",
+]
+
+
+def test_phimoe_routing_function():
+    from vllm.model_executor.models.phimoe import phimoe_routing_function
+    test_case = {
+        0: {
+            "hidden_states":
+            torch.tensor([1, 2, 3, 4, 5, 6, 7, 8],
+                         dtype=torch.float32,
+                         requires_grad=False).view(4, 2),
+            "gating_output":
+            torch.tensor([0.1, 0.2, 0.3, 0.4],
+                         dtype=torch.float32,
+                         requires_grad=False),
+            "topk":
+            2,
+            "renormalize":
+            False,
+        },
+        1: {
+            "hidden_states":
+            torch.tensor([1, 2, 3, 4, 5, 6, 7, 8],
+                         dtype=torch.float32,
+                         requires_grad=False).view(4, 2),
+            "gating_output":
+            torch.tensor([0.4, 0.2, 0.3, 0.4],
+                         dtype=torch.float32,
+                         requires_grad=False),
+            "topk":
+            2,
+            "renormalize":
+            False,
+        }
+    }
+
+    ground_truth = {
+        0: {
+            "topk_weights":
+            torch.tensor([1., 1.], dtype=torch.float32, requires_grad=False),
+            "topk_ids":
+            torch.tensor([3, 2], dtype=torch.long, requires_grad=False),
+        },
+        1: {
+            "topk_weights":
+            torch.tensor([0.5, 1.], dtype=torch.float32, requires_grad=False),
+            "topk_ids":
+            torch.tensor([0, 3], dtype=torch.long, requires_grad=False),
+        }
+    }
+
+    for test_id in test_case:
+        topk_weights, topk_ids = phimoe_routing_function(**test_case[test_id])
+        assert torch.allclose(topk_weights,
+                              ground_truth[test_id]["topk_weights"])
+        assert torch.equal(topk_ids, ground_truth[test_id]["topk_ids"])
+
+
+def get_gpu_memory():
+    try:
+        props = torch.cuda.get_device_properties(torch.cuda.current_device())
+        gpu_memory = props.total_memory / (1024**3)
+        return gpu_memory
+    except Exception:
+        return 0
+
+
+@pytest.mark.skipif(condition=is_cpu(),
+                    reason="This test takes a lot time to run on CPU, "
+                    "and vllm CI's disk space is not enough for this model.")
+@pytest.mark.skipif(condition=get_gpu_memory() < 100,
+                    reason="Skip this test if GPU memory is insufficient.")
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("max_tokens", [64])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_models(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs = hf_model.generate_greedy_logprobs_limit(
+            example_prompts, max_tokens, num_logprobs)
+
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, num_logprobs)
+    check_logprobs_close(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )