[tests] update v1 tests

tests/samplers/test_beam_search.py

@@ -6,10 +6,10 @@ Run `pytest tests/samplers/test_beam_search.py`.
 
 import pytest
 import os
-from ..utils import models_path_prefix
 from transformers import AutoModelForSeq2SeqLM
 
 from vllm.assets.audio import AudioAsset
+from ..utils import models_path_prefix
 
 
 @pytest.fixture(autouse=True)
@@ -83,7 +83,7 @@ def test_beam_search_passes_multimodal_data(
     # correctly. As such, we just need to check one extra modality to make
     # sure things pass through properly.
     audios = [AudioAsset("mary_had_lamb").audio_and_sample_rate]
-    model = "Qwen/Qwen2-Audio-7B-Instruct"
+    model = os.path.join(models_path_prefix, "Qwen/Qwen2-Audio-7B-Instruct")
     audio_seq = "<|audio_bos|><|AUDIO|><|audio_eos|>"
     prompts = [
         f"<|im_start|>user\n{audio_seq}Can you transcribe this?<|im_end|>\n<|im_start|>assistant\n"  #noqa: E501

tests/spec_decode/e2e/test_integration.py

@@ -69,7 +69,7 @@ def test_spec_decode_cuda_graph(vllm_runner, common_llm_kwargs,
         # Explicitly specify draft model quantization
         {
             "speculative_config": {
-                "model": "LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit",
+                "model": os.path.join(models_path_prefix, "LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit"),
                 "num_speculative_tokens": 5,
                 "quantization": "gptq",
             },
@@ -77,7 +77,7 @@ def test_spec_decode_cuda_graph(vllm_runner, common_llm_kwargs,
         # Explicitly specify GPTQ-based draft model to use marlin quantization
         {
             "speculative_config": {
-                "model": "LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit",
+                "model": os.path.join(models_path_prefix, "LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit"),
                 "num_speculative_tokens": 5,
                 "quantization": "marlin",
             },
@@ -85,7 +85,7 @@ def test_spec_decode_cuda_graph(vllm_runner, common_llm_kwargs,
         # Not explicitly specify draft model quantization
         {
             "speculative_config": {
-                "model": "LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit",
+                "model": os.path.join(models_path_prefix, "LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit"),
                 "num_speculative_tokens": 5,
                 "quantization": None,
             },
@@ -124,7 +124,7 @@ def test_speculative_model_quantization_config(vllm_runner, common_llm_kwargs,
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
 @pytest.mark.parametrize("test_llm_kwargs", [{
     "speculative_config": {
-        "model": "JackFram/llama-68m",
+        "model": os.path.join(models_path_prefix, "JackFram/llama-68m"),
         "num_speculative_tokens": 3,
         "disable_mqa_scorer": True,
     },

tests/v1/core/test_kv_cache_utils.py

 # SPDX-License-Identifier: Apache-2.0
 
+import os
 import pytest
 import torch
 
@@ -22,6 +23,7 @@ from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                         KVCacheGroupSpec, KVCacheTensor)
 from vllm.v1.metrics.stats import PrefixCacheStats
 from vllm.v1.request import Request
+from ...utils import models_path_prefix
 
 # yapf: enable
 
@@ -432,8 +434,8 @@ def test_unify_kv_cache_configs():
 
 @pytest.mark.parametrize(
     ("model_id", "max_model_len", "want_estimated_max_len"), [
-        ("Qwen/Qwen1.5-7B", 16385, 16384),
-        ("Qwen/Qwen1.5-7B", 16383, 16383),
+        (os.path.join(models_path_prefix, "Qwen/Qwen1.5-7B"), 16385, 16384),
+        (os.path.join(models_path_prefix, "Qwen/Qwen1.5-7B"), 16383, 16383),
     ])
 def test_estimate_max_model_len(model_id, max_model_len,
                                 want_estimated_max_len):

tests/v1/core/test_scheduler.py

@@ -2,6 +2,7 @@
 from typing import Optional
 from unittest.mock import Mock
 
+import os
 import pytest
 import torch
 
@@ -16,12 +17,13 @@ from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
 from vllm.v1.outputs import ModelRunnerOutput
 from vllm.v1.request import Request, RequestStatus
 from vllm.v1.structured_output import StructuredOutputManager
+from ...utils import models_path_prefix
 
 EOS_TOKEN_ID = 50256
 
 
 def create_scheduler(
-    model: str = "facebook/opt-125m",
+    model: str = os.path.join(models_path_prefix, "facebook/opt-125m"),
     max_num_seqs: int = 16,
     max_num_batched_tokens: int = 8192,
     enable_prefix_caching: Optional[bool] = None,
@@ -211,7 +213,7 @@ def test_schedule(enable_prefix_caching: Optional[bool],
 
 
 def test_schedule_multimodal_requests():
-    scheduler = create_scheduler(model="llava-hf/llava-1.5-7b-hf")
+    scheduler = create_scheduler(model=os.path.join(models_path_prefix, "llava-hf/llava-1.5-7b-hf"))
     mm_positions = [[PlaceholderRange(offset=i, length=100)]
                     for i in range(10)]
     requests = create_requests(
@@ -243,7 +245,7 @@ def test_schedule_partial_requests():
        there is insufficient encoder budget.
     """
     scheduler = create_scheduler(
-        model="llava-hf/llava-1.5-7b-hf",
+        model=os.path.join(models_path_prefix, "llava-hf/llava-1.5-7b-hf"),
         max_num_batched_tokens=1024,
     )
     mm_positions = [[PlaceholderRange(offset=100, length=600)]
@@ -303,7 +305,7 @@ def test_schedule_partial_requests():
 def test_no_mm_input_chunking():
     # Disable multimodal input chunking.
     scheduler = create_scheduler(
-        model="llava-hf/llava-1.5-7b-hf",
+        model=os.path.join(models_path_prefix, "llava-hf/llava-1.5-7b-hf"),
         max_num_batched_tokens=1024,
         disable_chunked_mm_input=True,
         max_model_len=2048,
@@ -347,7 +349,7 @@ def test_no_mm_input_chunking():
     # of a max_num_batched_tokens for the mm input.
     with pytest.raises(ValueError):
         _ = create_scheduler(
-            model="llava-hf/llava-1.5-7b-hf",
+            model=os.path.join(models_path_prefix, "llava-hf/llava-1.5-7b-hf"),
             max_num_batched_tokens=100,
             disable_chunked_mm_input=True,
         )
@@ -362,7 +364,7 @@ def test_schedule_concurrent_partial_requests(enable_prefix_caching: bool):
 
     """
     scheduler = create_scheduler(
-        model="facebook/opt-125m",
+        model=os.path.join(models_path_prefix, "facebook/opt-125m"),
         max_num_batched_tokens=1024,
         long_prefill_token_threshold=400,
         enable_prefix_caching=enable_prefix_caching,

tests/v1/core/test_scheduler_e2e.py

@@ -4,11 +4,12 @@ import os
 import pytest
 
 from vllm import LLM
+from ...utils import models_path_prefix
 
 if os.getenv("VLLM_USE_V1", "0") != "1":
     pytest.skip("Test package requires V1", allow_module_level=True)
 
-MODEL = "meta-llama/Llama-3.2-1B"
+MODEL = os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B")
 PROMPT = "Hello my name is Robert and I"
 
 

tests/v1/shutdown/test_delete.py

 # SPDX-License-Identifier: Apache-2.0
 """Test that we handle a startup Error and shutdown."""
 
+import os
 import pytest
 
 from tests.utils import wait_for_gpu_memory_to_clear
@@ -11,8 +12,9 @@ from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.sampling_params import RequestOutputKind
 from vllm.utils import cuda_device_count_stateless
 from vllm.v1.engine.async_llm import AsyncLLM
+from ...utils import models_path_prefix
 
-MODELS = ["meta-llama/Llama-3.2-1B"]
+MODELS = [os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B")]
 
 
 @pytest.mark.asyncio

tests/v1/shutdown/test_forward_error.py

@@ -3,6 +3,7 @@
 
 import asyncio
 
+import os
 import pytest
 
 from tests.utils import wait_for_gpu_memory_to_clear
@@ -14,8 +15,9 @@ from vllm.model_executor.models.llama import LlamaForCausalLM
 from vllm.utils import cuda_device_count_stateless
 from vllm.v1.engine.async_llm import AsyncLLM
 from vllm.v1.engine.exceptions import EngineDeadError
+from ...utils import models_path_prefix
 
-MODELS = ["meta-llama/Llama-3.2-1B"]
+MODELS = [os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B")]
 
 
 def evil_forward(self, *args, **kwargs):

tests/v1/shutdown/test_processor_error.py

@@ -3,6 +3,7 @@
 
 import asyncio
 
+import os
 import pytest
 
 from tests.v1.shutdown.utils import SHUTDOWN_TEST_TIMEOUT_SEC
@@ -12,8 +13,9 @@ from vllm.inputs.data import TokensPrompt
 from vllm.sampling_params import RequestOutputKind
 from vllm.v1.engine.async_llm import AsyncLLM
 from vllm.v1.engine.exceptions import EngineGenerateError
+from ...utils import models_path_prefix
 
-MODELS = ["meta-llama/Llama-3.2-1B"]
+MODELS = [os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B")]
 
 
 @pytest.mark.asyncio

tests/v1/shutdown/test_startup_error.py

 # SPDX-License-Identifier: Apache-2.0
 """Test that we handle a startup Error and shutdown."""
 
+import os
 import pytest
 
 from tests.utils import wait_for_gpu_memory_to_clear
@@ -12,8 +13,9 @@ from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.model_executor.models.llama import LlamaForCausalLM
 from vllm.utils import cuda_device_count_stateless
 from vllm.v1.engine.async_llm import AsyncLLM
+from ...utils import models_path_prefix
 
-MODELS = ["meta-llama/Llama-3.2-1B"]
+MODELS = [os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B")]
 
 
 def evil_method(self, *args, **kwargs):
@@ -69,7 +71,7 @@ def test_llm_startup_error(monkeypatch, model: str, tensor_parallel_size: int,
     Test profiling (forward()) and load weights failures.
     TODO(andy) - LLM without multiprocessing.
     """
-    if model != "meta-llama/Llama-3.2-1B":
+    if model != os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B"):
         pytest.skip(reason="Only test meta-llama/Llama-3.2-1B")
     if cuda_device_count_stateless() < tensor_parallel_size:
         pytest.skip(reason="Not enough CUDA devices")

tests/v1/spec_decode/test_max_len.py

 # SPDX-License-Identifier: Apache-2.0
 """Test whether spec decoding handles the max model length properly."""
 
+import os
 import pytest
 
 from vllm import LLM, SamplingParams
+from ...utils import models_path_prefix
 
 _PROMPTS = [
     "1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1",
@@ -21,7 +23,7 @@ def test_ngram_max_len(
         m.setenv("VLLM_USE_V1", "1")
 
         llm = LLM(
-            model="facebook/opt-125m",
+            model=os.path.join(models_path_prefix, "facebook/opt-125m"),
             max_model_len=100,
             enforce_eager=True,  # For faster initialization.
             speculative_config={
@@ -44,11 +46,11 @@ def test_eagle_max_len(
         m.setenv("VLLM_USE_V1", "1")
 
         llm = LLM(
-            model="meta-llama/Meta-Llama-3-8B-Instruct",
+            model=os.path.join(models_path_prefix, "meta-llama/Meta-Llama-3-8B-Instruct"),
             enforce_eager=True,  # For faster initialization.
             speculative_config={
                 "method": "eagle",
-                "model": "yuhuili/EAGLE-LLaMA3-Instruct-8B",
+                "model": os.path.join(models_path_prefix, "yuhuili/EAGLE-LLaMA3-Instruct-8B"),
                 "num_speculative_tokens": num_speculative_tokens,
             },
             max_model_len=100,

tests/v1/spec_decode/test_ngram.py

 # SPDX-License-Identifier: Apache-2.0
 
+import os
 import numpy as np
 
 from vllm.config import ModelConfig, SpeculativeConfig, VllmConfig
 from vllm.v1.spec_decode.ngram_proposer import (NgramProposer,
                                                 _find_subarray_kmp,
                                                 _kmp_lps_array)
+from ...utils import models_path_prefix
 
 
 def test_kmp_lps_array():
@@ -43,10 +45,10 @@ def test_ngram_proposer():
 
     def ngram_proposer(min_n: int, max_n: int, k: int) -> NgramProposer:
         # Dummy model config. Just to set max_model_len.
-        model_config = ModelConfig(model="facebook/opt-125m",
+        model_config = ModelConfig(model=os.path.join(models_path_prefix, "facebook/opt-125m"),
                                    task="generate",
                                    max_model_len=100,
-                                   tokenizer="facebook/opt-125m",
+                                   tokenizer=os.path.join(models_path_prefix, "facebook/opt-125m"),
                                    tokenizer_mode="auto",
                                    dtype="auto",
                                    seed=None,

tests/v1/test_async_llm_dp.py

@@ -14,9 +14,10 @@ from vllm.platforms import current_platform
 from vllm.sampling_params import RequestOutputKind
 from vllm.v1.engine.async_llm import AsyncLLM
 from vllm.v1.engine.core_client import DPAsyncMPClient
+from ..utils import models_path_prefix
 
 engine_args = AsyncEngineArgs(
-    model="ibm-research/PowerMoE-3b",
+    model=os.path.join(models_path_prefix, "ibm-research/PowerMoE-3b"),
     enforce_eager=True,
     disable_log_requests=True,
     tensor_parallel_size=int(os.getenv("TP_SIZE", 1)),

tests/v1/test_oracle.py

@@ -7,16 +7,17 @@ import vllm.envs as envs
 from vllm import LLM
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.engine.async_llm_engine import AsyncLLMEngine
+from ..utils import models_path_prefix
 
 UNSUPPORTED_MODELS_V1 = [
-    "openai/whisper-large-v3",  # transcription
-    "facebook/bart-large-cnn",  # encoder decoder
-    "mistralai/Mamba-Codestral-7B-v0.1",  # mamba
-    "ibm-ai-platform/Bamba-9B",  # hybrid
-    "BAAI/bge-m3",  # embedding
+    os.path.join(models_path_prefix, "openai/whisper-large-v3"),  # transcription
+    os.path.join(models_path_prefix, "facebook/bart-large-cnn"),  # encoder decoder
+    os.path.join(models_path_prefix, "mistralai/Mamba-Codestral-7B-v0.1"),  # mamba
+    os.path.join(models_path_prefix, "ibm-ai-platform/Bamba-9B"),  # hybrid
+    os.path.join(models_path_prefix, "BAAI/bge-m3"),  # embedding
 ]
 
-MODEL = "meta-llama/Llama-3.2-1B-Instruct"
+MODEL = os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B-Instruct")
 
 
 @pytest.mark.parametrize("model", UNSUPPORTED_MODELS_V1)

tests/v1/worker/test_gpu_model_runner.py

 # SPDX-License-Identifier: Apache-2.0
+import os
 import pytest
 
 from vllm.config import CacheConfig, ModelConfig, SchedulerConfig, VllmConfig
@@ -7,6 +8,7 @@ from vllm.v1.core.sched.output import (CachedRequestData, NewRequestData,
                                        SchedulerOutput)
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.worker.gpu_model_runner import GPUModelRunner
+from ...utils import models_path_prefix
 
 
 @pytest.fixture
@@ -17,9 +19,9 @@ def model_runner():
         max_model_len=512,
     )
     model_config = ModelConfig(
-        model="facebook/opt-125m",
+        model=os.path.join(models_path_prefix, "facebook/opt-125m"),
         task="generate",
-        tokenizer="facebook/opt-125m",
+        tokenizer=os.path.join(models_path_prefix, "facebook/opt-125m"),
         tokenizer_mode="auto",
         trust_remote_code=True,
         dtype="float16",