Merge branch 'v0.8.5.post1-dev' into v0.8.5-zero_overhead

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
@@ -140,4 +140,4 @@ def test_beam_search_passes_multimodal_data(
                 assert filtered_hf_output_ids[-1] == eos_token_id
                 filtered_hf_output_ids = filtered_hf_output_ids[:-1]
 
-            assert filtered_hf_output_ids == filtered_vllm_output_ids
+            assert filtered_hf_output_ids == filtered_vllm_output_ids
\ No newline at end of file

tests/spec_decode/e2e/test_compatibility.py

@@ -19,7 +19,7 @@ from ...utils import models_path_prefix
         {
             # Speculative max model len > overridden max model len should raise.
             "speculative_config": {
-                "model": "JackFram/llama-68m",
+                "model": os.path.join(models_path_prefix, "JackFram/llama-68m"),
                 "num_speculative_tokens": 5,
                 "max_model_len": 129,
             },
@@ -29,7 +29,7 @@ from ...utils import models_path_prefix
             # Speculative max model len > draft max model len should raise.
             # https://huggingface.co/JackFram/llama-68m/blob/3b606af5198a0b26762d589a3ee3d26ee6fa6c85/config.json#L12
             "speculative_config": {
-                "model": "JackFram/llama-68m",
+                "model": os.path.join(models_path_prefix, "JackFram/llama-68m"),
                 "num_speculative_tokens": 5,
                 "max_model_len": 2048 + 1,
             },
@@ -38,7 +38,7 @@ from ...utils import models_path_prefix
             # Speculative max model len > target max model len should raise.
             # https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/blob/9213176726f574b556790deb65791e0c5aa438b6/config.json#L18
             "speculative_config": {
-                "model": "JackFram/llama-68m",
+                "model": os.path.join(models_path_prefix, "JackFram/llama-68m"),
                 "num_speculative_tokens": 5,
                 "max_model_len": 131072 + 1,
             },
@@ -64,4 +64,4 @@ def test_spec_decode_xfail_spec_max_model_len(test_llm_generator):
 
     with pytest.raises(ValueError, match="cannot be larger than"):
         get_output_from_llm_generator(test_llm_generator, prompts,
-                                      sampling_params)
+                                      sampling_params)
\ No newline at end of file

tests/spec_decode/e2e/test_eagle_correctness.py

@@ -332,14 +332,14 @@ def test_eagle_disable_queue(vllm_runner, common_llm_kwargs,
         "dtype": "float16",
 
         # Main model
-        "model_name": "meta-llama/Llama-2-7b-chat-hf",
+        "model_name": os.path.join(models_path_prefix, "meta-llama/Llama-2-7b-chat-hf"),
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
 @pytest.mark.parametrize("test_llm_kwargs", [
     {
         "speculative_config": {
-            "model": "yuhuili/EAGLE-llama2-chat-7B",
+            "model": os.path.join(models_path_prefix, "yuhuili/EAGLE-llama2-chat-7B"),
             "num_speculative_tokens": MAX_SPEC_TOKENS,
         },
     },
@@ -382,14 +382,14 @@ def test_llama2_eagle_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
         "dtype": "float16",
 
         # Main model
-        "model_name": "meta-llama/Meta-Llama-3-8B-Instruct",
+        "model_name": os.path.join(models_path_prefix, "meta-llama/Meta-Llama-3-8B-Instruct"),
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
 @pytest.mark.parametrize("test_llm_kwargs", [
     {
         "speculative_config": {
-            "model": "yuhuili/EAGLE-LLaMA3-Instruct-8B",
+            "model": os.path.join(models_path_prefix, "yuhuili/EAGLE-LLaMA3-Instruct-8B"),
             "num_speculative_tokens": MAX_SPEC_TOKENS,
         },
     },
@@ -432,14 +432,14 @@ def test_llama3_eagle_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
         "dtype": "float16",
 
         # Main model
-        "model_name": "Qwen/Qwen2-7B-Instruct",
+        "model_name": os.path.join(models_path_prefix, "Qwen/Qwen2-7B-Instruct"),
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
 @pytest.mark.parametrize("test_llm_kwargs", [
     {
         "speculative_config": {
-            "model": "yuhuili/EAGLE-Qwen2-7B-Instruct",
+            "model": os.path.join(models_path_prefix, "yuhuili/EAGLE-Qwen2-7B-Instruct"),
             "num_speculative_tokens": MAX_SPEC_TOKENS,
         },
     },

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,
     },
@@ -151,4 +151,4 @@ def test_mqa_scorer(vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
                                   batch_size,
                                   max_output_len=output_len,
                                   seed=seed,
-                                  temperature=0.0)
+                                  temperature=0.0)
\ No newline at end of file

tests/spec_decode/e2e/test_mtp_correctness.py

@@ -20,12 +20,14 @@ With those tests, we can say at least, mtp would not break the
 correctess for the target model outputs.
 """
 
+import os
 import pytest
 
 from .conftest import run_equality_correctness_test
+from ...utils import models_path_prefix
 
 # main model
-MAIN_MODEL = "luccafong/deepseek_mtp_main_random"
+MAIN_MODEL = os.path.join(models_path_prefix, "luccafong/deepseek_mtp_main_random")
 
 # max. number of speculative tokens: this corresponds to
 # num_nextn_predict_layers in the config.json of the speculator model.
@@ -329,4 +331,4 @@ def test_mtp_disable_queue(vllm_runner, common_llm_kwargs,
 
 if __name__ == "__main__":
     import pytest
-    pytest.main([__file__])
+    pytest.main([__file__])
\ No newline at end of file

tests/spec_decode/e2e/test_ngram_correctness.py

@@ -334,7 +334,7 @@ def test_ngram_disable_queue(vllm_runner, common_llm_kwargs,
 @pytest.mark.parametrize(
     "common_llm_kwargs",
     [{
-        "model_name": "JackFram/llama-68m",
+        "model_name": os.path.join(models_path_prefix, "JackFram/llama-68m"),
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
@@ -372,4 +372,4 @@ def test_ngram_scorer(vllm_runner, common_llm_kwargs,
                                   batch_size,
                                   max_output_len=output_len,
                                   seed=seed,
-                                  temperature=0.0)
+                                  temperature=0.0)
\ No newline at end of file

tests/spec_decode/test_multi_step_worker.py

@@ -484,107 +484,107 @@ def test_multi_step_with_batch_expansion_incorrect_output():
     assert (num_mismatch > 0)
 
 
-@torch.inference_mode()
-@pytest.mark.parametrize('num_steps', [1, 2, 3, 4])
-# The choice of backends forces the multi_step_worker to choose between
-# the vanilla model_runner and TP1DraftModelRunner and that we can test
-# both code paths.
-@pytest.mark.parametrize('attn_backend',
-                         [_Backend.XFORMERS, _Backend.FLASH_ATTN])
-def test_multi_step_correct_kvcache(num_steps, attn_backend):
-    """Verify that the KV cache of the draft model 
-    is correctly updated for sequences with bonus token.
-    """
-    seed = 100
-    model_name = "JackFram/llama-68m"
-
-    block_size = 16
-    num_gpu_blocks = 2048 // block_size
-    batch_size = 1
-
-    with global_force_attn_backend_context_manager(attn_backend):
-        dtype = 'float16' if attn_backend == _Backend.FLASH_ATTN else 'float32'
-        multi_step_worker = create_worker(MultiStepWorker,
-                                          model_name,
-                                          block_size,
-                                          num_gpu_blocks,
-                                          seed,
-                                          model_runner_cls=TP1DraftModelRunner,
-                                          dtype=dtype)
-        multi_step_worker.set_include_gpu_probs_tensor()
-        worker = create_worker(Worker,
-                               model_name,
-                               block_size,
-                               num_gpu_blocks,
-                               seed,
-                               dtype=dtype)
-
-        prompts = [[0] for _ in range(batch_size)]
-        # Already generate two tokens for the sequence
-        # so that we can simulate the bonus token case
-        multi_step_continuations = [[
-            random.randint(0, 1000),
-            random.randint(0, 1000)
-        ] for _ in prompts]
-        final_prompt_lens = [len(prompt) + 2 + num_steps for prompt in prompts]
-
-        seq_ids_with_bonus_token_in_last_step = set(range(batch_size))
-        seq_group_metadata_list = create_seq_group_metadata_from_prompts(
-            prompts,
-            num_gpu_blocks,
-            block_size,
-            continuations=multi_step_continuations,
-            final_prompt_lens=final_prompt_lens)
-
-        # Run multi-step.
-        zero_kv_cache(multi_step_worker.cache_engine)
-        multi_step_worker.sampler_output(execute_model_req=ExecuteModelRequest(
-            seq_group_metadata_list=seq_group_metadata_list),
-                                         sample_len=num_steps,
-                                         seq_ids_with_bonus_token_in_last_step=
-                                         seq_ids_with_bonus_token_in_last_step)
-
-        # Run single-step repeatedly.
-        zero_kv_cache(worker.cache_engine)
-        # Generate the kv cache for the bonus token first
-        single_step_continuations = [c[:1] for c in multi_step_continuations]
-        seq_group_metadata_list = create_seq_group_metadata_from_prompts(
-            prompts,
-            num_gpu_blocks,
-            block_size,
-            continuations=single_step_continuations,
-            final_prompt_lens=final_prompt_lens)
-        single_step_output = worker.execute_model(
-            execute_model_req=ExecuteModelRequest(
-                seq_group_metadata_list=seq_group_metadata_list))
-        for _ in range(num_steps):
-            seq_group_metadata_list = create_seq_group_metadata_from_prompts(
-                prompts,
-                num_gpu_blocks,
-                block_size,
-                continuations=multi_step_continuations,
-                final_prompt_lens=final_prompt_lens)
-
-            single_step_output = worker.execute_model(
-                execute_model_req=ExecuteModelRequest(
-                    seq_group_metadata_list=seq_group_metadata_list))
-
-            for i, seq_group_output in enumerate(single_step_output[-1]):
-                multi_step_continuations[i].append(
-                    seq_group_output.samples[0].output_token)
-
-        # Verify that the KV cache of the single-step and
-        # multi-step workers are the same.
-        single_step_gpu_cache = worker.cache_engine[0].gpu_cache
-        multi_step_gpu_cache = multi_step_worker.cache_engine[0].gpu_cache
-        num_layers = len(single_step_gpu_cache)
-        allclose = lambda a, b: torch.allclose(
-            a.cuda(), b.cuda(), rtol=1e-2, atol=1e-2)
-        for i in range(num_layers):
-            assert allclose(single_step_gpu_cache[i][0],
-                            multi_step_gpu_cache[i][0])
-            assert allclose(single_step_gpu_cache[i][1],
-                            multi_step_gpu_cache[i][1])
+# @torch.inference_mode()
+# @pytest.mark.parametrize('num_steps', [1, 2, 3, 4])
+# # The choice of backends forces the multi_step_worker to choose between
+# # the vanilla model_runner and TP1DraftModelRunner and that we can test
+# # both code paths.
+# @pytest.mark.parametrize('attn_backend',
+#                          [_Backend.XFORMERS, _Backend.FLASH_ATTN])
+# def test_multi_step_correct_kvcache(num_steps, attn_backend):
+#     """Verify that the KV cache of the draft model 
+#     is correctly updated for sequences with bonus token.
+#     """
+#     seed = 100
+#     model_name = "JackFram/llama-68m"
+
+#     block_size = 16
+#     num_gpu_blocks = 2048 // block_size
+#     batch_size = 1
+
+#     with global_force_attn_backend_context_manager(attn_backend):
+#         dtype = 'float16' if attn_backend == _Backend.FLASH_ATTN else 'float32'
+#         multi_step_worker = create_worker(MultiStepWorker,
+#                                           model_name,
+#                                           block_size,
+#                                           num_gpu_blocks,
+#                                           seed,
+#                                           model_runner_cls=TP1DraftModelRunner,
+#                                           dtype=dtype)
+#         multi_step_worker.set_include_gpu_probs_tensor()
+#         worker = create_worker(Worker,
+#                                model_name,
+#                                block_size,
+#                                num_gpu_blocks,
+#                                seed,
+#                                dtype=dtype)
+
+#         prompts = [[0] for _ in range(batch_size)]
+#         # Already generate two tokens for the sequence
+#         # so that we can simulate the bonus token case
+#         multi_step_continuations = [[
+#             random.randint(0, 1000),
+#             random.randint(0, 1000)
+#         ] for _ in prompts]
+#         final_prompt_lens = [len(prompt) + 2 + num_steps for prompt in prompts]
+
+#         seq_ids_with_bonus_token_in_last_step = set(range(batch_size))
+#         seq_group_metadata_list = create_seq_group_metadata_from_prompts(
+#             prompts,
+#             num_gpu_blocks,
+#             block_size,
+#             continuations=multi_step_continuations,
+#             final_prompt_lens=final_prompt_lens)
+
+#         # Run multi-step.
+#         zero_kv_cache(multi_step_worker.cache_engine)
+#         multi_step_worker.sampler_output(execute_model_req=ExecuteModelRequest(
+#             seq_group_metadata_list=seq_group_metadata_list),
+#                                          sample_len=num_steps,
+#                                          seq_ids_with_bonus_token_in_last_step=
+#                                          seq_ids_with_bonus_token_in_last_step)
+
+#         # Run single-step repeatedly.
+#         zero_kv_cache(worker.cache_engine)
+#         # Generate the kv cache for the bonus token first
+#         single_step_continuations = [c[:1] for c in multi_step_continuations]
+#         seq_group_metadata_list = create_seq_group_metadata_from_prompts(
+#             prompts,
+#             num_gpu_blocks,
+#             block_size,
+#             continuations=single_step_continuations,
+#             final_prompt_lens=final_prompt_lens)
+#         single_step_output = worker.execute_model(
+#             execute_model_req=ExecuteModelRequest(
+#                 seq_group_metadata_list=seq_group_metadata_list))
+#         for _ in range(num_steps):
+#             seq_group_metadata_list = create_seq_group_metadata_from_prompts(
+#                 prompts,
+#                 num_gpu_blocks,
+#                 block_size,
+#                 continuations=multi_step_continuations,
+#                 final_prompt_lens=final_prompt_lens)
+
+#             single_step_output = worker.execute_model(
+#                 execute_model_req=ExecuteModelRequest(
+#                     seq_group_metadata_list=seq_group_metadata_list))
+
+#             for i, seq_group_output in enumerate(single_step_output[-1]):
+#                 multi_step_continuations[i].append(
+#                     seq_group_output.samples[0].output_token)
+
+#         # Verify that the KV cache of the single-step and
+#         # multi-step workers are the same.
+#         single_step_gpu_cache = worker.cache_engine[0].gpu_cache
+#         multi_step_gpu_cache = multi_step_worker.cache_engine[0].gpu_cache
+#         num_layers = len(single_step_gpu_cache)
+#         allclose = lambda a, b: torch.allclose(
+#             a.cuda(), b.cuda(), rtol=1e-2, atol=1e-2)
+#         for i in range(num_layers):
+#             assert allclose(single_step_gpu_cache[i][0],
+#                             multi_step_gpu_cache[i][0])
+#             assert allclose(single_step_gpu_cache[i][1],
+#                             multi_step_gpu_cache[i][1])
 
 
 @torch.inference_mode()

tests/spec_decode/test_spec_decode_worker.py

@@ -5,6 +5,7 @@ from collections import defaultdict
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 
+import os
 import pytest
 import torch
 
@@ -24,6 +25,7 @@ from vllm.worker.worker import Worker
 from .test_utils import mock_spec_decode_sampler
 from .utils import (create_batch, create_sampler_output_list, create_worker,
                     mock_worker)
+from ..utils import models_path_prefix
 
 
 @pytest.mark.parametrize('k', [1, 2, 6])
@@ -918,14 +920,14 @@ def test_correctly_load_weight_for_eagle():
     num_gpu_blocks = 8096 // block_size
     target_worker = create_worker(
         Worker,
-        "JackFram/llama-68m",
+        os.path.join(models_path_prefix, "JackFram/llama-68m"),
         block_size,
         num_gpu_blocks,
         seed,
     )
     draft_worker = create_worker(
         MultiStepWorker,
-        "abhigoyal/vllm-eagle-llama-68m-random",
+        os.path.join(models_path_prefix, "abhigoyal/vllm-eagle-llama-68m-random"),
         block_size,
         num_gpu_blocks,
         seed,
@@ -941,4 +943,4 @@ def test_correctly_load_weight_for_eagle():
         target_worker.model_runner.model.lm_head.weight.data)
     assert torch.allclose(
         worker.proposer_worker.worker.model_runner.model.lm_head.weight.data,
-        worker.scorer_worker.model_runner.model.lm_head.weight.data)
+        worker.scorer_worker.model_runner.model.lm_head.weight.data)
\ No newline at end of file

tests/tensorizer_loader/test_tensorizer.py

@@ -7,6 +7,7 @@ import pathlib
 import subprocess
 from functools import partial
 from unittest.mock import MagicMock, patch
+from typing import List, Tuple, Optional
 
 import openai
 import pytest
@@ -15,6 +16,7 @@ from huggingface_hub import snapshot_download
 
 from vllm import EngineArgs, LLMEngine, RequestOutput, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
+from vllm.lora.request import LoRARequest
 # yapf conflicts with isort for this docstring
 # yapf: disable
 from vllm.model_executor.model_loader.tensorizer import (TensorizerConfig,
@@ -243,7 +245,7 @@ def test_vllm_model_can_load_with_lora(vllm_runner, tmp_path):
         EXAMPLES_PATH / "offline_inference/multilora_inference.py",
     )
 
-    model_ref = "meta-llama/Llama-2-7b-hf"
+    model_ref = os.path.join(models_path_prefix, "meta-llama/Llama-2-7b-hf")
     # lora_path = snapshot_download(repo_id="yard1/llama-2-7b-sql-lora-test")
     lora_path = os.path.join(models_path_prefix, "yard1/llama-2-7b-sql-lora-test")
     test_prompts = multilora_inference.create_test_prompts(lora_path)
@@ -431,4 +433,4 @@ def test_vllm_tensorized_model_has_same_outputs(vllm_runner, tmp_path):
             prompts, sampling_params)
         # noqa: E501
 
-        assert outputs == deserialized_outputs
+        assert outputs == deserialized_outputs
\ No newline at end of file

tests/test_config.py

@@ -142,7 +142,7 @@ def test_get_sliding_window():
 @pytest.mark.skipif(current_platform.is_rocm(),
                     reason="Xformers backend is not supported on ROCm.")
 def test_get_pooling_config():
-    model_id = "sentence-transformers/all-MiniLM-L12-v2"
+    model_id = os.path.join(models_path_prefix, "sentence-transformers/all-MiniLM-L12-v2")
     model_config = ModelConfig(
         model_id,
         task="auto",
@@ -164,7 +164,7 @@ def test_get_pooling_config():
 @pytest.mark.skipif(current_platform.is_rocm(),
                     reason="Xformers backend is not supported on ROCm.")
 def test_get_pooling_config_from_args():
-    model_id = "sentence-transformers/all-MiniLM-L12-v2"
+    model_id = os.path.join(models_path_prefix, "sentence-transformers/all-MiniLM-L12-v2")
     model_config = ModelConfig(model_id,
                                task="auto",
                                tokenizer=model_id,
@@ -273,10 +273,10 @@ def test_rope_customization():
 @pytest.mark.skipif(current_platform.is_rocm(),
                     reason="Encoder Decoder models not supported on ROCm.")
 @pytest.mark.parametrize(("model_id", "is_encoder_decoder"), [
-    ("facebook/opt-125m", False),
-    ("facebook/bart-base", True),
-    ("meta-llama/Llama-3.2-1B-Instruct", False),
-    ("meta-llama/Llama-3.2-11B-Vision", True),
+    (os.path.join(models_path_prefix, "facebook/opt-125m"), False),
+    (os.path.join(models_path_prefix, "facebook/bart-base"), True),
+    (os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B-Instruct"), False),
+    (os.path.join(models_path_prefix, "meta-llama/Llama-3.2-11B-Vision"), True),
 ])
 def test_is_encoder_decoder(model_id, is_encoder_decoder):
     config = ModelConfig(
@@ -293,8 +293,8 @@ def test_is_encoder_decoder(model_id, is_encoder_decoder):
 
 
 @pytest.mark.parametrize(("model_id", "uses_mrope"), [
-    ("facebook/opt-125m", False),
-    ("Qwen/Qwen2-VL-2B-Instruct", True),
+    (os.path.join(models_path_prefix, "facebook/opt-125m"), False),
+    (os.path.join(models_path_prefix, "Qwen/Qwen2-VL-2B-Instruct"), True),
 ])
 def test_uses_mrope(model_id, uses_mrope):
     config = ModelConfig(
@@ -311,7 +311,7 @@ def test_uses_mrope(model_id, uses_mrope):
 
 
 def test_generation_config_loading():
-    model_id = "Qwen/Qwen2.5-1.5B-Instruct"
+    model_id = os.path.join(models_path_prefix, "Qwen/Qwen2.5-1.5B-Instruct")
 
     # When set generation_config to "vllm", the default generation config
     # will not be loaded.
@@ -377,4 +377,4 @@ def test_generation_config_loading():
         generation_config="vllm",
         override_generation_config=override_generation_config)
 
-    assert model_config.get_diff_sampling_param() == override_generation_config
+    assert model_config.get_diff_sampling_param() == override_generation_config
\ No newline at end of file

tests/test_regression.py

@@ -5,6 +5,7 @@ It should include tests that are reported by users and making sure they
 will never happen again.
 
 """
+import os
 import gc
 
 import pytest
@@ -12,8 +13,9 @@ import torch
 
 from vllm import LLM, SamplingParams
 
-from utils import models_path_prefix
-import os
+from .utils import models_path_prefix
+from vllm.utils import SUPPORT_TC, gpuname
+import vllm.envs as envs
 
 
 @pytest.mark.skip(reason="In V1, we reject tokens > max_seq_len")
@@ -23,7 +25,7 @@ def test_duplicated_ignored_sequence_group():
     sampling_params = SamplingParams(temperature=0.01,
                                      top_p=0.1,
                                      max_tokens=256)
-    llm = LLM(model="distilbert/distilgpt2",
+    llm = LLM(model=os.path.join(models_path_prefix, "distilbert/distilgpt2"),
               max_num_batched_tokens=4096,
               tensor_parallel_size=1)
     prompts = ["This is a short prompt", "This is a very long prompt " * 1000]
@@ -36,9 +38,15 @@ def test_max_tokens_none():
     sampling_params = SamplingParams(temperature=0.01,
                                      top_p=0.1,
                                      max_tokens=None)
-    llm = LLM(model="distilbert/distilgpt2",
-              max_num_batched_tokens=4096,
-              tensor_parallel_size=1)
+    if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND:
+        llm = LLM(model=os.path.join(models_path_prefix, "distilbert/distilgpt2"),
+                max_num_batched_tokens=4096,
+                tensor_parallel_size=1,
+                block_size=64)
+    else:
+        llm = LLM(model=os.path.join(models_path_prefix, "distilbert/distilgpt2"),
+                max_num_batched_tokens=4096,
+                tensor_parallel_size=1)
     prompts = ["Just say hello!"]
     outputs = llm.generate(prompts, sampling_params=sampling_params)
 
@@ -46,7 +54,7 @@ def test_max_tokens_none():
 
 
 def test_gc():
-    llm = LLM(model="distilbert/distilgpt2", enforce_eager=True)
+    llm = LLM(model=os.path.join(models_path_prefix, "distilbert/distilgpt2"), enforce_eager=True)
     del llm
 
     gc.collect()
@@ -63,7 +71,10 @@ def test_model_from_modelscope(monkeypatch: pytest.MonkeyPatch):
     # model: https://modelscope.cn/models/qwen/Qwen1.5-0.5B-Chat/summary
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_MODELSCOPE", "True")
-        llm = LLM(model=os.path.join(models_path_prefix, "qwen/Qwen1.5-0.5B-Chat"))
+        if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND:
+            llm = LLM(model=os.path.join(models_path_prefix, "qwen/Qwen1.5-0.5B-Chat"), block_size=64)
+        else:
+            llm = LLM(model=os.path.join(models_path_prefix, "qwen/Qwen1.5-0.5B-Chat"))
 
         prompts = [
             "Hello, my name is",
@@ -74,4 +85,4 @@ def test_model_from_modelscope(monkeypatch: pytest.MonkeyPatch):
         sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
 
         outputs = llm.generate(prompts, sampling_params)
-        assert len(outputs) == 4
+        assert len(outputs) == 4
\ No newline at end of file

tests/test_sampling_params.py

@@ -2,13 +2,15 @@
 """Tests for the SamplingParams class.
 """
 
+import os
 import pytest
 
 from vllm import SamplingParams
 from vllm.config import ModelConfig
 from vllm.entrypoints.openai.protocol import ChatCompletionRequest
+from utils import models_path_prefix
 
-MODEL_NAME = "Qwen/Qwen1.5-7B"
+MODEL_NAME = os.path.join(models_path_prefix, "Qwen/Qwen1.5-7B")
 
 
 def test_max_tokens_none():

tests/test_utils.py

@@ -8,6 +8,7 @@ import socket
 from collections.abc import AsyncIterator
 from unittest.mock import patch
 
+import os
 import pytest
 import torch
 from vllm_test_utils.monitor import monitor

tests/tool_use/utils.py

@@ -66,7 +66,7 @@ CONFIGS: dict[str, ServerConfig] = {
     },
     "llama": {
         "model":
-        "meta-llama/Meta-Llama-3.1-8B-Instruct",
+        os.path.join(models_path_prefix, "meta-llama/Meta-Llama-3.1-8B-Instruct"),
         "arguments": [
             "--enforce-eager", "--no-enable-prefix-caching",
             "--tool-call-parser", "llama3_json", "--chat-template",
@@ -77,7 +77,7 @@ CONFIGS: dict[str, ServerConfig] = {
     },
     "llama3.2": {
         "model":
-        "meta-llama/Llama-3.2-3B-Instruct",
+        os.path.join(models_path_prefix, "meta-llama/Llama-3.2-3B-Instruct"),
         "arguments": [
             "--enforce-eager", "--no-enable-prefix-caching",
             "--tool-call-parser", "llama3_json", "--chat-template",
@@ -88,7 +88,7 @@ CONFIGS: dict[str, ServerConfig] = {
     },
     "llama4": {
         "model":
-        "meta-llama/Llama-4-Scout-17B-16E-Instruct",
+        os.path.join(models_path_prefix, "meta-llama/Llama-4-Scout-17B-16E-Instruct"),
         "arguments": [
             "--enforce-eager", "--no-enable-prefix-caching",
             "--tool-call-parser", "pythonic", "--chat-template",
@@ -103,7 +103,7 @@ CONFIGS: dict[str, ServerConfig] = {
     },
     "llama4_json": {
         "model":
-        "meta-llama/Llama-4-Scout-17B-16E-Instruct",
+        os.path.join(models_path_prefix, "meta-llama/Llama-4-Scout-17B-16E-Instruct"),
         "arguments": [
             "--enforce-eager", "--no-enable-prefix-caching", "-tp", "4",
             "--distributed-executor-backend", "mp", "--tool-call-parser",
@@ -149,7 +149,7 @@ CONFIGS: dict[str, ServerConfig] = {
     # },
     "granite-3.0-8b": {
         "model":
-        "ibm-granite/granite-3.0-8b-instruct",
+        os.path.join(models_path_prefix, "ibm-granite/granite-3.0-8b-instruct"),
         "arguments": [
             "--enforce-eager", "--no-enable-prefix-caching",
             "--tool-call-parser", "granite", "--chat-template",
@@ -158,7 +158,7 @@ CONFIGS: dict[str, ServerConfig] = {
     },
     "granite-3.1-8b": {
         "model":
-        "ibm-granite/granite-3.1-8b-instruct",
+        os.path.join(models_path_prefix, "ibm-granite/granite-3.1-8b-instruct"),
         "arguments": [
             "--enforce-eager",
             "--no-enable-prefix-caching",
@@ -170,7 +170,7 @@ CONFIGS: dict[str, ServerConfig] = {
     },
     "internlm": {
         "model":
-        "internlm/internlm2_5-7b-chat",
+        os.path.join(models_path_prefix, "internlm/internlm2_5-7b-chat"),
         "arguments": [
             "--enforce-eager", "--no-enable-prefix-caching",
             "--tool-call-parser", "internlm", "--chat-template",
@@ -183,7 +183,7 @@ CONFIGS: dict[str, ServerConfig] = {
     },
     "toolACE": {
         "model":
-        "Team-ACE/ToolACE-8B",
+        os.path.join(models_path_prefix, "Team-ACE/ToolACE-8B"),
         "arguments": [
             "--enforce-eager", "--no-enable-prefix-caching",
             "--tool-call-parser", "pythonic", "--chat-template",
@@ -361,4 +361,4 @@ MESSAGES_WITH_PARALLEL_TOOL_RESPONSE: list[ChatCompletionMessageParam] = [{
     "content":
     "The weather in Orlando FL is 78 degrees fahrenheit with clear"
     "skies."
-}]
+}]
\ No newline at end of file

tests/tpu/test_quantization_accuracy.py → tests/tpu/untest_quantization_accuracy.py

@@ -4,6 +4,8 @@ from dataclasses import dataclass
 
 import lm_eval
 import pytest
+import os
+from ..utils import models_path_prefix
 
 TASK = "gsm8k"
 FILTER = "exact_match,strict-match"
@@ -23,7 +25,7 @@ class GSM8KAccuracyTestConfig:
 # NOTE: Accuracy scores measured on GPUs.
 ACCURACY_CONFIGS = [
     GSM8KAccuracyTestConfig(
-        model_name="neuralmagic/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+        model_name=os.path.join(models_path_prefix, "neuralmagic/Meta-Llama-3.1-8B-Instruct-quantized.w8a8"),
         excepted_value=0.76),  # no bias
     # NOTE(rob): We cannot re-initialize vLLM in the same process for TPU,
     # so only one of these tests can run in a single call to pytest. As
@@ -48,4 +50,4 @@ def test_gsm8k_correctness(config: GSM8KAccuracyTestConfig):
     measured_value = results["results"][TASK][FILTER]
     assert (measured_value - RTOL < EXPECTED_VALUE
             and measured_value + RTOL > EXPECTED_VALUE
-            ), f"Expected: {EXPECTED_VALUE} |  Measured: {measured_value}"
+            ), f"Expected: {EXPECTED_VALUE} |  Measured: {measured_value}"
\ No newline at end of file

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,
@@ -1241,4 +1243,4 @@ def test_memory_leak():
         scheduler.update_from_output(scheduler_output, model_runner_output)
 
     # Confirm no memory leak.
-    assert_scheduler_empty(scheduler)
+    assert_scheduler_empty(scheduler)
\ No newline at end of file

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"
 
 
@@ -26,4 +27,4 @@ def test_concurrent_partial_prefill(model):
     outputs = model.generate([PROMPT] * 3)
     assert len(outputs) == 3
     for output in outputs:
-        assert len(output.outputs) == 1
+        assert len(output.outputs) == 1
\ No newline at end of file