[tests] fix tests of core, engine and detokenizer

tests/compile/test_basic_correctness.py

@@ -29,18 +29,18 @@ class TestSetting:
     "test_setting",
     [
         # basic llama model
-        TestSetting(
-            model="meta-llama/Llama-3.2-1B-Instruct",
-            model_args=[],
-            pp_size=2,
-            tp_size=2,
-            attn_backend="FLASHINFER",
-            method="generate",
-            fullgraph=True,
-        ),
+        # TestSetting(
+        #     model=os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B-Instruct"),
+        #     model_args=[],
+        #     pp_size=2,
+        #     tp_size=2,
+        #     attn_backend="FLASHINFER",
+        #     method="generate",
+        #     fullgraph=True,
+        # ),
         # llama model with quantization
         TestSetting(
-            model="TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ",
+            model=os.path.join(models_path_prefix, "TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ"),
             model_args=["--quantization", "gptq"],
             pp_size=1,
             tp_size=1,
@@ -50,7 +50,7 @@ class TestSetting:
         ),
         # MoE model
         TestSetting(
-            model="ibm/PowerMoE-3b",
+            model=os.path.join(models_path_prefix, "ibm/PowerMoE-3b"),
             model_args=[],
             pp_size=1,
             tp_size=2,
@@ -60,7 +60,7 @@ class TestSetting:
         ),
         # embedding model
         TestSetting(
-            model="BAAI/bge-multilingual-gemma2",
+            model=os.path.join(models_path_prefix, "BAAI/bge-multilingual-gemma2"),
             model_args=["--task", "embed", "--dtype", "bfloat16"],
             pp_size=1,
             tp_size=1,
@@ -69,18 +69,18 @@ class TestSetting:
             fullgraph=True,
         ),
         # encoder-based embedding model (BERT)
-        TestSetting(
-            model="BAAI/bge-base-en-v1.5",
-            model_args=["--task", "embed"],
-            pp_size=1,
-            tp_size=1,
-            attn_backend="XFORMERS",
-            method="encode",
-            fullgraph=True,
-        ),
+        # TestSetting(
+        #     model=os.path.join(models_path_prefix, "BAAI/bge-base-en-v1.5"),
+        #     model_args=["--task", "embed"],
+        #     pp_size=1,
+        #     tp_size=1,
+        #     attn_backend="XFORMERS",
+        #     method="encode",
+        #     fullgraph=True,
+        # ),
         # vision language model
         TestSetting(
-            model="microsoft/Phi-3.5-vision-instruct",
+            model=os.path.join(models_path_prefix, "microsoft/Phi-3.5-vision-instruct"),
             model_args=["--trust-remote-code", "--max-model-len", "2048"],
             pp_size=2,
             tp_size=1,
@@ -146,4 +146,4 @@ def test_compile_correctness(
                 all_envs[-1][
                     "VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE"] = "0"  # type: ignore
 
-        compare_all_settings(model, all_args * 3, all_envs, method=method)
+        compare_all_settings(model, all_args * 3, all_envs, method=method)
\ No newline at end of file

tests/core/block/e2e/test_correctness.py

@@ -9,6 +9,8 @@ from vllm import SamplingParams
 from .conftest import get_token_ids_from_llm_generator
 import os
 from ....utils import models_path_prefix
+import vllm.envs as envs
+from vllm.utils import SUPPORT_TC, gpuname
 
 
 @pytest.mark.parametrize(
@@ -21,7 +23,7 @@ from ....utils import models_path_prefix
         "enforce_eager": True,
 
         # Allow only 5 sequences of ~1024 tokens in worst case.
-        "block_size": 16,
+        "block_size": 64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
         "num_gpu_blocks_override": 5 * (64 + 1),
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@@ -104,19 +106,19 @@ def test_block_manager_with_preemption(baseline_llm_generator,
     "per_test_common_llm_kwargs",
     [
         {
-            "block_size": 16,
+            "block_size": 64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
 
             # Allow only 2 sequences of ~128 tokens in worst case.
             # Note 8 = 128/block_size
             "num_gpu_blocks_override": 2 * (8 + 1),
         },
-        {
-            "block_size": 8,
+        # { 
+        #     "block_size": 8,
 
-            # Allow only 2 sequences of ~128 tokens in worst case.
-            # Note 16 = 128/block_size
-            "num_gpu_blocks_override": 2 * (16 + 2),
-        }
+        #     # Allow only 2 sequences of ~128 tokens in worst case.
+        #     # Note 16 = 128/block_size
+        #     "num_gpu_blocks_override": 2 * (16 + 2),
+        # }
     ])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{
     "num_lookahead_slots": 0,
@@ -197,15 +199,15 @@ def test_lookahead_greedy_equality_with_preemption(baseline_llm_generator,
     ])
 @pytest.mark.parametrize("per_test_common_llm_kwargs",
                          [{
-                             "block_size": 16,
+                             "block_size": 64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
                              "max_num_batched_tokens": 2,
                              "max_num_seqs": 2,
                          }, {
-                             "block_size": 16,
+                             "block_size": 64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
                              "max_num_batched_tokens": 3,
                              "max_num_seqs": 2,
                          }, {
-                             "block_size": 16,
+                             "block_size": 64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
                              "max_num_batched_tokens": 256,
                              "max_num_seqs": 10,
                          }])
@@ -271,7 +273,7 @@ def test_chunked_prefill_block_manager(baseline_llm_generator,
         "enforce_eager": True,
 
         # Allow only 5 sequences of ~1024 tokens in worst case.
-        "block_size": 16,
+        "block_size": 64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
         "num_gpu_blocks_override": 5 * (64 + 1),
 
         # Enable prefill cache
@@ -352,7 +354,7 @@ def test_block_manager_prefix_caching_enabled_with_preemption(
         "enforce_eager": True,
 
         # Allow only 5 sequences of ~1024 tokens in worst case.
-        "block_size": 16,
+        "block_size": 64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
         "num_gpu_blocks_override": 5 * (64 + 1),
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@@ -427,7 +429,7 @@ def test_auto_prefix_caching_with_preemption(baseline_llm_generator,
 
         # we keep the blocks small, so that hit eviction quickly
         "max_model_len": 48,
-        "block_size": 16,
+        "block_size": 64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
         "num_gpu_blocks_override": 3,
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@@ -477,4 +479,4 @@ def test_auto_prefix_caching_after_evition_start(baseline_llm_generator,
                                                     test_token_ids):
         assert expected_token_ids == actual_token_ids
 
-    assert baseline_token_ids == test_token_ids
+    assert baseline_token_ids == test_token_ids
\ No newline at end of file

tests/core/test_chunked_prefill_scheduler.py

@@ -2,6 +2,7 @@
 
 from unittest.mock import MagicMock
 
+import os
 import pytest  # noqa
 
 from vllm.config import CacheConfig, SchedulerConfig
@@ -12,6 +13,9 @@ from vllm.sampling_params import SamplingParams
 from vllm.sequence import Logprob, SequenceGroup
 
 from .utils import create_dummy_prompt
+from ..utils import models_path_prefix
+from vllm.utils import SUPPORT_TC, gpuname
+import vllm.envs as envs
 
 
 def get_sequence_groups(scheduler_output):
@@ -830,7 +834,7 @@ def test_prefix_caching_with_concurrent_partial_prefills():
     assert out.num_batched_tokens == 44
 
 
-@pytest.mark.parametrize("model", ["facebook/opt-125m"])
+@pytest.mark.parametrize("model", [os.path.join(models_path_prefix, "facebook/opt-125m")])
 @pytest.mark.parametrize("max_num_partial_prefills", [2, 4, 8])
 def test_chunked_prefill_with_actual_engine(model: str,
                                             max_num_partial_prefills: int):
@@ -847,6 +851,7 @@ def test_chunked_prefill_with_actual_engine(model: str,
         max_num_seqs=8,
         enable_chunked_prefill=True,
         gpu_memory_utilization=0.8,
+        block_size=64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
     )
 
     engine = LLMEngine.from_engine_args(engine_args)
@@ -858,4 +863,4 @@ def test_chunked_prefill_with_actual_engine(model: str,
     request_outputs = engine.step()
     # means all are prefilling
     assert len(request_outputs) == 0
-    assert len(engine.scheduler[0].running) == max_num_partial_prefills
+    assert len(engine.scheduler[0].running) == max_num_partial_prefills
\ No newline at end of file

tests/core/test_num_computed_tokens_update.py

@@ -9,6 +9,8 @@ from vllm.engine.llm_engine import LLMEngine
 from vllm.platforms import current_platform
 from vllm.sequence import SequenceGroup
 from ..utils import models_path_prefix
+from vllm.utils import SUPPORT_TC, gpuname
+import vllm.envs as envs
 
 MODEL = os.path.join(models_path_prefix, "JackFram/llama-160m")
 
@@ -37,7 +39,8 @@ def test_num_computed_tokens_update(num_scheduler_steps: int,
                         gpu_memory_utilization=0.7,
                         num_scheduler_steps=num_scheduler_steps,
                         enable_chunked_prefill=enable_chunked_prefill,
-                        enforce_eager=enforce_eager)
+                        enforce_eager=enforce_eager,
+                        block_size=64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16)
     engine: LLMEngine = runner.model.llm_engine
 
     # In multi-step + chunked-prefill there is no separate single prompt step.
@@ -81,4 +84,4 @@ def test_num_computed_tokens_update(num_scheduler_steps: int,
 
         # Test correctness of num_computed_tokens after the sequence finish.
         assert seq.data.get_num_computed_tokens(
-        ) == prompt_len + num_output_tokens - 1
+        ) == prompt_len + num_output_tokens - 1
\ No newline at end of file

tests/detokenizer/test_disable_detokenization.py

 # SPDX-License-Identifier: Apache-2.0
 
+import os
 import pytest
 
 from vllm.entrypoints.llm import LLM
 from vllm.sampling_params import SamplingParams
+from ..utils import models_path_prefix
+import vllm.envs as envs
+from vllm.utils import SUPPORT_TC, gpuname
 
 
 @pytest.mark.skip_v1
-@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", [os.path.join(models_path_prefix, "distilbert/distilgpt2")])
 def test_computed_prefix_blocks(model: str):
     # This test checks if the engine generates completions both with and
     # without optional detokenization, that detokenization includes text
@@ -18,7 +22,7 @@ def test_computed_prefix_blocks(model: str):
         "paper clips? Is there an easy to follow video tutorial available "
         "online for free?")
 
-    llm = LLM(model=model)
+    llm = LLM(model=model, block_size=64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16)
     sampling_params = SamplingParams(max_tokens=10,
                                      temperature=0.0,
                                      detokenize=False)
@@ -32,4 +36,4 @@ def test_computed_prefix_blocks(model: str):
     assert outputs_no_detokenization.text == ''
     assert outputs_with_detokenization.text != ''
     assert outputs_no_detokenization.token_ids == \
-        outputs_with_detokenization.token_ids
+        outputs_with_detokenization.token_ids
\ No newline at end of file

tests/detokenizer/test_stop_strings.py

@@ -2,11 +2,13 @@
 
 from typing import Any, Optional
 
+import os
 import pytest
 
 from vllm import LLM, SamplingParams, envs
+from ..utils import models_path_prefix
 
-MODEL = "meta-llama/llama-2-7b-hf"
+MODEL = os.path.join(models_path_prefix, "meta-llama/llama-2-7b-hf")
 MAX_TOKENS = 200
 
 
@@ -138,4 +140,4 @@ def test_stop_strings():
         _stop_token_id(vllm_model)
 
         _set_async_mode(vllm_model, False)
-        _stop_token_id(vllm_model)
+        _stop_token_id(vllm_model)
\ No newline at end of file

tests/engine/test_computed_prefix_blocks.py

 # SPDX-License-Identifier: Apache-2.0
 
+import os
 import pytest
 
 from vllm.engine.arg_utils import EngineArgs
 from vllm.engine.llm_engine import LLMEngine
 from vllm.sampling_params import SamplingParams
+from ..utils import models_path_prefix
+from vllm.utils import SUPPORT_TC, gpuname
+import vllm.envs as envs
 
 
-@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
-@pytest.mark.parametrize("block_size", [16])
+@pytest.mark.parametrize("model", [os.path.join(models_path_prefix, "distilbert/distilgpt2")])
+@pytest.mark.parametrize("block_size", [64] if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else [16])
 def test_computed_prefix_blocks(model: str, block_size: int):
     # This test checks if we are able to run the engine to completion
     # without triggering asserts.
@@ -33,4 +37,4 @@ def test_computed_prefix_blocks(model: str, block_size: int):
     engine.add_request("0", prompt + prompt2, sampling_params)
     engine.step()
     engine.add_request("1", prompt, sampling_params)
-    engine.step()
+    engine.step()
\ No newline at end of file

tests/engine/test_executor.py

@@ -13,6 +13,8 @@ from vllm.executor.uniproc_executor import UniProcExecutor
 from vllm.sampling_params import SamplingParams
 import os
 from ..utils import models_path_prefix
+from vllm.utils import SUPPORT_TC, gpuname
+import vllm.envs as envs
 
 
 class Mock:
@@ -57,6 +59,7 @@ def test_custom_executor(model, tmp_path):
             model=model,
             distributed_executor_backend=CustomUniExecutor,
             enforce_eager=True,  # reduce test time
+            block_size=64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
         )
         engine = LLMEngine.from_engine_args(engine_args)
         sampling_params = SamplingParams(max_tokens=1)
@@ -69,7 +72,7 @@ def test_custom_executor(model, tmp_path):
         os.chdir(cwd)
 
 
-@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", [os.path.join(models_path_prefix, "distilbert/distilgpt2")])
 def test_custom_executor_async(model, tmp_path):
     cwd = os.path.abspath(".")
     os.chdir(tmp_path)
@@ -80,6 +83,7 @@ def test_custom_executor_async(model, tmp_path):
             model=model,
             distributed_executor_backend=CustomUniExecutorAsync,
             enforce_eager=True,  # reduce test time
+            block_size=64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
         )
         engine = AsyncLLMEngine.from_engine_args(engine_args)
         sampling_params = SamplingParams(max_tokens=1)
@@ -96,7 +100,7 @@ def test_custom_executor_async(model, tmp_path):
         os.chdir(cwd)
 
 
-@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", [os.path.join(models_path_prefix, "distilbert/distilgpt2")])
 def test_respect_ray(model):
     # even for TP=1 and PP=1,
     # if users specify ray, we should use ray.
@@ -106,6 +110,7 @@ def test_respect_ray(model):
         model=model,
         distributed_executor_backend="ray",
         enforce_eager=True,  # reduce test time
+        block_size=64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
     )
     engine = LLMEngine.from_engine_args(engine_args)
-    assert engine.model_executor.uses_ray
+    assert engine.model_executor.uses_ray
\ No newline at end of file

tests/engine/test_short_mm_context.py

 # SPDX-License-Identifier: Apache-2.0
 
+import os
 import pytest
 
 from ..conftest import IMAGE_ASSETS
@@ -30,4 +31,4 @@ def test_context_length_too_short(vllm_runner, image_assets, model):
         with vllm_model:
             vllm_model.generate_greedy([HF_IMAGE_PROMPTS[0]],
                                        max_tokens=1,
-                                       images=[images[0]])
+                                       images=[images[0]])
\ No newline at end of file

tests/engine/test_skip_tokenizer_init.py

 # SPDX-License-Identifier: Apache-2.0
 
+import os
 import pytest
 
 from vllm.entrypoints.llm import LLM
 from vllm.sampling_params import SamplingParams
+from ..utils import models_path_prefix
+from vllm.utils import SUPPORT_TC, gpuname
+import vllm.envs as envs
 
 
-@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", [os.path.join(models_path_prefix, "distilbert/distilgpt2")])
 def test_skip_tokenizer_initialization(model: str):
     # This test checks if the flag skip_tokenizer_init skips the initialization
     # of tokenizer and detokenizer. The generated output is expected to contain
@@ -14,6 +18,7 @@ def test_skip_tokenizer_initialization(model: str):
     llm = LLM(
         model=model,
         skip_tokenizer_init=True,
+        block_size=64 if gpuname.startswith('BW') and envs.VLLM_FLASH_ATTN_BACKEND else 16,
     )
     sampling_params = SamplingParams(prompt_logprobs=True, detokenize=True)
 
@@ -26,4 +31,4 @@ def test_skip_tokenizer_initialization(model: str):
     completions = outputs[0].outputs
     assert len(completions) > 0
     assert completions[0].text == ""
-    assert completions[0].token_ids
+    assert completions[0].token_ids
\ No newline at end of file

tests/fastsafetensors_loader/test_fastsafetensors_loader.py

 # SPDX-License-Identifier: Apache-2.0
 
+import os
 from vllm import SamplingParams
 from vllm.config import LoadFormat
+from ..utils import models_path_prefix
 
-test_model = "openai-community/gpt2"
+test_model = os.path.join(models_path_prefix, "openai-community/gpt2")
 
 prompts = [
     "Hello, my name is",
@@ -19,4 +21,4 @@ def test_model_loader_download_files(vllm_runner):
     with vllm_runner(test_model,
                      load_format=LoadFormat.FASTSAFETENSORS) as llm:
         deserialized_outputs = llm.generate(prompts, sampling_params)
-        assert deserialized_outputs
+        assert deserialized_outputs
\ No newline at end of file

tests/fastsafetensors_loader/test_weight_utils.py

@@ -2,9 +2,11 @@
 
 import glob
 import tempfile
+import os
 
 import huggingface_hub.constants
 import torch
+from ..utils import models_path_prefix
 
 from vllm.model_executor.model_loader.weight_utils import (
     download_weights_from_hf, fastsafetensors_weights_iterator,
@@ -14,7 +16,7 @@ from vllm.model_executor.model_loader.weight_utils import (
 def test_fastsafetensors_model_loader():
     with tempfile.TemporaryDirectory() as tmpdir:
         huggingface_hub.constants.HF_HUB_OFFLINE = False
-        download_weights_from_hf("openai-community/gpt2",
+        download_weights_from_hf(os.path.join(models_path_prefix, "openai-community/gpt2"),
                                  allow_patterns=["*.safetensors"],
                                  cache_dir=tmpdir)
         safetensors = glob.glob(f"{tmpdir}/**/*.safetensors", recursive=True)
@@ -43,4 +45,4 @@ def test_fastsafetensors_model_loader():
 
 
 if __name__ == "__main__":
-    test_fastsafetensors_model_loader()
+    test_fastsafetensors_model_loader()
\ No newline at end of file