[CI] Try introducing isort. (#3495)

tests/core/test_block_manager.py

-import pytest
 import time
 from typing import List
 
+import pytest
+
 from vllm import SamplingParams
 from vllm.block import PhysicalTokenBlock
-from vllm.core.block_manager import (UncachedBlockAllocator, BlockSpaceManager,
-                                     AllocStatus)
+from vllm.core.block_manager import (AllocStatus, BlockSpaceManager,
+                                     UncachedBlockAllocator)
+from vllm.sequence import Logprob, Sequence, SequenceGroup, SequenceStatus
 from vllm.utils import Device
-from vllm.sequence import Sequence, SequenceGroup, SequenceStatus, Logprob
 
 from .utils import create_dummy_prompt
 

tests/core/test_scheduler.py

+import time
 from typing import List
+
 import pytest  # noqa
-import time
 
 from vllm.config import CacheConfig, SchedulerConfig
 from vllm.core.scheduler import Scheduler
-from vllm.sequence import SequenceGroup, Logprob
+from vllm.sequence import Logprob, SequenceGroup
 
 from .utils import create_dummy_prompt
 

tests/distributed/test_comm_ops.py

@@ -3,14 +3,12 @@
 Run `pytest tests/distributed/test_comm_ops.py --forked`.
 """
 import pytest
-import torch
 import ray
+import torch
 
 from vllm.model_executor.parallel_utils.communication_op import (
-    tensor_model_parallel_all_reduce,
-    tensor_model_parallel_all_gather,
-    broadcast_tensor_dict,
-)
+    broadcast_tensor_dict, tensor_model_parallel_all_gather,
+    tensor_model_parallel_all_reduce)
 from vllm.test_utils import (init_test_distributed_environment,
                              multi_process_tensor_parallel)
 

tests/distributed/test_custom_all_reduce.py

+import os
 import random
 
-import os
 import pytest
 import ray
 import torch

tests/entrypoints/test_guided_processors.py

 # This unit test should be moved to a new
 # tests/test_guided_decoding directory.
 
-from transformers import AutoTokenizer
 import torch
+from transformers import AutoTokenizer
 
-from vllm.model_executor.guided_logits_processors import (RegexLogitsProcessor,
-                                                          JSONLogitsProcessor)
+from vllm.model_executor.guided_logits_processors import (JSONLogitsProcessor,
+                                                          RegexLogitsProcessor)
 
 TEST_SCHEMA = {
     "type": "object",

tests/entrypoints/test_openai_server.py

+# imports for guided decoding tests
+import json
 import os
+import re
 import subprocess
+import sys
 import time
 
-import sys
+import jsonschema
+import openai  # use the official client for correctness check
 import pytest
-import requests
 # using Ray for overall ease of process management, parallel requests,
 # and debugging.
 import ray
-import openai  # use the official client for correctness check
+import requests
 # downloading lora to test lora requests
 from huggingface_hub import snapshot_download
 
-# imports for guided decoding tests
-import json
-import jsonschema
-import re
-
 from vllm.transformers_utils.tokenizer import get_tokenizer
 
 MAX_SERVER_START_WAIT_S = 600  # wait for server to start for 60 seconds

tests/kernels/conftest.py

 import pytest
+
 from vllm.utils import create_kv_caches_with_random
 
 

tests/kernels/test_activation.py

@@ -2,10 +2,10 @@ from typing import Type
 
 import pytest
 import torch
+from allclose_default import get_default_atol, get_default_rtol
 
 from vllm.model_executor.layers.activation import (FastGELU, GeluAndMul,
                                                    NewGELU, SiluAndMul)
-from allclose_default import get_default_atol, get_default_rtol
 
 DTYPES = [torch.half, torch.bfloat16, torch.float]
 NUM_TOKENS = [7, 83, 2048]  # Arbitrary values for testing

tests/kernels/test_attention.py

@@ -3,13 +3,12 @@ from typing import List, Optional, Tuple
 
 import pytest
 import torch
+from allclose_default import get_default_atol, get_default_rtol
 from xformers import ops as xops
 from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask
 
-from vllm._C import ops, cache_ops
-from vllm.utils import get_max_shared_memory_bytes
-from vllm.utils import is_hip
-from allclose_default import get_default_atol, get_default_rtol
+from vllm._C import cache_ops, ops
+from vllm.utils import get_max_shared_memory_bytes, is_hip
 
 FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
 # This will change depending on the compute capability.

tests/kernels/test_cache.py

 import random
+from typing import Tuple
 
 import pytest
 import torch
 
-from typing import Tuple
-
 from vllm._C import cache_ops
 
 COPYING_DIRECTION = [('cuda', 'cpu'), ('cuda', 'cuda'), ('cpu', 'cuda')]

tests/kernels/test_moe.py

@@ -7,8 +7,8 @@ import torch
 from transformers import MixtralConfig
 from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
 
-from vllm.model_executor.layers.fused_moe import fused_moe
 from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe import fused_moe
 from vllm.model_executor.models.mixtral import MixtralMoE
 
 

tests/kernels/test_pos_encoding.py

+from itertools import accumulate
 from typing import List, Optional
 
 import pytest
 import torch
 from allclose_default import get_default_atol, get_default_rtol
-from itertools import accumulate
+
 from vllm.model_executor.layers.rotary_embedding import get_rope
 
 IS_NEOX_STYLE = [True, False]

tests/kernels/test_prefix_prefill.py

 import random
-import pytest
 import time
 
+import pytest
 import torch
-from vllm.attention.ops.prefix_prefill import context_attention_fwd
 from xformers import ops as xops
 from xformers.ops.fmha.attn_bias import BlockDiagonalCausalFromBottomRightMask
 
+from vllm.attention.ops.prefix_prefill import context_attention_fwd
+
 NUM_HEADS = [64]
 NUM_QUERIES_PER_KV = [1, 8, 64]
 HEAD_SIZES = [128]

tests/kernels/test_rand.py

-import torch
-import pytest
 import random
 
+import pytest
+import torch
+
 from vllm.model_executor.layers.ops.rand import seeded_uniform
 from vllm.model_executor.utils import set_random_seed
 

tests/kernels/test_sampler.py

 import gc
 
-import torch
 import pytest
+import torch
 import triton
 import triton.language as tl
 
 from vllm.model_executor.layers.ops.sample import (
-    _uniform_to_exponential, sample, get_num_triton_sampler_splits,
-    MAX_TRITON_N_COLS)
-from vllm.model_executor.utils import set_random_seed
+    MAX_TRITON_N_COLS, _uniform_to_exponential, get_num_triton_sampler_splits,
+    sample)
 from vllm.model_executor.sampling_metadata import SamplingTensors
+from vllm.model_executor.utils import set_random_seed
 
 SINGLE_SPLIT_VOCAB_SIZE = 32000  # llama/mistral/mixtral vocab size
 MULTI_SPLIT_VOCAB_SIZE = MAX_TRITON_N_COLS + 100

tests/lora/conftest.py

@@ -2,7 +2,7 @@ import contextlib
 import gc
 import tempfile
 from collections import OrderedDict
-from unittest.mock import patch, MagicMock
+from unittest.mock import MagicMock, patch
 
 import pytest
 import ray
@@ -12,13 +12,13 @@ from huggingface_hub import snapshot_download
 
 import vllm
 from vllm.config import LoRAConfig
-from vllm.model_executor.layers.sampler import Sampler
-from vllm.model_executor.layers.logits_processor import LogitsProcessor
-from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                MergedColumnParallelLinear,
                                                RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.sampler import Sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
+from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.parallel_utils.parallel_state import (
     destroy_model_parallel, initialize_model_parallel)
 

tests/lora/test_layer_variation.py

+import tempfile
+from random import sample
 from typing import List, Optional
+
 import peft
 import pytest
-from random import sample
-import tempfile
 from transformers import AutoModelForCausalLM
 
 import vllm
 from vllm.lora.request import LoRARequest
+
 from .conftest import cleanup
 
 MODEL_PATH = "Felladrin/Llama-68M-Chat-v1"

tests/lora/test_layers.py

-import pytest
 import random
 from copy import deepcopy
 from dataclasses import dataclass
-from typing import List, Optional, Dict, Tuple
+from typing import Dict, List, Optional, Tuple
 
+import pytest
 import torch
 import torch.nn.functional as F
 
-from vllm.lora.layers import (
-    ColumnParallelLinearWithLoRA,
+from vllm.config import LoRAConfig
+from vllm.lora.layers import (BaseLayerWithLoRA, ColumnParallelLinearWithLoRA,
+                              LogitsProcessorWithLoRA, LoRAMapping,
                               MergedColumnParallelLinearWithLoRA,
                               QKVParallelLinearWithLora,
-    VocabParallelEmbeddingWithLoRA,
                               RowParallelLinearWithLoRA,
-    LogitsProcessorWithLoRA,
-    LoRAMapping,
-    BaseLayerWithLoRA,
-)
-from vllm.lora.models import (LoRALayerWeights, convert_mapping,
-                              PackedLoRALayerWeights)
-from vllm.config import LoRAConfig
-from vllm.model_executor.layers.logits_processor import LogitsProcessor
+                              VocabParallelEmbeddingWithLoRA)
+from vllm.lora.models import (LoRALayerWeights, PackedLoRALayerWeights,
+                              convert_mapping)
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                MergedColumnParallelLinear,
-                                               RowParallelLinear,
-                                               QKVParallelLinear)
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.vocab_parallel_embedding import (
-    VocabParallelEmbedding, ParallelLMHead)
+    ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.utils import set_random_seed
 
 from .utils import DummyLoRAManager

tests/lora/test_llama.py

@@ -3,6 +3,7 @@ import ray
 
 import vllm
 from vllm.lora.request import LoRARequest
+
 from .conftest import cleanup
 
 MODEL_PATH = "meta-llama/Llama-2-7b-hf"

tests/lora/test_lora_manager.py

@@ -8,11 +8,11 @@ from torch import nn
 
 from vllm.config import LoRAConfig
 from vllm.lora.layers import (ColumnParallelLinearWithLoRA,
-                              RowParallelLinearWithLoRA,
-                              MergedColumnParallelLinearWithLoRA)
+                              MergedColumnParallelLinearWithLoRA,
+                              RowParallelLinearWithLoRA)
 from vllm.lora.lora import LoRALayerWeights, PackedLoRALayerWeights
-from vllm.lora.models import (LoRAModel, LoRAModelManager,
-                              LRUCacheLoRAModelManager, LoRAMapping)
+from vllm.lora.models import (LoRAMapping, LoRAModel, LoRAModelManager,
+                              LRUCacheLoRAModelManager)
 from vllm.lora.request import LoRARequest
 from vllm.lora.worker_manager import (LRUCacheWorkerLoRAManager,
                                       WorkerLoRAManager)