[Core] Deprecating block manager v1 and make block manager v2 default (…

…vllm-project#8704)

Removing the block manager v1. This is the initial piece of prefix-caching-centric design. In order to achieve prefix-caching-centric design, we need to simplify the code path so that we only use v2 block manager (which has much higher performance on prefix caching).

.buildkite/test-pipeline.yaml

@@ -77,8 +77,8 @@ steps:
   - vllm/
   - tests/basic_correctness/test_chunked_prefill
   commands:
-  - VLLM_ATTENTION_BACKEND=XFORMERS VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1 pytest -v -s basic_correctness/test_chunked_prefill.py
-  - VLLM_ATTENTION_BACKEND=FLASH_ATTN VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1 pytest -v -s basic_correctness/test_chunked_prefill.py
+  - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_chunked_prefill.py
+  - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
 
 - label: Core Test # 10min
   mirror_hardwares: [amd]
@@ -88,11 +88,7 @@ steps:
   - vllm/distributed
   - tests/core
   commands:
-  - VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1  pytest -v -s core/test_scheduler.py
-  - VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1  pytest -v -s core core/test_chunked_prefill_scheduler.py
-  - VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1  pytest -v -s core core/block/e2e/test_correctness.py
-  - VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1  pytest -v -s core core/block/e2e/test_correctness_sliding_window.py
-  - pytest -v -s core --ignore=core/block/e2e/test_correctness.py --ignore=core/test_scheduler.py --ignore=core/test_chunked_prefill_scheduler.py --ignore=core/block/e2e/test_correctness.py --ignore=core/block/e2e/test_correctness_sliding_window.py
+  - pytest -v -s core
 
 - label: Entrypoints Test # 40min
   working_dir: "/vllm-workspace/tests"
@@ -192,8 +188,7 @@ steps:
   - vllm/
   - tests/prefix_caching
   commands:
-    - VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1 pytest -v -s prefix_caching/test_prefix_caching.py
-    - pytest -v -s prefix_caching --ignore=prefix_caching/test_prefix_caching.py
+    - pytest -v -s prefix_caching
 
 - label: Samplers Test # 36min
   source_file_dependencies:
@@ -217,8 +212,7 @@ steps:
   - tests/spec_decode
   commands:
     - pytest -v -s spec_decode/e2e/test_multistep_correctness.py
-    - VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1 pytest -v -s spec_decode/e2e/test_compatibility.py
-    - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py --ignore=spec_decode/e2e/test_compatibility.py
+    - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py
 
 - label: LoRA Test %N # 15min each
   mirror_hardwares: [amd]
@@ -405,7 +399,7 @@ steps:
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep -q 'Same node test passed'
-  - TARGET_TEST_SUITE=L4 VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1  pytest basic_correctness/ -v -s -m distributed_2_gpus
+  - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m distributed_2_gpus
   # Avoid importing model tests that cause CUDA reinitialization error
   - pytest models/encoder_decoder/language/test_bart.py -v -s -m distributed_2_gpus
   - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m distributed_2_gpus

benchmarks/benchmark_latency.py

@@ -38,7 +38,6 @@ def main(args: argparse.Namespace):
         quantization_param_path=args.quantization_param_path,
         device=args.device,
         ray_workers_use_nsight=args.ray_workers_use_nsight,
-        use_v2_block_manager=args.use_v2_block_manager,
         enable_chunked_prefill=args.enable_chunked_prefill,
         download_dir=args.download_dir,
         block_size=args.block_size,
@@ -221,9 +220,6 @@ def run_to_completion(profile_dir: Optional[str] = None):
     parser.add_argument("--enable-prefix-caching",
                         action='store_true',
                         help="Enable automatic prefix caching")
-    parser.add_argument('--use-v2-block-manager',
-                        action='store_true',
-                        default=EngineArgs.use_v2_block_manager)
     parser.add_argument(
         "--ray-workers-use-nsight",
         action='store_true',

benchmarks/benchmark_prefix_caching.py

@@ -33,7 +33,6 @@
 from transformers import PreTrainedTokenizerBase
 
 from vllm import LLM, SamplingParams
-from vllm.engine.arg_utils import EngineArgs
 from vllm.utils import FlexibleArgumentParser
 
 try:
@@ -134,7 +133,6 @@ def main(args):
               tokenizer_mode='auto',
               trust_remote_code=True,
               enforce_eager=True,
-              use_v2_block_manager=args.use_v2_block_manager,
               tensor_parallel_size=args.tensor_parallel_size,
               enable_prefix_caching=args.enable_prefix_caching)
 
@@ -176,10 +174,6 @@ def main(args):
     parser.add_argument('--enable-prefix-caching',
                         action='store_true',
                         help='enable prefix caching')
-    parser.add_argument('--use-v2-block-manager',
-                        action='store_true',
-                        default=EngineArgs.use_v2_block_manager,
-                        help='Use BlockSpaceMangerV2')
     parser.add_argument('--num-prompts',
                         type=int,
                         default=1,

benchmarks/benchmark_throughput.py

@@ -86,7 +86,6 @@ def run_vllm(
     distributed_executor_backend: Optional[str],
     gpu_memory_utilization: float = 0.9,
     num_scheduler_steps: int = 1,
-    use_v2_block_manager: bool = False,
     download_dir: Optional[str] = None,
     load_format: str = EngineArgs.load_format,
     disable_async_output_proc: bool = False,
@@ -113,7 +112,6 @@ def run_vllm(
         distributed_executor_backend=distributed_executor_backend,
         load_format=load_format,
         num_scheduler_steps=num_scheduler_steps,
-        use_v2_block_manager=use_v2_block_manager,
         disable_async_output_proc=disable_async_output_proc,
     )
 
@@ -176,7 +174,6 @@ async def run_vllm_async(
     distributed_executor_backend: Optional[str],
     gpu_memory_utilization: float = 0.9,
     num_scheduler_steps: int = 1,
-    use_v2_block_manager: bool = False,
     download_dir: Optional[str] = None,
     load_format: str = EngineArgs.load_format,
     disable_async_output_proc: bool = False,
@@ -204,7 +201,6 @@ async def run_vllm_async(
         distributed_executor_backend=distributed_executor_backend,
         load_format=load_format,
         num_scheduler_steps=num_scheduler_steps,
-        use_v2_block_manager=use_v2_block_manager,
         disable_async_output_proc=disable_async_output_proc,
         worker_use_ray=False,
         disable_log_requests=True,
@@ -341,8 +337,7 @@ def main(args: argparse.Namespace):
             args.enable_prefix_caching, args.enable_chunked_prefill,
             args.max_num_batched_tokens, args.distributed_executor_backend,
             args.gpu_memory_utilization, args.num_scheduler_steps,
-            args.use_v2_block_manager, args.download_dir, args.load_format,
-            args.disable_async_output_proc
+            args.download_dir, args.load_format, args.disable_async_output_proc
         ]
 
         if args.async_engine:
@@ -471,10 +466,6 @@ def main(args: argparse.Namespace):
         type=int,
         default=1,
         help="Maximum number of forward steps per scheduler call.")
-    parser.add_argument("--use-v2-block-manager",
-                        action='store_true',
-                        default=EngineArgs.use_v2_block_manager,
-                        help="Enable block manager v2.")
     parser.add_argument(
         "--enable-prefix-caching",
         action='store_true',

benchmarks/overheads/benchmark_hashing.py

@@ -16,7 +16,6 @@ def main(args):
         enforce_eager=True,
         enable_prefix_caching=True,
         tensor_parallel_size=args.tensor_parallel_size,
-        use_v2_block_manager=args.use_v2_block_manager,
     )
 
     sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
@@ -56,8 +55,5 @@ def main(args):
     parser.add_argument('--enable-prefix-caching',
                         action='store_true',
                         help='enable prefix caching')
-    parser.add_argument('--use-v2-block-manager',
-                        action='store_true',
-                        help='Use BlockSpaceMangerV2')
     args = parser.parse_args()
     main(args)

docs/source/models/spec_decode.rst

@@ -30,7 +30,6 @@ The following code configures vLLM in an offline mode to use speculative decodin
         tensor_parallel_size=1,
         speculative_model="facebook/opt-125m",
         num_speculative_tokens=5,
-        use_v2_block_manager=True,
     )
     outputs = llm.generate(prompts, sampling_params)
 
@@ -104,7 +103,6 @@ matching n-grams in the prompt. For more information read `this thread. <https:/
         speculative_model="[ngram]",
         num_speculative_tokens=5,
         ngram_prompt_lookup_max=4,
-        use_v2_block_manager=True,
     )
     outputs = llm.generate(prompts, sampling_params)
 
@@ -135,7 +133,6 @@ For more information see `this blog <https://pytorch.org/blog/hitchhikers-guide-
         tensor_parallel_size=4,
         speculative_model="ibm-fms/llama3-70b-accelerator",
         speculative_draft_tensor_parallel_size=1,
-        use_v2_block_manager=True,
     )
     outputs = llm.generate(prompts, sampling_params)
 

examples/offline_inference_mlpspeculator.py

@@ -50,8 +50,6 @@ def time_generation(llm: LLM, prompts: List[str],
     llm = LLM(
         model="meta-llama/Llama-2-13b-chat-hf",
         speculative_model="ibm-fms/llama-13b-accelerator",
-        # These are currently required for MLPSpeculator decoding
-        use_v2_block_manager=True,
     )
 
     print("With speculation")

tests/basic_correctness/test_chunked_prefill.py

@@ -12,20 +12,14 @@
 import pytest
 
 from ..models.utils import check_logprobs_close, check_outputs_equal
-from ..utils import check_deprecated_block_manager_usage, multi_gpu_test
+from ..utils import multi_gpu_test
 
 MODELS = [
     "facebook/opt-125m",
     "meta-llama/Llama-2-7b-hf",
 ]
 
 
-@pytest.fixture(scope="module", autouse=True)
-def check_deprecated_block_manager():
-    check_deprecated_block_manager_usage(
-        'tests/basic_correctness/test_chunked_prefill.py')
-
-
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["half"])
 @pytest.mark.parametrize("max_tokens", [32])
@@ -197,7 +191,6 @@ def test_models_with_fp8_kv_cache(
 @pytest.mark.parametrize("max_tokens", [16])
 @pytest.mark.parametrize("enforce_eager", [False])
 @pytest.mark.parametrize("chunk_size", [30, 32])
-@pytest.mark.parametrize("use_v2_block_manager", [False, True])
 # 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])
@@ -206,7 +199,6 @@ def test_with_prefix_caching(
     max_tokens: int,
     enforce_eager: bool,
     chunk_size: int,
-    use_v2_block_manager: bool,
     tensor_parallel_size: int,
 ) -> None:
     """
@@ -234,7 +226,6 @@ def test_with_prefix_caching(
                 enable_chunked_prefill=True,
                 enable_prefix_caching=enable,
                 tensor_parallel_size=tensor_parallel_size,
-                use_v2_block_manager=use_v2_block_manager,
                 enforce_eager=enforce_eager,
                 max_num_seqs=max_num_seqs,
         ) as vllm_model: