[Core] Deprecating block manager v1 and make block manager v2 default (#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).

tests/spec_decode/e2e/test_eagle_correctness.py

@@ -43,9 +43,6 @@ PRECISION = "float32"
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -86,9 +83,6 @@ def test_eagle_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -143,9 +137,6 @@ def test_eagle_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
     [{
         "enforce_eager": False,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -191,9 +182,6 @@ def test_eagle_e2e_greedy_correctness_cuda_graph(
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 
@@ -235,9 +223,6 @@ def test_eagle_e2e_greedy_correctness_with_preemption(
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 
@@ -283,9 +268,6 @@ def test_eagle_different_k(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 

tests/spec_decode/e2e/test_integration.py

@@ -12,8 +12,6 @@ MAIN_MODEL = "JackFram/llama-68m"
 @pytest.mark.parametrize(
     "common_llm_kwargs",
     [{
-        # Required for spec decode.
-        "use_v2_block_manager": True,
 
         # Verify equality when cuda graphs allowed.
         "enforce_eager": False,
@@ -57,9 +55,6 @@ def test_spec_decode_cuda_graph(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True,
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [
     {
@@ -111,9 +106,6 @@ def test_speculative_model_quantization_config(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True,
         "speculative_model": "JackFram/llama-68m",
         "num_speculative_tokens": 3,
     }])

tests/spec_decode/e2e/test_integration_dist_tp2.py

@@ -17,9 +17,6 @@ from .conftest import run_equality_correctness_test_tp
     [[
         # Skip cuda graph recording for fast test.
         "--enforce-eager",
-
-        # Required for spec decode.
-        "--use-v2-block-manager",
         "--tensor-parallel-size",
         "2"
     ]])
@@ -74,9 +71,6 @@ def test_target_model_tp_gt_1(common_llm_kwargs, per_test_common_llm_kwargs,
     [[
         # Skip cuda graph recording for fast test.
         "--enforce-eager",
-
-        # Required for spec decode.
-        "--use_v2_block_manager",
         "--tensor_parallel_size",
         "2",
 

tests/spec_decode/e2e/test_integration_dist_tp4.py

@@ -19,9 +19,6 @@ SPEC_MODEL = "JackFram/llama-68m"
     [[
         # Skip cuda graph recording for fast test.
         "--enforce_eager",
-
-        # Required for spec decode.
-        "--use-v2-block-manager",
         "--tensor-parallel-size",
         "4",
     ]])
@@ -71,9 +68,6 @@ def test_draft_model_tp_lt_target_model_tp4(common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "--enforce-eager",
-
-        # Required for spec decode.
-        "--use-v2-block-manager",
         "--tensor-parallel-size",
         "4",
     ]])

tests/spec_decode/e2e/test_logprobs.py

@@ -14,9 +14,6 @@ from .conftest import run_equality_correctness_test
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True,
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -67,9 +64,6 @@ def test_logprobs_equality(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -119,9 +113,6 @@ def test_logprobs_different_k(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -173,9 +164,6 @@ def test_logprobs_when_skip_speculation(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -251,8 +239,6 @@ def test_logprobs_temp_1(vllm_runner, common_llm_kwargs,
         "model_name": "JackFram/llama-160m",
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-        # Required for spec decode.
-        "use_v2_block_manager": True,
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])

tests/spec_decode/e2e/test_medusa_correctness.py

@@ -45,9 +45,6 @@ PRECISION = "float32"
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -93,9 +90,6 @@ def test_medusa_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -151,9 +145,6 @@ def test_medusa_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
     [{
         "enforce_eager": False,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -204,9 +195,6 @@ def test_medusa_e2e_greedy_correctness_cuda_graph(
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 
@@ -253,9 +241,6 @@ def test_medusa_e2e_greedy_correctness_with_preemption(
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 
@@ -306,9 +291,6 @@ def test_medusa_different_k(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 
@@ -356,9 +338,6 @@ def test_medusa_disable_queue(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 

tests/spec_decode/e2e/test_mlp_correctness.py

@@ -47,9 +47,6 @@ PRECISION = "float32"
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -94,9 +91,6 @@ def test_mlp_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -149,9 +143,6 @@ def test_mlp_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -195,9 +186,6 @@ def test_mlp_e2e_acceptance_rate(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
 
@@ -258,9 +246,6 @@ def test_mlp_e2e_seeded_correctness(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 
@@ -311,9 +296,6 @@ def test_mlp_e2e_greedy_correctness_with_preemption(
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 
@@ -366,9 +348,6 @@ def test_mlp_e2e_greedy_correctness_with_padding(
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 
@@ -419,9 +398,6 @@ def test_mlp_different_k(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Precision
         "dtype": PRECISION,
 
@@ -469,9 +445,6 @@ def test_mlp_disable_queue(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True,
         "speculative_model": SPEC_MODEL,
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])

tests/spec_decode/e2e/test_multistep_correctness.py

@@ -55,9 +55,6 @@ from .conftest import (get_output_from_llm_generator,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True,
     }])
 @pytest.mark.parametrize(
     "per_test_common_llm_kwargs",
@@ -124,9 +121,6 @@ def test_spec_decode_e2e_with_detokenization(test_llm_generator,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
     }])
@@ -190,9 +184,6 @@ def test_spec_decode_e2e_greedy_correctness_tiny_model_bs1(
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
     }])
@@ -246,9 +237,6 @@ def test_spec_decode_e2e_greedy_correctness_tiny_model_large_bs(
     [{
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize(
     "per_test_common_llm_kwargs",
@@ -303,9 +291,6 @@ def test_spec_decode_e2e_greedy_correctness_tiny_model_large_bs_diff_output_len(
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
     }])
@@ -353,9 +338,6 @@ def test_spec_decode_e2e_greedy_correctness_real_model_bs1(
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
     }])
@@ -404,9 +386,6 @@ def test_spec_decode_e2e_greedy_correctness_real_model_large_bs(
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [
     {
@@ -454,9 +433,6 @@ def test_spec_decode_e2e_greedy_correctness_with_preemption(
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize(
     "per_test_common_llm_kwargs",
@@ -514,9 +490,6 @@ def test_spec_decode_different_block_size(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -570,9 +543,6 @@ def test_skip_speculation(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -611,9 +581,6 @@ def test_disable_speculation(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -660,9 +627,6 @@ def test_many_k(vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])

tests/spec_decode/e2e/test_ngram_correctness.py

@@ -35,9 +35,6 @@ from .conftest import run_equality_correctness_test
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
     }])
@@ -82,9 +79,6 @@ def test_ngram_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # Print spec metrics.
         "disable_log_stats": False,
     }])
@@ -145,9 +139,6 @@ def test_ngram_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [
     {
@@ -195,9 +186,6 @@ def test_ngram_e2e_greedy_correctness_with_preemption(
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -254,9 +242,6 @@ def test_ngram_different_k(vllm_runner, common_llm_kwargs,
 
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
-
-        # Required for spec decode.
-        "use_v2_block_manager": True
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -303,7 +288,6 @@ def test_ngram_disable_queue(vllm_runner, common_llm_kwargs,
         "enforce_eager": True,
 
         # Required for spec decode.
-        "use_v2_block_manager": True,
         "speculative_model": "[ngram]",
         "num_speculative_tokens": 5,
         "ngram_prompt_lookup_max": 3,

tests/spec_decode/e2e/test_seed.py

@@ -17,9 +17,6 @@ SPEC_MODEL = "JackFram/llama-160m"
         # Skip cuda graph recording for fast test.
         "enforce_eager": True,
 
-        # Required for spec decode.
-        "use_v2_block_manager": True,
-
         # speculative model
         "speculative_model": "JackFram/llama-160m",
 

tests/utils.py

@@ -678,12 +678,3 @@ def get_client_text_logprob_generations(
     return [(text_generations, text,
              (None if x.logprobs is None else x.logprobs.top_logprobs))
             for completion in completions for x in completion.choices]
-
-
-def check_deprecated_block_manager_usage(test_name: str):
-    assert envs.VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1 is True, (
-        f"To allow the use of deprecated BlockSpaceManagerV1, set the "
-        f"environment variable VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1. "
-        f"You can run the tests with: "
-        f"`VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1 pytest {test_name}`"  #noqa
-    )

vllm/attention/backends/flash_attn.py

@@ -305,8 +305,6 @@ class FlashAttentionMetadataBuilder(
         self.runner = input_builder.runner
         self.sliding_window = input_builder.sliding_window
         self.block_size = input_builder.block_size
-        self.use_v2_block_manager = (
-            input_builder.scheduler_config.use_v2_block_manager)
 
     def _add_seq_group(
             self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
@@ -355,9 +353,9 @@ class FlashAttentionMetadataBuilder(
 
             # Compute slot mapping.
             is_profile_run = is_block_tables_empty(block_tables)
-            start_idx = compute_slot_mapping_start_idx(
-                is_prompt, query_len, context_len, self.sliding_window,
-                self.use_v2_block_manager)
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
             compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
                                  seq_len, context_len, start_idx,
                                  self.block_size, inter_data.block_tables)

vllm/attention/backends/flashinfer.py

@@ -475,8 +475,6 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
 
         self.sliding_window = input_builder.sliding_window
         self.block_size = input_builder.block_size
-        self.use_v2_block_manager = (
-            input_builder.scheduler_config.use_v2_block_manager)
 
         # Please follow https://docs.flashinfer.ai/tutorials/kv_layout.html#page-layout
         # for the precise definition of the following fields.
@@ -542,9 +540,9 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
             is_profile_run = is_block_tables_empty(block_tables)
 
             # Compute slot mapping.
-            start_idx = compute_slot_mapping_start_idx(
-                is_prompt, query_len, context_len, self.sliding_window,
-                self.use_v2_block_manager)
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
             compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
                                  seq_len, context_len, start_idx,
                                  self.block_size, inter_data.block_tables)

vllm/attention/backends/utils.py

@@ -38,18 +38,12 @@ def is_block_tables_empty(block_tables: Union[None, Dict]):
 
 
 def compute_slot_mapping_start_idx(is_prompt: bool, query_len: int,
-                                   context_len: int, sliding_window: int,
-                                   use_v2_block_manager: bool):
+                                   context_len: int, sliding_window: int):
     """
     Compute the start index of slot mapping.
     """
     start_idx = 0
     if is_prompt and sliding_window is not None:
-        assert use_v2_block_manager or context_len == 0, (
-            "Prefix caching is currently not supported with "
-            "sliding window attention in V1 block manager")
-        # When prefill, we use it to not write slots to kv cache
-        # to save memory.
         start_idx = max(0, query_len - sliding_window)
     return start_idx
 
@@ -138,8 +132,6 @@ class CommonMetadataBuilder(AttentionMetadataBuilder[TAttentionMetadata]):
 
         self.sliding_window = input_builder.sliding_window
         self.block_size = input_builder.block_size
-        self.use_v2_block_manager = (
-            input_builder.scheduler_config.use_v2_block_manager)
 
     def _add_seq_group(
             self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
@@ -180,9 +172,9 @@ class CommonMetadataBuilder(AttentionMetadataBuilder[TAttentionMetadata]):
 
             # Compute slot mapping.
             is_profile_run = is_block_tables_empty(block_tables)
-            start_idx = compute_slot_mapping_start_idx(
-                is_prompt, query_len, context_len, self.sliding_window,
-                self.use_v2_block_manager)
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
             compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
                                  seq_len, context_len, start_idx,
                                  self.block_size, inter_data.block_tables)

vllm/commit_id.py

+__commit__ = "93ec62b8556e279d2c050bdc1c3247831bd39466"

vllm/config.py

@@ -949,7 +949,6 @@ class SchedulerConfig:
             iteration.
         max_model_len: Maximum length of a sequence (including prompt
             and generated text).
-        use_v2_block_manager: Whether to use the BlockSpaceManagerV2 or not.
         num_lookahead_slots: The number of slots to allocate per sequence per
             step, beyond the known token ids. This is used in speculative
             decoding to store KV activations of tokens which may or may not be
@@ -976,7 +975,6 @@ class SchedulerConfig:
                  max_num_batched_tokens: Optional[int],
                  max_num_seqs: int,
                  max_model_len: int,
-                 use_v2_block_manager: bool = True,
                  num_lookahead_slots: int = 0,
                  delay_factor: float = 0.0,
                  enable_chunked_prefill: bool = False,
@@ -1026,7 +1024,6 @@ class SchedulerConfig:
 
         self.max_num_seqs = max_num_seqs
         self.max_model_len = max_model_len
-        self.use_v2_block_manager = use_v2_block_manager
         self.num_lookahead_slots = num_lookahead_slots
         self.delay_factor = delay_factor
         self.chunked_prefill_enabled = enable_chunked_prefill
@@ -1067,18 +1064,6 @@ class SchedulerConfig:
                 f"({self.num_scheduler_steps}) must be greater than or "
                 "equal to 1.")
 
-        if (not self.use_v2_block_manager \
-            and not envs.VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1):
-            raise ValueError(
-                "The use of BlockSpaceManagerV1 is deprecated and will "
-                "be removed in a future release. Please switch to "
-                "BlockSpaceManagerV2 by setting --use-v2-block-manager to "
-                "True. If you wish to suppress this error temporarily, "
-                "you can set the environment variable "
-                "`VLLM_ALLOW_DEPRECATED_BLOCK_MANAGER_V1=1. If your use "
-                "case is not supported in BlockSpaceManagerV2, please "
-                "file an issue with detailed information.")
-
     @property
     def is_multi_step(self) -> bool:
         return self.num_scheduler_steps > 1
@@ -1137,7 +1122,6 @@ class SpeculativeConfig:
         speculative_disable_mqa_scorer: Optional[bool],
         speculative_max_model_len: Optional[int],
         enable_chunked_prefill: bool,
-        use_v2_block_manager: bool,
         disable_log_stats: bool,
         speculative_disable_by_batch_size: Optional[int],
         ngram_prompt_lookup_max: Optional[int],
@@ -1178,9 +1162,6 @@ class SpeculativeConfig:
             enable_chunked_prefill (bool): Whether vLLM is configured to use
                 chunked prefill or not. Used for raising an error since its not
                 yet compatible with spec decode.
-            use_v2_block_manager (bool): Whether vLLM is configured to use the
-                v2 block manager or not. Used for raising an error since the v2
-                block manager is required with spec decode.
             speculative_disable_by_batch_size (Optional[int]): Disable
                 speculative decoding for new incoming requests when the number
                 of enqueue requests  is larger than this value, if provided.
@@ -1231,11 +1212,6 @@ class SpeculativeConfig:
                 "Speculative decoding and chunked prefill are "
                 f"currently mutually exclusive ({enable_chunked_prefill=}).")
 
-        if not use_v2_block_manager:
-            raise ValueError(
-                "Speculative decoding requires usage of the V2 "
-                "block manager. Enable it with --use-v2-block-manager.")
-
         # TODO: The user should be able to specify revision/max model len
         # for the draft model. It is not currently supported.
         draft_revision = None

vllm/core/block/utils.py

@@ -4,28 +4,6 @@ from vllm.utils import (STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE,
                         STR_NOT_IMPL_ENC_DEC_SWA)
 
 
-def _get_block_mgr_sliding_window_attr(block_mgr):
-    '''
-    BlockManagerV1 and BlockManagerV2 have slightly different
-    members related to sliding window attention (SWA). This
-    function extracts the appropriate member to use for determining
-    whether SWA is enabled.
-
-    Arguments:
-
-    * block_mgr: BlockManagerV1 or BlockManagerV2 instance
-    '''
-
-    if hasattr(block_mgr, 'block_sliding_window'):
-        return block_mgr.block_sliding_window
-    if hasattr(block_mgr, 'max_block_sliding_window'):
-        return block_mgr.max_block_sliding_window
-
-    raise AttributeError("Block manager instance has neither " + \
-                         "block_sliding_window nor " + \
-                         "max_block_sliding_window attributes.")
-
-
 def check_no_caching_or_swa_for_blockmgr_encdec(
         block_mgr, seq_group: SequenceGroup) -> None:
     '''
@@ -41,7 +19,7 @@ def check_no_caching_or_swa_for_blockmgr_encdec(
     '''
 
     if seq_group.is_encoder_decoder():
-        if _get_block_mgr_sliding_window_attr(block_mgr) is not None:
+        if block_mgr.max_block_sliding_window is not None:
             raise NotImplementedError(STR_NOT_IMPL_ENC_DEC_SWA)
 
         if block_mgr.enable_caching:

vllm/core/block_manager_v2.py → vllm/core/block_manager.py

@@ -17,7 +17,7 @@ SeqId = int
 EncoderSeqId = str
 
 
-class BlockSpaceManagerV2(BlockSpaceManager):
+class SelfAttnBlockSpaceManager(BlockSpaceManager):
     """BlockSpaceManager which manages the allocation of KV cache.
 
     It owns responsibility for allocation, swapping, allocating memory for

vllm/core/interfaces.py

@@ -28,13 +28,9 @@ class BlockSpaceManager(ABC):
     def get_block_space_manager_class(version: str):
         version = version.lower()
 
-        if version == "v1":
-            from vllm.core.block_manager_v1 import BlockSpaceManagerV1
-            return BlockSpaceManagerV1
-
-        if version == "v2":
-            from vllm.core.block_manager_v2 import BlockSpaceManagerV2
-            return BlockSpaceManagerV2
+        if version == "selfattn":
+            from vllm.core.block_manager import SelfAttnBlockSpaceManager
+            return SelfAttnBlockSpaceManager
 
         if version == "placeholder":
             from vllm.core.placeholder_block_space_manager import (