[mypy] Enable mypy type checking for `vllm/core` (#7229)

.github/workflows/mypy.yaml

@@ -35,7 +35,6 @@ jobs:
         mypy
         mypy tests --follow-imports skip
         mypy vllm/attention --follow-imports skip
-        mypy vllm/core --follow-imports skip
         mypy vllm/distributed --follow-imports skip
         mypy vllm/engine  --follow-imports skip
         mypy vllm/executor --follow-imports skip

format.sh

@@ -99,7 +99,6 @@ echo 'vLLM mypy:'
 mypy --follow-imports skip  # Note that this is less strict than CI
 mypy tests --follow-imports skip
 mypy vllm/attention --follow-imports skip
-mypy vllm/core --follow-imports skip
 mypy vllm/distributed --follow-imports skip
 mypy vllm/engine  --follow-imports skip
 mypy vllm/executor --follow-imports skip

pyproject.toml

@@ -58,6 +58,7 @@ files = [
     "vllm/adapter_commons",
     "vllm/assets",
     "vllm/entrypoints",
+    "vllm/core",
     "vllm/inputs",
     "vllm/logging",
     "vllm/multimodal",

vllm/block.py

 """Token blocks."""
-from typing import List, Optional
+from typing import TYPE_CHECKING, Iterator, List, Optional
 
 from vllm.utils import Device
 
-DEFAULT_LAST_ACCESSED_TIME = -1
+DEFAULT_LAST_ACCESSED_TIME: float = -1
 
 
 class PhysicalTokenBlock:
@@ -59,6 +59,11 @@ class BlockTable:
     def __getitem__(self, key):
         return self._blocks[key]
 
+    if TYPE_CHECKING:
+
+        def __iter__(self) -> Iterator[PhysicalTokenBlock]:
+            raise RuntimeError("Method should be automatically generated")
+
     def __setitem__(self, key, value):
         if isinstance(key, slice):
             blocks = value

vllm/core/block/cpu_gpu_block_allocator.py

@@ -132,7 +132,7 @@ class CpuGpuBlockAllocator(DeviceAwareBlockAllocator):
 
     def allocate_immutable_blocks(self, prev_block: Optional[Block],
                                   block_token_ids: List[List[int]],
-                                  device: Optional[Device]) -> List[Block]:
+                                  device: Device) -> List[Block]:
         """Allocates a new group of immutable blocks with the provided block 
         token IDs on the specified device.
 

vllm/core/block_manager_v1.py

@@ -278,7 +278,7 @@ class BlockSpaceManagerV1(BlockSpaceManager):
         # request ID
         self.cross_block_tables: Dict[str, BlockTable] = {}
 
-    def _get_seq_num_required_blocks(self, seq: Sequence) -> int:
+    def _get_seq_num_required_blocks(self, seq: Optional[Sequence]) -> int:
         return 0 if seq is None else seq.n_blocks
 
     def can_allocate(self, seq_group: SequenceGroup) -> AllocStatus:
@@ -310,13 +310,14 @@ class BlockSpaceManagerV1(BlockSpaceManager):
             return AllocStatus.LATER
 
     def _allocate_sequence(self, \
-                           seq: Sequence, \
+                           seq: Optional[Sequence], \
                            ref_count: int, \
                            is_encoder_decoder: bool = True) -> BlockTable:
         # Allocate new physical token blocks that will store the prompt tokens.
-        num_prompt_blocks = seq.n_blocks
+        num_prompt_blocks = self._get_seq_num_required_blocks(seq)
 
         block_table: BlockTable = BlockTable()
+        assert seq is not None
         for logical_idx in range(num_prompt_blocks):
             if (self.block_sliding_window is not None
                     and logical_idx >= self.block_sliding_window):

vllm/core/block_manager_v2.py

@@ -120,8 +120,10 @@ class BlockSpaceManagerV2(BlockSpaceManager):
         )
 
         if seq_group.is_encoder_decoder():
+            encoder_seq = seq_group.get_encoder_seq()
+            assert encoder_seq is not None
             num_required_blocks += BlockTable.get_num_required_blocks(
-                seq_group.get_encoder_seq().get_token_ids(),
+                encoder_seq.get_token_ids(),
                 block_size=self.block_size,
             )
 
@@ -189,7 +191,9 @@ class BlockSpaceManagerV2(BlockSpaceManager):
         check_no_caching_or_swa_for_blockmgr_encdec(self, seq_group)
 
         if seq_group.is_encoder_decoder():
-            block_table = self._allocate_sequence(seq_group.get_encoder_seq())
+            encoder_seq = seq_group.get_encoder_seq()
+            assert encoder_seq is not None
+            block_table = self._allocate_sequence(encoder_seq)
             self.cross_block_tables[request_id] = block_table
 
     def can_append_slots(self, seq_group: SequenceGroup,

vllm/core/embedding_model_block_manager.py

@@ -77,8 +77,8 @@ class EmbeddingModelBlockSpaceManager(BlockSpaceManager):
         pass
 
     def get_common_computed_block_ids(self,
-                                      seq_group: SequenceGroup) -> List[int]:
-        return None  # type: ignore
+                                      seq_group: List[Sequence]) -> List[int]:
+        return []
 
     def mark_blocks_as_computed(self, seq_group: SequenceGroup):
         pass

vllm/core/scheduler.py

@@ -221,10 +221,10 @@ class SchedulerSwappedInOutputs:
     """
     # Selected sequences that are going to be swapped in and is in a
     # decoding phase.
-    decode_seq_groups: List[SequenceGroup]
+    decode_seq_groups: List[ScheduledSequenceGroup]
     # Selected sequences that are going to be swapped in and in a prefill
     # phase. I.e., it means the prefill has been chunked.
-    prefill_seq_groups: List[SequenceGroup]
+    prefill_seq_groups: List[ScheduledSequenceGroup]
     # The blocks to swap in.
     blocks_to_swap_in: List[Tuple[int, int]]
     # The blocks to copy.
@@ -254,7 +254,7 @@ class SchedulerPrefillOutputs:
     to be recomputed from scratch.
     """
     # Selected sequences for prefill.
-    seq_groups: List[SequenceGroup]
+    seq_groups: List[ScheduledSequenceGroup]
     # Ignored sequence groups.
     ignored_seq_groups: List[SequenceGroup]
     num_lookahead_slots: int
@@ -289,7 +289,9 @@ def scheduler_running_outputs_builder():
 
 
 def scheduled_seq_group_builder():
-    return ScheduledSequenceGroup(seq_group=None, token_chunk_size=0)
+    return ScheduledSequenceGroup(SequenceGroup("", [], -1),
+                                  token_chunk_size=0)
+    # return ScheduledSequenceGroup(seq_group=None, token_chunk_size=0)
 
 
 class Scheduler:
@@ -791,7 +793,7 @@ class Scheduler:
             SchedulerPrefillOutputs.
         """
         ignored_seq_groups: List[SequenceGroup] = []
-        seq_groups: List[SequenceGroup] = []
+        seq_groups: List[ScheduledSequenceGroup] = []
 
         waiting_queue = self.waiting
 
@@ -1130,7 +1132,9 @@ class Scheduler:
 
             if seq_group.is_encoder_decoder():
                 # Encoder associated with SequenceGroup
-                encoder_seq_data = seq_group.get_encoder_seq().data
+                encoder_seq = seq_group.get_encoder_seq()
+                assert encoder_seq is not None
+                encoder_seq_data = encoder_seq.data
                 # Block table for cross-attention
                 # Also managed at SequenceGroup level
                 cross_block_table = self.block_manager.get_cross_block_table(