Removed padding block from a list of available blocks in allocators (#…

…312)

Block 0 is used for padding. This PR removes the padding block from a
list of available blocks in block allocators v1 and v2

vllm/core/block/cpu_gpu_block_allocator.py

@@ -4,7 +4,7 @@
                                         DeviceAwareBlockAllocator)
 from vllm.core.block.naive_block import NaiveBlock, NaiveBlockAllocator
 from vllm.core.block.prefix_caching_block import PrefixCachingBlockAllocator
-from vllm.utils import Device
+from vllm.utils import Device, is_hpu
 
 
 class CpuGpuBlockAllocator(DeviceAwareBlockAllocator):
@@ -52,7 +52,11 @@ def create(
             - The block IDs are assigned contiguously, with GPU block IDs coming
                 before CPU block IDs.
         """
-        block_ids = list(range(num_gpu_blocks + num_cpu_blocks))
+        # For HPU, block id 0 is used only for padding
+        reserved_blocks = 1 if is_hpu() else 0
+        block_ids = list(
+            range(reserved_blocks, num_gpu_blocks + num_cpu_blocks))
+        num_gpu_blocks -= reserved_blocks
         gpu_block_ids = block_ids[:num_gpu_blocks]
         cpu_block_ids = block_ids[num_gpu_blocks:]
 

vllm/core/block_manager_v1.py

@@ -13,7 +13,7 @@
 from vllm.core.interfaces import AllocStatus, BlockSpaceManager
 from vllm.logger import init_logger
 from vllm.sequence import Sequence, SequenceGroup, SequenceStatus
-from vllm.utils import Device
+from vllm.utils import Device, is_hpu
 
 logger = init_logger(__name__)
 
@@ -171,7 +171,9 @@ def __init__(
 
         # Initialize the free blocks.
         self.free_blocks: BlockTable = []
-        for i in range(num_blocks):
+        # For HPU, block id 0 is used only for padding
+        reserved_blocks = 1 if is_hpu() else 0
+        for i in range(reserved_blocks, num_blocks):
             block = PhysicalTokenBlock(device=device,
                                        block_number=i,
                                        block_size=block_size,