[Bugfix] Use null block (0) for padded block table entries (#35431)

Signed-off-by: SandishKumarHN <sandish@fb.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
SandishKumarHN
2026-03-30 14:02:51 -07:00
committed by GitHub
co-authored by Lucas Wilkinson Matthew Bonanni
parent 1fc69f59bb
commit bcc6f67447
15 changed files with 125 additions and 100 deletions
+4 -3
View File
@@ -122,6 +122,7 @@ from vllm.v1.attention.backend import (
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder
from vllm.v1.attention.backends.utils import (
NULL_BLOCK_ID,
create_fast_prefill_custom_backend,
get_dcp_local_seq_lens,
reorder_batch_to_split_decodes_and_prefills,
@@ -2135,9 +2136,9 @@ class GPUModelRunner(
blk_table = self.input_batch.block_table[kv_cache_gid]
blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
# Fill unused with -1. Needed for reshape_and_cache in full cuda
# graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
blk_table_tensor[num_reqs:num_reqs_padded].fill_(-1)
# Fill unused block table entries with NULL_BLOCK_ID (null block)
# for CUDAGraph padding. Block 0 is reserved for padding.
blk_table_tensor[num_reqs:num_reqs_padded].fill_(NULL_BLOCK_ID)
return blk_table_tensor
assert slot_mappings is not None