[Bugfix] Align block table for TRTLLM MLA edge-case (#39324)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
This commit is contained in:
@@ -257,6 +257,13 @@ class MultiGroupBlockTable:
|
|||||||
f"must match block_sizes length ({len(block_sizes)})"
|
f"must match block_sizes length ({len(block_sizes)})"
|
||||||
)
|
)
|
||||||
|
|
||||||
|
# Align to a multiple of (128 / block_size) as required
|
||||||
|
# by some attention backends such as TRTLLM (#39324)
|
||||||
|
max_num_blocks = [
|
||||||
|
cdiv(n, 128 // bs) * (128 // bs) if bs <= 128 else n
|
||||||
|
for n, bs in zip(max_num_blocks, block_sizes)
|
||||||
|
]
|
||||||
|
|
||||||
self.block_tables = [
|
self.block_tables = [
|
||||||
BlockTable(
|
BlockTable(
|
||||||
block_size,
|
block_size,
|
||||||
|
|||||||
@@ -41,6 +41,11 @@ class BlockTables:
|
|||||||
# As a result, one block on the current rank covers `block_size * cp_size`
|
# As a result, one block on the current rank covers `block_size * cp_size`
|
||||||
# tokens in the full, global (unsharded) sequence.
|
# tokens in the full, global (unsharded) sequence.
|
||||||
max_num_blocks = cdiv(self.max_model_len, block_size * self.cp_size)
|
max_num_blocks = cdiv(self.max_model_len, block_size * self.cp_size)
|
||||||
|
# Align to a multiple of (128 / block_size) as required
|
||||||
|
# by some attention backends such as TRTLLM (#39324)
|
||||||
|
if block_size <= 128:
|
||||||
|
alignment = 128 // block_size
|
||||||
|
max_num_blocks = cdiv(max_num_blocks, alignment) * alignment
|
||||||
block_table = StagedWriteTensor(
|
block_table = StagedWriteTensor(
|
||||||
(self.max_num_reqs, max_num_blocks),
|
(self.max_num_reqs, max_num_blocks),
|
||||||
dtype=torch.int32,
|
dtype=torch.int32,
|
||||||
|
|||||||
Reference in New Issue
Block a user