[Bugfix] Support non-uniform page sizes in KVBlockZeroer (#49704)
Signed-off-by: Elvir Crncevic <elvircrn@gmail.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.6
parent
2279575cd9
commit
972848f276
@@ -20,9 +20,10 @@ def test_block_ids_are_not_overwritten_while_copy_is_in_flight():
|
||||
zeroer.device = device
|
||||
zeroer._meta = (
|
||||
torch.tensor([storage.data_ptr()], dtype=torch.uint64, device=device),
|
||||
page_size_el,
|
||||
page_size_el,
|
||||
1,
|
||||
torch.tensor([page_size_el], dtype=torch.int64, device=device),
|
||||
page_size_el // page_size_el, # max_chunks = 1
|
||||
page_size_el, # blk_size
|
||||
1, # n_segs
|
||||
)
|
||||
|
||||
stream = torch.cuda.Stream()
|
||||
@@ -39,3 +40,49 @@ def test_block_ids_are_not_overwritten_while_copy_is_in_flight():
|
||||
assert torch.all(storage[1] == 0)
|
||||
assert torch.all(storage[2] == 0)
|
||||
assert torch.all(storage[3] == 1)
|
||||
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
|
||||
def test_non_uniform_page_sizes():
|
||||
"""Two segments with different page sizes (e.g. MLA + DSA indexer)."""
|
||||
device = torch.device("cuda")
|
||||
num_blocks = 4
|
||||
page_size_a = 10496 # int32 elements
|
||||
page_size_b = 2112
|
||||
|
||||
storage_a = torch.ones((num_blocks, page_size_a), dtype=torch.int32, device=device)
|
||||
storage_b = torch.ones((num_blocks, page_size_b), dtype=torch.int32, device=device)
|
||||
|
||||
zeroer = KVBlockZeroer.__new__(KVBlockZeroer)
|
||||
zeroer.device = device
|
||||
|
||||
seg_page_sizes = [page_size_a, page_size_b]
|
||||
max_ps = max(seg_page_sizes)
|
||||
|
||||
def largest_power_of_2_divisor(n):
|
||||
return n & -n
|
||||
|
||||
blk_size = min(min(largest_power_of_2_divisor(ps) for ps in seg_page_sizes), 1024)
|
||||
|
||||
zeroer._meta = (
|
||||
torch.tensor(
|
||||
[storage_a.data_ptr(), storage_b.data_ptr()],
|
||||
dtype=torch.uint64,
|
||||
device=device,
|
||||
),
|
||||
torch.tensor(seg_page_sizes, dtype=torch.int64, device=device),
|
||||
max_ps // blk_size,
|
||||
blk_size,
|
||||
2,
|
||||
)
|
||||
|
||||
stream = torch.cuda.Stream()
|
||||
with torch.cuda.stream(stream):
|
||||
zeroer.zero_block_ids([1, 2])
|
||||
stream.synchronize()
|
||||
|
||||
for storage in (storage_a, storage_b):
|
||||
assert torch.all(storage[0] == 1)
|
||||
assert torch.all(storage[1] == 0)
|
||||
assert torch.all(storage[2] == 0)
|
||||
assert torch.all(storage[3] == 1)
|
||||
|
||||
Reference in New Issue
Block a user