DCP supports hybrid attention (#40996)

Signed-off-by: YanXu <yancey.yx@alibaba-inc.com>
Signed-off-by: Jingyi Yang <girasoleyang@gmail.com>
Co-authored-by: Jingyi Yang <girasoleyang@gmail.com>
This commit is contained in:
Yan Xu
2026-07-09 21:34:45 -07:00
committed by GitHub
co-authored by Jingyi Yang
parent 2d814a0082
commit 95ed0feaa5
26 changed files with 783 additions and 125 deletions
+6
View File
@@ -238,6 +238,7 @@ def test_sampling_metadata_in_input_batch(device: str, batch_size: int):
vocab_size=1024,
block_sizes=[1],
kernel_block_sizes=[1],
max_num_blocks_per_req=[1024],
)
reqs: list[CachedRequestState] = []
req_id_reqs = {}
@@ -332,6 +333,7 @@ def test_swap_states_in_input_batch(device: str, batch_size: int, swap_list: lis
vocab_size=1024,
block_sizes=[1],
kernel_block_sizes=[1],
max_num_blocks_per_req=[1024],
)
ref_input_batch: InputBatch = InputBatch(
max_num_reqs=batch_size,
@@ -341,6 +343,7 @@ def test_swap_states_in_input_batch(device: str, batch_size: int, swap_list: lis
vocab_size=1024,
block_sizes=[1],
kernel_block_sizes=[1],
max_num_blocks_per_req=[1024],
)
reqs: list[CachedRequestState] = []
@@ -409,6 +412,7 @@ def test_pooling_prompt_lens_not_aliased(device: str):
vocab_size=VOCAB_SIZE,
block_sizes=[16],
kernel_block_sizes=[16],
max_num_blocks_per_req=[64],
is_pooling_model=True,
)
@@ -444,6 +448,7 @@ def test_placeholder_spec_token_ids_written_verbatim():
vocab_size=VOCAB_SIZE,
block_sizes=[16],
kernel_block_sizes=[16],
max_num_blocks_per_req=[1],
)
req = CachedRequestState(
req_id="req",
@@ -491,6 +496,7 @@ def test_pooling_metadata_token_id_buffers(
vocab_size=VOCAB_SIZE,
block_sizes=[16],
kernel_block_sizes=[16],
max_num_blocks_per_req=[64],
is_pooling_model=True,
)
req = _construct_pooling_request(0, PoolingParams(**pooling_params))