[Perf] Cache token_to_req_indices for dsv4, 5x~6x kernel performance improvement (#47474)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -104,12 +104,9 @@ class CompressorMetadataBuilder(AttentionMetadataBuilder):
|
||||
common_attn_metadata: CommonAttentionMetadata,
|
||||
fast_build: bool = False,
|
||||
) -> CompressorMetadata:
|
||||
query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
|
||||
num_reqs = common_attn_metadata.num_reqs
|
||||
query_lens = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
|
||||
x = torch.repeat_interleave(torch.arange(num_reqs), query_lens).pin_memory()
|
||||
token_to_req_indices = self.token_to_req_indices[: x.shape[0]]
|
||||
token_to_req_indices.copy_(x, non_blocking=True)
|
||||
token_to_req_indices = common_attn_metadata.token_to_req_indices(
|
||||
self.token_to_req_indices
|
||||
)
|
||||
return CompressorMetadata(
|
||||
block_table=common_attn_metadata.block_table_tensor.clamp_(min=0),
|
||||
slot_mapping=common_attn_metadata.slot_mapping,
|
||||
|
||||
Reference in New Issue
Block a user