[Deprecation] Deprecate items scheduled for 0.29 (#55353)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -43,10 +43,12 @@ def _prefer_two_stage_compressor() -> bool:
|
||||
|
||||
|
||||
def _get_c128_boundary(metadata: CommonAttentionMetadata) -> bool | None:
|
||||
starts = metadata._num_computed_tokens_cpu
|
||||
if starts is None:
|
||||
seq_lens_cpu = metadata.seq_lens_cpu_upper_bound
|
||||
if seq_lens_cpu is None:
|
||||
return None
|
||||
|
||||
query_lens = metadata.query_start_loc_cpu[1:] - metadata.query_start_loc_cpu[:-1]
|
||||
starts = seq_lens_cpu - query_lens
|
||||
starts_list = starts.tolist()
|
||||
query_start_loc = metadata.query_start_loc_cpu.tolist()
|
||||
return any(
|
||||
@@ -116,6 +118,10 @@ class CompressorMetadataBuilder(AttentionMetadataBuilder):
|
||||
_, _, num_decode_tokens, _ = split_decodes_and_prefills(
|
||||
common_attn_metadata, decode_threshold=1
|
||||
)
|
||||
async_spec_decode = (
|
||||
self.vllm_config.scheduler_config.async_scheduling
|
||||
and self.vllm_config.speculative_config is not None
|
||||
)
|
||||
return CompressorMetadata(
|
||||
block_table=common_attn_metadata.block_table_tensor.clamp_(min=0),
|
||||
slot_mapping=common_attn_metadata.slot_mapping,
|
||||
@@ -124,7 +130,7 @@ class CompressorMetadataBuilder(AttentionMetadataBuilder):
|
||||
num_decode_tokens=num_decode_tokens,
|
||||
c128_boundary=(
|
||||
_get_c128_boundary(common_attn_metadata)
|
||||
if self.block_size == 8
|
||||
if self.block_size == 8 and not async_spec_decode
|
||||
else None
|
||||
),
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user