[Attention][DCP] Enable FlashInfer MLA for DSpark drafting (#54277)
Signed-off-by: Summer Yang <girasoleyang@gmail.com>
This commit is contained in:
@@ -18,7 +18,7 @@ from vllm.v1.attention.backends.utils import PAD_SLOT_ID
|
||||
from vllm.v1.kv_cache_interface import KVCacheConfig
|
||||
from vllm.v1.worker.gpu.attn_utils import build_slot_mappings_by_layer
|
||||
from vllm.v1.worker.gpu.block_table import BlockTables
|
||||
from vllm.v1.worker.gpu.cp_utils import cp_local_slot, prepare_dcp_local_seq_lens
|
||||
from vllm.v1.worker.gpu.cp_utils import cp_local_slot
|
||||
from vllm.v1.worker.gpu.dp_utils import DPSyncState, dispatch_cg_and_sync_dp
|
||||
from vllm.v1.worker.gpu.input_batch import InputBatch, InputBuffers
|
||||
from vllm.v1.worker.gpu.model_states.interface import ModelState
|
||||
@@ -291,16 +291,6 @@ class DFlashSpeculator(DraftModelSpeculator):
|
||||
if not self.draft_attn_layer_names:
|
||||
return None
|
||||
assert num_query_per_req is None # Omitted for DFlash, read from self instead
|
||||
if dcp_local_seq_lens is None and self.block_tables.cp_size > 1:
|
||||
prepare_dcp_local_seq_lens(
|
||||
self.input_buffers.dcp_local_seq_lens,
|
||||
self.input_buffers.seq_lens,
|
||||
num_reqs,
|
||||
self.block_tables.cp_size,
|
||||
self.block_tables.cp_rank,
|
||||
self.block_tables.cp_interleave,
|
||||
)
|
||||
dcp_local_seq_lens = self.input_buffers.dcp_local_seq_lens
|
||||
return super()._build_draft_attn_metadata(
|
||||
num_reqs,
|
||||
num_reqs_padded,
|
||||
|
||||
@@ -21,6 +21,7 @@ from vllm.v1.worker.gpu.attn_utils import (
|
||||
init_attn_backend,
|
||||
)
|
||||
from vllm.v1.worker.gpu.block_table import BlockTables
|
||||
from vllm.v1.worker.gpu.cp_utils import prepare_dcp_local_seq_lens
|
||||
from vllm.v1.worker.gpu.dp_utils import DPSyncState
|
||||
from vllm.v1.worker.gpu.input_batch import InputBatch, InputBuffers
|
||||
from vllm.v1.worker.gpu.model_states.interface import ModelState
|
||||
@@ -292,6 +293,18 @@ class DraftModelSpeculator(BaseSpeculator):
|
||||
out=draft_seq_lens_cpu_upper_bound[:num_reqs],
|
||||
)
|
||||
draft_seq_lens_cpu_upper_bound[:num_reqs].clamp_(max=self.max_model_len)
|
||||
if dcp_local_seq_lens is None and self.block_tables.cp_size > 1:
|
||||
# Draft steps advance and rewind their own global sequence lengths,
|
||||
# so the target model's DCP-local lengths may already be stale.
|
||||
prepare_dcp_local_seq_lens(
|
||||
self.input_buffers.dcp_local_seq_lens,
|
||||
self.input_buffers.seq_lens,
|
||||
num_reqs,
|
||||
self.block_tables.cp_size,
|
||||
self.block_tables.cp_rank,
|
||||
self.block_tables.cp_interleave,
|
||||
)
|
||||
dcp_local_seq_lens = self.input_buffers.dcp_local_seq_lens
|
||||
attn_metadata = build_attn_metadata(
|
||||
attn_groups=self.attn_groups,
|
||||
num_reqs=num_reqs_padded,
|
||||
|
||||
Reference in New Issue
Block a user