[Bugfix][MRV2] Initialize DCP metadata after batch partitioning (#55212)

Signed-off-by: Raphael Rialland <raphael.rialland@mistral.ai>
Signed-off-by: Ziyang Zhang <hafuhafu@qq.com>
Signed-off-by: Sy03 <1370724210@qq.com>
Co-authored-by: Raphael Rialland <raphael.rialland@mistral.ai>
Co-authored-by: Ziyang Zhang <hafuhafu@qq.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Sy03
2026-09-10 06:50:09 +00:00
committed by GitHub
co-authored by Raphael Rialland Ziyang Zhang Codex mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 912d2e79fe
commit 83990f5bcc
12 changed files with 460 additions and 77 deletions
+10 -12
View File
@@ -38,7 +38,7 @@ from vllm.v1.kv_cache_interface import KVCacheConfig
from vllm.v1.spec_decode.dynamic.utils import build_dynamic_sd_schedule_lookup
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 prepare_dcp_local_seq_lens
from vllm.v1.worker.gpu.cp_utils import maybe_prepare_dcp_local_seq_lens
from vllm.v1.worker.gpu.input_batch import InputBatch, InputBuffers
from vllm.v1.worker.gpu.model_states.interface import ModelState
from vllm.v1.worker.utils import AttentionGroup, clear_layer_kv_caches
@@ -699,17 +699,15 @@ def prepare_inputs_to_capture(
slot_mappings, kv_cache_config
)
# HACK(woosuk): Special handling for DCP.
if block_tables.cp_size > 1:
prepare_dcp_local_seq_lens(
input_buffers.dcp_local_seq_lens,
input_batch.seq_lens,
num_reqs,
block_tables.cp_size,
block_tables.cp_rank,
block_tables.cp_interleave,
)
input_batch.dcp_local_seq_lens = input_buffers.dcp_local_seq_lens[:num_reqs]
input_batch.dcp_local_seq_lens = maybe_prepare_dcp_local_seq_lens(
input_buffers.dcp_local_seq_lens,
input_batch.seq_lens,
input_batch.num_reqs,
block_tables.cp_size,
block_tables.cp_rank,
block_tables.cp_interleave,
num_reqs_padded=input_batch.num_reqs_after_padding,
)
# NOTE(woosuk): Attention metadata is required not just by standard attention
# kernels, but also by specialized attention-like operations (e.g., Inkling's sconv,