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
+4 -3
View File
@@ -28,7 +28,7 @@ from vllm.v1.sample.thinking_budget_state import (
maybe_create_thinking_budget_state_holder,
)
from vllm.v1.utils import copy_slice
from vllm.v1.worker.block_table import MultiGroupBlockTable
from vllm.v1.worker.block_table import MultiGroupBlockTable, SlotMappingMode
@dataclass
@@ -99,13 +99,14 @@ class InputBatch:
vocab_size: int,
block_sizes: list[int], # The block_size of each kv cache group
kernel_block_sizes: list[int],
max_num_blocks_per_req: list[int] | None = None,
max_num_blocks_per_req: list[int],
logitsprocs: LogitsProcessors | None = None,
logitsprocs_need_output_token_ids: bool = False,
num_spec_tokens: int = 0,
is_pooling_model: bool = False,
cp_kv_cache_interleave_size: int = 1,
reasoning_config: ReasoningConfig | None = None,
slot_mapping_modes: list[SlotMappingMode] | None = None,
):
self.thinking_budget_state_holder = maybe_create_thinking_budget_state_holder(
reasoning_config,
@@ -171,7 +172,6 @@ class InputBatch:
# Block table.
self.block_table = MultiGroupBlockTable(
max_num_reqs=max_num_reqs,
max_model_len=max_model_len,
max_num_batched_tokens=max_num_batched_tokens,
pin_memory=PIN_MEMORY,
device=device,
@@ -179,6 +179,7 @@ class InputBatch:
kernel_block_sizes=kernel_block_sizes,
max_num_blocks=max_num_blocks_per_req,
cp_kv_cache_interleave_size=cp_kv_cache_interleave_size,
slot_mapping_modes=slot_mapping_modes,
)
# Sampling-related.