[Attention][Spec Decode] Enable varlen Mamba2 decode for adaptive verification (#53036)

Signed-off-by: Guan-Ming (Wesley) Chiu <105915352+guan404ming@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
This commit is contained in:
Guan-Ming Chiu
2026-10-16 11:23:59 -07:00
committed by GitHub
co-authored by vllm-agent
parent e7a5bdb454
commit 3dc6def360
2 changed files with 131 additions and 2 deletions
+19 -2
View File
@@ -6,7 +6,7 @@ from typing import Any
import torch
from vllm.config import VllmConfig
from vllm.config import VllmConfig, get_current_vllm_config_or_none
from vllm.model_executor.layers.mamba.checkpoint import (
MambaPrefillCheckpointBuilder,
MambaPrefillCheckpointMetadata,
@@ -15,12 +15,13 @@ from vllm.utils.torch_utils import async_tensor_h2d
from vllm.v1.attention.backend import (
AttentionBackend,
CommonAttentionMetadata,
max_decode_query_len,
)
from vllm.v1.attention.backends.mamba_attn import (
BaseMambaAttentionMetadata,
BaseMambaAttentionMetadataBuilder,
)
from vllm.v1.kv_cache_interface import MambaSpec
from vllm.v1.kv_cache_interface import KVCacheSpec, MambaSpec
def compute_varlen_chunk_metadata(
@@ -104,6 +105,14 @@ class Mamba2AttentionBackend(AttentionBackend):
def is_ssm(cls) -> bool:
return True
@classmethod
def supports_device_cpu_query_lens_mismatch(cls) -> bool:
# Decode kernels take device offsets; ReplaySSM plans off CPU counts.
vllm_config = get_current_vllm_config_or_none()
if vllm_config is None:
return True
return not vllm_config.cache_config.use_replayssm
@dataclass
class Mamba2AttentionMetadata(BaseMambaAttentionMetadata):
@@ -124,6 +133,14 @@ class Mamba2AttentionMetadataBuilder(
):
metadata_cls = Mamba2AttentionMetadata
@classmethod
def get_varlen_cudagraph_max_query_len(
cls,
vllm_config: VllmConfig,
kv_cache_spec: KVCacheSpec,
) -> int | None:
return max_decode_query_len(vllm_config)
def __init__(
self,
kv_cache_spec: MambaSpec,