[Diffusion] Add native paged KV backend for diffusion workers (#6102)

Signed-off-by: Acerak01-fy <wfy2003324@163.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
Wang fuyinandHongsheng Liu authored and GitHub committed 2026-08-22 21:45:45 +08:00
1 parent c2daa471c3
commit 72ee535fcd
34 files changed
+4595 -71

No files matched your search

+22
View File
@@ -184,6 +184,28 @@ class OmniPlatform(Platform):
"""
return "vllm_omni.diffusion.worker.diffusion_model_runner.DiffusionModelRunner"
@classmethod
def get_diffusion_kv_block_tables_cls(cls) -> type:
"""Return the platform's native paged-KV BlockTables implementation."""
from vllm.v1.worker.gpu.block_table import BlockTables
return BlockTables
@classmethod
def build_diffusion_kv_attn_metadata(cls, **kwargs: Any) -> dict[str, Any]:
"""Build native attention metadata for the diffusion paged path.
The common CUDA/ROCm/XPU path uses vLLM's builder. Platforms with a
backend-specific metadata contract can override this hook without
making the diffusion adapter import their optional runtime package.
``seq_lens_cpu`` is an adapter-only convenience value and is removed
before calling the upstream builder.
"""
from vllm.v1.worker.gpu.attn_utils import build_attn_metadata
kwargs.pop("seq_lens_cpu", None)
return build_attn_metadata(**kwargs)
@classmethod
def init_diffusion_worker_vllm_config(
cls,