[Diffusion] Add paged KV cache support for HunyuanImage3 DiT on NPU and GPU (#6563)
Signed-off-by: Acerak01-fy <wfy2003324@163.com> Signed-off-by: zwhzzz0821 <zwhzzz0821@users.noreply.github.com> Signed-off-by: zwhzzz0821 <2831474076@qq.com> Co-authored-by: zwhzzz0821 <zwhzzz0821@users.noreply.github.com> Co-authored-by: zwhzzz0821 <2831474076@qq.com>
This commit is contained in:
28 files changed
+1789
-205
No files matched your search
@@ -1,5 +1,5 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
|
||||
|
||||
from contextlib import nullcontext
|
||||
from enum import Enum
|
||||
@@ -150,6 +150,12 @@ class OmniPlatform(Platform):
|
||||
)
|
||||
backend_cls.validate_available()
|
||||
|
||||
@classmethod
|
||||
def supports_diffusion_dense_flash_attention(cls) -> bool:
|
||||
"""Whether the platform's dense ``FLASH_ATTN`` dependencies exist."""
|
||||
|
||||
return True
|
||||
|
||||
@classmethod
|
||||
def supports_torch_inductor(cls) -> bool:
|
||||
"""Check if the platform supports torch.compile with inductor backend."""
|
||||
@@ -195,17 +201,36 @@ class OmniPlatform(Platform):
|
||||
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.
|
||||
This default is the GPU/common path and uses vLLM's builder. NPU
|
||||
overrides it to build Ascend attention metadata without making the
|
||||
shared diffusion adapter import ``vllm_ascend``. ``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 get_diffusion_paged_kv_attn_backend(cls, attn_backend: type, *, ulysses_degree: int) -> type:
|
||||
"""Specialize a native paged backend for diffusion execution."""
|
||||
|
||||
del ulysses_degree
|
||||
return attn_backend
|
||||
|
||||
@classmethod
|
||||
def requires_diffusion_paged_kv_prewrite(cls) -> bool:
|
||||
"""Whether paged attention must write K/V before native execution.
|
||||
|
||||
The default GPU path keeps cache-update ownership in its native
|
||||
attention call. Ascend overrides this because piecewise FIA should
|
||||
write the complete K/V span once, then read it from cache for every
|
||||
segment.
|
||||
"""
|
||||
|
||||
return False
|
||||
|
||||
@classmethod
|
||||
def init_diffusion_worker_vllm_config(
|
||||
cls,
|
||||
@@ -214,6 +239,15 @@ class OmniPlatform(Platform):
|
||||
"""Initialize platform-specific state for diffusion worker VllmConfig."""
|
||||
return None
|
||||
|
||||
@classmethod
|
||||
def configure_diffusion_vllm_config(
|
||||
cls,
|
||||
vllm_config: Any,
|
||||
od_config: Any,
|
||||
) -> None:
|
||||
"""Apply platform-specific native cache geometry for diffusion."""
|
||||
return None
|
||||
|
||||
@classmethod
|
||||
def init_diffusion_model_runner_runtime(
|
||||
cls,
|
||||
|
||||
Reference in new issue
Block a user