[Blackwell] Add CUDNN_ATTN and FLASHINFER_ATTN backends for diffusion (auto-route) (#3079)

Signed-off-by: lishunyang <lishunyang12@163.com>
Signed-off-by: Zhang <jianmusings@gmail.com>
Signed-off-by: Zhang Jian <jianmusings@gmail.com>
Co-authored-by: frauttauteffasu <247219357+frauttauteffasu@users.noreply.github.com>
Co-authored-by: Zhang Jian <jianmusings@gmail.com>
This commit is contained in:
authored and GitHub committed 2026-05-28 11:09:25 +08:00
1 parent 1903cee56a
commit 373d3c7476
11 files changed
+859 -41

No files matched your search

+10 -2
View File
@@ -94,10 +94,18 @@ class RotaryEmbedding(CustomOp):
self.interleaved = not is_neox_style
self.apply_rotary_emb_flash_attn = None
self.has_mindie = False
# ``find_spec("flash_attn")`` is True as long as *any* package publishes
# the ``flash_attn`` namespace — including ``flash-attn-4``, which ships
# only ``flash_attn.cute`` and no ``flash_attn.ops``. Guard the import
# so a partial namespace doesn't crash the RoPE init; the CUDA forward
# path uses ``vllm.vllm_flash_attn.layers.rotary`` anyway.
if find_spec("flash_attn") is not None:
from flash_attn.ops.triton.rotary import apply_rotary
try:
from flash_attn.ops.triton.rotary import apply_rotary
self.apply_rotary_emb_flash_attn = apply_rotary
self.apply_rotary_emb_flash_attn = apply_rotary
except ImportError:
pass
if find_spec("mindiesd") is not None:
self.has_mindie = True