The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Enable PyTorch SDPA attention and FP8 ops for gfx1170 (#16014)
Tijana Vukoviccf7aedc1e6d6
diff --git a/comfy/model_management.py b/comfy/model_management.pyindex ffc4f129d..2cd1afa67 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -519,13 +519,13 @@ try:if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:if aotriton_supported(): # AMD efficient attention implementation depends on aotriton.if torch_version_numeric >= (2, 7): # works on 2.6 but doesn't actually seem to improve much- if any((a in arch) for a in ["gfx90a", "gfx942", "gfx950", "gfx1100", "gfx1101", "gfx1150", "gfx1151"]): # TODO: more arches, TODO: gfx950+ if any((a in arch) for a in ["gfx90a", "gfx942", "gfx950", "gfx1100", "gfx1101", "gfx1150", "gfx1151", "gfx1170"]): # TODO: more arches, TODO: gfx950ENABLE_PYTORCH_ATTENTION = Trueif rocm_version >= (7, 0):if any((a in arch) for a in ["gfx1200", "gfx1201"]):ENABLE_PYTORCH_ATTENTION = Trueif torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):- if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950"]): # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0+ if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950", "gfx1170"]): # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0SUPPORT_FP8_OPS = Trueexcept: