Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Not sure if AMD actually support fp16 acc but it doesn't crash. (#9258)

comfyanonymous5828607ccfef
diff --git a/comfy/model_management.py b/comfy/model_management.py
index dc5b4711d..c08f759e5 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -340,7 +340,7 @@ if ENABLE_PYTORCH_ATTENTION:
PRIORITIZE_FP16 = False # TODO: remove and replace with something that shows exactly which dtype is faster than the other
try:
- if is_nvidia() and PerformanceFeature.Fp16Accumulation in args.fast:
+ if (is_nvidia() or is_amd()) and PerformanceFeature.Fp16Accumulation in args.fast:
torch.backends.cuda.matmul.allow_fp16_accumulation = True
PRIORITIZE_FP16 = True # TODO: limit to cards where it actually boosts performance
logging.info("Enabled fp16 accumulation.")