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.

Allow FP16 accumulation with `--fast` (#6453)

catboxanon43a74c0de175
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 225a83e05..ca84f2064 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -241,6 +241,12 @@ if ENABLE_PYTORCH_ATTENTION:
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_mem_efficient_sdp(True)
+try:
+ if is_nvidia() and args.fast:
+ torch.backends.cuda.matmul.allow_fp16_accumulation = True
+except:
+ pass
+
try:
if int(torch_version[0]) == 2 and int(torch_version[2]) >= 5:
torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)