The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
CPU device doesn't support fp8, nvfp4, mxfp8 or int8. (#17170)
comfyanonymousd4056d50c3c5
diff --git a/comfy/model_management.py b/comfy/model_management.pyindex 13b690bdb..6e6aaefa3 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -2018,6 +2018,9 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, mareturn Falsedef supports_fp8_compute(device=None):+ if is_device_cpu(device):+ return False+if SUPPORT_FP8_OPS:return True@@ -2042,6 +2045,9 @@ def supports_fp8_compute(device=None):return Truedef supports_nvfp4_compute(device=None):+ if is_device_cpu(device):+ return False+if not is_nvidia():return False@@ -2052,6 +2058,9 @@ def supports_nvfp4_compute(device=None):return Truedef supports_mxfp8_compute(device=None):+ if is_device_cpu(device):+ return False+if not is_nvidia():return False@@ -2080,6 +2089,9 @@ def supports_fp64(device=None):return Truedef supports_int8_compute(device=None):+ if is_device_cpu(device):+ return False+# The eager comfy_kitchen backend implements int8 weight-only quantized# matmul via torch._int_mm, which PyTorch does not implement for MPS.# https://github.com/pytorch/pytorch/issues/141287