The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Support ascend npu (#5436)
Huazhong Jic4bfdba3301e
diff --git a/README.md b/README.mdindex 371421617..000d76801 100644--- a/README.md+++ b/README.md@@ -224,6 +224,16 @@ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS ve```pip install torch-directml``` Then you can launch ComfyUI with: ```python main.py --directml```+#### Ascend NPUs++For models compatible with Ascend Extension for PyTorch (torch_npu). To get started, ensure your environment meets the prerequisites outlined on the [installation](https://ascend.github.io/docs/sources/ascend/quick_install.html) page. Here's a step-by-step guide tailored to your platform and installation method:++1. Begin by installing the recommended or newer kernel version for Linux as specified in the Installation page of torch-npu, if necessary.+2. Proceed with the installation of Ascend Basekit, which includes the driver, firmware, and CANN, following the instructions provided for your specific platform.+3. Next, install the necessary packages for torch-npu by adhering to the platform-specific instructions on the [Installation](https://ascend.github.io/docs/sources/pytorch/install.html#pytorch) page.+4. Finally, adhere to the [ComfyUI manual installation](#manual-install-windows-linux) guide for Linux. Once all components are installed, you can run ComfyUI as described earlier.++# Running```python main.py```diff --git a/comfy/model_management.py b/comfy/model_management.pyindex db2a61395..c36c52ffd 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -86,6 +86,13 @@ try:except:pass+try:+ import torch_npu+ _ = torch.npu.device_count()+ npu_available = torch.npu.is_available()+except:+ npu_available = False+if args.cpu:cpu_state = CPUState.CPU@@ -97,6 +104,12 @@ def is_intel_xpu():return Truereturn False+def is_ascend_npu():+ global npu_available+ if npu_available:+ return True+ return False+def get_torch_device():global directml_enabledglobal cpu_state@@ -110,6 +123,8 @@ def get_torch_device():else:if is_intel_xpu():return torch.device("xpu", torch.xpu.current_device())+ elif is_ascend_npu():+ return torch.device("npu", torch.npu.current_device())else:return torch.device(torch.cuda.current_device())@@ -130,6 +145,12 @@ def get_total_memory(dev=None, torch_total_too=False):mem_reserved = stats['reserved_bytes.all.current']mem_total_torch = mem_reservedmem_total = torch.xpu.get_device_properties(dev).total_memory+ elif is_ascend_npu():+ stats = torch.npu.memory_stats(dev)+ mem_reserved = stats['reserved_bytes.all.current']+ _, mem_total_npu = torch.npu.mem_get_info(dev)+ mem_total_torch = mem_reserved+ mem_total = mem_total_npuelse:stats = torch.cuda.memory_stats(dev)mem_reserved = stats['reserved_bytes.all.current']@@ -209,7 +230,7 @@ try:if int(torch_version[0]) >= 2:if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:ENABLE_PYTORCH_ATTENTION = True- if is_intel_xpu():+ if is_intel_xpu() or is_ascend_npu():if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:ENABLE_PYTORCH_ATTENTION = Trueexcept:@@ -274,6 +295,8 @@ def get_torch_device_name(device):return "{}".format(device.type)elif is_intel_xpu():return "{} {}".format(device, torch.xpu.get_device_name(device))+ elif is_ascend_npu():+ return "{} {}".format(device, torch.npu.get_device_name(device))else:return "CUDA {}: {}".format(device, torch.cuda.get_device_name(device))@@ -860,6 +883,8 @@ def xformers_enabled():return Falseif is_intel_xpu():return False+ if is_ascend_npu():+ return Falseif directml_enabled:return Falsereturn XFORMERS_IS_AVAILABLE@@ -884,6 +909,8 @@ def pytorch_attention_flash_attention():return Trueif is_intel_xpu():return True+ if is_ascend_npu():+ return Truereturn Falsedef mac_version():@@ -923,6 +950,13 @@ def get_free_memory(dev=None, torch_free_too=False):mem_free_torch = mem_reserved - mem_activemem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reservedmem_free_total = mem_free_xpu + mem_free_torch+ elif is_ascend_npu():+ stats = torch.npu.memory_stats(dev)+ mem_active = stats['active_bytes.all.current']+ mem_reserved = stats['reserved_bytes.all.current']+ mem_free_npu, _ = torch.npu.mem_get_info(dev)+ mem_free_torch = mem_reserved - mem_active+ mem_free_total = mem_free_npu + mem_free_torchelse:stats = torch.cuda.memory_stats(dev)mem_active = stats['active_bytes.all.current']@@ -984,6 +1018,9 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, maif is_intel_xpu():return True+ if is_ascend_npu():+ return True+if torch.version.hip:return True@@ -1081,6 +1118,8 @@ def soft_empty_cache(force=False):torch.mps.empty_cache()elif is_intel_xpu():torch.xpu.empty_cache()+ elif is_ascend_npu():+ torch.npu.empty_cache()elif torch.cuda.is_available():if force or is_nvidia(): #This seems to make things worse on ROCm so I only do it for cudatorch.cuda.empty_cache()