From ec124607f47371a6cfd61a795f86a7f1cbd44651 Mon Sep 17 00:00:00 2001
From: wangshuai09 <391746016@qq.com>
Date: Sat, 27 Jan 2024 17:21:32 +0800
Subject: Add NPU Support

---
 modules/devices.py                             |  9 +++++--
 modules/initialize.py                          |  6 ++++-
 modules/npu_specific.py                        | 34 ++++++++++++++++++++++++++
 modules/textual_inversion/textual_inversion.py |  4 +++
 requirements.txt                               |  4 +++
 requirements_versions.txt                      |  4 +++
 webui.sh                                       |  4 +++
 7 files changed, 62 insertions(+), 3 deletions(-)
 create mode 100644 modules/npu_specific.py

diff --git a/modules/devices.py b/modules/devices.py
index ea1f712f..f1e56501 100644
--- a/modules/devices.py
+++ b/modules/devices.py
@@ -3,7 +3,7 @@ import contextlib
 from functools import lru_cache
 
 import torch
-from modules import errors, shared
+from modules import errors, shared, npu_specific
 
 if sys.platform == "darwin":
     from modules import mac_specific
@@ -40,6 +40,9 @@ def get_optimal_device_name():
     if has_xpu():
         return xpu_specific.get_xpu_device_string()
 
+    if npu_specific.has_npu:
+        return npu_specific.get_npu_device_string()
+
     return "cpu"
 
 
@@ -67,6 +70,9 @@ def torch_gc():
     if has_xpu():
         xpu_specific.torch_xpu_gc()
 
+    if npu_specific.has_npu:
+        npu_specific.torch_npu_gc()
+
 
 def enable_tf32():
     if torch.cuda.is_available():
@@ -164,4 +170,3 @@ def first_time_calculation():
     x = torch.zeros((1, 1, 3, 3)).to(device, dtype)
     conv2d = torch.nn.Conv2d(1, 1, (3, 3)).to(device, dtype)
     conv2d(x)
-
diff --git a/modules/initialize.py b/modules/initialize.py
index ac95fc6f..3285cc3c 100644
--- a/modules/initialize.py
+++ b/modules/initialize.py
@@ -143,13 +143,17 @@ def initialize_rest(*, reload_script_modules=False):
         its optimization may be None because the list of optimizaers has neet been filled
         by that time, so we apply optimization again.
         """
+        from modules import devices
+        # Work around due to bug in torch_npu, revert me after fixed, @see https://gitee.com/ascend/pytorch/issues/I8KECW?from=project-issue
+        if devices.npu_specific.has_npu:
+            import torch
+            torch.npu.set_device(0)
 
         shared.sd_model  # noqa: B018
 
         if sd_hijack.current_optimizer is None:
             sd_hijack.apply_optimizations()
 
-        from modules import devices
         devices.first_time_calculation()
     if not shared.cmd_opts.skip_load_model_at_start:
         Thread(target=load_model).start()
diff --git a/modules/npu_specific.py b/modules/npu_specific.py
new file mode 100644
index 00000000..d8aebf9c
--- /dev/null
+++ b/modules/npu_specific.py
@@ -0,0 +1,34 @@
+import importlib
+import torch
+
+from modules import shared
+
+
+def check_for_npu():
+    if importlib.util.find_spec("torch_npu") is None:
+        return False
+    import torch_npu
+    torch_npu.npu.set_device(0)
+
+    try:
+        # Will raise a RuntimeError if no NPU is found
+        _ = torch.npu.device_count()
+        return torch.npu.is_available()
+    except RuntimeError:
+        return False
+
+
+def get_npu_device_string():
+    if shared.cmd_opts.device_id is not None:
+        return f"npu:{shared.cmd_opts.device_id}"
+    return "npu:0"
+
+
+def torch_npu_gc():
+    # Work around due to bug in torch_npu, revert me after fixed, @see https://gitee.com/ascend/pytorch/issues/I8KECW?from=project-issue
+    torch.npu.set_device(0)
+    with torch.npu.device(get_npu_device_string()):
+        torch.npu.empty_cache()
+
+
+has_npu = check_for_npu()
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index 04dda585..9c062503 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -151,6 +151,10 @@ class EmbeddingDatabase:
         return embedding
 
     def get_expected_shape(self):
+        # workaround
+        if devices.npu_specific.has_npu:
+            import torch
+            torch.npu.set_device(0)
         vec = shared.sd_model.cond_stage_model.encode_embedding_init_text(",", 1)
         return vec.shape[1]
 
diff --git a/requirements.txt b/requirements.txt
index 80b43845..4537402b 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -5,6 +5,8 @@ accelerate
 basicsr
 blendmodes
 clean-fid
+cloudpickle
+decorator
 einops
 fastapi>=0.90.1
 gfpgan
@@ -26,9 +28,11 @@ resize-right
 
 safetensors
 scikit-image>=0.19
+synr==0.5.0
 timm
 tomesd
 torch
 torchdiffeq
 torchsde
+tornado
 transformers==4.30.2
diff --git a/requirements_versions.txt b/requirements_versions.txt
index cb7403a9..95515b55 100644
--- a/requirements_versions.txt
+++ b/requirements_versions.txt
@@ -4,6 +4,8 @@ accelerate==0.21.0
 basicsr==1.4.2
 blendmodes==2022
 clean-fid==0.1.35
+cloudpickle==3.0.0
+decorator==5.1.1
 einops==0.4.1
 fastapi==0.94.0
 gfpgan==1.3.8
@@ -23,10 +25,12 @@ realesrgan==0.3.0
 resize-right==0.0.2
 safetensors==0.3.1
 scikit-image==0.21.0
+synr==0.5.0
 timm==0.9.2
 tomesd==0.1.3
 torch
 torchdiffeq==0.2.3
 torchsde==0.2.6
+tornado==6.4
 transformers==4.30.2
 httpx==0.24.1
diff --git a/webui.sh b/webui.sh
index cff43327..3f6e87fd 100755
--- a/webui.sh
+++ b/webui.sh
@@ -159,6 +159,10 @@ then
     if echo "$gpu_info" | grep -q "AMD" && [[ -z "${TORCH_COMMAND}" ]]
     then
         export TORCH_COMMAND="pip install torch==2.0.1+rocm5.4.2 torchvision==0.15.2+rocm5.4.2 --index-url https://download.pytorch.org/whl/rocm5.4.2"
+    elif echo "$gpu_info" | grep -q "Huawei" && [[ -z "${TORCH_COMMAND}" ]]
+    then
+        export TORCH_COMMAND="pip install torch==2.1.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu; pip install torch_npu"
+    
     fi
 fi
 
-- 
cgit v1.2.1


From cc3f604310458eed7d26456c1b3934d582283ffe Mon Sep 17 00:00:00 2001
From: wangshuai09 <391746016@qq.com>
Date: Wed, 31 Jan 2024 10:46:53 +0800
Subject: Update

---
 modules/devices.py                             | 7 +++++++
 modules/initialize.py                          | 5 +----
 modules/launch_utils.py                        | 8 ++++++++
 modules/npu_specific.py                        | 5 +----
 modules/textual_inversion/textual_inversion.py | 5 +----
 requirements.txt                               | 4 ----
 requirements_npu.txt                           | 4 ++++
 requirements_versions.txt                      | 4 ----
 8 files changed, 22 insertions(+), 20 deletions(-)
 create mode 100644 requirements_npu.txt

diff --git a/modules/devices.py b/modules/devices.py
index c737162a..28c0c54d 100644
--- a/modules/devices.py
+++ b/modules/devices.py
@@ -88,9 +88,16 @@ def torch_gc():
         xpu_specific.torch_xpu_gc()
 
     if npu_specific.has_npu:
+        torch_npu_set_device()
         npu_specific.torch_npu_gc()
 
 
+def torch_npu_set_device():
+    # Work around due to bug in torch_npu, revert me after fixed, @see https://gitee.com/ascend/pytorch/issues/I8KECW?from=project-issue
+    if npu_specific.has_npu:
+        torch.npu.set_device(0)
+
+
 def enable_tf32():
     if torch.cuda.is_available():
 
diff --git a/modules/initialize.py b/modules/initialize.py
index cc34fd6f..f7313ff4 100644
--- a/modules/initialize.py
+++ b/modules/initialize.py
@@ -143,10 +143,7 @@ def initialize_rest(*, reload_script_modules=False):
         by that time, so we apply optimization again.
         """
         from modules import devices
-        # Work around due to bug in torch_npu, revert me after fixed, @see https://gitee.com/ascend/pytorch/issues/I8KECW?from=project-issue
-        if devices.npu_specific.has_npu:
-            import torch
-            torch.npu.set_device(0)
+        devices.torch_npu_set_device()
 
         shared.sd_model  # noqa: B018
 
diff --git a/modules/launch_utils.py b/modules/launch_utils.py
index 3ff4576a..107c72b0 100644
--- a/modules/launch_utils.py
+++ b/modules/launch_utils.py
@@ -338,6 +338,7 @@ def prepare_environment():
             torch_index_url = os.environ.get('TORCH_INDEX_URL', "https://pytorch-extension.intel.com/release-whl/stable/xpu/us/")
             torch_command = os.environ.get('TORCH_COMMAND', f"pip install torch==2.0.0a0 intel-extension-for-pytorch==2.0.110+gitba7f6c1 --extra-index-url {torch_index_url}")
     requirements_file = os.environ.get('REQS_FILE', "requirements_versions.txt")
+    requirements_file_for_npu = os.environ.get('REQS_FILE_FOR_NPU', "requirements_npu.txt")
 
     xformers_package = os.environ.get('XFORMERS_PACKAGE', 'xformers==0.0.23.post1')
     clip_package = os.environ.get('CLIP_PACKAGE', "https://github.com/openai/CLIP/archive/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1.zip")
@@ -421,6 +422,13 @@ def prepare_environment():
         run_pip(f"install -r \"{requirements_file}\"", "requirements")
         startup_timer.record("install requirements")
 
+    if not os.path.isfile(requirements_file_for_npu):
+        requirements_file_for_npu = os.path.join(script_path, requirements_file_for_npu)
+
+    if "torch_npu" in torch_command and not requirements_met(requirements_file_for_npu):
+        run_pip(f"install -r \"{requirements_file_for_npu}\"", "requirements_for_npu")
+        startup_timer.record("install requirements_for_npu")
+
     if not args.skip_install:
         run_extensions_installers(settings_file=args.ui_settings_file)
 
diff --git a/modules/npu_specific.py b/modules/npu_specific.py
index d8aebf9c..94100691 100644
--- a/modules/npu_specific.py
+++ b/modules/npu_specific.py
@@ -8,11 +8,10 @@ def check_for_npu():
     if importlib.util.find_spec("torch_npu") is None:
         return False
     import torch_npu
-    torch_npu.npu.set_device(0)
 
     try:
         # Will raise a RuntimeError if no NPU is found
-        _ = torch.npu.device_count()
+        _ = torch_npu.npu.device_count()
         return torch.npu.is_available()
     except RuntimeError:
         return False
@@ -25,8 +24,6 @@ def get_npu_device_string():
 
 
 def torch_npu_gc():
-    # Work around due to bug in torch_npu, revert me after fixed, @see https://gitee.com/ascend/pytorch/issues/I8KECW?from=project-issue
-    torch.npu.set_device(0)
     with torch.npu.device(get_npu_device_string()):
         torch.npu.empty_cache()
 
diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index d16e3b9a..6d815c0b 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -150,10 +150,7 @@ class EmbeddingDatabase:
         return embedding
 
     def get_expected_shape(self):
-        # workaround
-        if devices.npu_specific.has_npu:
-            import torch
-            torch.npu.set_device(0)
+        devices.torch_npu_set_device()
         vec = shared.sd_model.cond_stage_model.encode_embedding_init_text(",", 1)
         return vec.shape[1]
 
diff --git a/requirements.txt b/requirements.txt
index d1e4ede9..731a1be7 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -4,8 +4,6 @@ accelerate
 
 blendmodes
 clean-fid
-cloudpickle
-decorator
 einops
 facexlib
 fastapi>=0.90.1
@@ -26,10 +24,8 @@ resize-right
 
 safetensors
 scikit-image>=0.19
-synr==0.5.0
 tomesd
 torch
 torchdiffeq
 torchsde
-tornado
 transformers==4.30.2
diff --git a/requirements_npu.txt b/requirements_npu.txt
new file mode 100644
index 00000000..5e6a4364
--- /dev/null
+++ b/requirements_npu.txt
@@ -0,0 +1,4 @@
+cloudpickle
+decorator
+synr==0.5.0
+tornado
diff --git a/requirements_versions.txt b/requirements_versions.txt
index 1c66cd8c..5e30b5ea 100644
--- a/requirements_versions.txt
+++ b/requirements_versions.txt
@@ -3,8 +3,6 @@ Pillow==9.5.0
 accelerate==0.21.0
 blendmodes==2022
 clean-fid==0.1.35
-cloudpickle==3.0.0
-decorator==5.1.1
 einops==0.4.1
 facexlib==0.3.0
 fastapi==0.94.0
@@ -23,12 +21,10 @@ pytorch_lightning==1.9.4
 resize-right==0.0.2
 safetensors==0.4.2
 scikit-image==0.21.0
-synr==0.5.0
 spandrel==0.1.6
 tomesd==0.1.3
 torch
 torchdiffeq==0.2.3
 torchsde==0.2.6
-tornado==6.4
 transformers==4.30.2
 httpx==0.24.1
-- 
cgit v1.2.1