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 +++
 4 files changed, 50 insertions(+), 3 deletions(-)
 create mode 100644 modules/npu_specific.py

(limited to 'modules')

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]
 
-- 
cgit v1.2.3