Upate OpenVINO to PyTorch 2.6 and fix mismatched shapes error on too …

…many resolution changes

CHANGELOG.md

@@ -11,13 +11,13 @@
   - set in *settings -> user interface -> language*
   - [localization](https://github.com/vladmandic/wiki/Locale) documentation  
 - **Torch**:  
-  - for **zluda** set default to `torch==2.6.0+cu126`  
+  - for **zluda** set default to `torch==2.6.0+cu118`  
+  - for **openvino** set default to `torch==2.6.0+cpu` and `openvino==2025.0.0`  
 - **Other**:  
   - asymmetric tiling  
     allows for configurable image tiling for x/y axis separately  
     enable in *scripts -> asymmetric tiling*  
     *note*: traditional symmetric tiling is achieved by setting circular mode for both x and y  
-  - update openvino to `2025.0.0`  
 - **UI**:  
   - force browser cache-invalidate on page load  
   - use correct timezone for log display  

installer.py

@@ -726,13 +726,12 @@ def install_ipex(torch_command):
 
 def install_openvino(torch_command):
     t_start = time.time()
-    # Python 3.12: RuntimeError: Dynamo is not supported on Python 3.12+
-    check_python(supported_minors=[9, 10, 11], reason='OpenVINO backend requires a Python version from 3.9, 3.10 or 3.11')
+    check_python(supported_minors=[9, 10, 11, 12], reason='OpenVINO backend requires a Python version between 3.9 and 3.12')
     log.info('OpenVINO: selected')
     if sys.platform == 'darwin':
-        torch_command = os.environ.get('TORCH_COMMAND', 'torch==2.3.1 torchvision==0.18.1')
+        torch_command = os.environ.get('TORCH_COMMAND', 'torch==2.6.0 torchvision==0.21.0')
     else:
-        torch_command = os.environ.get('TORCH_COMMAND', 'torch==2.3.1+cpu torchvision==0.18.1+cpu --index-url https://download.pytorch.org/whl/cpu')
+        torch_command = os.environ.get('TORCH_COMMAND', 'torch==2.6.0+cpu torchvision==0.21.0+cpu --index-url https://download.pytorch.org/whl/cpu')
 
     install(os.environ.get('OPENVINO_COMMAND', 'openvino==2025.0.0'), 'openvino')
     install(os.environ.get('NNCF_COMMAND', 'nncf==2.15.0'), 'nncf')

modules/control/proc/zoe/zoedepth/models/base_models/midas_repo/midas/model_loader.py

@@ -42,7 +42,7 @@ def load_model(device, model_path, model_type="dpt_large_384", optimize=True, he
         network input
     """
     if "openvino" in model_type:
-        from openvino.runtime import Core
+        from openvino import Core
 
     keep_aspect_ratio = not square
 

modules/intel/openvino/__init__.py

@@ -3,10 +3,10 @@
 import torch
 import nncf
 
-from openvino.frontend import FrontEndManager
-from openvino.frontend.pytorch.fx_decoder import TorchFXPythonDecoder
 from openvino.frontend.pytorch.torchdynamo.partition import Partitioner
-from openvino.runtime import Core, Type, PartialShape, serialize
+from openvino.frontend.pytorch.fx_decoder import TorchFXPythonDecoder
+from openvino.frontend import FrontEndManager
+from openvino import Core, Type, PartialShape, serialize
 from openvino.properties import hint as ov_hints
 
 from torch._dynamo.backends.common import fake_tensor_unsupported
@@ -23,6 +23,11 @@
 from modules import shared, devices, sd_models
 
 
+torch._dynamo.eval_frame.check_if_dynamo_supported = lambda: True # pylint: disable=protected-access
+if hasattr(torch._dynamo.config, "inline_inbuilt_nn_modules"):
+    torch._dynamo.config.inline_inbuilt_nn_modules = False # pylint: disable=protected-access
+
+
 DEFAULT_OPENVINO_PYTHON_CONFIG = MappingProxyType(
     {
         "use_python_fusion_cache": True,
@@ -114,9 +119,9 @@ def cached_model_name(model_hash_str, device, args, cache_root, reversed = False
     for input_data in args:
         if isinstance(input_data, torch.SymInt):
             if reversed:
-                inputs_str = "_" + "torch.SymInt1" + inputs_str
+                inputs_str = "_" + "torch.SymInt[]" + inputs_str
             else:
-                inputs_str += "_" + "torch.SymInt1"
+                inputs_str += "_" + "torch.SymInt[]"
         elif isinstance(input_data, int):
             pass
         else:
@@ -176,7 +181,7 @@ def openvino_compile(gm: GraphModule, *example_inputs, model_hash_str: str = Non
         for input_data in example_inputs:
             if isinstance(input_data, torch.SymInt):
                 input_types.append(torch.SymInt)
-                input_shapes.append(torch.Size([1]))
+                input_shapes.append(torch.Size([]))
             elif isinstance(input_data, int):
                 pass
             else:
@@ -426,9 +431,8 @@ def get_subgraph_type(tensor):
     return tensor
 
 
-@register_backend
 @fake_tensor_unsupported
-def openvino_fx(subgraph, example_inputs):
+def openvino_fx(subgraph, example_inputs, options=None):
     global dont_use_4bit_nncf
     global dont_use_nncf
     global dont_use_quant
@@ -528,6 +532,8 @@ def _call(*args):
     for node in model.graph.nodes:
         if node.target == torch.ops.aten.mul_.Tensor:
             node.target = torch.ops.aten.mul.Tensor
+        elif node.target == torch.ops.aten._unsafe_index.Tensor:
+            node.target = torch.ops.aten.index.Tensor
     with devices.inference_context():
         model.eval()
     partitioner = Partitioner(options=None)
@@ -543,3 +549,7 @@ def _call(*args):
         res = execute(compiled_model, *args, executor="openvino", executor_parameters=executor_parameters, file_name=maybe_fs_cached_name)
         return res
     return _call
+
+
+if "openvino_fx" not in torch.compiler.list_backends():
+    register_backend(compiler_fn=openvino_fx, name="openvino_fx")

modules/sd_models_compile.py

@@ -63,7 +63,6 @@ def ipex_optimize_model(model, op=None, sd_model=None): # pylint: disable=unused
 def optimize_openvino(sd_model):
     try:
         from modules.intel.openvino import openvino_fx # pylint: disable=unused-import
-        torch._dynamo.eval_frame.check_if_dynamo_supported = lambda: True # pylint: disable=protected-access
         if shared.compiled_model_state is not None:
             shared.compiled_model_state.compiled_cache.clear()
             shared.compiled_model_state.req_cache.clear()
@@ -164,13 +163,15 @@ def torch_compile_model(model, op=None, sd_model=None): # pylint: disable=unused
                 model = torch.compile(model.to(devices.device),
                     mode=shared.opts.cuda_compile_mode,
                     backend=shared.opts.cuda_compile_backend,
-                    fullgraph=shared.opts.cuda_compile_fullgraph
+                    fullgraph=shared.opts.cuda_compile_fullgraph,
+                    dynamic=None if shared.opts.cuda_compile_backend != "openvino_fx" else False,
                 ).to(return_device)
             else:
                 model = torch.compile(model,
                     mode=shared.opts.cuda_compile_mode,
                     backend=shared.opts.cuda_compile_backend,
-                    fullgraph=shared.opts.cuda_compile_fullgraph
+                    fullgraph=shared.opts.cuda_compile_fullgraph,
+                    dynamic=None if shared.opts.cuda_compile_backend != "openvino_fx" else False,
                 )
             devices.torch_gc()
             return model

modules/upscaler.py

@@ -187,7 +187,6 @@ def compile_upscaler(model):
 
             if shared.opts.cuda_compile_backend == "openvino_fx":
                 from modules.intel.openvino import openvino_fx # pylint: disable=unused-import
-                torch._dynamo.eval_frame.check_if_dynamo_supported = lambda: True # pylint: disable=protected-access
 
             log_level = logging.WARNING if shared.opts.cuda_compile_verbose else logging.CRITICAL # pylint: disable=protected-access
             if hasattr(torch, '_logging'):
@@ -206,7 +205,12 @@ def compile_upscaler(model):
                 shared.log.error(f"Torch inductor config error: {e}")
 
             t0 = time.time()
-            model = torch.compile(model, mode=shared.opts.cuda_compile_mode, backend=shared.opts.cuda_compile_backend, fullgraph=shared.opts.cuda_compile_fullgraph) # pylint: disable=attribute-defined-outside-init
+            model = torch.compile(model,
+                mode=shared.opts.cuda_compile_mode,
+                backend=shared.opts.cuda_compile_backend,
+                fullgraph=shared.opts.cuda_compile_fullgraph,
+                dynamic=None if shared.opts.cuda_compile_backend != "openvino_fx" else False,
+            ) # pylint: disable=attribute-defined-outside-init
             setup_logging() # compile messes with logging so reset is needed
             t1 = time.time()
             shared.log.info(f"Upscaler compile: time={t1-t0:.2f}")