OpenVINO fix cache loading

CHANGELOG.md

@@ -78,6 +78,7 @@ And it also includes fixes for all reported issues so far
     enable *Compress Model weights with NNCF* from *Compute Settings* and set a 4-bit NNCF mode  
     4-bit and 8-bit with OpenVINO is CPU only for now  
   - reduce system memory usage after compile  
+  - fix cache loading with multiple models  
 - **Fixes**  
   - ipadapter: allow changing of model/image on-the-fly  
   - ipadapter: fix fallback of cross-attention on unload  

modules/intel/openvino/__init__.py

@@ -342,6 +342,11 @@ def partition_graph(gm: GraphModule, use_python_fusion_cache: bool, model_hash_s
 
     return gm
 
+def generate_subgraph_str(tensor):
+    if hasattr(tensor, "weight"):
+        shared.compiled_model_state.model_str = shared.compiled_model_state.model_str + str(tensor.weight)
+    return tensor
+
 @register_backend
 @fake_tensor_unsupported
 def openvino_fx(subgraph, example_inputs):
@@ -351,7 +356,9 @@ def openvino_fx(subgraph, example_inputs):
     if not shared.opts.openvino_disable_model_caching:
         os.environ.setdefault('OPENVINO_TORCH_MODEL_CACHING', "1")
         # Create a hash to be used for caching
-        model_hash_str = sha256(subgraph.code.encode('utf-8')).hexdigest()
+        shared.compiled_model_state.model_str = ""
+        subgraph.apply(generate_subgraph_str)
+        model_hash_str = sha256(shared.compiled_model_state.model_str.encode('utf-8')).hexdigest()
         if (shared.compiled_model_state.cn_model != [] and shared.compiled_model_state.partition_id == 0):
             model_hash_str = model_hash_str + str(shared.compiled_model_state.cn_model)
 

modules/sd_models_compile.py

@@ -8,6 +8,7 @@ from installer import setup_logging
 #Used by OpenVINO, can be used with TensorRT or Olive
 class CompiledModelState:
     def __init__(self):
+        self.model_str = ""
         self.first_pass = True
         self.first_pass_refiner = True
         self.first_pass_vae = True
@@ -76,6 +77,8 @@ def optimize_openvino():
                 shared.compiled_model_state.lora_model = []
             shared.compiled_model_state.compiled_cache.clear()
             shared.compiled_model_state.partitioned_modules.clear()
+            shared.compiled_model_state.partition_id = 0
+            shared.compiled_model_state.model_str = ""
         shared.compiled_model_state.first_pass = True if not shared.opts.cuda_compile_precompile else False
         shared.compiled_model_state.first_pass_vae = True if not shared.opts.cuda_compile_precompile else False
         shared.compiled_model_state.first_pass_refiner = True if not shared.opts.cuda_compile_precompile else False