models (ViT backbone) iterate extract_patches over batch dim…

`Patches.call`: use `tf.map_fn` instead of running
entire batch through `tf.image.extract_patches`
(faster, less VRAM, allows ONNX conversion to work)

src/eynollah/model_zoo/model_zoo.py

@@ -19,7 +19,7 @@ MODEL_VRAM_LIMITS = {
     "enhancement": 980, # due to bs 3
     "col_classifier": 210,
     "page": 618,
-    "textline": 1680, # 954 for bs 1
+    "textline": 1880, # 954 for bs 1
     "region_1_2": 1580,
     "region_fl_np": 1756,
     "table": 1818,
@@ -338,6 +338,14 @@ class EynollahModelZoo:
                     # 'cudnn_conv_algo_search': 'EXHAUSTIVE',
                     #'cudnn_conv_use_max_workspace': 0,
                     # 'do_copy_in_default_stream': True,
+                    # enable_cuda_graph
+                    # cudnn_conv1d_pad_to_nc1d
+                    # prefer_nhwc
+                    # tunable_op_enable
+                    # tunable_op_tuning_enable
+                    # tunable_op_max_tuning_duration_ms
+                    # use_ep_level_unified_stream
+                    # enable_skip_layer_norm_strict_mode
                     # ...
                 })] + providers
         if 'TensorrtExecutionProvider' in providers:
@@ -347,9 +355,28 @@ class EynollahModelZoo:
                     'device_id': gpu,
                     'trt_max_workspace_size': MODEL_VRAM_LIMITS[model_category] * 1024 * 1024,
                     # 'trt_fp16_enable': True,
-                    # 'trt_engine_cache_enable': True,
-                    # 'trt_timing_cache_enable': True,
+                    # trt_bf16_enable
+                    'trt_engine_cache_enable': True,
+                    'trt_timing_cache_enable': True,
                     # ...
+                    # trt_engine_hw_compatible
+                    # trt_engine_cache_path
+                    # trt_engine_cache_prefix
+                    # trt_timing_cache_path
+                    # trt_onnx_model_folder_path
+                    # trt_ep_context_file_path
+                    # trt_cuda_graph_enable
+                    # trt_profile_opt_shapes
+                    # trt_profile_min_shapes
+                    # trt_profile_max_shapes
+                    # trt_builder_optimization_level
+                    # trt_build_heuristics_enable
+                    # trt_sparsity_enable
+                    # trt_weight_stripped_engine_enable
+                    # trt_dla_core
+                    # trt_dla_enable
+                    # trt_min_subgraph_size
+                    # trt_ep_context_embed_mode
                 })] + providers
         provider0 = providers[0]
         if isinstance(provider0, tuple):

src/eynollah/patch_encoder.py

@@ -29,7 +29,13 @@ class Patches(layers.Layer):
         self.patch_size_y = patch_size_y
 
     def call(self, images):
-        batch_size = tf.shape(images)[0]
+        #batch_size = tf.shape(images)[0]
+        return tf.map_fn(self.call_single, images)
+
+    def call_single(self, image):
+        # avoid batched extract_patches: too much memory,
+        # and variable batch dim not supported by ONNX implementation
+        images = tf.expand_dims(image, axis=0)
         patches = tf.image.extract_patches(
             images=images,
             sizes=[1, self.patch_size_y, self.patch_size_x, 1],
@@ -37,8 +43,9 @@ class Patches(layers.Layer):
             rates=[1, 1, 1, 1],
             padding="VALID",
         )
-        patch_dims = patches.shape[-1]
-        return tf.reshape(patches, [batch_size, -1, patch_dims])
+        _, n_rows, n_cols, patch_dims = patches.shape
+        n_tiles = patches.shape[1] * patches.shape[2] #-1
+        return tf.reshape(patches, [1, n_tiles, patch_dims])
 
     def get_config(self):
         return dict(patch_size_x=self.patch_size_x,