upload to hf

machine-learning/export/immich_model_exporter/exporters/constants.py

@@ -21,3 +21,34 @@ SOURCE_TO_METADATA = {
         "InsightFace", "https://github.com/deepinsight/insightface/tree/master", "facial recognition"
     ),
 }
+
+RKNN_SOCS = ["rk3566", "rk3576", "rk3588"]
+
+# the builder hangs when using flash attention with these models
+RKNN_VISUAL_FLASH_ATTENTION_BLACKLIST = {
+    "ViT-H-14-378-quickgelu__dfn5b",
+    "ViT-L-16-SigLIP-384__webli",
+    "ViT-L-16-SigLIP2-384__webli",
+    "ViT-L-16-SigLIP2-512__webli",
+    "ViT-SO400M-14-SigLIP-384__webli",
+    "ViT-SO400M-14-SigLIP2-378__webli",
+    "ViT-SO400M-16-SigLIP2-378__webli",
+    "ViT-SO400M-16-SigLIP2-512__webli",
+}
+
+
+# glob to delete old UUID blobs when reuploading models
+_uuid_char = "[a-fA-F0-9]"
+_uuid_glob = _uuid_char * 8 + "-" + _uuid_char * 4 + "-" + _uuid_char * 4 + "-" + _uuid_char * 4 + "-" + _uuid_char * 12
+DELETE_PATTERNS = [
+    "**/*onnx*",
+    "**/Constant*",
+    "**/*.weight",
+    "**/*.bias",
+    "**/*.proj",
+    "**/*in_proj_bias",
+    "**/*.npy",
+    "**/*.latent",
+    "**/*.pos_embed",
+    f"**/{_uuid_glob}",
+]

machine-learning/export/immich_model_exporter/exporters/onnx/models/mclip.py

@@ -38,7 +38,9 @@ def to_onnx(
         _export_text_encoder(model, textual_path, opset_version)
     else:
         print(f"Model {textual_path} already exists, skipping")
-    visual_path, _ = openclip_to_onnx(_MCLIP_TO_OPENCLIP[model_name], output_dir_visual, no_cache=no_cache)
+    visual_path, _ = openclip_to_onnx(
+        _MCLIP_TO_OPENCLIP[model_name], opset_version, output_dir_visual, no_cache=no_cache
+    )
     assert visual_path is not None, "Visual model export failed"
     return visual_path, textual_path
 

machine-learning/export/immich_model_exporter/exporters/onnx/models/openclip.py

@@ -105,13 +105,14 @@ def _export_image_encoder(
         assert isinstance(output, torch.Tensor)
         return output
 
+    model.forward = encode_image
+
     args = (torch.randn(1, 3, model_cfg.image_size, model_cfg.image_size),)
-    traced = torch.jit.trace(encode_image, args)  # type: ignore[no-untyped-call]
 
     with warnings.catch_warnings():
         warnings.simplefilter("ignore", UserWarning)
         torch.onnx.export(
-            traced,
+            model,
             args,
             output_path.as_posix(),
             input_names=["image"],
@@ -133,13 +134,14 @@ def _export_text_encoder(
         assert isinstance(output, torch.Tensor)
         return output
 
+    model.forward = encode_text
+
     args = (torch.ones(1, model_cfg.sequence_length, dtype=torch.int32),)
-    traced = torch.jit.trace(encode_text, args)  # type: ignore[no-untyped-call]
 
     with warnings.catch_warnings():
         warnings.simplefilter("ignore", UserWarning)
         torch.onnx.export(
-            traced,
+            model,
             args,
             output_path.as_posix(),
             input_names=["text"],

machine-learning/export/immich_model_exporter/exporters/rknn.py

@@ -1,6 +1,6 @@
 from pathlib import Path
 
-RKNN_SOCS = ["rk3566", "rk3576", "rk3588"]
+from .constants import RKNN_SOCS, RKNN_VISUAL_FLASH_ATTENTION_BLACKLIST
 
 
 def _export_platform(
@@ -22,11 +22,12 @@ def _export_platform(
 
     rknn = RKNN(verbose=False)
 
+    # flash_attention = model_dir.name != "visual" or model_dir.parent.name not in RKNN_VISUAL_FLASH_ATTENTION_BLACKLIST
     rknn.config(
         target_platform=target_platform,
         dynamic_input=dynamic_input,
         disable_rules=["fuse_matmul_softmax_matmul_to_sdpa"] if not fuse_matmul_softmax_matmul_to_sdpa else [],
-        enable_flash_attention=True,
+        enable_flash_attention=False,
         model_pruning=True,
     )
     ret = rknn.load_onnx(model=input_path.as_posix())
@@ -49,13 +50,13 @@ def _export_platforms(model_dir: Path, dynamic_input=None, no_cache: bool = Fals
     fuse_matmul_softmax_matmul_to_sdpa = True
     for soc in RKNN_SOCS:
         try:
-            _export_platform(model_dir, soc, dynamic_input, fuse_matmul_softmax_matmul_to_sdpa)
+            _export_platform(model_dir, soc, dynamic_input, fuse_matmul_softmax_matmul_to_sdpa, no_cache=no_cache)
         except Exception as e:
             print(f"Failed to export model for {soc}: {e}")
             if "inputs or 'outputs' must be set" in str(e):
                 print("Retrying without fuse_matmul_softmax_matmul_to_sdpa")
                 fuse_matmul_softmax_matmul_to_sdpa = False
-                _export_platform(model_dir, soc, dynamic_input, fuse_matmul_softmax_matmul_to_sdpa)
+                _export_platform(model_dir, soc, dynamic_input, fuse_matmul_softmax_matmul_to_sdpa, no_cache=no_cache)
 
 
 def export(model_dir: Path, no_cache: bool = False):