Add conversion script

ModernBERT-base-4bit.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9a9d1c57cce5d2cdb960bbeb3cfd97443a0752e0542ab958eb636c6a63a87774
-size 279063
+oid sha256:b0e5498f10d54a0351e27f59884ece8901418a639c1545f130acfd7ed9de1a6e
+size 278896

ModernBERT-base-4bit.mlpackage/Manifest.json

@@ -1,18 +1,18 @@
 {
     "fileFormatVersion": "1.0.0",
     "itemInfoEntries": {
-        "E15EBF2F-BF39-4526-8AB8-51E73AFEC7FD": {
+        "5679C35F-3674-42B5-A1EB-CF36A3309400": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Specification",
             "name": "model.mlmodel",
             "path": "com.apple.CoreML/model.mlmodel"
         },
-        "FD5E8367-55E6-413C-8437-713B00D64144": {
+        "F51070FD-D894-4A2E-BEBB-774355701FB7": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Weights",
             "name": "weights",
             "path": "com.apple.CoreML/weights"
         }
     },
-    "rootModelIdentifier": "E15EBF2F-BF39-4526-8AB8-51E73AFEC7FD"
+    "rootModelIdentifier": "5679C35F-3674-42B5-A1EB-CF36A3309400"
 }

ModernBERT-base.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:67cc80bf6060a2d08059fe98c081fc9ebc34b074252143ca4dfa03d3e4f277df
-size 267080
+oid sha256:117a44931a590725359233dcf82eb1df1a2ce609c39da433cc0c12c613f33ce0
+size 266913

ModernBERT-base.mlpackage/Manifest.json

@@ -1,18 +1,18 @@
 {
     "fileFormatVersion": "1.0.0",
     "itemInfoEntries": {
-        "A373F851-2CE5-43A4-8511-ECFD46CE8A03": {
+        "273F953B-8A06-40EB-96D9-D2F027393D23": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Specification",
             "name": "model.mlmodel",
             "path": "com.apple.CoreML/model.mlmodel"
         },
-        "C8A72ADC-A8C6-48A6-8B70-5F91F4815174": {
+        "D92B9527-3316-412B-B7A2-8B412B463912": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Weights",
             "name": "weights",
             "path": "com.apple.CoreML/weights"
         }
     },
-    "rootModelIdentifier": "A373F851-2CE5-43A4-8511-ECFD46CE8A03"
+    "rootModelIdentifier": "273F953B-8A06-40EB-96D9-D2F027393D23"
 }

ModernBERT-large-4bit.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dbe210fd61ab3ced4e7041c483472e46a3383d4e44849a6d5cdba2c4eea3614b
-size 350974
+oid sha256:b4907ed377d3c6808a53c2287cad2298b15d25d0c66de2ca49ae81cd3e7e562c
+size 350807

ModernBERT-large-4bit.mlpackage/Manifest.json

@@ -1,18 +1,18 @@
 {
     "fileFormatVersion": "1.0.0",
     "itemInfoEntries": {
-        "5C57D828-673E-436E-ABD8-F8FAF5D3A45E": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        },
-        "BDC69D2B-C6B6-4E19-A9C2-B2333CA42723": {
+        "2A4DCB67-0E33-4D7D-93B5-415E9AF6EEAB": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Specification",
             "name": "model.mlmodel",
             "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "EA3A02AD-4E6B-4B0D-B3A2-0F39B5630C04": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
         }
     },
-    "rootModelIdentifier": "BDC69D2B-C6B6-4E19-A9C2-B2333CA42723"
+    "rootModelIdentifier": "2A4DCB67-0E33-4D7D-93B5-415E9AF6EEAB"
 }

ModernBERT-large.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:420ef83f761a8cc1c6baa280486ea89dd2892bce55a32d471af9a28a27b28c38
-size 335838
+oid sha256:da340e916321bea2e8bf96121154b0f4bd17cf87fe3e330dc891f5aa702cfab7
+size 335671

ModernBERT-large.mlpackage/Manifest.json

@@ -1,18 +1,18 @@
 {
     "fileFormatVersion": "1.0.0",
     "itemInfoEntries": {
-        "20D16BF3-9CD7-4631-8373-8CA41044B1AE": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        },
-        "652E02D8-970C-42B2-806B-A5B2911B653D": {
+        "033ACE82-9893-4FCC-8758-C4A2366AB463": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Specification",
             "name": "model.mlmodel",
             "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "4B48085F-08D1-45D5-B120-69ED6033119C": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
         }
     },
-    "rootModelIdentifier": "652E02D8-970C-42B2-806B-A5B2911B653D"
+    "rootModelIdentifier": "033ACE82-9893-4FCC-8758-C4A2366AB463"
 }

README.md

@@ -47,3 +47,20 @@ print("Result:")
 print(text.replacingOccurrences(of: "[MASK]", with: predictedTokenText.trimmingCharacters(in: .whitespaces)))
 // The capital of Ireland is Dublin.
 ```
+
+### Conversion
+
+```
+uv run https://hf.co/finnvoorhees/ModernBERT-CoreML/raw/main/convert.py
+```
+
+```
+usage: convert.py [-h] [--model MODEL] [--quantize]
+
+Convert ModernBERT to CoreML
+
+options:
+  -h, --help     show this help message and exit
+  --model MODEL  Model name
+  --quantize     Linear quantize model
+```

convert.py

@@ -0,0 +1,83 @@
+# /// script
+# requires-python = ">=3.11"
+# dependencies = [
+#     "coremltools",
+#     "torch",
+#     "transformers",
+# ]
+#
+# [tool.uv.sources]
+# transformers = { git = "https://github.com/huggingface/transformers.git", branch = "main" }
+# ///
+
+from transformers import AutoModelForMaskedLM
+import torch
+import coremltools as ct
+import numpy as np
+import argparse
+
+
+def log(text):
+    print(f"\033[92m\033[1m{text}\033[0m")
+
+
+parser = argparse.ArgumentParser(
+    prog="convert.py", description="Convert ModernBERT to CoreML"
+)
+parser.add_argument("--model", type=str, default="ModernBERT-base", help="Model name")
+parser.add_argument("--quantize", action="store_true", help="Linear quantize model")
+args = parser.parse_args()
+
+
+class Model(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.model = AutoModelForMaskedLM.from_pretrained(f"answerdotai/{args.model}")
+
+    def forward(self, input_ids):
+        attention_mask = torch.ones_like(input_ids)
+        return self.model(input_ids=input_ids, attention_mask=attention_mask).logits
+
+
+log("Loading model…")
+model = Model().eval()
+
+log("Tracing model…")
+example_input = (torch.zeros((1, 1), dtype=torch.int32),)
+traced_model = torch.jit.trace(model, example_input)
+
+log("Converting model…")
+input_shape = (
+    1,
+    ct.RangeDim(
+        lower_bound=1, upper_bound=model.model.config.max_position_embeddings, default=1
+    ),
+)
+mlmodel = ct.convert(
+    traced_model,
+    inputs=[ct.TensorType(name="input_ids", shape=input_shape, dtype=np.int32)],
+    outputs=[ct.TensorType(name="logits")],
+    minimum_deployment_target=ct.target.macOS15,
+)
+
+if args.quantize:
+    log("Quantizing model…")
+    op_config = ct.optimize.coreml.OpLinearQuantizerConfig(
+        mode="linear_symmetric",
+        dtype="int4",
+        granularity="per_block",
+        block_size=32,
+    )
+    config = ct.optimize.coreml.OptimizationConfig(global_config=op_config)
+    mlmodel = ct.optimize.coreml.linear_quantize_weights(mlmodel, config=config)
+
+mlmodel.author = "Finn Voorhees"
+mlmodel.short_description = "https://hf.co/finnvoorhees/ModernBERT-CoreML"
+
+log("Saving mlmodel…")
+if args.quantize:
+    mlmodel.save(f"{args.model}-4bit.mlpackage")
+else:
+    mlmodel.save(f"{args.model}.mlpackage")
+
+log("Done!")