Temporarily patch transformers.js convert script to fix the model conversion

app.py

@@ -86,6 +86,8 @@ class ModelConverter:
         try:
             urlretrieve(archive_url, archive_path)
             self._extract_archive(archive_path)
+            self._patch_convert_script()
+            self._install_scripts_requirements()
             logger.info("Repository downloaded and extracted successfully")
         except Exception as e:
             raise RuntimeError(f"Failed to setup repository: {e}")
@@ -101,6 +103,60 @@ class ModelConverter:
             extracted_folder = next(Path(tmp_dir).iterdir())
             extracted_folder.rename(self.config.repo_path)
 
+    def _install_scripts_requirements(self) -> None:
+        req_path = self.config.repo_path / "scripts" / "requirements.txt"
+        if req_path.exists():
+            subprocess.run(
+                [
+                    sys.executable,
+                    "-m",
+                    "pip",
+                    "install",
+                    "--no-cache-dir",
+                    "-r",
+                    str(req_path),
+                ],
+                check=True,
+            )
+
+    def _patch_convert_script(self) -> None:
+        """Patch transformers.js convert script to force eager attention for better ONNX compatibility."""
+        path = self.config.repo_path / "scripts" / "convert.py"
+        if not path.exists():
+            return
+        try:
+            text = path.read_text(encoding="utf-8")
+            marker = "export_kwargs = dict("
+            if marker in text:
+                lines = text.splitlines()
+                start_idx = None
+                paren = 0
+                for i, line in enumerate(lines):
+                    if marker in line:
+                        start_idx = i
+                        break
+                if start_idx is not None:
+                    for k in range(start_idx, len(lines)):
+                        paren += lines[k].count("(")
+                        paren -= lines[k].count(")")
+                        if paren <= 0:
+                            insert_at = k + 1
+                            lines.insert(
+                                insert_at,
+                                "    export_kwargs.setdefault('model_kwargs', {})",
+                            )
+                            lines.insert(
+                                insert_at + 1,
+                                "    export_kwargs['model_kwargs']['attn_implementation'] = 'eager'",
+                            )
+                            patched = "\n".join(lines) + (
+                                "\n" if text.endswith("\n") else ""
+                            )
+                            path.write_text(patched, encoding="utf-8")
+                            break
+        except Exception:
+            pass
+
     def _run_conversion_subprocess(
         self, input_model_id: str, extra_args: List[str] = None
     ) -> subprocess.CompletedProcess:
@@ -110,8 +166,6 @@ class ModelConverter:
             "-m",
             "scripts.convert",
             "--quantize",
-            "--opset",
-            "18",
             "--model_id",
             input_model_id,
         ]
@@ -126,6 +180,8 @@ class ModelConverter:
             text=True,
             env={
                 "HF_TOKEN": self.config.hf_token,
+                "TRANSFORMERS_ATTENTION_IMPLEMENTATION": "eager",
+                "PYTORCH_SDP_KERNEL": "math",
             },
         )
 

requirements.txt

@@ -1,7 +1,10 @@
 huggingface_hub==0.35.3
 streamlit==1.50.0
 onnxscript==0.5.4
-transformers[torch]==4.49.0
+onnxconverter_common==1.16.0
+onnx_graphsurgeon==0.5.8
+torch==2.5.1
+transformers==4.49.0
 onnxruntime==1.20.1
 optimum@git+https://github.com/huggingface/optimum.git@b04feaea78cda58d79b8da67dca3fd0c4ab33435
 onnx==1.17.0