Ensure tensor alignment is preserved by adding dummy padding metadata if necessary

_hf_gguf.py

@@ -187,7 +187,9 @@ class HuggingGGUFstream:
         self.header = {}
         self.metadata = {}
         self.endian = '<'
+        self.alignment = 32
         self.metaend = 0
+        self.offset = 0
         self.filesize = fp.details.get('size')
 
         if (data := self.fp.read(4)) != b'GGUF':
@@ -356,7 +358,36 @@ class HuggingGGUFstream:
 
                 yield key.value, val
 
+            if (alignment := self.metadata.get('general.alignment')) is not None:
+                self.alignment = alignment.value
+
             self.metaend = self.fp.loc
+            self.offset = self.metaend % self.alignment
+
+    def adjust_padding(
+        self,
+    ) -> None:
+        if self.header['tensors'].value == 0:
+            return
+
+        dummy_key = 'dummy.padding'
+        cur_metaend = 0
+
+        for data in self.header.values():
+            cur_metaend += len(data.data)
+
+        for k, v in self.metadata.items():
+            if k != dummy_key:
+                cur_metaend += len(v.data)
+
+        if (cur_metaend % self.alignment) != self.offset:
+            dummy_len = 8 + len(dummy_key) + 4 + 4 + 8
+            dummy_offset = (cur_metaend + dummy_len) % self.alignment
+            dummy_padding = (32 + (self.offset - dummy_offset)) % self.alignment
+
+            self.add_metadata(dummy_key, GGUFValueType.UINT8, [0] * dummy_padding)
+        else:
+            self.remove_metadata(dummy_key)
 
     def add_metadata(
         self,

app.py

@@ -1279,6 +1279,7 @@ def stream_repo_file(
                         k[2] = []
 
                 gguf.add_metadata(*k)
+        gguf.adjust_padding()
 
         yield gguf.filesize