diff --git a/patch.py b/patch.py
new file mode 100644
index 0000000..e8109ca
--- /dev/null
+++ b/patch.py
@@ -0,0 +1,19 @@
+path = '/usr/local/lib/python3.12/dist-packages/transformers/tokenization_utils_base.py' 
+with open(path, 'r') as f:
+    content = f.read()
+
+old = ('        self.SPECIAL_TOKENS_ATTRIBUTES = self.SPECIAL_TOKENS_ATTRIBUTES'
+       ' + list(special_tokens.keys())')
+new = ('        # PATCH: some models have extra_special_tokens as list instead of dict\n'      
+       '        if isinstance(special_tokens, list):\n'
+       '            special_tokens = {t: t for t in special_tokens}\n'
+       '        self.SPECIAL_TOKENS_ATTRIBUTES = self.SPECIAL_TOKENS_ATTRIBUTES'
+       ' + list(special_tokens.keys())')
+
+if old in content:
+    content = content.replace(old, new)
+    with open(path, 'w') as f:
+        f.write(content)
+    print('Patch applied successfully')
+else:
+    print('WARNING: pattern not found')
\ No newline at end of file
diff --git a/patch_triton.py b/patch_triton.py
new file mode 100644
index 0000000..8fb3579
--- /dev/null
+++ b/patch_triton.py
@@ -0,0 +1,24 @@
+path = '/usr/local/lib/python3.12/dist-packages/vllm/v1/attention/backends/triton_attn.py'
+with open(path, 'r') as f:
+    content = f.read()
+
+old = '''    def validate_head_size(cls, head_size: int) -> None:
+        # Triton Attention supports any head size above 32
+        if head_size < 32:
+            raise ValueError(
+                f"Head size {head_size} is not supported by TritonAttention."
+                f"Head sizes need to be larger or equal 32 for this backend. "
+                "Set VLLM_ATTENTION_BACKEND=FLEX_ATTENTION to use "
+                "FlexAttention backend which supports all head sizes.")'''
+
+new = '''    def validate_head_size(cls, head_size: int) -> None:
+        # PATCH: allow all head sizes (Triton compiles at runtime)
+        return'''
+
+if old in content:
+    content = content.replace(old, new)
+    with open(path, 'w') as f:
+        f.write(content)
+    print('patch_triton: validate_head_size bypassed successfully')
+else:
+    print('patch_triton: WARNING - pattern not found, patch skipped')