fix missing parameter in config for qwen (#365)

* fix config for qwen

* fixes

Author: xadupre

onnx_diagnostic/ci_models/export_qwen25_vl.py

@@ -280,6 +280,10 @@ def main(
         ).eval()
         data = dict(model=model)
         config = model.config
+        if not hasattr(config, "bos_token_id") or not config.bos_token_id:
+            config.bos_token_id = 151643
+        if not hasattr(config, "eos_token_id") or not config.eos_token_id:
+            config.eos_token_id = 151645
     else:
         print("-- random model")
         data = get_untrained_model(model_id, second_input=second_input, verbose=1)

onnx_diagnostic/helpers/log_helper.py

@@ -1921,9 +1921,7 @@ def first_err(df: pandas.DataFrame) -> pandas.Series:
             return lambdas[formula]
 
         if formula == "onnx_n_nodes_no_cst":
-            return lambda df: gdf(df, "onnx_n_nodes", 0) - gdf(
-                df, "op_onnx__Constant", 0
-            ).fillna(0)
+            return lambda df: gdf(df, "onnx_n_nodes", 0) - gdf(df, "op_onnx__Constant", 0)
         if formula == "peak_gpu_torch":
             return lambda df: gdf(df, "mema_gpu_5_after_export") - gdf(df, "mema_gpu_4_reset")
         if formula == "peak_gpu_nvidia":

onnx_diagnostic/torch_export_patches/patches/_patch_transformers_qwen2_5.py

@@ -256,10 +256,21 @@ def qwen_sdpa_attention(
         return attn_output
 
     def qwen_version_selector(opset: int, *args: torch.Tensor) -> Tuple[str, torch.dtype]:
+        import onnx_ir
+
         first_float_tensor = next(
             a
             for a in args
-            if a is not None and a.dtype in {torch.float16, torch.float32, torch.bfloat16}
+            if a is not None
+            and a.dtype
+            in {
+                torch.float16,
+                torch.float32,
+                torch.bfloat16,
+                onnx_ir.DataType.BFLOAT16,
+                onnx_ir.DataType.FLOAT16,
+                onnx_ir.DataType.FLOAT,
+            }
         )
         dtype = first_float_tensor.dtype
         strategy = patched_Qwen2_5_VLVisionAttention.STRATEGY_FOR_ATTENTION()