[Auto-Paralllel] fix intermediate api use

paddlenlp/transformers/model_utils.py

@@ -32,6 +32,7 @@
 import aistudio_sdk
 import numpy as np
 import paddle
+import paddle.distributed as dist
 import paddle.nn as nn
 import six
 from huggingface_hub import (
@@ -2988,14 +2989,40 @@ def _generate_auto_dist_config(self, auto_dist_degree):
             "pp_config": None,
             "cp_config": None,
         }
+        has_auto_dist_config = False
         for name, layer in self.named_sublayers(include_self=True):
             if hasattr(layer, "auto_dist_config"):
+                has_auto_dist_config = True
                 if name != "":
                     prefix = name + "."
                 else:
                     prefix = ""
                 layer_config = layer.auto_dist_config(prefix)
                 merged_config = self.merge_auto_dist_configs([merged_config, layer_config])
+        if not has_auto_dist_config:
+            model_file = inspect.getfile(self.__class__)
+            model_dir = os.path.dirname(model_file)
+            config_path = os.path.join(model_dir, "intermediate_api_config.json")
+            assert os.path.exists(config_path), (
+                f"intermediate api config file not found at {config_path}. "
+                "Please ensure the file exists or implement auto_dist_config in layers."
+            )
+            with open(config_path, "r") as f:
+                iconfig = json.load(f)
+
+            def process_config_value(value):
+                if isinstance(value, dict):
+                    return {k: process_config_value(v) for k, v in value.items()}
+                elif isinstance(value, list):
+                    return [getattr(dist, v)() for v in value if hasattr(dist, v)]
+                elif isinstance(value, str) and hasattr(dist, value):
+                    return getattr(dist, value)()
+                return value
+
+            iconfig = {k: process_config_value(v) for k, v in iconfig.items()}
+            layer_config = {f"{k}": v for k, v in iconfig.items()}
+            merged_config = self.merge_auto_dist_configs([merged_config, layer_config])
+
         final_config = {
             "dp_config": None,
             "mp_config": None,

paddlenlp/transformers/qwen/intermediate_api_config.json

@@ -0,0 +1,36 @@
+{
+    "sp_config": {
+        "parallelize_plan": {
+            "qwen.wte": [
+                "RowWiseParallel",
+                "SequenceParallelBegin"
+            ],
+            "qwen.h.*.attn.c_attn": "ColWiseParallel",
+            "qwen.h.*.attn.c_proj": "RowWiseParallel",
+            "qwen.h.*.attn": "SequenceParallelDisable",
+            "qwen.h.*.mlp.gate_up_fused_proj": "ColWiseParallel",
+            "qwen.h.*.mlp.w1": "ColWiseParallel",
+            "qwen.h.*.mlp.w2": "ColWiseParallel",
+            "qwen.h.*.mlp.c_proj": "RowWiseParallel",
+            "qwen.h.*.mlp": "SequenceParallelDisable",
+            "lm_head.weight": "ColWiseParallel",
+            "lm_head": "SequenceParallelEnd"
+        }
+    },
+    "mp_config": {
+        "parallelize_plan": {
+            "qwen.wte": "RowWiseParallel",
+            "qwen.h.*.attn.c_attn": "ColWiseParallel",
+            "qwen.h.*.attn.c_proj": "RowWiseParallel",
+            "qwen.h.*.mlp.gate_up_fused_proj": "ColWiseParallel",
+            "qwen.h.*.mlp.w1": "ColWiseParallel",
+            "qwen.h.*.mlp.w2": "ColWiseParallel",
+            "qwen.h.*.mlp.c_proj": "RowWiseParallel",
+            "lm_head.weight": "ColWiseParallel"
+        }
+    },
+    "pp_config": {
+        "split_spec": "qwen.h",
+        "global_spec": "qwen.global_layer"
+    }
+}

paddlenlp/transformers/qwen/modeling_network.py

@@ -17,7 +17,6 @@
 import warnings
 
 import paddle
-import paddle.distributed as dist
 import paddle.nn.functional as F
 from paddle import nn
 from paddle.distributed.fleet.utils import recompute
@@ -660,48 +659,6 @@ def forward(
 
         return lm_logits
 
-    def auto_dist_config(self, prefix=""):
-        if prefix != "":
-            assert prefix.endswith(".")
-        config = {
-            "sp_config": {
-                "parallelize_plan": {
-                    f"{prefix}qwen.wte": [
-                        dist.RowWiseParallel(),
-                        dist.SequenceParallelBegin(),
-                    ],
-                    f"{prefix}qwen.h.*.attn.c_attn": dist.ColWiseParallel(),
-                    f"{prefix}qwen.h.*.attn.c_proj": dist.RowWiseParallel(),
-                    f"{prefix}qwen.h.*.attn": dist.SequenceParallelDisable(),
-                    f"{prefix}qwen.h.*.mlp.gate_up_fused_proj": dist.ColWiseParallel(),
-                    f"{prefix}qwen.h.*.mlp.w1": dist.ColWiseParallel(),
-                    f"{prefix}qwen.h.*.mlp.w2": dist.ColWiseParallel(),
-                    f"{prefix}qwen.h.*.mlp.c_proj": dist.RowWiseParallel(),
-                    f"{prefix}qwen.h.*.mlp": dist.SequenceParallelDisable(need_transpose=False),
-                    f"{prefix}lm_head.weight": dist.ColWiseParallel(),
-                    f"{prefix}lm_head": dist.SequenceParallelEnd(),
-                }
-            },
-            "mp_config": {
-                "parallelize_plan": {
-                    f"{prefix}qwen.wte": dist.RowWiseParallel(),
-                    f"{prefix}qwen.h.*.attn.c_attn": dist.ColWiseParallel(),
-                    f"{prefix}qwen.h.*.attn.c_proj": dist.RowWiseParallel(),
-                    f"{prefix}qwen.h.*.mlp.gate_up_fused_proj": dist.ColWiseParallel(),
-                    f"{prefix}qwen.h.*.mlp.w1": dist.ColWiseParallel(),
-                    f"{prefix}qwen.h.*.mlp.w2": dist.ColWiseParallel(),
-                    f"{prefix}qwen.h.*.mlp.c_proj": dist.RowWiseParallel(),
-                    f"{prefix}lm_head.weight": dist.ColWiseParallel(),
-                }
-            },
-            "pp_config": {
-                "split_spec": f"{prefix}qwen.h",
-                "global_spec": f"{prefix}qwen.global_layer",
-            },
-        }
-
-        return config
-
 
 class GlobalNet(nn.Layer):
     def __init__(self, config) -> None: