feat(model): add support for Qwen model in RL PipelineParallel (#11135)

Jason233333 · web-flow · commit 5dd2855a526c · 2025-10-11T10:35:33.000+08:00
diff --git a/paddlenlp/transformers/qwen2/modeling_pp.py b/paddlenlp/transformers/qwen2/modeling_pp.py
@@ -232,6 +232,7 @@ class Qwen2ForCausalLMPipe(PipelinePretrainedModel, PipelineLayer):
     config_class = Qwen2Config
 
     _get_tensor_parallel_mappings = Qwen2PretrainedModel._get_tensor_parallel_mappings
+    _get_fuse_or_split_param_mappings = Qwen2PretrainedModel._get_fuse_or_split_param_mappings
     _init_weights = Qwen2PretrainedModel._init_weights
     _keys_to_ignore_on_load_unexpected = Qwen2PretrainedModel._keys_to_ignore_on_load_unexpected
     _get_model_flops = Qwen2PretrainedModel._get_model_flops
diff --git a/paddlenlp/transformers/qwen3/modeling_pp.py b/paddlenlp/transformers/qwen3/modeling_pp.py
@@ -232,6 +232,7 @@ class Qwen3ForCausalLMPipe(PipelinePretrainedModel, PipelineLayer):
     config_class = Qwen3Config
 
     _get_tensor_parallel_mappings = Qwen3PretrainedModel._get_tensor_parallel_mappings
+    _get_fuse_or_split_param_mappings = Qwen3PretrainedModel._get_fuse_or_split_param_mappings
     _init_weights = Qwen3PretrainedModel._init_weights
     _keys_to_ignore_on_load_unexpected = Qwen3PretrainedModel._keys_to_ignore_on_load_unexpected
     _get_model_flops = Qwen3PretrainedModel._get_model_flops
