support qwen3 name_mapping (#3180)

This commit is contained in:
gaoziyuan
2025-08-04 16:37:34 +08:00
committed by GitHub
parent eca8fc7ca6
commit a592d17615

View File

@@ -429,4 +429,24 @@ class Qwen3ForCausalLMRL(Qwen3ForCausalLM, BaseRLModel):
return "Qwen3ForCausalLMRL" return "Qwen3ForCausalLMRL"
def get_name_mappings_to_training(self, trainer_degree=None) -> Dict[str, str]: def get_name_mappings_to_training(self, trainer_degree=None) -> Dict[str, str]:
pass # Prepare placeholders
place_holders = ["weight"]
# Initialize mapping dictionary
self._update_base_mappings("model")
base_name = "model.layers"
# Helper function to add layer mappings
def _add_layer_mappings(layer_idx):
# FFN mappings
for ph in place_holders:
self.infer_to_train_mapping[f"{base_name}.{layer_idx}.mlp.up_gate_proj.{ph}"] = (
f"{base_name}.{layer_idx}.mlp.gate_up_fused_proj.{ph}"
)
for layer_idx in range(self.fd_config.model_config.num_hidden_layers):
_add_layer_mappings(layer_idx)
self._complete_missing_mappings()
return self.infer_to_train_mapping