Qwen3 finetune recipes

src/megatron/bridge/peft/utils.py

@@ -430,6 +430,12 @@ def __init__(
         model_parallel_config.sequence_parallel = False  # SP is irrelevant for the lora linear layer
         self.config = model_parallel_config
 
+        # Ensure adapter parameters are initialized when creating adapter layers.
+        # In some flows (e.g., after import), perform_initialization may be False to skip heavy init.
+        _prev_perform_initialization = getattr(model_parallel_config, "perform_initialization", True)
+        if hasattr(model_parallel_config, "perform_initialization"):
+            model_parallel_config.perform_initialization = True
+
         if input_is_parallel:
             self.linear_in = RowParallelLinear(
                 in_features,

src/megatron/bridge/recipes/qwen/__init__.py

@@ -29,11 +29,17 @@
 
 # Qwen3 models
 from .qwen3 import (
+    qwen3_1p7b_finetune_config,
     qwen3_1p7b_pretrain_config,
+    qwen3_4b_finetune_config,
     qwen3_4b_pretrain_config,
+    qwen3_8b_finetune_config,
     qwen3_8b_pretrain_config,
+    qwen3_14b_finetune_config,
     qwen3_14b_pretrain_config,
+    qwen3_32b_finetune_config,
     qwen3_32b_pretrain_config,
+    qwen3_600m_finetune_config,
     qwen3_600m_pretrain_config,
 )
 
@@ -64,6 +70,12 @@
     "qwen3_8b_pretrain_config",
     "qwen3_14b_pretrain_config",
     "qwen3_32b_pretrain_config",
+    "qwen3_600m_finetune_config",
+    "qwen3_1p7b_finetune_config",
+    "qwen3_4b_finetune_config",
+    "qwen3_8b_finetune_config",
+    "qwen3_14b_finetune_config",
+    "qwen3_32b_finetune_config",
     # Qwen3 MoE models
     "qwen3_30b_a3b_pretrain_config",
     "qwen3_235b_a22b_pretrain_config",