[v1] add init on rank0 for fsdp2 (#10264)

examples/v1/train_freeze/train_freeze_sft.yaml

@@ -1,5 +1,4 @@
 model: Qwen/Qwen3-4B
-trust_remote_code: true
 model_class: llm
 
 template: qwen3_nothink

examples/v1/train_full/train_full_deepspeed.yaml

@@ -1,5 +1,4 @@
 model: Qwen/Qwen3-0.6B
-
 model_class: llm
 
 template: qwen3_nothink

examples/v1/train_full/train_full_fsdp2.yaml

@@ -1,5 +1,4 @@
 model: Qwen/Qwen3-0.6B
-trust_remote_code: true
 model_class: llm
 
 template: qwen3_nothink

examples/v1/train_lora/train_lora_sft.yaml

@@ -1,5 +1,4 @@
 model: Qwen/Qwen3-4B
-trust_remote_code: true
 model_class: llm
 
 template: qwen3_nothink
@@ -28,7 +27,6 @@ train_dataset: data/v1_sft_demo.yaml
 ### training
 output_dir: ./outputs/test_lora
 micro_batch_size: 1
-global_batch_size: 4
 cutoff_len: 2048
 learning_rate: 1.0e-4
 bf16: true

examples/v1/train_lora/train_lora_sft_rank0.yaml

@@ -0,0 +1,40 @@
+model: Qwen/Qwen3-4B
+model_class: llm
+
+template: qwen3_nothink
+
+# PEFT Configuration
+peft_config:
+  name: lora
+  r: 16
+  lora_alpha: 32
+  lora_dropout: 0.05
+  target_modules: all
+
+# Kernel Config
+kernel_config:
+  name: auto
+  include_kernels: auto
+
+# FSDP Config
+dist_config:
+  name: fsdp2
+  dcp_path: null
+
+init_config:
+  name: init_on_rank0
+
+### data
+train_dataset: data/v1_sft_demo.yaml
+
+### training
+output_dir: ./outputs/test_lora
+micro_batch_size: 1
+cutoff_len: 2048
+learning_rate: 1.0e-4
+bf16: true
+max_steps: 10
+
+### sample
+sample_backend: hf
+max_new_tokens: 128

examples/v1/train_qlora/quantization.yaml

@@ -1,5 +1,4 @@
 model: Qwen/Qwen3-0.6B
-trust_remote_code: true
 model_class: llm
 
 template: qwen3_nothink