[v1] add deepspeed zero3 trigger for low memory usage weight loading (#10300)

2026-04-22 04:46:00 +08:00 · 2026-04-21 14:09:52 +08:00
parent f5d739b132
commit 28a6ea1cdc
13 changed files with 160 additions and 18 deletions
--- a/examples/v1/train_freeze/train_freeze_sft.yaml
+++ b/examples/v1/train_freeze/train_freeze_sft.yaml
@@ -29,7 +29,6 @@ micro_batch_size: 1
 global_batch_size: 4
 cutoff_len: 2048
 learning_rate: 2.0e-5
-bf16: false
 max_steps: 10

 ### sample
--- a/examples/v1/train_full/train_full_deepspeed.yaml
+++ b/examples/v1/train_full/train_full_deepspeed.yaml
@@ -19,5 +19,4 @@ output_dir: outputs/Qwen3-0.6B-deepspeed
 micro_batch_size: 1
 cutoff_len: 2048
 learning_rate: 1.0e-4
-bf16: true
 max_steps: 10
--- a/examples/v1/train_full/train_full_fsdp2.yaml
+++ b/examples/v1/train_full/train_full_fsdp2.yaml
@@ -21,7 +21,6 @@ output_dir: outputs/test_fsdp2
 micro_batch_size: 1
 cutoff_len: 2048
 learning_rate: 1.0e-4
-bf16: false
 max_steps: 10

 ### sample
--- a/examples/v1/train_lora/train_lora_sft.yaml
+++ b/examples/v1/train_lora/train_lora_sft.yaml
@@ -29,7 +29,6 @@ output_dir: ./outputs/test_lora
 micro_batch_size: 1
 cutoff_len: 2048
 learning_rate: 1.0e-4
-bf16: true
 max_steps: 10

 ### sample
--- a/examples/v1/train_qlora/quantization.yaml
+++ b/examples/v1/train_qlora/quantization.yaml
@@ -34,7 +34,6 @@ output_dir: outputs/test_quantization
 micro_batch_size: 1
 cutoff_len: 2048
 learning_rate: 1.0e-4
-bf16: false
 max_steps: 10

 ### sample