.

ds_config_zero3.json

@@ -6,12 +6,15 @@
     "stage": 3,
     "overlap_comm": true,
     "contiguous_gradients": true,
-    "reduce_bucket_size": 200000000,
-    "stage3_prefetch_bucket_size": 100000000,
+
+    "reduce_bucket_size": 150000000,
+    "stage3_prefetch_bucket_size": 75000000,
     "stage3_param_persistence_threshold": 1000000,
 
-    "offload_optimizer": { "device": "cpu", "pin_memory": true },
-    "offload_param":     { "device": "none" }
+    "offload_optimizer": { "device": "none" },
+    "offload_param":     { "device": "none" },
+
+    "stage3_gather_16bit_weights_on_model_save": false
   },
 
   "optimizer": {

ds_config_zero3_cpu_offload.json

@@ -0,0 +1,27 @@
+{
+  "train_micro_batch_size_per_gpu": 1,
+  "gradient_accumulation_steps": 1,
+
+  "zero_optimization": {
+    "stage": 3,
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "reduce_bucket_size": 200000000,
+    "stage3_prefetch_bucket_size": 100000000,
+    "stage3_param_persistence_threshold": 1000000,
+
+    "offload_optimizer": { "device": "cpu", "pin_memory": true },
+    "offload_param":     { "device": "none" }
+  },
+
+  "optimizer": {
+    "type": "AdamW",
+    "params": { "lr": 2e-5, "betas": [0.9, 0.999], "eps": 1e-8, "weight_decay": 0.1 }
+  },
+
+  "bf16": { "enabled": true },
+  "fp16": { "enabled": false },
+
+  "gradient_clipping": 1.0,
+  "wall_clock_breakdown": false
+}