deepspeed doesn't work with flash-attn, and the gpu savings w flash attn are better than the deepspeed headaches

ds_config.json

@@ -1,6 +1,6 @@
 {
   "bf16": {
-    "enabled": "auto",
+    "enabled": "auto"
   },
   "fp16": {
     "enabled": "auto",
@@ -10,15 +10,6 @@
     "hysteresis": 2,
     "min_loss_scale": 1
   },
-  "optimizer": {
-    "type": "AdamW",
-    "params": {
-      "lr": "auto",
-      "betas": "auto",
-      "eps": "auto",
-      "weight_decay": "auto"
-    }
-  },
   "scheduler": {
     "type": "WarmupLR",
     "params": {
@@ -28,29 +19,19 @@
     }
   },
   "zero_optimization": {
-    "stage": 3,
-    "offload_optimizer": {
-      "device": "cpu",
-      "pin_memory": true
-    },
-    "offload_param": {
-      "device": "cpu",
-      "pin_memory": true
-    },
+    "stage": 2,
     "overlap_comm": true,
+    "allgather_partitions": true,
+    "allgather_bucket_size": 5e8,
     "contiguous_gradients": true,
-    "sub_group_size": 1e9,
     "reduce_bucket_size": "auto",
-    "stage3_prefetch_bucket_size": "auto",
-    "stage3_param_persistence_threshold": "auto",
-    "stage3_max_live_parameters": 1e9,
-    "stage3_max_reuse_distance": 1e9,
-    "stage3_gather_16bit_weights_on_model_save": true
+    "reduce_scatter": true
   },
   "gradient_accumulation_steps": "auto",
   "gradient_clipping": "auto",
   "steps_per_print": 5,
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",
-  "wall_clock_breakdown": false
+  "wall_clock_breakdown": false,
+  "round_robin_gradients": true
 }