micro_batch_size: 4

README.md

@@ -65,6 +65,31 @@ CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=0 PYTORCH_CUDA_ALLOC_CONF=expandable
 ```
 
 ```
+Seed set to 23
+Time to instantiate model: 0.23 seconds.
+Total parameters: 138,084,864
+Verifying settings ...
+Measured TFLOPs: 6972.54
+Epoch 1 | iter 256 step 1 | loss train: 10.530, val: n/a | iter time: 3230.47 ms (step) remaining time: 3 days, 5:34:13
+Epoch 1 | iter 512 step 2 | loss train: 10.520, val: n/a | iter time: 589.19 ms (step) remaining time: 3 days, 0:40:40
+Epoch 1 | iter 768 step 3 | loss train: 10.485, val: n/a | iter time: 591.81 ms (step) remaining time: 2 days, 23:01:54
+Epoch 1 | iter 1024 step 4 | loss train: 10.447, val: n/a | iter time: 589.35 ms (step) remaining time: 2 days, 22:11:32
+Epoch 1 | iter 1280 step 5 | loss train: 10.350, val: n/a | iter time: 589.38 ms (step) remaining time: 2 days, 21:40:13
+Epoch 1 | iter 1536 step 6 | loss train: 10.241, val: n/a | iter time: 593.75 ms (step) remaining time: 2 days, 21:18:19
+Epoch 1 | iter 1792 step 7 | loss train: 10.134, val: n/a | iter time: 592.92 ms (step) remaining time: 2 days, 21:01:58
+Epoch 1 | iter 2048 step 8 | loss train: 10.049, val: n/a | iter time: 590.74 ms (step) remaining time: 2 days, 20:49:12
+Epoch 1 | iter 2304 step 9 | loss train: 9.869, val: n/a | iter time: 594.27 ms (step) remaining time: 2 days, 20:39:10
+Epoch 1 | iter 2560 step 10 | loss train: 9.771, val: n/a | iter time: 590.04 ms (step) remaining time: 2 days, 20:30:14
+Epoch 1 | iter 2816 step 11 | loss train: 9.643, val: n/a | iter time: 588.32 ms (step) remaining time: 2 days, 20:22:22
+Epoch 1 | iter 3072 step 12 | loss train: 9.557, val: n/a | iter time: 588.95 ms (step) remaining time: 2 days, 20:15:26
+Epoch 1 | iter 3328 step 13 | loss train: 9.487, val: n/a | iter time: 589.32 ms (step) remaining time: 2 days, 20:09:05
+Epoch 1 | iter 3584 step 14 | loss train: 9.413, val: n/a | iter time: 588.95 ms (step) remaining time: 2 days, 20:03:24
+Epoch 1 | iter 3840 step 15 | loss train: 9.322, val: n/a | iter time: 591.62 ms (step) remaining time: 2 days, 19:58:18
+Epoch 1 | iter 4096 step 16 | loss train: 9.241, val: n/a | iter time: 593.65 ms (step) remaining time: 2 days, 19:53:30
+Epoch 1 | iter 4352 step 17 | loss train: 9.163, val: n/a | iter time: 593.89 ms (step) remaining time: 2 days, 19:49:00
+Epoch 1 | iter 4608 step 18 | loss train: 9.122, val: n/a | iter time: 590.63 ms (step) remaining time: 2 days, 19:44:42
+Epoch 1 | iter 4864 step 19 | loss train: 9.077, val: n/a | iter time: 590.87 ms (step) remaining time: 2 days, 19:40:47
+Epoch 1 | iter 5120 step 20 | loss train: 9.018, val: n/a | iter time: 588.44 ms (step) remaining time: 2 days, 19:36:59
 # ...
 ```
 

scripts/pretrain-core-model.yaml

@@ -67,7 +67,8 @@ train:
   # global_batch_size: 256
 
   # Number of samples per data-parallel rank (type: int, default: 4)
-  micro_batch_size: 2
+  micro_batch_size: 4
+  # micro_batch_size: 2
   # micro_batch_size: 1
 
   # Number of iterations with learning rate warmup active (type: int, default: 2000)