[KUNLUNXIN] minor fix for llama3-8b, remove cpu-initialization

FlagOpen · Oct 4, 2024 · c6c25fd · c6c25fd
1 parent ceb9e7a
commit c6c25fd
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/training/kunlunxin/llama3_8B-megatron/config/training_adapter.sh b/training/kunlunxin/llama3_8B-megatron/config/training_adapter.sh
@@ -15,7 +15,7 @@ VENDOR_ARGS=" \
     --use-mcore-models \
     --use-flash-attn \
     --disable-bias-linear \
-    --use-cpu-initialization --hidden-dropout 0 --attention-dropout 0 \
+    --hidden-dropout 0 --attention-dropout 0 \
     --no-async-tensor-model-parallel-allreduce --no-gradient-accumulation-fusion
 "