update for v4 so we don't crash

stanford-crfm · Oct 23, 2024 · f0ca163 · f0ca163
1 parent 1706803
commit f0ca163
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/config/llama_7b_tulu.yaml b/config/llama_7b_tulu.yaml
@@ -14,7 +14,7 @@ model:  # 7B class model
   num_heads: 32
   num_kv_heads: 32
   use_flash_attention: True
-  flash_attention_block_size: 1024
+  flash_attention_block_size: 512
   use_bias: false
   use_layer_norm_weight: false
 trainer: