optimize E8: precompute causal mask buffer, add eval early exit, and enable compile_model in e8.csv

1 files changed, 7 insertions, 6 deletions

diff --git a/jobs/e8.csv b/jobs/e8.csv
index 0c393a0..153b373 100644
--- a/jobs/e8.csv
+++ b/jobs/e8.csv
@@ -1,7 +1,8 @@
 job,model,seed,flags
-e8-armA-none,transformer,0,--range_end 1000 --max_steps 32 --device cuda --weight_decay 0.1 --max_train_steps 200000 --scratch_mode none
-e8-armB-structured,transformer,0,--range_end 1000 --max_steps 32 --device cuda --weight_decay 0.1 --max_train_steps 200000 --scratch_mode structured
-e8-armC-filler,transformer,0,--range_end 1000 --max_steps 32 --device cuda --weight_decay 0.1 --max_train_steps 200000 --scratch_mode filler --scratch_len 16
-e8-armD1-randlearn,transformer,0,--range_end 1000 --max_steps 32 --device cuda --weight_decay 0.1 --max_train_steps 200000 --scratch_mode random_learned --scratch_len 16
-e8-armD2-randfroz,transformer,0,--range_end 1000 --max_steps 32 --device cuda --weight_decay 0.1 --max_train_steps 200000 --scratch_mode random_frozen --scratch_len 16
-e8-armD3-randnoise,transformer,0,--range_end 1000 --max_steps 32 --device cuda --weight_decay 0.1 --max_train_steps 200000 --scratch_mode random_noise --scratch_len 16
+e8-armA-none,transformer,0,--range_end 1000 --max_steps 32 --device cuda --compile_model True --weight_decay 0.1 --max_train_steps 200000 --scratch_mode none
+e8-armB-structured,transformer,0,--range_end 1000 --max_steps 32 --device cuda --compile_model True --weight_decay 0.1 --max_train_steps 200000 --scratch_mode structured
+e8-armC-filler,transformer,0,--range_end 1000 --max_steps 32 --device cuda --compile_model True --weight_decay 0.1 --max_train_steps 200000 --scratch_mode filler --scratch_len 16
+e8-armD1-randlearn,transformer,0,--range_end 1000 --max_steps 32 --device cuda --compile_model True --weight_decay 0.1 --max_train_steps 200000 --scratch_mode random_learned --scratch_len 16
+e8-armD2-randfroz,transformer,0,--range_end 1000 --max_steps 32 --device cuda --compile_model True --weight_decay 0.1 --max_train_steps 200000 --scratch_mode random_frozen --scratch_len 16
+e8-armD3-randnoise,transformer,0,--range_end 1000 --max_steps 32 --device cuda --compile_model True --weight_decay 0.1 --max_train_steps 200000 --scratch_mode random_noise --scratch_len 16
+