data exp + packing

scripts/gemma_data_exp/gemma_qa.sh

@@ -6,18 +6,20 @@
 #SBATCH --output=outputs/%x-%j.out
 #SBATCH --error=outputs/%x-%j.out
 
-uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=16 --gradient_clipping=1.0 \
 --gpu_ids all --main_process_port 29561 intx_sft.py configs/qa.yaml \
---model_name_or_path=google/gemma-2-2b-it --num_train_epochs=4 --per_device_train_batch_size=32 \
---gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=2.2 --per_device_train_batch_size=-1 \
+--gradient_accumulation_steps=16 --per_device_eval_batch_size=4 --exp_setup=hyper_lora --aggregator_type=perceiver \
 --target_modules=down_proj \
---num_self_attends_per_block=4 --num_latent_factor=1 \
+--num_self_attends_per_block=8 --num_latent_factor=2 \
 --lora_r=8 \
---eval_steps=5000 --save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
+--eval_steps=1000000000000000 --do_eval=no --do_predict=no --eval_on_start=False \
+--save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
 --neftune_noise_alpha=5 --use_light_weight_lora=False \
 --add_negative_prompt=False \
 --add_repeat_prompt=False \
---use_sequence_packing=True --per_rank_gen=True \
+--use_sequence_packing=True --max_packed_inp_len=8192 --max_packed_ctx_len=16384 \
+--per_rank_gen=True \
 --per_layer_processing=True \
 --gen_lora_l1_reg_coef=0.1 \
 

scripts/gemma_data_exp/gemma_qa_8_gpu.sh

@@ -0,0 +1,25 @@
+#!/bin/bash
+#SBATCH --job-name=ctxlora_medium
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=8
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+uv run accelerate launch --num_processes=8 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29561 intx_sft.py configs/qa.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=2.2 --per_device_train_batch_size=-1 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=4 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_self_attends_per_block=8 --num_latent_factor=2 \
+--lora_r=16 \
+--eval_steps=1000000000000000 --do_eval=no --do_predict=no --eval_on_start=False \
+--save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --max_packed_inp_len=8192 --max_packed_ctx_len=16384 \
+--per_rank_gen=True \
+--per_layer_processing=True \
+--gen_lora_l1_reg_coef=0.1 \
+

scripts/gemma_data_exp/gemma_self_gen_qa.sh

@@ -8,16 +8,17 @@
 
 uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=16 --gradient_clipping=1.0 \
 --gpu_ids all --main_process_port 29562 intx_sft.py configs/self_gen_3_small.yaml \
---model_name_or_path=google/gemma-2-2b-it --num_train_epochs=3 --per_device_train_batch_size=8 \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=2.2 --per_device_train_batch_size=-1 \
 --gradient_accumulation_steps=16 --per_device_eval_batch_size=4 --exp_setup=hyper_lora --aggregator_type=perceiver \
 --target_modules=down_proj \
---num_self_attends_per_block=4 --num_latent_factor=1 \
+--num_self_attends_per_block=8 --num_latent_factor=2 \
 --lora_r=8 \
---eval_steps=5000 --save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
+--eval_steps=1000000000000000 --do_eval=no --do_predict=no --eval_on_start=False \
+--save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
 --neftune_noise_alpha=5 --use_light_weight_lora=False \
 --add_negative_prompt=False \
 --add_repeat_prompt=False \
---use_sequence_packing=True --per_rank_gen=True \
+--use_sequence_packing=True --max_packed_inp_len=8192 --max_packed_ctx_len=16384 \
+--per_rank_gen=True \
 --per_layer_processing=True \
 --gen_lora_l1_reg_coef=0.1 \
-

scripts/gemma_data_exp/gemma_self_gen_qa_mini.sh

@@ -11,13 +11,13 @@ uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=16 --gr
 --model_name_or_path=google/gemma-2-2b-it --num_train_epochs=3 --per_device_train_batch_size=8 \
 --gradient_accumulation_steps=16 --per_device_eval_batch_size=4 --exp_setup=hyper_lora --aggregator_type=perceiver \
 --target_modules=down_proj \
---num_self_attends_per_block=8 --num_latent_factor=1 \
+--num_self_attends_per_block=8 --num_latent_factor=2 \
 --lora_r=8 \
 --eval_steps=5000 --save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
 --neftune_noise_alpha=5 --use_light_weight_lora=False \
 --add_negative_prompt=False \
 --add_repeat_prompt=False \
---use_sequence_packing=True --max_packed_inp_len=16384 --max_packed_ctx_len=32768 \
+--use_sequence_packing=True --max_packed_inp_len=8192 --max_packed_ctx_len=16384 \
 --per_rank_gen=True \
 --per_layer_processing=True \
 --gen_lora_l1_reg_coef=0.1 \

scripts/gemma_data_exp/gemma_self_gen_qa_repeat.sh

@@ -3,21 +3,24 @@
 #SBATCH --partition=a3
 #SBATCH --nodes=1
 #SBATCH --gpus=4
+# NO ! #SBATCH --begin=now+6hour                   # delay scheduling
 #SBATCH --output=outputs/%x-%j.out
 #SBATCH --error=outputs/%x-%j.out
 
-uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=32 --gradient_clipping=1.0 \
---gpu_ids all --main_process_port 29563 intx_sft.py configs/self_gen_3_and_pretrain_small.yaml \
---model_name_or_path=google/gemma-2-2b-it --num_train_epochs=2 --per_device_train_batch_size=4 \
---gradient_accumulation_steps=32 --per_device_eval_batch_size=4 --exp_setup=hyper_lora --aggregator_type=perceiver \
+uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=16 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/self_gen_3_small.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=2 --per_device_train_batch_size=-1 \
+--gradient_accumulation_steps=16 --per_device_eval_batch_size=4 --exp_setup=hyper_lora --aggregator_type=perceiver \
 --target_modules=down_proj \
---num_self_attends_per_block=4 --num_latent_factor=1 \
+--num_self_attends_per_block=8 --num_latent_factor=2 \
 --lora_r=8 \
---eval_steps=5000 --save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
+--eval_steps=1000000000000000 --do_eval=no --do_predict=no --eval_on_start=False \
+--save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
 --neftune_noise_alpha=5 --use_light_weight_lora=False \
 --add_negative_prompt=False \
---add_repeat_prompt=False \
---use_sequence_packing=True --per_rank_gen=True \
+--add_repeat_prompt=True --repeat_prob=0.1 \
+--use_sequence_packing=True --max_packed_inp_len=8192 --max_packed_ctx_len=16384 \
+--per_rank_gen=True \
 --per_layer_processing=True \
 --gen_lora_l1_reg_coef=0.1 \
 

scripts/gemma_data_exp/gemma_self_gen_qa_repeat_and_pretrain.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+#SBATCH --job-name=ctxlora_medium
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+# NO ! #SBATCH --begin=now+6hour                   # delay scheduling
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=16 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29564 intx_sft.py configs/data_exp/self_gen_3_and_pretrain_small.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=1.5 \
+--gradient_accumulation_steps=16 --per_device_eval_batch_size=4 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_self_attends_per_block=8 --num_latent_factor=2 \
+--lora_r=8 \
+--eval_steps=1000000000000000 --do_eval=no --do_predict=no --eval_on_start=False \
+--save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--add_negative_prompt=False \
+--add_repeat_prompt=True --repeat_prob=0.1 \
+--use_sequence_packing=True --max_packed_inp_len=8192 --max_packed_ctx_len=16384 \
+--per_rank_gen=True \
+--per_layer_processing=True \
+--gen_lora_l1_reg_coef=0.1 \
+

scripts/gemma_data_exp/gemma_self_gen_qa_repeat_and_pretrain_augment.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+#SBATCH --job-name=ctxlora_medium
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+# NO ! #SBATCH --begin=now+6hour                   # delay scheduling
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=16 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29565 intx_sft.py configs/data_exp/self_gen_3_and_pretrain_small_aug.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=1 \
+--gradient_accumulation_steps=16 --per_device_eval_batch_size=4 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_self_attends_per_block=8 --num_latent_factor=2 \
+--lora_r=8 \
+--eval_steps=1000000000000000 --do_eval=no --do_predict=no --eval_on_start=False \
+--save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--add_negative_prompt=False \
+--add_repeat_prompt=True --repeat_prob=0.1 \
+--use_sequence_packing=True --max_packed_inp_len=8192 --max_packed_ctx_len=16384 \
+--per_rank_gen=True \
+--per_layer_processing=True \
+--gen_lora_l1_reg_coef=0.1 \
+