add per rank latent query + per layer hyper head

scripts/repeat_text/gemma_llama_xl.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29562 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=16 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=True \
+--use_sequence_packing=True --ctx_encoder_model_name_or_path=meta-llama/Llama-3.2-3B-Instruct

scripts/scale_modules_fixed_latents/gemma_llama_xl.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29561 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=True --light_weight_latent_size=512 \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_fixed_latents/gemma_llama_xl_down_ln.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29566 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj --extra_modules=input_layernorm,post_attention_layernorm \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=2 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=True --light_weight_latent_size=512 \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_fixed_latents/gemma_llama_xl_ln_only.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=4 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=4 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=None --extra_modules=input_layernorm,post_attention_layernorm \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=4 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=True --light_weight_latent_size=512 \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_fixed_latents/gemma_llama_xl_up_down.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29562 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj,up_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=4 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=True --light_weight_latent_size=512 \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_fixed_latents/gemma_llama_xl_up_down_ln.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=10.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj,up_proj --extra_modules=input_layernorm,post_attention_layernorm \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=2 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=True --light_weight_latent_size=512 \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_no_lightweight/gemma_llama_2.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29562 intx_sft.py configs/pretrain_all_2.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=1.0 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=16 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --ctx_encoder_model_name_or_path=meta-llama/Llama-3.2-3B-Instruct

scripts/scale_modules_no_lightweight/gemma_llama_xl.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29562 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=16 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --ctx_encoder_model_name_or_path=meta-llama/Llama-3.2-3B-Instruct

scripts/scale_modules_no_lightweight/gemma_llama_xl_per_rank_gen.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29562 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=16 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=16 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --per_rank_gen=True

scripts/scale_modules_no_lightweight/gemma_llama_xl_per_rank_gen_latent_fac_1.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=4 \
+--lora_r=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --per_rank_gen=True

scripts/scale_modules_no_lightweight/gemma_llama_xl_per_rank_gen_latent_fac_1_llama.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=2 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --per_rank_gen=True \
+--ctx_encoder_model_name_or_path=meta-llama/Llama-3.2-3B-Instruct
+

scripts/scale_modules_no_lightweight/gemma_llama_xl_per_rank_gen_latent_fac_1_r_16.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=1 \
+--lora_r=16 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --per_rank_gen=True

scripts/scale_modules_no_lightweight/gemma_llama_xl_r_16.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=16 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29564 intx_sft.py configs/pretrain_all_xl_and_sum.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=16 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=16 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_r=16 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_no_lightweight/gemma_llama_xl_r_16_ln.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=16 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/pretrain_all_xl_and_sum.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=16 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj --extra_modules=input_layernorm,post_attention_layernorm \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=16 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_r=16 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_no_lightweight/gemma_llama_xl_up_down.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=16 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/pretrain_all_xl_and_sum.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=16 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=up_proj,down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_r=8 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=4 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29571 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=4 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=16 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=3e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl_down_ln.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=4 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29565 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=4 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj --extra_modules=input_layernorm,post_attention_layernorm \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=True --light_weight_latent_size=512 \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl_ln_only.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=4 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29564 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=4 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=None --extra_modules=input_layernorm,post_attention_layernorm \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=True --light_weight_latent_size=512 \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl_up_down.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=4 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29562 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=4 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj,up_proj \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=True --light_weight_latent_size=512 \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True

scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl_up_down_ln.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=gemma_llama_instruct
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+# module load
+# module load cuda/12.1
+# module load cudnn/8.9.7
+# module load nccl/cuda-12.1/2.18.3
+# module load hpcx/2.20
+
+# export OMP_NUM_THREADS=24
+# export TRITON_CACHE_DIR=/tmp/.triton/
+. ~/miniconda3/etc/profile.d/conda.sh
+conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
+# eval "$@"
+
+accelerate launch --num_processes=4 --gradient_accumulation_steps=4 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/pretrain_all_xl.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=5.1 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=4 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj,up_proj --extra_modules=input_layernorm,post_attention_layernorm \
+--num_blocks=1 --num_self_attends_per_block=8 --num_latent_factor=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=True --light_weight_latent_size=512 \
+--load_best_model_at_end=True --metric_for_best_model=pwc_loss --add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True