apunkt/doc-to-lora
1
0
Fork 0
mirror of https://github.com/SakanaAI/doc-to-lora.git synced 2026-07-23 17:01:04 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

clean up configs + scripts + chat_templates (+ %generation% tag)

Browse source
This commit is contained in:
51616 2025-06-06 04:56:45 +00:00
parent 6c43c6a329
commit a9a164e2de
76 changed files with 72 additions and 3237 deletions
Download patch file Download diff file Expand all files Collapse all files

33
scripts/ablate_arch/gemma_xl_per_rank_gen_fac2.sh
View file

@ -1,33 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#SBATCH --partition=a3
#SBATCH --nodes=1
#SBATCH --exclude=slurm0-a3nodeset-2
#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 29560 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_self_attends_per_block=8 --num_latent_factor=2 \
--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

32
scripts/ablate_arch/gemma_xl_per_rank_gen_fac2_neft1.sh
View file

@ -1,32 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#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 29560 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_self_attends_per_block=8 --num_latent_factor=2 \
--lora_r=8 \
--eval_steps=5000 --save_steps=5000 --learning_rate=2e-5 --lora_dropout=0.0 \
--neftune_noise_alpha=1 --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

34
scripts/ablate_arch/gemma_xl_per_rank_gen_fac2_per_layer_processing.sh
View file

@ -1,34 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#SBATCH --partition=a3
#SBATCH --nodes=1
#SBATCH --exclude=slurm0-a3nodeset-2
#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_self_attends_per_block=8 --num_latent_factor=2 \
--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 \
--per_layer_processing=True

34
scripts/ablate_arch/gemma_xl_per_rank_gen_fac2_per_layer_processing_4e-5_lr.sh
View file

@ -1,34 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#SBATCH --partition=a3
#SBATCH --nodes=1
#SBATCH --exclude=slurm0-a3nodeset-2
#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_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 \
--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 \
--per_layer_processing=True

35
scripts/ablate_arch/gemma_xl_per_rank_gen_fac2_per_layer_processing_decoder_2.sh
View file

@ -1,35 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#SBATCH --partition=a3
#SBATCH --nodes=1
#SBATCH --exclude=slurm0-a3nodeset-2
#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_self_attends_per_block=8 --num_latent_factor=2 \
--decoder_depth=2 \
--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 \
--per_layer_processing=True

34
scripts/ablate_arch/gemma_xl_per_rank_gen_fac2_per_layer_processing_r16.sh
View file

@ -1,34 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#SBATCH --partition=a3
#SBATCH --nodes=1
#SBATCH --exclude=slurm0-a3nodeset-2
#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_self_attends_per_block=8 --num_latent_factor=2 \
--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 \
--per_layer_processing=True

32
scripts/ablate_arch/gemma_xl_per_rank_gen_fac4.sh
View file

@ -1,32 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#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_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

33
scripts/ablate_arch/gemma_xl_per_rank_gen_fac4_per_layer_processing.sh
View file

@ -1,33 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#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_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 --per_layer_processing=True \
--gen_lora_l1_reg_coef=0.1

33
scripts/ablate_arch/gemma_xl_per_rank_gen_fac4_per_layer_processing_decoder_2.sh
View file

@ -1,33 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#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_self_attends_per_block=8 --num_latent_factor=4 \
--decoder_depth=2 \
--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 --per_layer_processing=True

33
scripts/ablate_arch/gemma_xl_per_rank_gen_fac4_per_layer_processing_no_reg.sh
View file

@ -1,33 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#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_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 --per_layer_processing=True \
--gen_lora_l1_reg_coef=0.0

33
scripts/ablate_arch/gemma_xl_per_rank_gen_fac4_per_layer_processing_token_mixing.sh
View file

@ -1,33 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#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_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 --per_layer_processing=True \
--use_token_mixing=True

34
scripts/fw_qa_3/gemma_per_rank_fac2_per_layer.sh
View file

@ -1,34 +0,0 @@
#!/bin/bash
#SBATCH --job-name=ctxlora
#SBATCH --partition=a3
#SBATCH --nodes=1
#SBATCH --exclude=slurm0-a3nodeset-2
#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_3.yaml \
--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=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_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 \
--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 \
--per_layer_processing=True

33
scripts/fw_qa_3/medium_gemma_per_rank_fac2_per_layer.sh
View file

@ -1,33 +0,0 @@
#!/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
# 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=8 --gradient_accumulation_steps=4 --gradient_clipping=1.0 \
--gpu_ids all --main_process_port 29568 intx_sft.py configs/pretrain_all_xl_3_medium.yaml \
--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=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_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 \
--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 \
--per_layer_processing=True

18
scripts/fw_qa_3/medium_gemma_per_rank_fac2_per_layer_4_gpus.sh → scripts/fw_qa_pretrain/gemma_pretrain.sh
View file

@ -14,20 +14,22 @@
# 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
# . ~/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 29564 intx_sft.py configs/pretrain_all_xl_3_medium.yaml \
--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=1 --per_device_train_batch_size=32 \
--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=32 --gradient_clipping=1.0 \
--gpu_ids all --main_process_port 29560 intx_sft.py configs/fw_qa_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=8 --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=5000 --save_steps=5000 --learning_rate=4e-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_negative_prompt=False \
--add_repeat_prompt=False \
--use_sequence_packing=True --per_rank_gen=True \
--per_layer_processing=True
--per_layer_processing=True \
--gen_lora_l1_reg_coef=0.1 \

17
scripts/fw_qa_3/medium_gemma_per_rank_fac2_per_layer_4_gpus_gen_lora_reg.sh → scripts/fw_qa_pretrain/gemma_pretrain_and_pwc.sh
View file

@ -14,21 +14,22 @@
# 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
# . ~/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 29564 intx_sft.py configs/pretrain_all_xl_3_medium.yaml \
--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=1 --per_device_train_batch_size=32 \
--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --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 29560 intx_sft.py configs/fw_qa_pretrain_small_and_pwc.yaml \
--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=1 --per_device_train_batch_size=4 \
--gradient_accumulation_steps=16 --per_device_eval_batch_size=8 --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=5000 --save_steps=5000 --learning_rate=4e-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_negative_prompt=False \
--add_repeat_prompt=False \
--use_sequence_packing=True --per_rank_gen=True \
--per_layer_processing=True \
--gen_lora_l1_reg_coef=0.1
--gen_lora_l1_reg_coef=0.1 \

31
scripts/repeat_text/gemma_llama_xl.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_fixed_latents/gemma_llama_xl.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_fixed_latents/gemma_llama_xl_down_ln.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_fixed_latents/gemma_llama_xl_ln_only.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_fixed_latents/gemma_llama_xl_up_down.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_fixed_latents/gemma_llama_xl_up_down_ln.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_no_lightweight/gemma_llama_2.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_no_lightweight/gemma_llama_xl.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_no_lightweight/gemma_llama_xl_per_rank_gen.sh
View file

@ -1,31 +0,0 @@
#!/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

32
scripts/scale_modules_no_lightweight/gemma_llama_xl_per_rank_gen_latent_fac_1.sh
View file

@ -1,32 +0,0 @@
#!/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

33
scripts/scale_modules_no_lightweight/gemma_llama_xl_per_rank_gen_latent_fac_1_llama.sh
View file

@ -1,33 +0,0 @@
#!/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

32
scripts/scale_modules_no_lightweight/gemma_llama_xl_per_rank_gen_latent_fac_1_r_16.sh
View file

@ -1,32 +0,0 @@
#!/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

31
scripts/scale_modules_no_lightweight/gemma_llama_xl_r_16.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_no_lightweight/gemma_llama_xl_r_16_ln.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_no_lightweight/gemma_llama_xl_up_down.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl_down_ln.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl_ln_only.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl_up_down.sh
View file

@ -1,31 +0,0 @@
#!/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

31
scripts/scale_modules_scale_latents_no_lightweight/gemma_llama_xl_up_down_ln.sh
View file

@ -1,31 +0,0 @@
#!/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

0
scripts/fw_qa_3_mini/medium_gemma_per_rank_fac2_per_layer_4_gpus_gen_lora_reg.sh → scripts/self_gen_fw_qa_3_mini/medium_gemma_per_rank_fac2_per_layer_4_gpus_gen_lora_reg.sh
View file

0
scripts/fw_qa_3_mini/medium_gemma_per_rank_fac2_per_layer_4_gpus_gen_lora_reg_repeat.sh → scripts/self_gen_fw_qa_3_mini/medium_gemma_per_rank_fac2_per_layer_4_gpus_gen_lora_reg_repeat.sh
View file