fw_qa_3_mini

configs/pretrain_all_xl_3.yaml

@@ -0,0 +1,59 @@
+output_dir: "" # just a placeholder
+bf16: true
+model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+label_names: ["labels"]
+# eval_on_start: True
+# eval_strategy: "steps"
+# eval_steps: 500
+# save_strategy: "no"
+# # save_steps: 500
+# logging_strategy: "steps"
+# logging_steps: 100
+# use_liger_kernel: true
+# remove_unused_columns: false
+
+# needed to avoid OOM by compute the metrics batch by batch
+# w/o this the trainer stores logits of all sample in memory...
+# batch_eval_metrics: true
+
+per_device_train_batch_size: 8
+per_device_eval_batch_size: 8
+max_val_samples_per_ds: 1000
+# optim: schedule_free_adamw
+
+learning_rate: 0.00004
+# lr_scheduler_type: "constant_with_warmup"
+neftune_noise_alpha: 5
+weight_decay: 0.01
+#
+warmup_steps: 100
+
+dataloader_prefetch_factor: 8
+dataloader_num_workers: 8
+# LoRA
+lora_r: 8
+lora_dropout: 0.0
+target_modules:
+  - down_proj
+# data
+train_ds_names:
+  - fw_qa_3_mini
+  - ctx_qa
+  - pwc
+  - hotpot_qa
+  - squad
+  - drop
+  - narrativeqa
+  - quoref
+  - ropes
+  - synthetic_convqa
+
+val_ds_names:
+  - fw_qa_xl
+  - ctx_qa
+  - pwc
+  - hotpot_qa
+  - squad
+
+load_best_model_at_end: true
+metric_for_best_model: eval_pwc_loss

scripts/ablate_arch/gemma_xl_per_rank_gen_fac2_per_layer_processing_4e-5_lr.sh

@@ -0,0 +1,34 @@
+#!/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

scripts/ablate_arch/gemma_xl_per_rank_gen_fac2_per_layer_processing_decoder_2.sh

@@ -0,0 +1,35 @@
+#!/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

scripts/ablate_arch/gemma_xl_per_rank_gen_fac2_per_layer_processing_r16.sh

@@ -0,0 +1,34 @@
+#!/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

scripts/fw_qa_3/gemma_per_rank_fac2_per_layer.sh

@@ -0,0 +1,34 @@
+#!/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=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

src/ctx_to_lora/data_utils.py

@@ -98,6 +98,13 @@ DS_KWARGS = {
             split="train",
         ),
     ),
+    "fw_qa_3_mini": dict(
+        train=dict(
+            path="parquet",
+            data_files=glob("data/raw_datasets/fw_qa_3/*[!val].parquet"),
+            split="train[:100000]",
+        ),
+    ),
     "ctx_qa": dict(
         train=dict(
             path="parquet",
@@ -250,7 +257,7 @@ def add_repeat_prompt_fn(samples):
     ):
         # Only process if the context is not already in the set
         if ctx in unique_contexts:
-           continue
+            continue
         # remove too long contexts
         if len(ctx) > 1000:
             continue

src/ctx_to_lora/modeling_utils.py

@@ -756,14 +756,14 @@ class HyperLoRA(nn.Module):
             if self.config.per_layer_processing:
                 layers = [
                     # nn.Linear(self.d_latent, self.d_latent),
-                    Mix(
-                        "bs n_layers n_modules r d -> bs n_layers n_modules r d0",
-                        weight_shape="n_layers d d0",
-                        bias_shape="n_layers d0",
-                        n_layers=self.n_layers,
-                        d=self.d_latent,
-                        d0=self.d_latent,
-                    ),
+                    # Mix(
+                    #     "bs n_layers n_modules r d -> bs n_layers n_modules r d0",
+                    #     weight_shape="n_layers d d0",
+                    #     bias_shape="n_layers d0",
+                    #     n_layers=self.n_layers,
+                    #     d=self.d_latent,
+                    #     d0=self.d_latent,
+                    # ),
                     ResMLPBlockPerLayer(
                         self.n_layers,
                         self.d_latent,
@@ -772,27 +772,29 @@ class HyperLoRA(nn.Module):
                         # self.num_modules,
                         # self.lora_config.r,
                         # self.d_latent,
-                    ),
+                    )
+                    for _ in range(4)
                 ]
             else:
                 layers = [
-                    nn.Linear(self.d_latent, self.d_latent),
+                    # nn.Linear(self.d_latent, self.d_latent),
                     MLPResidualBlock(
                         input_size=self.config.latent_size,
                         hidden_size=self.config.latent_size * 4,
                         output_size=self.config.latent_size,
                         dropout_rate=getattr(self.config, "dropout_rate", 0),
-                    ),
-                ]
-            if self.config.use_token_mixing:
-                layers = [
-                    ResMLPTokenMixingPerLayer(
-                        self.n_layers,
-                        self.num_modules,
-                        self.lora_config.r,
-                        self.d_latent,
                     )
-                ] + layers
+                    for _ in range(4)
+                ]
+            # if self.config.use_token_mixing:
+            #     layers = [
+            #         ResMLPTokenMixingPerLayer(
+            #             self.n_layers,
+            #             self.num_modules,
+            #             self.lora_config.r,
+            #             self.d_latent,
+            #         )
+            #     ] + layers
 
             self.layers = nn.Sequential(*layers)