roughly match the base arch (not yet per layer)

scripts/ablate_arch/gemma_xl_per_rank_gen_fac2.sh

@@ -0,0 +1,32 @@
+#!/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=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/ablate_arch/gemma_xl_per_rank_gen_fac2_neft1.sh

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

scripts/ablate_arch/gemma_xl_per_rank_gen_fac4.sh

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

scripts/ablate_arch/gemma_xl_per_rank_gen_fac4_per_layer_processing.sh

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

scripts/ablate_arch/gemma_xl_per_rank_gen_fac4_per_layer_processing_decoder_2.sh

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

scripts/ablate_arch/gemma_xl_per_rank_gen_fac4_per_layer_processing_no_reg.sh

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

scripts/ablate_arch/gemma_xl_per_rank_gen_fac4_per_layer_processing_token_mixing.sh

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

src/ctx_to_lora/modeling_idefics2.py

@@ -80,7 +80,7 @@ class Idefics2PerceiverConfig(PretrainedConfig):
         self,
         input_size: int,
         intermediate_size_factor: int = 1,
-        hidden_act="gelu_pytorch_tanh",
+        hidden_act="silu",
         hidden_size=4096,
         rms_norm_eps=1e-06,
         resampler_n_latents=64,

src/ctx_to_lora/modeling_utils.py

@@ -288,7 +288,7 @@ class Mixer(nn.Module):
         self.gate_proj = nn.Linear(input_size, intermediate_emb_size, bias=False)
         self.up_proj = nn.Linear(input_size, intermediate_emb_size, bias=False)
         self.down_proj = nn.Linear(intermediate_emb_size, output_size, bias=False)
-        self.act_fn = nn.GELU(approximate="tanh")
+        self.act_fn = nn.SiLU()
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
@@ -309,14 +309,14 @@ class MLPResidualBlock(nn.Module):
         # if pre_layer_norm:
         # layers.append(nn.LayerNorm(input_size))
         layers = [
-            Gemma3RMSNorm(input_size),
+            nn.LayerNorm(input_size),
             nn.Dropout(dropout_rate),
             nn.Linear(input_size, hidden_size),
-            nn.GELU(approximate="tanh"),
+            nn.SiLU(),
             nn.Dropout(dropout_rate),
             nn.Linear(hidden_size, output_size),
-            Gemma3RMSNorm(output_size),
-            # nn.GELU(approximate="tanh"),
+            nn.LayerNorm(output_size),
+            # nn.SiLU(),
         ]
         # if post_dropout:
         #     layers.append(nn.Dropout(dropout_rate))
@@ -576,26 +576,64 @@ class MixerPerLayer(nn.Module):
         self.up_proj = UpMixPerLayer(n_layers, n_modules, r, d_in, d_in * 4)
         self.gate_proj = UpMixPerLayer(n_layers, n_modules, r, d_in, d_in * 4)
         self.down_proj = DownMixPerLayer(n_layers, n_modules, r, d_in * 4, d_in)
-        self.act_fn = nn.GELU(approximate="tanh")
+        self.act_fn = nn.SiLU()
 
     def forward(self, x):
         return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
 
 
+# class ResMLPBlockPerLayer(nn.Module):
+#     def __init__(self, n_layers: int, n_modules: int, r: int, d_in: int):
+#         super().__init__()
+#         # input shape: [bs, n_layers, n_modules, feature_dim]
+#         self.pre_norm = nn.LayerNorm(d_in)
+#         self.post_norm = nn.LayerNorm(d_in)
+#         self.mixer = MixerPerLayer(n_layers, n_modules, r, d_in)
+
+#     def forward(self, x):
+#         inp = x
+#         x = self.pre_norm(x)
+#         x = self.mixer(x)
+#         x = self.post_norm(x)
+#         return x + inp
+
+
 class ResMLPBlockPerLayer(nn.Module):
-    def __init__(self, n_layers: int, n_modules: int, r: int, d_in: int):
+    def __init__(
+        self,
+        n_layers: int,
+        input_size: int,
+        hidden_size: int,
+        output_size: int,
+    ):
         super().__init__()
-        # input shape: [bs, n_layers, n_modules, feature_dim]
-        self.pre_norm = Gemma3RMSNorm(d_in)
-        self.post_norm = Gemma3RMSNorm(d_in)
-        self.mixer = MixerPerLayer(n_layers, n_modules, r, d_in)
+        layers = [
+            nn.LayerNorm(input_size),
+            # nn.Linear(input_size, hidden_size),
+            Mix(
+                "bs n_layers n_modules r d_in -> bs n_layers n_modules r d_hid",
+                weight_shape="n_layers d_in d_hid",
+                bias_shape="n_layers d_hid",
+                n_layers=n_layers,
+                d_in=input_size,
+                d_hid=hidden_size,
+            ),
+            nn.SiLU(),
+            # nn.Linear(hidden_size, output_size),
+            Mix(
+                "bs n_layers n_modules r d_hid -> bs n_layers n_modules r d_out",
+                weight_shape="n_layers d_hid d_out",
+                bias_shape="n_layers d_out",
+                n_layers=n_layers,
+                d_hid=hidden_size,
+                d_out=output_size,
+            ),
+            nn.LayerNorm(output_size),
+        ]
+        self.layers = nn.Sequential(*layers)
 
     def forward(self, x):
-        inp = x
-        x = self.pre_norm(x)
-        x = self.mixer(x)
-        x = self.post_norm(x)
-        return x + inp
+        return x + self.layers(x)
 
 
 class ResMLPTokenMixingPerLayer(nn.Module):
@@ -603,8 +641,8 @@ class ResMLPTokenMixingPerLayer(nn.Module):
         super().__init__()
         # input shape: [bs, n_layers, n_modules, feature_dim]
         d_hid = d_in // 4
-        self.pre_norm = Gemma3RMSNorm(d_in)
-        self.post_norm = Gemma3RMSNorm(d_in)
+        self.pre_norm = nn.LayerNorm(d_in)
+        self.post_norm = nn.LayerNorm(d_in)
         # self.proj_down = Mix(
         #     "bs n_layers n_modules r d_in -> bs n_layers n_modules r d_hid",
         #     weight_shape="n_layers d_in d_hid",
@@ -624,7 +662,7 @@ class ResMLPTokenMixingPerLayer(nn.Module):
         #     d_hid=d_hid,
         #     d_in=d_in,
         # )
-        # act_fn = nn.GELU(approximate="tanh")
+        # act_fn = nn.SiLU()
         # out_proj = Mix(
         #     "bs n_layers n_modules r d_in -> bs n_layers n_modules r0 d_in0",
         #     weight_shape="n_layers r d_in r0 d_in0",
@@ -652,7 +690,7 @@ class ResMLPTokenMixingPerLayer(nn.Module):
             r=r,
             r0=r * n_token_factors,
         )
-        act_fn = nn.GELU(approximate="tanh")
+        act_fn = nn.SiLU()
         token_mix_down = Mix(
             "bs n_layers n_modules r0 d_in -> bs n_layers n_modules r d_in",
             weight_shape="n_layers r0 r",
@@ -717,17 +755,29 @@ class HyperLoRA(nn.Module):
             # )
             if self.config.per_layer_processing:
                 layers = [
-                    nn.Linear(self.d_latent, self.d_latent),
+                    # 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,
+                    # ),
                     ResMLPBlockPerLayer(
                         self.n_layers,
-                        self.num_modules,
-                        self.lora_config.r,
                         self.d_latent,
+                        self.d_latent * 4,
+                        self.d_latent,
+                        # self.num_modules,
+                        # self.lora_config.r,
+                        # self.d_latent,
                     ),
                 ]
             else:
                 layers = [
-                    nn.Linear(self.d_latent, self.d_latent),
+                    # nn.Linear(self.d_latent, self.d_latent),
+                    # nn.LayerNorm(self.d_latent),
                     MLPResidualBlock(
                         input_size=self.config.latent_size,
                         hidden_size=self.config.latent_size * 4,
@@ -819,9 +869,9 @@ class HyperLoRA(nn.Module):
                     if self.config.per_layer_processing:
                         self.head = Mix(
                             "bs n_layers n_modules r d_latent -> bs n_layers n_modules r d_lora",
-                            weight_shape="n_layers n_modules r d_latent d_lora",
+                            weight_shape="n_layers n_modules d_latent d_lora",
                             # bias_shape=None,  # no bias
-                            bias_shape="n_layers n_modules r d_lora",
+                            bias_shape="n_layers n_modules d_lora",
                             n_layers=len(self.layer_indices),
                             n_modules=n_modules,
                             d_latent=self.config.latent_size,
@@ -831,9 +881,9 @@ class HyperLoRA(nn.Module):
                     else:
                         self.head = Mix(
                             "bs n_layers n_modules r d_latent -> bs n_layers n_modules r d_lora",
-                            weight_shape="n_modules r d_latent d_lora",
+                            weight_shape="n_modules d_latent d_lora",
                             # bias_shape=None,  # no bias
-                            bias_shape="n_modules r d_lora",
+                            bias_shape="n_modules d_lora",
                             n_layers=len(self.layer_indices),
                             n_modules=n_modules,
                             d_latent=self.config.latent_size,
@@ -852,15 +902,16 @@ class HyperLoRA(nn.Module):
                             r=self.config.lora_config.r,
                             d_lora=d_lora,
                         )
-                    self.head = Mix(
-                        "bs n_layers n_modules d_latent -> bs n_layers n_modules r d_lora",
-                        weight_shape="d_latent r d_lora",
-                        # bias_shape=None,  # no bias
-                        bias_shape="r d_lora",
-                        d_latent=self.config.latent_size,
-                        r=self.config.lora_config.r,
-                        d_lora=d_lora,
-                    )
+                    else:
+                        self.head = Mix(
+                            "bs n_layers n_modules d_latent -> bs n_layers n_modules r d_lora",
+                            weight_shape="d_latent r d_lora",
+                            # bias_shape=None,  # no bias
+                            bias_shape="r d_lora",
+                            d_latent=self.config.latent_size,
+                            r=self.config.lora_config.r,
+                            d_lora=d_lora,
+                        )
                 else:
                     # each module processes d -> r d_out independently
                     if self.config.per_layer_processing:

src/ctx_to_lora/training_utils.py

@@ -114,7 +114,7 @@ def train_model(
         logger.info(f"Resuming from the checkpoint: {checkpoint}")
 
     is_modulated_model = isinstance(model, ModulatedPretrainedModel)
-    trainer_cls = Trainer if not is_modulated_model else ModulatedModelTrainer
+    trainer_cls = Trainer  # if not is_modulated_model else ModulatedModelTrainer
     trainer_kwargs = dict(
         model=model,
         args=training_args,
@@ -123,10 +123,10 @@ def train_model(
         data_collator=train_collator,
         compute_metrics=compute_metrics,
     )
-    if is_modulated_model:
-        logger.info(f"Training with modulated model. Using CustomTrainer.")
-        trainer_kwargs["gen_lora_l1_reg_coef"] = training_args.gen_lora_l1_reg_coef
-        del training_args.gen_lora_l1_reg_coef
+    # if is_modulated_model:
+    #     logger.info(f"Training with modulated model. Using CustomTrainer.")
+    #     trainer_kwargs["gen_lora_l1_reg_coef"] = training_args.gen_lora_l1_reg_coef
+    #     del training_args.gen_lora_l1_reg_coef
 
     trainer = trainer_cls(**trainer_kwargs)