fusion-bench 0.2.6__py3-none-any.whl → 0.2.7__py3-none-any.whl

fusion_bench_config/llama_full_finetune.yaml

@@ -0,0 +1,19 @@
+defaults:
+  - hydra: default
+  - fabric: llama_fsdp
+  # --- Model, Method, Task ---
+  - method: lm_finetune/fullfinetune_sft.yaml
+  - modelpool: CausalLMPool/llama_alpaca_cleaned.yaml
+  - taskpool: dummy
+  - _self_
+
+_target_: fusion_bench.programs.FabricModelFusionProgram
+_recursive_: false
+
+fast_dev_run: false # Run a single batch of data to test the model or method
+# Run the script without actually running the experiment, use with `print_config=true`.
+# You can also use `--cfg` or `-c` to show the configuration instead of running.
+dry_run: false
+print_config: true # Print the configuration to the console
+report_save_path: null # path to save the result report
+print_function_call: true # set to false if you don't want to print the details of instantiate calls

fusion_bench_config/method/lm_finetune/bradley_terry_rm.yaml

@@ -0,0 +1,47 @@
+_target_: fusion_bench.method.BradleyTerryRewardModeling
+_recursive_: False
+
+optimizer:
+  _target_: torch.optim.AdamW
+  lr: 1e-5
+  weight_decay: 0.01
+  fused: null
+
+lr_scheduler:
+  _target_: fusion_bench.optim.lr_scheduler.CosineDecayWithWarmup
+  T_max: _T_max_ # this will be replaced by the expected number of training steps
+  init_lr: 0
+  warmup_steps: 100
+  max_lr: ${..optimizer.lr}
+  min_lr: 1e-6
+
+dataloader_kwargs:
+  # per-gpu batch size
+  batch_size: 1
+  num_workers: 0
+  pin_memory: True
+
+# Training hyperparameters
+# if max_epochs=-1, max_steps will be used to determine the number of training steps
+max_epochs: 3
+max_steps: -1
+max_steps_per_epoch: -1
+accumulate_grad_batches: 1
+lr_scheduler_interval: step
+lr_scheduler_frequency: 1
+# Checkpointing may be done by epoch or step, and at the end of training
+# `checkpoint_save_interval` can be 'epoch' or 'step'
+checkpoint_save_interval: epoch
+checkpoint_save_frequency: 1
+# Whether to use gradient clipping, and if so, the value and algorithm
+gradient_clip_val: null
+gradient_clip_algorithm: norm
+save_optimizer_state: false
+# save_full_model must be true when using shared FSDP
+save_full_model: true
+# save_ckpt_type can be 'hf' or 'lightning'
+save_ckpt_type: lightning
+# Path to checkpoint to load from, used for resuming training
+ckpt_path: null
+max_length: 4096
+fix_token_embedding: true

fusion_bench_config/method/lm_finetune/fullfinetune_sft.yaml

@@ -3,14 +3,17 @@ _recursive_: False
 
 optimizer:
   _target_: torch.optim.AdamW
-  fused: True
+  lr: 1e-5
   weight_decay: 0.01
-  lr: 5e-5
+  fused: null
 
 lr_scheduler:
-  _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+  _target_: fusion_bench.optim.lr_scheduler.CosineDecayWithWarmup
   T_max: _T_max_ # this will be replaced by the expected number of training steps
-  eta_min: 1e-6
+  init_lr: 0
+  warmup_steps: 100
+  max_lr: ${..optimizer.lr}
+  min_lr: 1e-6
 
 dataloader_kwargs:
   # per-gpu batch size
@@ -36,5 +39,9 @@ gradient_clip_algorithm: norm
 save_optimizer_state: false
 # save_full_model must be true when using shared FSDP
 save_full_model: true
+# save_ckpt_type can be 'hf' or 'lightning'
+save_ckpt_type: lightning
 # Path to checkpoint to load from, used for resuming training
 ckpt_path: null
+max_length: 4096
+fix_token_embedding: true

fusion_bench_config/method/lm_finetune/peftfinetune_sft.yaml

@@ -3,9 +3,9 @@ _recursive_: False
 
 optimizer:
   _target_: torch.optim.AdamW
-  fused: True
+  lr: 1e-4
   weight_decay: 0.01
-  lr: 5e-5
+  fused: null
 
 lr_scheduler:
   _target_: torch.optim.lr_scheduler.CosineAnnealingLR
@@ -56,6 +56,8 @@ gradient_clip_algorithm: norm
 save_optimizer_state: false
 # save_full_model must be true when using shared FSDP
 save_full_model: false
+# save_ckpt_type can be 'peft' or 'lightning'
+save_ckpt_type: lightning
 # Path to checkpoint to load from, used for resuming training
 ckpt_path: null
 max_length: 4096

fusion_bench_config/method/surgery/adamerging_surgery.yaml

@@ -0,0 +1,27 @@
+# this option can be "clip_task_wise_adamerging"
+name: clip_layer_wise_adamerging_surgery
+# this weights can be a list of float, or a string that points to a *.np, *.pt file containing the weights
+# if weights is specified, skip the test-time adaptation training
+weights: null
+# learning rate
+optimizer: adam
+lr: 1e-3
+init_values: 0.3
+# if `clamp_weights` is true, the weights will be clamped to [0, 1]
+clamp_weights: false
+# arguments of `functional_call`
+tie_weights: true
+strict: false
+# this is overrided by `fabric.devices` if launched from the `fusion_bench` CLI.
+devices: 1
+batch_size: 16
+num_workers: 8
+max_steps: 1000
+fast_dev_run: ${fast_dev_run}
+# the path for saving the merging weights
+save_merging_weights: 'merging_weights.pt'
+cache_dir: outputs
+
+# parameters of Surgery
+eval_iterations: 200
+surgery_steps: 1000

fusion_bench_config/modelpool/CausalLMPool/llama_alpaca_cleaned.yaml

@@ -0,0 +1,21 @@
+_target_: fusion_bench.modelpool.CausalLMPool
+
+pretrained_model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+
+models:
+  _pretrained_:
+    _target_: transformers.AutoModelForCausalLM.from_pretrained
+    pretrained_model_name_or_path: ${...pretrained_model_name_or_path}
+    torch_dtype: bfloat16
+
+tokenizer:
+  _target_: transformers.AutoTokenizer.from_pretrained
+  pretrained_model_name_or_path: ${..pretrained_model_name_or_path}
+
+train_datasets:
+  alpaca-cleaned:
+    _target_: fusion_bench.dataset.llama.alpaca.load_tokenized_alpaca_dataset
+    tokenizer: ${...tokenizer}
+    path: "yahma/alpaca-cleaned"
+    split: train
+    cache_path: null

fusion_bench_config/modelpool/CausalLMPool/llama_codealpaca.yaml

@@ -0,0 +1,21 @@
+_target_: fusion_bench.modelpool.CausalLMPool
+
+pretrained_model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+
+models:
+  _pretrained_:
+    _target_: transformers.AutoModelForCausalLM.from_pretrained
+    pretrained_model_name_or_path: ${...pretrained_model_name_or_path}
+    torch_dtype: bfloat16
+
+tokenizer:
+  _target_: transformers.AutoTokenizer.from_pretrained
+  pretrained_model_name_or_path: ${..pretrained_model_name_or_path}
+
+train_datasets:
+  codealpaca:
+    _target_: fusion_bench.dataset.llama.alpaca.load_tokenized_alpaca_dataset
+    tokenizer: ${...tokenizer}
+    path: sahil2801/CodeAlpaca-20k
+    split: train
+    cache_path: null

fusion_bench_config/modelpool/CausalLMPool/llama_metamathqa.yaml

@@ -0,0 +1,19 @@
+_target_: fusion_bench.modelpool.CausalLMPool
+
+pretrained_model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+
+models:
+  _pretrained_:
+    _target_: transformers.AutoModelForCausalLM.from_pretrained
+    pretrained_model_name_or_path: ${...pretrained_model_name_or_path}
+    torch_dtype: bfloat16
+
+tokenizer:
+  _target_: transformers.AutoTokenizer.from_pretrained
+  pretrained_model_name_or_path: ${..pretrained_model_name_or_path}
+
+train_datasets:
+  metamathqa:
+    _target_: fusion_bench.dataset.llama.metamathqa.load_tokenized_metamathqa
+    tokenizer: ${...tokenizer}
+    cache_path: null

fusion_bench_config/modelpool/CausalLMPool/llama_ultrachat.yaml

@@ -0,0 +1,18 @@
+_target_: fusion_bench.modelpool.CausalLMPool
+
+pretrained_model_name_or_path: meta-llama/Llama-3-1B-Instruct
+
+models:
+  _pretrained_:
+    _target_: transformers.AutoModelForCausalLM.from_pretrained
+    pretrained_model_name_or_path: ${...pretrained_model_name_or_path}
+    torch_dtype: bfloat16
+
+tokenizer:
+  _target_: transformers.AutoTokenizer.from_pretrained
+  pretrained_model_name_or_path: ${..pretrained_model_name_or_path}
+
+train_datasets:
+  ultrachat-200k:
+    _target_: fusion_bench.dataset.llama.ultrachat.load_tokenized_ultrachat_200k
+    tokenizer: ${...tokenizer}

fusion_bench_config/modelpool/SeqenceClassificationModelPool/llama_preference700k.yaml

@@ -0,0 +1,23 @@
+_target_: fusion_bench.modelpool.SeqenceClassificationModelPool
+
+pretrained_model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+
+models:
+  _pretrained_:
+    _target_: fusion_bench.modelpool.seq_classification_lm.create_reward_model_from_pretrained
+    pretrained_model_name_or_path: ${...pretrained_model_name_or_path}
+    torch_dtype: bfloat16
+    use_flash_attention_2: true
+
+tokenizer:
+  _target_: transformers.AutoTokenizer.from_pretrained
+  pretrained_model_name_or_path: ${..pretrained_model_name_or_path}
+  pad_token: <|end_of_text|> # do not use eos token (<|eos_id|>) as padding token because it is used as the end of each content
+
+train_datasets:
+  preference_700k:
+    _target_: fusion_bench.dataset.llama.preference_700k.load_tokenized_preference_700k_for_rlhf
+    tokenizer: ${...tokenizer}
+    path: hendrydong/preference_700K
+    split: train
+    cache_path: null

fusion_bench_config/modelpool/SeqenceClassificationModelPool/single_reward_model.yaml

@@ -0,0 +1,14 @@
+_target_: fusion_bench.modelpool.SeqenceClassificationModelPool
+
+pretrained_model_name_or_path: fusion-bench/Llama-3.2-1B-Instruct_Bradly-Terry-RM_Preference-700k
+
+models:
+  _pretrained_:
+    _target_: transformers.AutoModelForSequenceClassification.from_pretrained
+    pretrained_model_name_or_path: ${...pretrained_model_name_or_path}
+    torch_dtype: bfloat16
+
+tokenizer:
+  _target_: transformers.AutoTokenizer.from_pretrained
+  pretrained_model_name_or_path: ${..pretrained_model_name_or_path}
+  pad_token: <|end_of_text|> # do not use eos token (<|eos_id|>) as padding token because it is used as the end of each content

fusion_bench_config/nyuv2_config.yaml

@@ -1,13 +1,17 @@
 defaults:
   - hydra: default
+  - fabric: auto
   - modelpool: nyuv2_modelpool
   - method: simple_average
   - taskpool: nyuv2_taskpool
   - _self_
+
+_target_: fusion_bench.programs.FabricModelFusionProgram
+_recursive_: false
+
 fast_dev_run: false # Run a single batch of data to test the model or method
 use_lightning: true # Use the fabric to run the experiment
 print_config: true # Print the configuration to the console
 save_report: false # path to save the result report
-fabric: null
 trainer:
   devices: 1

fusion_bench_config/taskpool/reward_model_evaluation.yaml

@@ -0,0 +1,18 @@
+_target_: fusion_bench.taskpool.llama.reward_model.RewardModelEvaluationTaskPool
+
+test_datasets:
+  preference_700k:
+    _target_: fusion_bench.dataset.llama.preference_700k.load_tokenized_preference_700k_for_rlhf
+    tokenizer: ${...tokenizer}
+    path: hendrydong/preference_700K
+    split: train
+    cache_path: null
+
+dataloader_kwargs:
+  shuffle: False
+  batch_size: 16
+
+tokenizer: ${..modelpool.tokenizer}
+
+max_num_samples: 1000
+seed: 42

fusion_bench_config/llama_weighted_average.yaml

@@ -1,26 +0,0 @@
-defaults:
-  - example_config
-  - override method: weighted_average_for_llama
-  - override modelpool: llama_for_causallm
-  - _self_
-modelpool:
-  models:
-    # the pre-trained model (base model) is optional
-    # if not provided, the first model will be used as the base model
-    - name: _pretrained_
-      path: meta-llama/Meta-Llama-3-8B
-    - name: expert_1
-      path: meta-llama/Meta-Llama-3-8B
-    - name: expert_2
-      path: meta-llama/Meta-Llama-3-8B-Instruct
-method:
-  normalize: true # if true, the weights will be normalized before merging
-  weights: # List of weights for each model
-    - 0.5
-    - 0.5
-  # if true, only the backbone of the model will be merged and the head will be keeped as the pre-trained model (if the pre-trained model is provided, otherwise the head of the first model will be used)
-  # if false, the whole model will be merged
-  backbone_only: true
-  merged_model_save_path: null
-  save_tokenizer: true
-  push_to_hub: false