fusion-bench 0.2.11__py3-none-any.whl → 0.2.13__py3-none-any.whl

fusion_bench_config/method/lm_finetune/fullfinetune_sft.yaml

@@ -1,12 +1,10 @@
 _target_: fusion_bench.method.FullFinetuneSFT
 _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
@@ -14,13 +12,11 @@ lr_scheduler:
   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

fusion_bench_config/method/lm_finetune/peftfinetune_sft.yaml

@@ -1,23 +1,19 @@
 _target_: fusion_bench.method.PeftFinetuneSFT
 _recursive_: False
-
 optimizer:
   _target_: torch.optim.AdamW
   lr: 1e-4
   weight_decay: 0.01
   fused: null
-
 lr_scheduler:
   _target_: torch.optim.lr_scheduler.CosineAnnealingLR
   T_max: _T_max_ # this will be replaced by the expected number of training steps
   eta_min: 1e-6
-
 dataloader_kwargs:
   # per-gpu batch size
   batch_size: 1
   num_workers: 0
   pin_memory: True
-
 peft_config:
   _target_: peft.LoraConfig
   task_type: peft.TaskType.CAUSAL_LM
@@ -33,11 +29,9 @@ peft_config:
   lora_alpha: 16
   lora_dropout: 0
   bias: none
-
 adapter_name: default
 # whether to merge and unload the adapter after training
 merge_and_unload: false
-
 # Training hyperparameters
 # if max_epochs=-1, max_steps will be used to determine the number of training steps
 max_epochs: 3

fusion_bench_config/method/mixtral_moe_upscaling.yaml

@@ -1,6 +1,5 @@
 # or fusion_bench.method.MixtralUpscalingAlgorithm
-_target_: fusion_bench.method.MixtralForCausalLMUpscalingAlgorithm 
-
+_target_: fusion_bench.method.MixtralForCausalLMUpscalingAlgorithm
 num_experts: 4
 experts_per_token: 2
 # path to save the upscaled model

fusion_bench_config/method/model_recombination.yaml

@@ -1,4 +1,3 @@
 _target_: fusion_bench.method.ModelRecombinationAlgorithm
-
 # if `return_model_pool` is not null, the argument `return_modelpool` passed to the `run` method will be ignored.
 return_modelpool: null

fusion_bench_config/method/opcm/opcm.yaml

@@ -1,5 +1,4 @@
 _target_: fusion_bench.method.opcm.opcm.OPCMForCLIP
-
 # shuffle the order of the models
 shuffle_order: true
 # the scaling factor for the SVD projection

fusion_bench_config/method/opcm/task_arithmetic.yaml

@@ -1,7 +1,5 @@
 _target_: fusion_bench.method.opcm.task_arithmetic.ContinualTaskArithmeticForCLIP
-
 scaling_factor: 0.3
-
 # shuffle the order of the models
 shuffle_order: true
 # the random seed to use

fusion_bench_config/method/opcm/ties_merging.yaml

@@ -1,5 +1,4 @@
 _target_: fusion_bench.method.opcm.ties_merging.ContinualTiesMergingForCLIP
-
 # Scaling factor $\lambda$
 scaling_factor: 0.5
 threshold: 20
@@ -7,7 +6,6 @@ threshold: 20
 remove_keys: []
 # Function to merge the models, default is sum. Options are 'sum', 'mean', and 'max'
 merge_func: sum
-
 # shuffle the order of the models
 shuffle_order: true
 # the random seed to use

fusion_bench_config/method/opcm/weight_average.yaml

@@ -1,5 +1,4 @@
 _target_: fusion_bench.method.opcm.weight_average.ContinualWeightAverageForCLIP
-
 # shuffle the order of the models
 shuffle_order: true
 # the random seed to use

fusion_bench_config/method/pwe_moe/epo_for_openclip.yaml

@@ -0,0 +1,30 @@
+_target_: fusion_bench.method.pwe_moe.openclip_pwe_moe.PWEMoEExactParetoOptimalForOpenCLIP
+#! === Model Architecture Arguments ===
+# if true, then we only apply the weight ensembling MoE to MLPs, else, we apply it to all layers
+partial: true
+# weight-ensembling MoE arguments
+# initial outputs for the routing gates and the merging weights for the remaining layers
+init_lambda: 0.3
+# number of hidden layers in the routing gate
+router_hidden_layers: 2
+# path to the checkpoint file, if not provided, then the training is performed
+checkpoint_path: null
+#! === Training Arguments ===
+# if false, the training is skipped
+run_train: true
+num_steps: 2000
+save_interval: 1000
+# learning rate
+lr: 1e-2
+alpha: 1 # alpha for dirichlet, if alpha=1, then it is uniform
+# dataloader arguments
+dataloader_kwargs:
+  # per-device batch size
+  batch_size: 16
+  num_workers: 0
+#! === Evaluation Arguments ===
+# if false, the evaluation is skipped
+run_eval: false
+# if true, then we only evaluate the model on the first 20 batches of the test dataset
+quick_evaluation: false
+num_evaluation_samples: equal_weight

fusion_bench_config/method/pwe_moe/ls_for_openclip.yaml

@@ -0,0 +1,30 @@
+_target_: fusion_bench.method.pwe_moe.openclip_pwe_moe.PWEMoELinearScalarizationForOpenCLIP
+#! === Model Architecture Arguments ===
+# if true, then we only apply the weight ensembling MoE to MLPs, else, we apply it to all layers
+partial: true
+# weight-ensembling MoE arguments
+# initial outputs for the routing gates and the merging weights for the remaining layers
+init_lambda: 0.3
+# number of hidden layers in the routing gate
+router_hidden_layers: 2
+# path to the checkpoint file, if not provided, then the training is performed
+checkpoint_path: null
+#! === Training Arguments ===
+# if false, the training is skipped
+run_train: true
+num_steps: 2000
+save_interval: 1000
+# learning rate
+lr: 1e-2
+alpha: 1 # alpha for dirichlet, if alpha=1, then it is uniform
+# dataloader arguments
+dataloader_kwargs:
+  # per-device batch size
+  batch_size: 16
+  num_workers: 0
+#! === Evaluation Arguments ===
+# if false, the evaluation is skipped
+run_eval: false
+# if true, then we only evaluate the model on the first 20 batches of the test dataset
+quick_evaluation: false
+num_evaluation_samples: equal_weight

fusion_bench_config/method/{pwe_moe_ls_for_clip.yaml → pwe_moe/pwe_moe_ls_for_clip.yaml}

@@ -1,22 +1,23 @@
 _target_: fusion_bench.method.PWEMoELinearScalarizationForCLIP # or PWEMoExactParetoOptimalForCLIP
+#! === Model Architecture Arguments ===
 upscale_mlp: true
 upscale_attn: true
 # scaling factor for the remaining parameters
 init_lambda: 0.3
 router_hidden_layers: 2
+#! === Training Arguments ===
 lr: 1e-5
 num_steps: 8000
 save_interval: 2000
 alpha: 1 # alpha for dirichlet, if alpha=1, then it is uniform
 # load model from this checkpoint
+dataloader_kwargs:
+  # per-device batch size
+  batch_size: 16
+  num_workers: 4
 checkpoint_path: null
-
+#! === Evaluation Arguments ===
 # evaluation grid
 eval_grid: true
 eval_grid_n: 8
 eval_grid_m: 2
-
-dataloader_kwargs:
-  # per-device batch size
-  batch_size: 16
-  num_workers: 4

fusion_bench_config/method/rankone_moe/rankone_moe.yaml

@@ -6,12 +6,10 @@ save_checkpoint: False
 router_hidden_layers: 1
 init_lambda: 0.3
 batch_reduce: true
-
 # device to compute svd
 svd_accelerator: cuda
 rank_k: 32 # How many experts are added to the pool per task?
-select_k: -1  # How many experts are selected from the pool to merge? Range is (1, rank_k*task_num). In particular -1: All the experts in the pool
-
+select_k: -1 # How many experts are selected from the pool to merge? Range is (1, rank_k*task_num). In particular -1: All the experts in the pool
 # learning rate
 lr: 1e-4
 optimizer: adam

fusion_bench_config/method/regmean/gpt2_regmean.yaml

@@ -1,5 +1,4 @@
 _target_: fusion_bench.method.RegMeanAlgorithmForGPT2
-
 # list, regular expression of names of parameters that need to be excluded
 exclude_param_names_regex: []
 # numbers of examples to compute regmean weights

fusion_bench_config/method/slerp/slerp.yaml

@@ -1,6 +1,4 @@
 _target_: fusion_bench.method.SlerpMergeAlgorithm
-
 t: 0.5 # interpolation factor
-
 DOT_THRESHOLD: 0.9995
 epsilon: 1e-8

fusion_bench_config/method/sparselo_pruning/llama_iterative_sparselo.yaml

@@ -17,4 +17,4 @@ sparsity_ratio: 0.5
 n: 2
 m: 4
 # string to specify the path to where the pruned model is saved
-model_save_path: null
+model_save_path: null

fusion_bench_config/method/sparselo_pruning/llama_pcp_sparselo.yaml

@@ -17,4 +17,4 @@ sparsity_ratio: 0.5
 n: 2
 m: 4
 # string to specify the path to where the pruned model is saved
-model_save_path: null
+model_save_path: null

fusion_bench_config/method/sparselo_pruning/llama_sparselo.yaml

@@ -16,4 +16,4 @@ sparsity_ratio: 0.5
 n: 2
 m: 4
 # string to specify the path to where the pruned model is saved
-model_save_path: null
+model_save_path: null

fusion_bench_config/method/surgery/adamerging_surgery.yaml

@@ -21,7 +21,6 @@ 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
+surgery_steps: 1000

fusion_bench_config/method/task_arithmetic.yaml

@@ -1,2 +1,2 @@
 _target_: fusion_bench.method.TaskArithmeticAlgorithm
-scaling_factor: 0.5
+scaling_factor: 0.3

fusion_bench_config/method/task_singular_vector/TaskSingularVectorMerging.yaml

@@ -1,6 +1,5 @@
 _target_: fusion_bench.method.TaskSingularVectorMerging
 remove_keys: null
-
 # alpha is a float or a list of floats
 # example:
 # alpha: 1

fusion_bench_config/method/ties_merging.yaml

@@ -1,6 +1,6 @@
 _target_: fusion_bench.method.TiesMergingAlgorithm
 # Scaling factor $\lambda$
-scaling_factor: 0.5
+scaling_factor: 0.3
 threshold: 20
 # List of keys to remove from the state dict, default is empty
 remove_keys: []

fusion_bench_config/method/trust_region/clip_task_arithmetic.yaml

@@ -1,5 +1,4 @@
 _target_: fusion_bench.method.trust_region.clip_task_arithmetic.TaskArithmeticWithTrustRegionForCLIP
-
 scaling_factor: 0.3
 threshold_quantile: 0.99
 max_samples: 128

fusion_bench_config/method/wemoe/sparse_weight_ensembling_moe.yaml

@@ -1,32 +1,25 @@
 name: ??? # this can be sparse_clip_weight_ensembling_moe
-
 # the path for loading the model weights, if specified, skip the test-time adaptation training
 #checkpoint: /home/enneng/fusion_bench/outputs/sparse_we_moe/shared_gate/1routerlayer_clip-vit-base-patch32_TA8_sparse_weight_ensembling_moe_checkpoint_0_0.ckpt
 checkpoint: False
 # the path for saving the model weights.
 save_checkpoint: False
-
 # router
 router_hidden_layers: 2
 init_lambda: 0.3
 batch_reduce: true
-
 # sparse task vectors
 tv_prune_ratio: 0.9
-
 # sparse gate module
 post_sparse_gate: False
 gate_prune_ratio: 0.0
-
 # shared gate
 shared_gate: true
 position_encoding: false
 position_encoding_dim: 8
-
 # tta learning rate
 lr: 1e-4
 optimizer: adam
-
 # this is overrided by `fabric.devices` if launched from the `fusion_bench` CLI.
 devices: 1
 batch_size: 16
@@ -34,6 +27,5 @@ num_workers: 16
 max_steps: 1000
 # if true, we will use the gradient accumulation across tasks to save memory
 use_grad_accumulate: false
-
 cache_dir: outputs
 fast_dev_run: ${fast_dev_run}

fusion_bench_config/model/clip-vit/clip-vit-base-patch16_cifar10.yaml

@@ -1 +1 @@
-cifar10: tanganke/clip-vit-base-patch16_cifar10
+cifar10: tanganke/clip-vit-base-patch16_cifar10

fusion_bench_config/model/clip-vit/clip-vit-large-patch14.yaml

@@ -1 +1 @@
-_pretrained_: openai/clip-vit-large-patch14
+_pretrained_: openai/clip-vit-large-patch14

fusion_bench_config/model/clip-vit/clip-vit-large-patch14_oxford-iiit-pet.yaml

@@ -1 +1 @@
-oxford-iiit-pet: tanganke/clip-vit-large-patch14_oxford-iiit-pet
+oxford-iiit-pet: tanganke/clip-vit-large-patch14_oxford-iiit-pet

fusion_bench_config/model/clip-vit/clip-vit-large-patch14_oxford_flowers102.yaml

@@ -1 +1 @@
-oxford_flowers102: tanganke/clip-vit-large-patch14_oxford_flowers102
+oxford_flowers102: tanganke/clip-vit-large-patch14_oxford_flowers102

fusion_bench_config/model/clip-vit/clip-vit-large-patch14_pcam.yaml

@@ -1 +1 @@
-pcam: tanganke/clip-vit-large-patch14_pcam
+pcam: tanganke/clip-vit-large-patch14_pcam

fusion_bench_config/model/clip-vit/clip-vit-large-patch14_rendered-sst2.yaml

@@ -1 +1 @@
-rendered-sst2: tanganke/clip-vit-large-patch14_rendered-sst2
+rendered-sst2: tanganke/clip-vit-large-patch14_rendered-sst2

fusion_bench_config/model/clip-vit/clip-vit-large-patch14_resisc45.yaml

@@ -1 +1 @@
-resisc45: tanganke/clip-vit-large-patch14_resisc45
+resisc45: tanganke/clip-vit-large-patch14_resisc45

fusion_bench_config/model/clip-vit/clip-vit-large-patch14_stanford-cars.yaml

@@ -1 +1 @@
-stanford-cars: tanganke/clip-vit-large-patch14_stanford-cars
+stanford-cars: tanganke/clip-vit-large-patch14_stanford-cars

fusion_bench_config/model/clip-vit/clip-vit-large-patch14_stl10.yaml

@@ -1 +1 @@
-stl10: tanganke/clip-vit-large-patch14_stl10
+stl10: tanganke/clip-vit-large-patch14_stl10

fusion_bench_config/model/clip-vit/clip-vit-large-patch14_sun397.yaml

@@ -1 +1 @@
-sun397: tanganke/clip-vit-large-patch14_sun397
+sun397: tanganke/clip-vit-large-patch14_sun397

fusion_bench_config/model/clip-vit/clip-vit-large-patch14_svhn.yaml

@@ -1 +1 @@
-svhn: tanganke/clip-vit-large-patch14_svhn
+svhn: tanganke/clip-vit-large-patch14_svhn

fusion_bench_config/modelpool/CLIPVisionModelPool/clip-vit-base-patch16_TA8_lora.yaml

@@ -1,5 +1,4 @@
 _target_: fusion_bench.modelpool.CLIPVisionModelPool
-
 models:
   _pretrained_:
     _target_: fusion_bench.models.linearized.vision_model.load_fft_vision_model_hf
@@ -44,10 +43,8 @@ models:
     base_model_name: openai/clip-vit-base-patch16
     peft_name: tanganke/clip-vit-base-patch16_dtd_lora-16
     merge_and_unload: true
-
 processor:
   _target_: transformers.CLIPProcessor.from_pretrained
   pretrained_model_name_or_path: openai/clip-vit-base-patch16
-
 train_datasets: null
 test_datasets: null

fusion_bench_config/modelpool/CLIPVisionModelPool/clip-vit-base-patch16_individual.yaml

@@ -6,14 +6,11 @@
 #   ...
 defaults:
   - CLIPVisionModelPool@: _template
-
 models:
   _pretrained_:
     _target_: transformers.CLIPVisionModel.from_pretrained
     pretrained_model_name_or_path: ${...base_model}
-
 processor:
   _target_: transformers.CLIPProcessor.from_pretrained
   pretrained_model_name_or_path: ${..base_model}
-
 base_model: openai/clip-vit-base-patch16

fusion_bench_config/modelpool/CLIPVisionModelPool/clip-vit-base-patch16_individual_lora.yaml

@@ -1,14 +1,11 @@
 _target_: fusion_bench.modelpool.CLIPVisionModelPool
-
 models:
   sun397:
     _target_: fusion_bench.models.linearized.vision_model.load_lora_vision_model_hf
     base_model_name: openai/clip-vit-base-patch16
     peft_name: tanganke/clip-vit-base-patch16_sun397_lora-16
-
 processor:
   _target_: transformers.CLIPProcessor.from_pretrained
   pretrained_model_name_or_path: openai/clip-vit-base-patch16
-
 train_datasets: null
 test_datasets: null

fusion_bench_config/modelpool/CLIPVisionModelPool/clip-vit-base-patch32_TA8_control_task.yaml

@@ -12,13 +12,10 @@ defaults:
       - clip-vit-base-patch32_dtd
   - /dataset/image_classification/train@train_datasets:
       - tiny-imagenet
-
 _target_: fusion_bench.modelpool.CLIPVisionModelPool
 _recursive_: false
-
 models: ???
 train_datasets: ???
-
 processor:
   _target_: transformers.CLIPProcessor.from_pretrained
   pretrained_model_name_or_path: openai/clip-vit-base-patch32

fusion_bench_config/modelpool/CLIPVisionModelPool/clip-vit-base-patch32_individual.yaml

@@ -1,13 +1,10 @@
 defaults:
   - CLIPVisionModelPool@: _template
-
 models:
   _pretrained_:
     _target_: transformers.CLIPVisionModel.from_pretrained
     pretrained_model_name_or_path: ${...base_model}
-
 processor:
   _target_: transformers.CLIPProcessor.from_pretrained
   pretrained_model_name_or_path: ${..base_model}
-
 base_model: openai/clip-vit-base-patch32

fusion_bench_config/modelpool/CLIPVisionModelPool/clip-vit-base-patch32_single_task_projection.yaml

@@ -3,13 +3,10 @@ defaults:
       - clip-vit-base-patch32
       - clip-vit-base-patch32_sun397
       - clip-vit-base-patch32_stanford-cars
-
 _target_: fusion_bench.modelpool.CLIPVisionModelPool
 _recursive_: false
-
 train_datasets: null
 test_datasets: null
-
 processor:
   _target_: transformers.CLIPProcessor.from_pretrained
   pretrained_model_name_or_path: openai/clip-vit-base-patch32

fusion_bench_config/modelpool/CLIPVisionModelPool/clip-vit-base-patch32_two_tasks_control_task.yaml

@@ -2,17 +2,13 @@ defaults:
   - _self_
   - /dataset/image_classification/train@train_datasets:
       - tiny-imagenet
-
 _target_: fusion_bench.modelpool.CLIPVisionModelPool
 _recursive_: false
-
 models:
   _pretrained_: openai/clip-vit-base-patch32
   model_1: tanganke/clip-vit-base-patch32_sun397
   model_2: tanganke/clip-vit-base-patch32_stanford-cars
-  
 train_datasets: ???
-
 processor:
   _target_: transformers.CLIPProcessor.from_pretrained
   pretrained_model_name_or_path: openai/clip-vit-base-patch32

fusion_bench_config/modelpool/CLIPVisionModelPool/clip-vit-large-patch14_individual.yaml

@@ -6,14 +6,11 @@
 #   ...
 defaults:
   - CLIPVisionModelPool@: _template
-
 models:
   _pretrained_:
     _target_: transformers.CLIPVisionModel.from_pretrained
     pretrained_model_name_or_path: ${...base_model}
-
 processor:
   _target_: transformers.CLIPProcessor.from_pretrained
   pretrained_model_name_or_path: ${..base_model}
-
 base_model: openai/clip-vit-large-patch14

fusion_bench_config/modelpool/CausalLMPool/llama_alpaca_cleaned.yaml

@@ -1,17 +1,13 @@
 _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

fusion_bench_config/modelpool/CausalLMPool/llama_codealpaca.yaml

@@ -1,17 +1,13 @@
 _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

fusion_bench_config/modelpool/CausalLMPool/llama_for_causallm.yaml

@@ -12,7 +12,6 @@ models:
   expert_2:
     _target_: transformers.LlamaForCausalLM.from_pretrained
     pretrained_model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
-
 model_kwargs:
   torch_dtype: float16
 tokenizer:

fusion_bench_config/modelpool/CausalLMPool/llama_metamathqa.yaml

@@ -1,17 +1,13 @@
 _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

fusion_bench_config/modelpool/CausalLMPool/llama_ultrachat.yaml

@@ -1,17 +1,13 @@
 _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

fusion_bench_config/modelpool/CausalLMPool/simle_mixtral_exp_v4.yaml

@@ -13,7 +13,6 @@ models:
   expert_3:
     _target_: transformers.AutoModelForCausalLM.from_pretrained
     pretrained_model_name_or_path: uukuguy/speechless-code-mistral-7b-v1.0
-
 model_kwargs:
   torch_dtype: float16
 tokenizer:

fusion_bench_config/modelpool/CausalLMPool/single_llama_model.yaml

@@ -6,12 +6,9 @@ models:
   _pretrained_:
     _target_: transformers.LlamaForCausalLM.from_pretrained
     pretrained_model_name_or_path: ${...base_model}
-
 model_kwargs:
   torch_dtype: float16
-
 tokenizer:
   _target_: transformers.AutoTokenizer.from_pretrained
   pretrained_model_name_or_path: ${..base_model}
-
 base_model: decapoda-research/llama-7b-hf