move unmaintained examples to archive (#2903) [skip ci]

examples/archived/README.md

@@ -0,0 +1,5 @@
+# Archived Examples
+
+This directory contains examples that are no longer maintained and may no longer be functional.
+
+We keep them around for archival purposes in case they are useful to others.

examples/archived/cerebras/btlm-ft.yml

@@ -0,0 +1,82 @@
+base_model: cerebras/btlm-3b-8k-base
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: GPT2Tokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+tokenizer_use_fast: true
+tokenizer_legacy: true
+push_dataset_to_hub:
+hf_use_auth_token: true
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_prepared_run
+val_set_size: 0.05
+
+adapter:
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len:
+sample_packing: false
+sample_packing_eff_est:
+sample_packing_seq_len_multiplier:
+total_num_tokens:
+
+lora_r:
+lora_alpha:
+lora_dropout:
+lora_target_modules:
+lora_target_linear:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+output_dir: ./outputs/btlm-out
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_fused
+adam_beta2: 0.95
+adam_eps: 0.000000001
+max_grad_norm: 1.0
+
+torchdistx_path:
+lr_scheduler: cosine
+lr_quadratic_warmup: true
+learning_rate: 0.000085
+train_on_inputs: true
+group_by_length: false
+bf16: auto
+tf32: true
+
+gradient_checkpointing: false
+resume_from_checkpoint:
+logging_steps: 1
+
+flash_attention: true
+sdp_attention:
+flash_optimum:
+
+gptq_groupsize:
+gptq_model_v1:
+
+warmup_steps: 32
+evals_per_epoch: 4
+saves_per_epoch: 1
+save_total_limit:
+
+weight_decay: 0.1
+special_tokens:
+  pad_token: "<|endoftext|>"
+fsdp:
+#  - full_shard
+#  - auto_wrap
+fsdp_config:
+#  fsdp_state_dict_type: FULL_STATE_DICT
+#  fsdp_transformer_layer_cls_to_wrap: BTLMBlock

examples/archived/cerebras/qlora.yml

@@ -0,0 +1,51 @@
+base_model: cerebras/Cerebras-GPT-1.3B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+push_dataset_to_hub:
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+adapter: qlora
+lora_model_dir:
+sequence_len: 2048
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - c_fc
+  - c_attn
+  - c_proj
+lora_target_linear:
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/qlora-out
+batch_size: 4
+micro_batch_size: 4
+num_epochs: 2
+optimizer: paged_adamw_8bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.0002
+bf16: auto
+tf32: true
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+xformers_attention: true
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.1
+special_tokens:
+  pad_token: "<|endoftext|>"

examples/archived/code-llama/13b/lora.yml

@@ -0,0 +1,57 @@
+base_model: codellama/CodeLlama-13b-hf
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/code-llama/13b/qlora.yml

@@ -0,0 +1,58 @@
+base_model: codellama/CodeLlama-13b-hf
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/code-llama/34b/lora.yml

@@ -0,0 +1,57 @@
+base_model: codellama/CodeLlama-34b-hf
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/code-llama/34b/qlora.yml

@@ -0,0 +1,58 @@
+base_model: codellama/CodeLlama-34b-hf
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/code-llama/7b/lora.yml

@@ -0,0 +1,57 @@
+base_model: codellama/CodeLlama-7b-hf
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/code-llama/7b/qlora.yml

@@ -0,0 +1,58 @@
+base_model: codellama/CodeLlama-7b-hf
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/code-llama/README.md

@@ -0,0 +1,22 @@
+# Overview
+
+This is an example of CodeLLaMA configuration for 7b, 13b and 34b.
+
+The 7b variant fits on any 24GB VRAM GPU and will take up about 17 GB of VRAM during training if using qlora and 20 GB if using lora. On a RTX 4090 it trains 3 epochs of the default dataset in about 15 minutes.
+
+The 13b variant will fit if you change these settings to these values:
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+
+The 34b variant does not fit on 24GB of VRAM - you will need something with +40 gb VRAM that also supports flash attention v2 - A6000 or A100 are good choices.
+
+```shell
+accelerate launch scripts/finetune.py examples/code-llama/[MODEL_SIZE]/qlora.yml
+
+```
+or
+
+```shell
+accelerate launch scripts/finetune.py examples/code-llama/[MODEL_SIZE]/lora.yml
+
+```

examples/archived/dbrx/16bit-lora.yaml

@@ -0,0 +1,74 @@
+base_model: LnL-AI/dbrx-base-converted-v2
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 512
+sample_packing: false
+pad_to_sequence_len: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+adapter: lora
+lora_model_dir:
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+# w1, w2, & v1 will hang the trainer
+lora_target_modules:
+  - q_proj # attn
+  - k_proj # attn
+  - v_proj # attn
+  - out_proj # attn
+  - layer # router
+#  - w1
+#  - w2
+#  - v1
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: false  # don't use with fsdp_activation_checkpointing
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch:
+saves_per_epoch: 1
+
+weight_decay: 0.0
+fsdp:
+  - full_shard
+  - auto_wrap
+fsdp_config:
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: false
+  fsdp_use_orig_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: DbrxBlock
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_activation_checkpointing: true

examples/archived/dbrx/8bit-lora.yaml

@@ -0,0 +1,77 @@
+base_model: LnL-AI/dbrx-base-converted-v2
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 512
+sample_packing: false
+pad_to_sequence_len: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+adapter: lora
+lora_model_dir:
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+# w1, w2, & v1 will hang the trainer
+lora_target_modules:
+  - q_proj # attn
+  - k_proj # attn
+  - v_proj # attn
+  - out_proj # attn
+  - layer # router
+#  - w1
+#  - w2
+#  - v1
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: false  # don't use with fsdp_activation_checkpointing
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch:
+saves_per_epoch: 1
+
+weight_decay: 0.0
+fsdp:
+  - full_shard
+  - auto_wrap
+fsdp_config:
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: false
+  fsdp_use_orig_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: DbrxBlock
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_activation_checkpointing: true

examples/archived/dbrx/README.md

@@ -0,0 +1,26 @@
+# DBRX MoE
+
+Currently, for LoRA, only the `q_proj`, `k_proj`, `v_proj` `out_proj` and `layer` Linear layers are trainable.
+
+We are using the "converted" base models based on [this issue](https://huggingface.co/databricks/dbrx-instruct/discussions/10)
+where the Experts are fused as an `nn.Parameter` rather than a `nn.Linear` layer. However, the implementation
+is still a bit buggy and attempting to train a LoRA adapter over those `w1`, `w2` and `v1` layers
+results in the trainer hanging.
+
+
+### FSDP
+We've tested using the [`LnL-AI/dbrx-base-converted-v2`](https://huggingface.co/LnL-AI/dbrx-base-converted-v2) model as the base model for FSDP.
+
+The high memory usage seen w/ FSDP is due to FSDP not supporting 8bit optimizers.
+
+- 16-bit LoRA w/ FSDP
+  - ✅ w/o CPU Offload - 8x80GB uses ~80GiB/gpu
+  - ❌ w/ CPU Offload - `paged_adamw_8bit` optimizer errors from being on cpu
+- ✅ 8-bit LoRA w/ FSDP
+- ❌ 4-bit QLoRA w/ FSDP - errors w/: `Error an illegal memory access was encountered at line 90 in file /src/csrc/ops.cu`
+- ✅ bf16 full finetune w/ FSDP, freezing all but first 8 layers (8x80GB uses ~78GiB/gpu)
+
+
+### Deepspeed
+
+WIP

examples/archived/dbrx/fft-ds-zero3.yaml

@@ -0,0 +1,49 @@
+base_model: LnL-AI/dbrx-base-converted-v2
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 512
+sample_packing: false
+pad_to_sequence_len: false
+
+unfrozen_parameters:
+  - transformer.blocks.[0-7].
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch:
+saves_per_epoch: 1
+
+weight_decay: 0.0
+deepspeed: deepspeed_configs/zero3_bf16.json

examples/archived/deepcoder/deepcoder-14B-preview-lora.yml

@@ -0,0 +1,58 @@
+base_model: agentica-org/DeepCoder-14B-Preview
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/falcon/config-7b-lora.yml

@@ -0,0 +1,57 @@
+base_model: tiiuae/falcon-7b
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
+trust_remote_code: true
+
+load_in_8bit: true
+load_in_4bit: false
+gptq: false
+push_dataset_to_hub:
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca:chat
+dataset_prepared_path:
+val_set_size: 0.05
+adapter: lora
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len:
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.0
+lora_target_linear: true
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/falcon-7b
+batch_size: 2
+micro_batch_size: 1
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.00003
+bf16: auto
+tf32: true
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+xformers_attention: true
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 40
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  pad_token: "<|endoftext|>"
+  bos_token: "<|endoftext|>"
+  eos_token: "<|endoftext|>"

examples/archived/falcon/config-7b-qlora.yml

@@ -0,0 +1,87 @@
+# 1b: tiiuae/falcon-rw-1b
+# 40b: tiiuae/falcon-40b
+base_model: tiiuae/falcon-7b
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
+trust_remote_code: true
+
+
+load_in_8bit: false
+# enable 4bit for QLoRA
+load_in_4bit: true
+gptq: false
+push_dataset_to_hub:
+datasets:
+  - path: QingyiSi/Alpaca-CoT
+    data_files:
+      - Chain-of-Thought/formatted_cot_data/gsm8k_train.json
+    type: "alpaca:chat"
+dataset_prepared_path:
+val_set_size: 0.05
+# enable QLoRA
+adapter: qlora
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len:
+
+# hyperparameters from QLoRA paper Appendix B.2
+# "We find hyperparameters to be largely robust across datasets"
+lora_r: 64
+lora_alpha: 16
+# 0.1 for models up to 13B
+# 0.05 for 33B and 65B models
+lora_dropout: 0.05
+# add LoRA modules on all linear layers of the base model
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/qlora-out
+
+# QLoRA paper Table 9
+# - 16 for 7b & 13b
+# - 32 for 33b, 64 for 64b
+# Max size tested on A6000
+# - 7b: 40
+# - 40b: 4
+# decrease if OOM, increase for max VRAM utilization
+micro_batch_size: 1
+gradient_accumulation_steps: 2
+num_epochs: 4
+# Optimizer for QLoRA
+optimizer: paged_adamw_32bit
+torchdistx_path:
+lr_scheduler: cosine
+# QLoRA paper Table 9
+# - 2e-4 for 7b & 13b
+# - 1e-4 for 33b & 64b
+learning_rate: 0.0002
+bf16: auto
+tf32: true
+gradient_checkpointing: true
+# stop training after this many evaluation losses have increased in a row
+# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
+early_stopping_patience: 3
+resume_from_checkpoint:
+auto_resume_from_checkpoints: true
+logging_steps: 1
+xformers_attention: true
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.000001
+special_tokens:
+  pad_token: "<|endoftext|>"
+  bos_token: "<|endoftext|>"
+  eos_token: "<|endoftext|>"

examples/archived/falcon/config-7b.yml

@@ -0,0 +1,54 @@
+base_model: tiiuae/falcon-7b
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
+trust_remote_code: true
+gptq: false
+push_dataset_to_hub:
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca:chat
+dataset_prepared_path:
+val_set_size: 0.05
+adapter:
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len:
+lora_r: 64
+lora_alpha: 32
+lora_dropout: 0.0
+lora_target_linear: true
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/falcon-7b
+batch_size: 2
+micro_batch_size: 1
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.00003
+bf16: auto
+tf32: true
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+xformers_attention: true
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 40
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  pad_token: "<|endoftext|>"
+  bos_token: "<|endoftext|>"
+  eos_token: "<|endoftext|>"

examples/archived/gemma/qlora.yml

@@ -0,0 +1,56 @@
+# use google/gemma-7b if you have access
+base_model: mhenrichsen/gemma-7b
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+
+# huggingface repo
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+val_set_size: 0.1
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+
+gradient_accumulation_steps: 3
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/gptj/qlora.yml

@@ -0,0 +1,48 @@
+base_model: EleutherAI/gpt-j-6b
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+push_dataset_to_hub:
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+adapter: qlora
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len:
+lora_r: 8
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_linear: true
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/qlora-out
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 2
+optimizer: paged_adamw_8bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.0001
+bf16: auto
+tf32: true
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+xformers_attention: true
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.1
+special_tokens:
+  pad_token: "<|endoftext|>"

examples/archived/jeopardy-bot/config.yml

@@ -0,0 +1,51 @@
+base_model: huggyllama/llama-7b
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+datasets:
+  - path: openaccess-ai-collective/jeopardy
+    type: jeopardy
+dataset_prepared_path:
+val_set_size: 0.02
+adapter:
+lora_model_dir:
+sequence_len: 512
+max_packed_sequence_len:
+lora_r:
+lora_alpha:
+lora_dropout:
+lora_target_modules:
+lora_fan_in_fan_out: false
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/jeopardy-bot-7b
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.00003
+bf16: auto
+tf32: true
+resume_from_checkpoint:
+logging_steps: 5
+xformers_attention: true
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 20
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.1
+tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/mpt-7b/README.md

@@ -0,0 +1,6 @@
+# MPT-7B
+
+```shell
+accelerate launch scripts/finetune.py examples/mpt-7b/config.yml
+
+```

examples/archived/mpt-7b/config.yml

@@ -0,0 +1,53 @@
+base_model: mosaicml/mpt-7b
+# optionally might have model_type or tokenizer_type
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true  # required for mpt as their model class is not merged into transformers yet
+load_in_8bit: false
+datasets:
+  - path: vicgalle/alpaca-gpt4
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.02
+adapter:
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len:
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+  - q_proj
+  - v_proj
+lora_fan_in_fan_out: false
+wandb_project: mpt-alpaca-7b
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/mpt-alpaca-7b
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.0000002
+bf16: auto
+tf32: true
+resume_from_checkpoint:
+logging_steps: 5
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 20
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0001
+tokens:
+  pad_token: "<|padding|>"
+  bos_token: "<|endoftext|>"
+  eos_token: "<|endoftext|>"
+  unk_token: "<|endoftext|>"

examples/archived/openllama-3b/README.md

@@ -0,0 +1,16 @@
+# openllama-3b
+
+Basic full tune
+```shell
+accelerate launch scripts/finetune.py examples/openllama-3b/config.yml
+```
+
+LoRA
+```shell
+accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml
+```
+
+QLoRA
+```shell
+accelerate launch scripts/finetune.py examples/openllama-3b/qlora.yml
+```

examples/archived/openllama-3b/config.yml

@@ -0,0 +1,52 @@
+base_model: openlm-research/open_llama_3b_v2
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+push_dataset_to_hub:
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.02
+adapter:
+lora_model_dir:
+sequence_len: 1024
+sample_packing: true
+lora_r:
+lora_alpha:
+lora_dropout:
+lora_target_modules:
+lora_target_linear:
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/openllama-out
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.000003
+float16: true
+bf16: false
+fp16: false
+tf32: false
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 20
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.1
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/openllama-3b/lora.yml

@@ -0,0 +1,60 @@
+base_model: openlm-research/open_llama_3b_v2
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+push_dataset_to_hub:
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.02
+adapter: lora
+lora_model_dir:
+sequence_len: 1024
+sample_packing: true
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.0
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/lora-out
+gradient_accumulation_steps: 1
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.0002
+bf16: false
+fp16: true
+tf32: false
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 20
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.1
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/openllama-3b/qlora.yml

@@ -0,0 +1,53 @@
+base_model: openlm-research/open_llama_3b_v2
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+push_dataset_to_hub:
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+adapter: qlora
+lora_model_dir:
+sequence_len: 1024
+sample_packing: true
+lora_r: 8
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_linear: true
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/qlora-out
+gradient_accumulation_steps: 1
+micro_batch_size: 2
+num_epochs: 4
+optimizer: paged_adamw_32bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.0002
+bf16: false
+fp16: true
+tf32: false
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 20
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.1
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/archived/pythia-12b/README.md

@@ -0,0 +1,9 @@
+# Pythia 12B
+
+- Single-GPU A100 only (?)
+
+```shell
+python scripts/finetune.py examples/pythia-12b/config.yml
+```
+
+⚠️ Multiple-GPU A100 - Doesn't seem to work with multi-gpu without causing OOM! ⚠️

examples/archived/pythia-12b/config.yml

@@ -0,0 +1,42 @@
+base_model: EleutherAI/pythia-12b-deduped
+base_model_ignore_patterns: pytorch*  # prefer safetensors
+# optionally might have model_type or tokenizer_type
+model_type: GPTNeoXForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+gptq: false
+device_map: auto
+datasets:
+  - path: vicgalle/alpaca-gpt4
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+adapter:
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len: 2048
+lora_r: 64
+lora_alpha: 32
+lora_dropout: 0.0
+lora_target_linear: true
+lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/pythia-12b
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 5
+learning_rate: 0.00003
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+bf16: false
+fp16: false
+float16: true
+tf32: true
+flash_optimum: true
+resume_from_checkpoint:
+gradient_checkpointing: true

examples/archived/pythia/lora.yml

@@ -0,0 +1,36 @@
+base_model: EleutherAI/pythia-1.4b-deduped
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+adapter: lora
+lora_model_dir:
+sequence_len: 512
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - query_key_value
+lora_target_linear:
+lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/lora-alpaca-pythia
+gradient_accumulation_steps: 1
+micro_batch_size: 4
+num_epochs: 4
+learning_rate: 0.00001
+bf16: auto
+tf32: true
+resume_from_checkpoint:
+weight_decay: 0.1
+evals_per_epoch: 4
+logging_steps: 1

examples/archived/qwen/README.md

@@ -0,0 +1,10 @@
+# Qwen
+
+TODO
+
+# Qwen2 MoE
+
+✅ multipack
+✅ qwen2_moe 4-bit QLoRA
+✅ qwen2_moe 16-bit LoRA
+❓ qwen2_moe 8-bit LoRA

examples/archived/qwen/lora.yml

@@ -0,0 +1,56 @@
+base_model: Qwen/Qwen-7B
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 2048  # supports up to 8192
+sample_packing: false
+pad_to_sequence_len:
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention:
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/qwen/qlora.yml

@@ -0,0 +1,56 @@
+base_model: Qwen/Qwen-7B
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 2048  # supports up to 8192
+sample_packing: false
+pad_to_sequence_len:
+
+adapter: qlora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention:
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/qwen/qwen2-moe-lora.yaml

@@ -0,0 +1,52 @@
+base_model: Qwen/Qwen1.5-MoE-A2.7B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 1024  # supports up to 32k
+sample_packing: false
+pad_to_sequence_len: false
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 4
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/qwen/qwen2-moe-qlora.yaml

@@ -0,0 +1,55 @@
+base_model: Qwen/Qwen1.5-MoE-A2.7B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 1024  # supports up to 32k
+sample_packing: false
+pad_to_sequence_len: false
+
+adapter: qlora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 4
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/redpajama/README.md

@@ -0,0 +1,6 @@
+# RedPajama 3B preview release
+
+```shell
+accelerate launch scripts/finetune.py examples/redpajama/config-3b.yml
+
+```

examples/archived/redpajama/config-3b.yml

@@ -0,0 +1,54 @@
+base_model: togethercomputer/RedPajama-INCITE-Chat-3B-v1
+# optionally might have model_type or tokenizer_type
+model_type: GPTNeoXForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code:
+load_in_8bit: false
+datasets:
+  - path: vicgalle/alpaca-gpt4
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.02
+adapter:
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len:
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+  - q_proj
+  - v_proj
+lora_fan_in_fan_out: false
+wandb_project: redpajama-alpaca-3b
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/redpajama-alpaca-3b
+batch_size: 4
+micro_batch_size: 1
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+torchdistx_path:
+lr_scheduler: cosine
+learning_rate: 0.0000002
+bf16: auto
+tf32: true
+resume_from_checkpoint:
+logging_steps: 5
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 20
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0001
+tokens:
+  pad_token: "<|padding|>"
+  bos_token: "<|endoftext|>"
+  eos_token: "<|endoftext|>"
+  unk_token: "<|endoftext|>"

examples/archived/replit-3b/config-lora.yml

@@ -0,0 +1,48 @@
+base_model: replit/replit-code-v1-3b
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+load_in_8bit: false
+datasets:
+  - path: vicgalle/alpaca-gpt4
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+adapter: lora
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len:
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+  - Wqkv
+  - mlp_up
+  - mlp_down
+wandb_project: lora-replit
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/lora-replit
+batch_size: 8
+micro_batch_size: 1
+num_epochs: 4
+optimizer:
+torchdistx_path:
+lr_scheduler:
+learning_rate: 0.00001
+bf16: auto
+tf32: true
+gradient_checkpointing:
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 20
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0
+#special_tokens:

examples/archived/stablelm-2/1.6b/fft.yml

@@ -0,0 +1,59 @@
+base_model: stabilityai/stablelm-2-1_6b
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter:
+lora_model_dir:
+lora_r:
+lora_alpha:
+lora_dropout:
+lora_target_linear:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+flash_attn_cross_entropy: false
+flash_attn_rms_norm: true
+flash_attn_fuse_qkv: false
+flash_attn_fuse_mlp: true
+
+warmup_steps: 100
+evals_per_epoch: 4
+saves_per_epoch: 1
+
+deepspeed: #deepspeed_configs/zero2.json # multi-gpu only
+weight_decay: 0.1
+special_tokens:

examples/archived/stablelm-2/1.6b/lora.yml

@@ -0,0 +1,58 @@
+base_model: stabilityai/stablelm-2-1_6b
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+flash_attn_cross_entropy: false
+flash_attn_rms_norm: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/stablelm-2/README.md

@@ -0,0 +1,36 @@
+# StableLM 2
+
+This repository contains examples for training and processing using StableLM-2. It also includes a section to help you estimate the GPU requirements for your specific use case.
+
+## Estimating GPU Requirements
+
+| type          | deepspeed | batch size | context length | vRAM GPU (GBs) |
+|---------------|-----------|------------|----------------|----------------|
+| full finetune | N/A       | 1          | 4096           | ~21.5GBs       |
+| full finetune | zero2     | 1          | 4096           | ~20GBs         |
+| lora          | N/A       | 1          | 4096           | ~16.6GBs       |
+
+The above are estimates and might differ slight depending on the setup for example whether you pack your sequence lengths or not (the above assumes you do to length 4096).
+
+This blog post from Hamel Husain was a great resource for estimating these numbers: https://hamel.dev/notes/llm/03_estimating_vram.html
+
+## Training
+We have example scripts here for both full finetuning and lora using the popular alpaca dataset:
+
+```shell
+# preprocess the dataset
+CUDA_VISIBLE_DEVICES="" python -m axolotl.cli.preprocess examples/stablelm-2/1.6b/lora.yml
+```
+
+Single GPU Training:
+```shell
+python -m axolotl.cli.train examples/stablelm-2/fft.yml --deepspeed deepspeed_configs/zero2.json
+# OR
+python -m axolotl.cli.train examples/stablelm-2/1.6b/lora.yml
+```
+
+Multinode GPU Training with `accelerate`:
+```shell
+# make sure you've configured accelerate properly
+accelerate launch -m axolotl.cli.train examples/stablelm-2/1.6b/fft.yml --deepspeed deepspeed_configs/zero2.json
+```

examples/archived/starcoder2/qlora.yml

@@ -0,0 +1,58 @@
+base_model: bigcode/starcoder2-3b
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+
+
+dataset_prepared_path:
+val_set_size: 0.2
+output_dir: ./outputs/qlora
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 8
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: auto
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 20
+evals_per_epoch: 4
+eval_steps:
+saves_per_epoch: 4
+save_steps:
+save_total_limit: 2
+weight_decay:
+special_tokens:

examples/archived/tiny-llama/README.md

@@ -0,0 +1,17 @@
+# Overview
+
+This is a simple example of how to finetune TinyLlama1.1B using either lora or qlora:
+
+LoRa:
+
+```
+accelerate launch -m axolotl.cli.train examples/tiny-llama/lora.yml
+```
+
+qLoRa:
+
+```
+accelerate launch -m axolotl.cli.train examples/tiny-llama/qlora.yml
+```
+
+Both take about 10 minutes to complete on a 4090.

examples/archived/tiny-llama/lora-mps.yml

@@ -0,0 +1,56 @@
+base_model: TinyLlama/TinyLlama_v1.1
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+eval_sample_packing: false
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+fp16: false
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: false
+
+warmup_steps: 10
+evals_per_epoch: 0
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/tiny-llama/lora.yml

@@ -0,0 +1,54 @@
+base_model: TinyLlama/TinyLlama_v1.1
+# optionally might have model_type or tokenizer_type
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/tiny-llama/pretrain.yml

@@ -0,0 +1,45 @@
+base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+max_steps: 200
+pretraining_dataset:
+  - path: allenai/c4
+    name: en
+    type: pretrain
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./outputs/model-out
+
+sequence_len: 2048
+sample_packing: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch:
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/tiny-llama/qlora.yml

@@ -0,0 +1,56 @@
+base_model: TinyLlama/TinyLlama_v1.1
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/archived/xgen-7b/xgen-7b-8k-qlora.yml

@@ -0,0 +1,86 @@
+# An example finetuning Saleforce's XGen-7b model with 8k context using qlora
+# on Tim Dettmer's Guanaco dataset.
+base_model: Salesforce/xgen-7b-8k-base
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+load_in_8bit: false
+# enable 4bit for QLoRA
+load_in_4bit: true
+gptq: false
+push_dataset_to_hub:
+datasets:
+  - path: timdettmers/openassistant-guanaco
+    data_files:
+      - openassistant_best_replies_train.jsonl
+    type: "completion"
+dataset_prepared_path:
+val_set_size: 0.05
+# enable QLoRA
+adapter: qlora
+lora_model_dir:
+sequence_len: 8192
+max_packed_sequence_len:
+
+# hyperparameters from QLoRA paper Appendix B.2
+# "We find hyperparameters to be largely robust across datasets"
+lora_r: 64
+lora_alpha: 16
+# 0.1 for models up to 13B
+# 0.05 for 33B and 65B models
+lora_dropout: 0.05
+# add LoRA modules on all linear layers of the base model
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+output_dir: ./outputs/qlora-out
+
+# QLoRA paper Table 9
+# - 16 for 7b & 13b
+# - 32 for 33b, 64 for 64b
+# Max size tested on A6000
+# - 7b: 40
+# - 40b: 4
+# decrease if OOM, increase for max VRAM utilization
+micro_batch_size: 1
+gradient_accumulation_steps: 1
+num_epochs: 4
+# Optimizer for QLoRA
+optimizer: paged_adamw_32bit
+torchdistx_path:
+lr_scheduler: cosine
+# QLoRA paper Table 9
+# - 2e-4 for 7b & 13b
+# - 1e-4 for 33b & 64b
+learning_rate: 0.00002
+bf16: auto
+tf32: false
+gradient_checkpointing: true
+# stop training after this many evaluation losses have increased in a row
+# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
+early_stopping_patience: 3
+resume_from_checkpoint:
+auto_resume_from_checkpoints: true
+logging_steps: 1
+xformers_attention: true
+flash_attention:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  eos_token: "<|endoftext|>"
+  bos_token: "<|endoftext|>"
+  unk_token: "<|endoftext|>"
+  pad_token: "<|endoftext|>"

examples/archived/yi-34B-chat/README.md

@@ -0,0 +1,5 @@
+# Overview
+
+This is an example of a Yi-34B-Chat configuration. It demonstrates that it is possible to finetune a 34B model on a GPU with 24GB of VRAM.
+
+Tested on an RTX 4090 with `python -m axolotl.cli.train examples/mistral/qlora.yml`, a single epoch of finetuning on the alpaca dataset using qlora runs in 47 mins, using 97% of available memory.

examples/archived/yi-34B-chat/qlora.yml

@@ -0,0 +1,64 @@
+base_model: 01-ai/Yi-34B-Chat
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+sequence_len: 1024
+bf16: auto
+tf32: false
+flash_attention: true
+special_tokens:
+  bos_token: "<|startoftext|>"
+  eos_token: "<|endoftext|>"
+  unk_token: "<unk>"
+
+# Data
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+warmup_steps: 10
+
+# Iterations
+num_epochs: 1
+
+# Evaluation
+val_set_size: 0.1
+evals_per_epoch: 5
+eval_sample_packing: false
+eval_batch_size: 1
+
+# LoRA
+output_dir: ./outputs/qlora-out
+adapter: qlora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+
+# Sampling
+sample_packing: false
+pad_to_sequence_len: false
+
+# Batching
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+gradient_checkpointing: true
+
+# wandb
+wandb_project:
+
+# Optimizer
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+# Misc
+resume_from_checkpoint:
+logging_steps: 1
+weight_decay: 0