wip for new datasets abstractions

* auto gptq support

* more tweaks and add yml

* remove old gptq docker

* don't need explicit peft install for tests

* fix setup.py to use extra index url

install torch for tests
fix cuda version for autogptq index
set torch in requirements so that it installs properly
move gptq install around to work with github cicd

* gptq doesn't play well with sample packing

* address pr feedback

* remove torch install for now

* set quantization_config from model config

* Fix the implementation for getting quant config from model config

.github/workflows/main.yml

@@ -23,11 +23,6 @@ jobs:
             python_version: "3.10"
             pytorch: 2.0.1
             axolotl_extras:
-          - cuda: 118
-            cuda_version: 11.8.0
-            python_version: "3.9"
-            pytorch: 2.0.1
-            axolotl_extras: gptq
     runs-on: self-hosted
     steps:
       - name: Checkout
@@ -73,11 +68,6 @@ jobs:
             pytorch: 2.0.1
             axolotl_extras:
             is_latest: true
-          - cuda: 118
-            cuda_version: 11.8.0
-            python_version: "3.9"
-            pytorch: 2.0.1
-            axolotl_extras: gptq
     runs-on: self-hosted
     steps:
       - name: Checkout

.github/workflows/tests.yml

@@ -24,7 +24,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip install -e .[peft]
+          pip install -e .
           pip install -r requirements-tests.txt
 
       - name: Run tests

README.md

@@ -163,6 +163,8 @@ accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml \
   ```
   </details>
 
+- Windows: Please use WSL or Docker!
+
 ### Dataset
 
 Axolotl supports a variety of dataset formats. Below are some of the formats you can use.
@@ -328,6 +330,15 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
       name: enron_emails
       type: completion # format from earlier
 
+  # huggingface repo with multiple named configurations/subsets
+  datasets:
+    - path: bigcode/commitpackft
+      name:
+        - ruby
+        - python
+        - typescript
+      type: ... # unimplemented custom format
+
   # local
   datasets:
     - path: data.jsonl # or json
@@ -407,6 +418,10 @@ fp16: true
 # Use CUDA tf32
 tf32: true # require >=ampere
 
+# No AMP (automatic mixed precision)
+bfloat16: true # require >=ampere
+float16: true
+
 # a list of one or more datasets to finetune the model with
 datasets:
   # hf dataset repo | "json" for local dataset, make sure to fill data_files
@@ -459,6 +474,9 @@ dataset_shard_idx:
 # the maximum length of an input to train with, this should typically be less than 2048
 # as most models have a token/context limit of 2048
 sequence_len: 2048
+# pad inputs so each step uses constant sized buffers
+# this will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently
+pad_to_sequence_len:
 # max sequence length to concatenate training samples together up to
 # inspired by StackLLaMA. see https://huggingface.co/blog/stackllama#supervised-fine-tuning
 # FutureWarning: This will soon be DEPRECATED
@@ -493,6 +511,12 @@ lora_modules_to_save:
 lora_out_dir:
 lora_fan_in_fan_out: false
 
+# ReLoRA configuration
+# must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
+relora_steps: # number of steps per ReLoRA restart
+relora_warmup_steps: # number of per-restart warmup steps
+relora_cpu_offload: # true to perform lora weight merges on cpu during restarts, for modest gpu memory savings
+
 # wandb configuration if you're using it
 wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
 wandb_project: # your wandb project name
@@ -515,7 +539,7 @@ lr_quadratic_warmup:
 logging_steps:
 save_strategy: # set to `no` to skip checkpoint saves
 save_steps: # leave empty to save at each epoch
-eval_steps:
+eval_steps: # leave empty to eval at each epoch
 save_total_limit: # checkpoints saved at a time
 max_steps:
 
@@ -601,12 +625,14 @@ fsdp_config:
 # Deepspeed config path
 deepspeed:
 
+# Advanced DDP Arguments
+ddp_timeout:
+ddp_bucket_cap_mb:
+ddp_broadcast_buffers:
+
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:
 
-# Set padding for data collator to 'longest'
-collator_pad_to_longest:
-
 # Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize
 pretraining_dataset:
 
@@ -626,7 +652,7 @@ strict:
 
 Run
 ```bash
-accelerate launch scripts/finetune.py configs/your_config.yml
+accelerate launch scripts/finetune.py your_config.yml
 ```
 
 #### Multi-GPU

deepspeed/zero2.json

@@ -0,0 +1,46 @@
+{
+    "zero_optimization": {
+      "stage": 2,
+      "offload_optimizer": {
+        "device": "cpu"
+      },
+      "contiguous_gradients": true,
+      "overlap_comm": true
+    },
+    "bf16": {
+      "enabled": "auto"
+    },
+    "fp16": {
+      "enabled": "auto",
+      "auto_cast": false,
+      "loss_scale": 0,
+      "initial_scale_power": 32,
+      "loss_scale_window": 1000,
+      "hysteresis": 2,
+      "min_loss_scale": 1
+    },
+    "optimizer": {
+      "type": "AdamW",
+      "params": {
+        "lr": "auto",
+        "betas": [
+          0.9,
+          0.999
+        ],
+        "eps": 1e-8,
+        "weight_decay": "auto"
+      }
+    },
+    "scheduler": {
+      "type": "WarmupDecayLR",
+      "params": {
+        "warmup_min_lr": "auto",
+        "warmup_max_lr": "auto",
+        "warmup_num_steps": "auto",
+        "total_num_steps": "auto"
+      }
+    },
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}

deepspeed/zero3.json

@@ -35,10 +35,7 @@
     "type": "AdamW",
     "params": {
       "lr": "auto",
-      "betas": [
-        0.9,
-        0.95
-      ],
+      "betas": "auto",
       "eps": 1e-8,
       "weight_decay": "auto"
     }

docker/Dockerfile

@@ -11,14 +11,13 @@ RUN apt-get update && \
 
 WORKDIR /workspace
 
-RUN pip3 install --force-reinstall "peft @ git+https://github.com/huggingface/peft.git@main"
 RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN cd axolotl && \
     if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[flash-attn,$AXOLOTL_EXTRAS]; \
+        pip install -e .[flash-attn,gptq,$AXOLOTL_EXTRAS]; \
     else \
-        pip install -e .[flash-attn]; \
+        pip install -e .[flash-attn,gptq]; \
     fi
 
 # fix so that git fetch/pull from remote works

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

@@ -0,0 +1,67 @@
+base_model: codellama/CodeLlama-13b-hf
+base_model_config: codellama/CodeLlama-13b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 100000
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,69 @@
+base_model: codellama/CodeLlama-13b-hf
+base_model_config: codellama/CodeLlama-13b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 100000
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,67 @@
+base_model: codellama/CodeLlama-34b-hf
+base_model_config: codellama/CodeLlama-34b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 100000
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,69 @@
+base_model: codellama/CodeLlama-34b-hf
+base_model_config: codellama/CodeLlama-34b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 100000
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,67 @@
+base_model: codellama/CodeLlama-7b-hf
+base_model_config: codellama/CodeLlama-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 100000
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,69 @@
+base_model: codellama/CodeLlama-7b-hf
+base_model_config: codellama/CodeLlama-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 100000
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/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/gptq-lora-7b/README.md

@@ -1,8 +0,0 @@
-# LLaMa 7B using LoRA
-
-This is a good place to start for beginners. This will run on an NVIDIA RTX4090 with no other changes needed.
-
-```shell
-accelerate launch scripts/finetune.py examples/gptq-lora-7b/config.yml
-
-```

examples/gptq-lora-7b/config.yml

@@ -1,63 +0,0 @@
-base_model: Neko-Institute-of-Science/LLaMA-7B-4bit-128g
-base_model_config: Neko-Institute-of-Science/LLaMA-7B-4bit-128g
-model_type: LlamaForCausalLM
-tokenizer_type: LlamaTokenizer
-trust_remote_code:
-load_in_8bit: true
-gptq: true
-datasets:
-  - path: vicgalle/alpaca-gpt4
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-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: llama-7b-lora-int4
-wandb_entity:
-wandb_watch:
-wandb_run_id:
-wandb_log_model:
-output_dir: ./llama-7b-lora-int4
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 3
-optimizer: adamw_bnb_8bit
-torchdistx_path:
-lr_scheduler: cosine
-learning_rate: 0.0000002
-train_on_inputs: false
-group_by_length: false
-fp16: true
-bf16: false
-tf32: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 5
-xformers_attention:
-flash_attention:
-gradient_checkpointing: true
-gptq_groupsize: 128
-gptq_model_v1: false
-warmup_steps: 20
-eval_steps: 110
-save_steps: 660
-debug:
-deepspeed:
-weight_decay: 0.0001
-fsdp:
-fsdp_config:
-tokens:
-  pad_token: "[PAD]"
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/llama-2/gptq-lora.yml

@@ -0,0 +1,76 @@
+base_model: TheBloke/Llama-2-7B-GPTQ
+base_model_config: TheBloke/Llama-2-7B-GPTQ
+is_llama_derived_model: false
+gptq: true
+gptq_bits: 4
+model_type: AutoModelForCausalLM
+tokenizer_type: LlamaTokenizer
+tokenizer_use_fast: true
+tokenizer_legacy: true
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+push_dataset_to_hub:
+hf_use_auth_token: true
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+adapter: lora
+lora_model_dir:
+sequence_len: 4096
+sample_packing:
+lora_r: 8
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - k_proj
+  - o_proj
+  - q_proj
+  - v_proj
+lora_target_linear:
+lora_fan_in_fan_out:
+wandb_project:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+output_dir: ./model-out
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 3
+optimizer: adamw_torch
+adam_beta2: 0.95
+adam_eps: 0.00001
+max_grad_norm: 1.0
+torchdistx_path:
+lr_scheduler: cosine
+lr_quadratic_warmup: true
+learning_rate: 0.000017
+train_on_inputs: false
+group_by_length: false
+bf16: false
+fp16: false
+float16: true
+tf32: true
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention:
+sdp_attention:
+flash_optimum:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 100
+eval_steps:
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.1
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/llama-2/relora.yml

@@ -0,0 +1,73 @@
+base_model: meta-llama/Llama-2-7b-hf
+base_model_config: meta-llama/Llama-2-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./relora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+relora_steps: 150
+relora_warmup_steps: 10
+relora_cpu_offload: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 4
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps: 50
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/pythia-12b/config.yml

@@ -47,4 +47,3 @@ local_rank:
 gradient_checkpointing: true
 fsdp:
 fsdp_config:
-collator_pad_to_longest: true

requirements.txt

@@ -1,12 +1,18 @@
+--extra-index-url https://download.pytorch.org/whl/cu118
+--extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
+torch==2.0.1
+auto-gptq
+packaging
 peft @ git+https://github.com/huggingface/peft.git
 transformers @ git+https://github.com/huggingface/transformers.git
 bitsandbytes>=0.41.1
 accelerate @ git+https://github.com/huggingface/accelerate@2a289f6108e77a77a4efffb3f6316bc98538413b
 addict
+evaluate
 fire
-PyYAML==6.0
+PyYAML>=6.0
 datasets
-flash-attn==2.0.8
+flash-attn>=2.0.8
 sentencepiece
 wandb
 einops
@@ -15,7 +21,7 @@ optimum
 hf_transfer
 colorama
 numba
-numpy==1.24.4
+numpy>=1.24.4
 # qlora things
 bert-score==0.3.13
 evaluate==0.4.0
@@ -23,3 +29,4 @@ rouge-score==0.1.2
 scipy
 scikit-learn==1.2.2
 pynvml
+art

scripts/finetune.py

@@ -4,27 +4,28 @@ import importlib
 import logging
 import os
 import random
-import signal
 import sys
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Union
 
 import fire
 import torch
+import transformers
 import yaml
 
 # add src to the pythonpath so we don't need to pip install this
-from optimum.bettertransformer import BetterTransformer
+from art import text2art
 from transformers import GenerationConfig, TextStreamer
 
+from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
 from axolotl.logging_config import configure_logging
+from axolotl.train import TrainDatasetMeta, train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.data import prepare_dataset
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process
-from axolotl.utils.models import load_model, load_tokenizer
+from axolotl.utils.models import load_tokenizer
 from axolotl.utils.tokenization import check_dataset_labels
-from axolotl.utils.trainer import setup_trainer
 from axolotl.utils.wandb import setup_wandb_env_vars
 
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
@@ -37,15 +38,12 @@ LOG = logging.getLogger("axolotl.scripts")
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 
 
-def print_axolotl_text_art():
-    ascii_art = """
-                           dP            dP   dP
-                           88            88   88
-.d8888b. dP.  .dP .d8888b. 88 .d8888b. d8888P 88
-88'  `88  `8bd8'  88'  `88 88 88'  `88   88   88
-88.  .88  .d88b.  88.  .88 88 88.  .88   88   88
-`88888P8 dP'  `dP `88888P' dP `88888P'   dP   dP
-"""
+def print_axolotl_text_art(suffix=None):
+    font = "nancyj"
+    ascii_text = "  axolotl"
+    if suffix:
+        ascii_text += f"  x  {suffix}"
+    ascii_art = text2art(" axolotl", font=font)
 
     if is_main_process():
         print(ascii_art)
@@ -60,7 +58,45 @@ def get_multi_line_input() -> Optional[str]:
     return instruction
 
 
-def do_inference(cfg, model, tokenizer, prompter: Optional[str]):
+def do_merge_lora(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
+):
+    model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
+    safe_serialization = cfg.save_safetensors is True
+
+    LOG.info("running merge of LoRA with base model")
+    model = model.merge_and_unload()
+    model.to(dtype=torch.float16)
+
+    if cfg.local_rank == 0:
+        LOG.info("saving merged model")
+        model.save_pretrained(
+            str(Path(cfg.output_dir) / "merged"),
+            safe_serialization=safe_serialization,
+        )
+        tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
+
+
+def shard(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
+):
+    model, _ = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
+    safe_serialization = cfg.save_safetensors is True
+    LOG.debug("Re-saving model w/ sharding")
+    model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
+
+
+def do_inference(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
+):
+    model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
+    prompter = cli_args.prompter
     default_tokens = {"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>"}
 
     for token, symbol in default_tokens.items():
@@ -82,6 +118,8 @@ def do_inference(cfg, model, tokenizer, prompter: Optional[str]):
             max_seq_len=255, mem_freq=50, top_k=5, max_cache_size=None
         )
 
+    model = model.to(cfg.device)
+
     while True:
         print("=" * 80)
         # support for multiline inputs
@@ -133,6 +171,10 @@ def choose_config(path: Path):
             "No YAML config files found in the specified directory. Are you using a .yml extension?"
         )
 
+    if len(yaml_files) == 1:
+        print(f"Using default YAML file '{yaml_files[0]}'")
+        return yaml_files[0]
+
     print("Choose a YAML file:")
     for idx, file in enumerate(yaml_files):
         print(f"{idx + 1}. {file}")
@@ -155,12 +197,7 @@ def check_not_in(list1: List[str], list2: Union[Dict[str, Any], List[str]]) -> b
     return not any(el in list2 for el in list1)
 
 
-def train(
-    config: Path = Path("configs/"),
-    prepare_ds_only: bool = False,
-    **kwargs,
-):
-    print_axolotl_text_art()
+def load_cfg(config: Path = Path("examples/"), **kwargs):
     if Path(config).is_dir():
         config = choose_config(config)
 
@@ -184,132 +221,58 @@ def train(
     normalize_config(cfg)
 
     setup_wandb_env_vars(cfg)
+    return cfg
 
-    # load the tokenizer first
-    LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
+
+def load_datasets(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
+) -> TrainDatasetMeta:
     tokenizer = load_tokenizer(cfg)
 
-    if (
-        check_not_in(["shard", "merge_lora"], kwargs) and not cfg.inference
-    ):  # don't need to load dataset for these
-        train_dataset, eval_dataset, total_num_steps = prepare_dataset(cfg, tokenizer)
+    train_dataset, eval_dataset, total_num_steps = prepare_dataset(cfg, tokenizer)
 
-    if cfg.debug or "debug" in kwargs:
+    if cli_args.debug or cfg.debug:
         LOG.info("check_dataset_labels...")
         check_dataset_labels(
             train_dataset.select(
-                [random.randrange(0, len(train_dataset) - 1) for _ in range(5)]  # nosec
+                [
+                    random.randrange(0, len(train_dataset) - 1)  # nosec
+                    for _ in range(cli_args.debug_num_examples)
+                ]
             ),
             tokenizer,
+            num_examples=cli_args.debug_num_examples,
+            text_only=cli_args.debug_text_only,
         )
 
-    if prepare_ds_only:
-        LOG.info("Finished preparing dataset. Exiting...")
-        return
-
-    # Load the model and tokenizer
-    LOG.info("loading model and (optionally) peft_config...")
-    model, peft_config = load_model(cfg, tokenizer)
-
-    safe_serialization = cfg.save_safetensors is True
-
-    if "merge_lora" in kwargs and cfg.adapter is not None:
-        LOG.info("running merge of LoRA with base model")
-        model = model.merge_and_unload()
-        model.to(dtype=torch.float16)
-
-        if cfg.local_rank == 0:
-            LOG.info("saving merged model")
-            model.save_pretrained(
-                str(Path(cfg.output_dir) / "merged"),
-                safe_serialization=safe_serialization,
-            )
-            tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
-        return
-
-    if cfg.inference:
-        LOG.info("calling do_inference function")
-        prompter: Optional[str] = "AlpacaPrompter"
-        if "prompter" in kwargs:
-            if kwargs["prompter"] == "None":
-                prompter = None
-            else:
-                prompter = kwargs["prompter"]
-        do_inference(cfg, model, tokenizer, prompter=prompter)
-        return
-
-    if "shard" in kwargs:
-        model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
-        return
-
-    trainer = setup_trainer(
-        cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
+    return TrainDatasetMeta(
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        total_num_steps=total_num_steps,
     )
 
-    model.config.use_cache = False
 
-    if torch.__version__ >= "2" and sys.platform != "win32":
-        LOG.info("Compiling torch model")
-        model = torch.compile(model)
-
-    # go ahead and presave, so we have the adapter config available to inspect
-    if peft_config:
-        LOG.info(f"Pre-saving adapter config to {cfg.output_dir}")
-        peft_config.save_pretrained(cfg.output_dir)
-
-    # In case we want to stop early with ctrl+c, this is a nice to have to save the pretrained model
-    if cfg.local_rank == 0:
-
-        def terminate_handler(_, __, model):
-            if cfg.flash_optimum:
-                model = BetterTransformer.reverse(model)
-            model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
-            sys.exit(0)
-
-        signal.signal(
-            signal.SIGINT, lambda signum, frame: terminate_handler(signum, frame, model)
-        )
-
-    LOG.info("Starting trainer...")
-    if cfg.group_by_length:
-        LOG.info("hang tight... sorting dataset for group_by_length")
-    resume_from_checkpoint = cfg.resume_from_checkpoint
-    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
-        possible_checkpoints = [
-            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
-        ]
-        if len(possible_checkpoints) > 0:
-            sorted_paths = sorted(
-                possible_checkpoints,
-                key=lambda path: int(path.split("-")[-1]),
-            )
-            resume_from_checkpoint = sorted_paths[-1]
-            LOG.info(
-                f"Using Auto-resume functionality to start with checkpoint at {resume_from_checkpoint}"
-            )
-
-    if not Path(cfg.output_dir).is_dir():
-        os.makedirs(cfg.output_dir, exist_ok=True)
-    tokenizer.save_pretrained(cfg.output_dir)
-    if cfg.flash_optimum:
-        with torch.backends.cuda.sdp_kernel(
-            enable_flash=True, enable_math=True, enable_mem_efficient=True
-        ):
-            trainer.train(resume_from_checkpoint=resume_from_checkpoint)
+def do_cli(config: Path = Path("examples/"), **kwargs):
+    print_axolotl_text_art()
+    parsed_cfg = load_cfg(config, **kwargs)
+    parser = transformers.HfArgumentParser((TrainerCliArgs))
+    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
+        return_remaining_strings=True
+    )
+    if parsed_cli_args.inference:
+        do_inference(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    elif parsed_cli_args.merge_lora:
+        do_merge_lora(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    elif parsed_cli_args.shard:
+        shard(cfg=parsed_cfg, cli_args=parsed_cli_args)
     else:
-        trainer.train(resume_from_checkpoint=resume_from_checkpoint)
-
-    LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
-
-    # TODO do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
-    # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple processes attempt to write the same file
-    if cfg.fsdp:
-        trainer.save_model(cfg.output_dir)
-    elif cfg.local_rank == 0:
-        if cfg.flash_optimum:
-            model = BetterTransformer.reverse(model)
-        model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
+        dataset_meta = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+        if parsed_cli_args.prepare_ds_only:
+            return
+        train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
 
 
 if __name__ == "__main__":
-    fire.Fire(train)
+    fire.Fire(do_cli)

setup.py

@@ -2,15 +2,27 @@
 
 from setuptools import find_packages, setup
 
-install_requires = []
-with open("./requirements.txt", encoding="utf-8") as requirements_file:
-    # don't include peft yet until we check the int4
-    # need to manually install peft for now...
-    reqs = [r.strip() for r in requirements_file.readlines() if "peft" not in r]
-    reqs = [r for r in reqs if "flash-attn" not in r]
-    reqs = [r for r in reqs if r and r[0] != "#"]
-    for r in reqs:
-        install_requires.append(r)
+
+def parse_requirements():
+    _install_requires = []
+    _dependency_links = []
+    with open("./requirements.txt", encoding="utf-8") as requirements_file:
+        lines = [
+            r.strip() for r in requirements_file.readlines() if "auto-gptq" not in r
+        ]
+        for line in lines:
+            if line.startswith("--extra-index-url"):
+                # Handle custom index URLs
+                _, url = line.split()
+                _dependency_links.append(url)
+            elif "flash-attn" not in line and line and line[0] != "#":
+                # Handle standard packages
+                _install_requires.append(line)
+    return _install_requires, _dependency_links
+
+
+install_requires, dependency_links = parse_requirements()
+
 
 setup(
     name="axolotl",
@@ -19,12 +31,10 @@ setup(
     package_dir={"": "src"},
     packages=find_packages(),
     install_requires=install_requires,
+    dependency_links=dependency_links,
     extras_require={
         "gptq": [
-            "alpaca_lora_4bit @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip",
-        ],
-        "gptq_triton": [
-            "alpaca_lora_4bit[triton] @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip",
+            "auto-gptq",
         ],
         "flash-attn": [
             "flash-attn==2.0.8",
@@ -32,8 +42,5 @@ setup(
         "extras": [
             "deepspeed",
         ],
-        "peft": [
-            "peft @ git+https://github.com/huggingface/peft.git",
-        ],
     },
 )

src/axolotl/common/cli.py

@@ -0,0 +1,43 @@
+"""
+shared module for cli specific things
+"""
+
+import logging
+from dataclasses import dataclass, field
+from typing import Optional
+
+from axolotl.logging_config import configure_logging
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.models import load_model, load_tokenizer
+
+configure_logging()
+LOG = logging.getLogger("axolotl.common.cli")
+
+
+@dataclass
+class TrainerCliArgs:
+    """
+    dataclass representing the various non-training arguments
+    """
+
+    debug: bool = field(default=False)
+    debug_text_only: bool = field(default=False)
+    debug_num_examples: int = field(default=5)
+    inference: bool = field(default=False)
+    merge_lora: bool = field(default=False)
+    prepare_ds_only: bool = field(default=False)
+    prompter: Optional[str] = field(default=None)
+    shard: bool = field(default=False)
+
+
+def load_model_and_tokenizer(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
+):
+    LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
+    tokenizer = load_tokenizer(cfg)
+    LOG.info("loading model and (optionally) peft_config...")
+    model, _ = load_model(cfg, tokenizer, inference=cli_args.inference)
+
+    return model, tokenizer

src/axolotl/core/datasets.py

@@ -0,0 +1,144 @@
+import logging
+from dataclasses import dataclass, field
+from enum import Enum
+from pathlib import Path
+from typing import Any, Dict, Generator, List, Optional, Union
+
+from datasets import Dataset as Dataset_ds
+from datasets import DatasetDict, IterableDataset, load_dataset, load_from_disk
+from huggingface_hub import hf_hub_download
+
+logger = logging.getLogger("axolotl")
+
+
+class DsType(Enum):
+    JSON = "json"
+    ARROW = "arrow"
+    PARQUET = "parquet"
+
+
+@dataclass
+class DatasetConfiguration:
+    path: str
+    type: str
+    name: Optional[str] = field(
+        default=None,
+        metadata={"help": "the name of the dataset configuration to load."},
+    )
+    ds_type: Optional[DsType] = None
+    data_files: Optional[Union[str, List[str]]] = None
+    shards: Optional[int] = None
+    test_size: Optional[float] = None
+
+    @staticmethod
+    def from_dict(d: Dict[str, Any]) -> Generator["DatasetConfiguration", None, None]:
+        if "name" in d and isinstance(d["name"], list):
+            name = d.pop("name")
+            for n in name:
+                yield DatasetConfiguration(
+                    **d,
+                    name=n,
+                )
+
+
+def load_dataset_from_local(config: DatasetConfiguration) -> Optional[Dataset_ds]:
+    local_path = Path(config.path)
+    if not local_path.exists():
+        return None
+    ds = None
+    if local_path.is_dir():
+        if config.ds_type:
+            # TODO dirs with arrow or parquet files could be loaded with `load_from_disk`
+            ds = load_from_disk(config.path)
+        else:
+            ds = load_dataset(
+                config.path,
+                name=config.name,
+                data_files=config.data_files,
+                streaming=False,
+                split=None,
+            )
+    elif local_path.is_file():
+        ds_type = "json"
+        if config.ds_type:
+            ds_type = config.ds_type.value
+        elif "parquet" in config.path:
+            ds_type = "parquet"
+        elif "arrow" in config.path:
+            ds_type = "arrow"
+        ds = load_dataset(
+            ds_type,
+            name=config.name,
+            data_files=config.path,
+            streaming=False,
+            split=None,  # is this correct?
+        )
+    if not ds:
+        raise ValueError(
+            "unhandled dataset load: local path exists, but is neither a directory or a file"
+        )
+    return ds
+
+
+# TODO should this be a DatasetDict?
+class Dataset(Dataset_ds):
+    _config: DatasetConfiguration
+
+    def __init__(self, *args, config: DatasetConfiguration = None, **kwargs):
+        self._config = config
+        super().__init__(*args, **kwargs)
+
+    @staticmethod
+    def from_config(
+        config: DatasetConfiguration,
+        token: bool = False,
+        default_test_size: float = 0.1,
+    ):
+        ds = load_dataset_from_local(config)
+        if not ds:
+            try:
+                ds = load_dataset(
+                    config.path,
+                    name=config.name,
+                    data_files=config.data_files,
+                    token=token,
+                )
+            except FileNotFoundError:
+                pass
+        if not ds:
+            fp = hf_hub_download(
+                repo_id=config.path,
+                repo_type="dataset",
+                filename=config.data_files,
+                token=token,
+            )
+            ds = load_dataset(
+                "json", name=config.name, data_files=fp, streaming=False, split=None
+            )
+        if not ds:
+            raise ValueError("unhandled dataset load")
+        test_size = config.test_size if config.test_size else default_test_size
+        # determine if the dataset is pre-tokenized
+        check_ds = ds["train"] if isinstance(ds, DatasetDict) and "train" in ds else ds
+        is_ds_tokenized = False
+        if "input_ids" in check_ds.features:
+            is_ds_tokenized = True
+            if "attention_mask" not in check_ds.features:
+                logger.warning("`attention_mask` missing from pre-tokenized dataset")
+            if "labels" not in check_ds.features:
+                logger.warning("`labels` missing from pre-tokenized dataset")
+        if test_size and (not isinstance(ds, DatasetDict) or "test" not in ds):
+            ds.train_test_split(test_size=test_size, shuffle=False)
+            pass
+
+
+class DatasetCollection:
+    datasets: List[Dataset] = []
+
+    def __init__(self, datasets: Union[Dataset, List[Dataset]]):
+        self.datasets = datasets if isinstance(datasets, list) else [datasets]
+
+    def __iter__(self):
+        for ds in self.datasets:
+            for d in ds:
+                yield d

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -2,7 +2,9 @@
 
 # copied from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/llama_flash_attn_monkey_patch.py
 
+import logging
 import warnings
+from functools import partial
 from typing import List, Optional, Tuple, Union
 
 import torch
@@ -33,6 +35,9 @@ except ImportError:
     )
 
 
+LOG = logging.getLogger("axolotl")
+
+
 def replace_llama_attn_with_flash_attn(packed: Optional[bool] = False):
     transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
         _prepare_decoder_attention_mask
@@ -44,6 +49,34 @@ def replace_llama_attn_with_flash_attn(packed: Optional[bool] = False):
             llama_model_forward
         )
 
+    try:
+        from flash_attn.losses.cross_entropy import CrossEntropyLoss
+
+        LOG.info("patching with flash_attn.losses.cross_entropy")
+        transformers.models.llama.modeling_llama.CrossEntropyLoss = partial(
+            CrossEntropyLoss, inplace_backward=True
+        )
+    except ImportError:
+        LOG.info(
+            "optimized flash-attention CrossEntropyLoss not found (run `pip install 'git+https://github.com/Dao-AILab/flash-attention.git#egg=xentropy_cuda_lib&subdirectory=csrc/xentropy'`)"
+        )
+
+    try:
+        from flash_attn.ops.rms_norm import RMSNorm
+
+        class LlamaRMSNorm(RMSNorm):
+            """Patched LLamaRMSNorm"""
+
+            def __init__(self, hidden_size, eps=1e-6):
+                super().__init__(hidden_size, eps=eps)
+
+        LOG.info("patching with flash_attn.ops.rms_norm")
+        transformers.models.llama.modeling_llama.LlamaRMSNorm = LlamaRMSNorm
+    except ImportError:
+        LOG.info(
+            "optimized flash-attention RMSNorm not found (run `pip install 'git+https://github.com/Dao-AILab/flash-attention.git#egg=dropout_layer_norm&subdirectory=csrc/layer_norm'`)"
+        )
+
 
 # Disable the transformation of the attention mask in LlamaModel as the flash attention
 # requires the attention mask to be the same as the key_padding_mask

src/axolotl/monkeypatch/relora.py

@@ -0,0 +1,393 @@
+"""Implements the ReLoRA training procedure from https://arxiv.org/abs/2307.05695, minus the initial full fine-tune."""
+import glob
+import json
+import logging
+import os.path
+import shutil
+from pathlib import Path
+from typing import Dict, List, Sequence
+
+import bitsandbytes as bnb
+import peft
+import safetensors.torch as st
+import torch
+from huggingface_hub import snapshot_download
+from torch.optim.lr_scheduler import LRScheduler
+from torch.optim.optimizer import Optimizer
+from transformers import (
+    TrainerCallback,
+    TrainerControl,
+    TrainerState,
+    TrainingArguments,
+)
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.distributed import is_main_process
+
+LOG = logging.getLogger("axolotl.relora")
+
+
+def reset_optimizer(optimizer: torch.optim.Optimizer):
+    for group in optimizer.param_groups:
+        for param in group["params"]:
+            param_state = optimizer.state[param]
+            for key in param_state:
+                if "qmap" in key:
+                    continue
+
+                if key == "step" and isinstance(param_state[key], int):
+                    param_state[key] = 0
+                else:
+                    param_state[key] = torch.zeros_like(param_state[key])
+
+
+class ReLoRACallback(TrainerCallback):
+    """Callback to merge LoRA weights into the base model and save full-weight checkpoints"""
+
+    def __init__(self, cfg: DictDefault):
+        self.relora_steps = cfg.relora_steps
+        self.cpu_offload = cfg.relora_cpu_offload
+        self.quantized = cfg.load_in_4bit or cfg.load_in_8bit
+        self.last_full_model = cfg.base_model
+        self.resume_from_checkpoint = cfg.resume_from_checkpoint
+
+        if not os.path.exists(self.last_full_model):
+            self.last_full_model = str(Path(snapshot_download(cfg.base_model)))
+
+        assert os.path.exists(
+            self.last_full_model
+        ), "for ReLORA base_model must be a local path"
+
+        self.num_lora_restarts = 0
+        self.need_full_save = False
+
+    def on_train_begin(
+        self,
+        _args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        if self.resume_from_checkpoint:
+            weight_path = os.path.join(self.resume_from_checkpoint, "relora")
+            if not os.path.exists(weight_path):
+                LOG.warning(
+                    "Resuming ReLoRA from checkpoint, but no full-weight save found"
+                )
+            else:
+                LOG.info(f"Loading adjusted base weights from {weight_path}")
+                load_weight_checkpoint(model, weight_path)
+        return control
+
+    def on_step_begin(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        optimizer: torch.optim.Optimizer,
+        **_kwargs,
+    ):
+        if state.global_step > 0 and state.global_step % self.relora_steps == 0:
+            checkpoint_folder = os.path.join(
+                args.output_dir,
+                f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}",
+                "relora",
+            )
+
+            with torch.no_grad():
+                merge_and_save(
+                    model,
+                    self.last_full_model,
+                    checkpoint_folder,
+                    reinit=True,
+                    quantized=self.quantized,
+                    actually_save=is_main_process(),
+                    cpu_offload=self.cpu_offload,
+                )
+                reset_optimizer(optimizer)
+
+            if self.quantized:
+                self.last_full_model = checkpoint_folder
+            self.num_lora_restarts += 1
+
+        return control
+
+    def on_save(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        checkpoint_folder = os.path.join(
+            args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}", "relora"
+        )
+        if (
+            state.global_step >= self.relora_steps
+            and state.global_step % self.relora_steps != 0
+        ):
+            if self.quantized:
+                if is_main_process() and self.last_full_model != checkpoint_folder:
+                    # ensure the latest full parameter save is in the latest checkpoint
+                    # folder, so that automatic pruning of checkpoints does not remove it
+                    LOG.info(f"moving last full parameter save to {checkpoint_folder}")
+                    os.makedirs(checkpoint_folder, exist_ok=True)
+                    chunks = glob.glob(
+                        f"{self.last_full_model}/model*.safetensors"
+                    ) + glob.glob(f"{self.last_full_model}/model*.index.json")
+                    for path in chunks:
+                        new_path = os.path.abspath(shutil.move(path, checkpoint_folder))
+                        try:
+                            os.symlink(new_path, path)
+                        except OSError:
+                            # probably on windows without permission to symlink
+                            pass
+
+                    self.last_full_model = checkpoint_folder
+            else:
+                model.model.save_pretrained(checkpoint_folder, safe_serialization=True)
+
+        return control
+
+    def on_log(
+        self,
+        _args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        logs: Dict[str, float],
+        **_kwargs,
+    ):
+        logs["num_lora_restarts"] = self.num_lora_restarts
+        return control
+
+    def on_train_end(
+        self,
+        args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        if self.quantized:
+            # perform final merge and save
+            with torch.no_grad():
+                merge_and_save(
+                    model,
+                    self.last_full_model,
+                    args.output_dir,
+                    reinit=False,
+                    quantized=self.quantized,
+                    actually_save=is_main_process(),
+                    cpu_offload=self.cpu_offload,
+                )
+        # no need to save if unquantized, as finetune.py will call merge_and_unload()
+        return control
+
+
+class ReLoRAScheduler(LRScheduler):
+    """Wraps another scheduler to apply per-lora-restart learning rate warmups."""
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        inner_schedule: LRScheduler,
+        relora_steps: int,
+        warmup_steps: int,
+        min_lr_scale: float = 0.001,
+    ) -> None:
+        self.inner_schedule = inner_schedule
+        self.relora_steps = relora_steps
+        self.warmup_steps = warmup_steps
+        self.min_lr_scale = min_lr_scale
+        super().__init__(optimizer, inner_schedule.last_epoch, inner_schedule.verbose)
+
+    def get_lr(self) -> float:
+        self.inner_schedule.last_epoch = self.last_epoch
+
+        original = self.inner_schedule.get_lr()
+        step = self.last_epoch
+        if step < self.relora_steps:
+            scale = 1
+        else:
+            cycle_t = min(1.0, (step % self.relora_steps) / self.warmup_steps)
+            scale = cycle_t * (1 - self.min_lr_scale) + self.min_lr_scale
+
+        if isinstance(original, Sequence):
+            return [lr * scale for lr in original]
+        return original * scale
+
+
+def sharded_paths(path: str, module_names: List[str]) -> Dict[str, str]:
+    model_name = "model.safetensors"
+    if not os.path.exists(str(Path(path) / model_name)) and not os.path.exists(
+        str(Path(path) / f"{model_name}.index.json")
+    ):
+        model_name = "pytorch_model.bin"
+
+    index_path = str(Path(path) / f"{model_name}.index.json")
+    if os.path.exists(index_path):
+        with open(index_path, "r", encoding="utf-8") as file:
+            data = json.load(file)
+        return data["weight_map"]
+    return {(module_name + ".weight"): model_name for module_name in module_names}
+
+
+def lora_delta_weight(layer: peft.tuners.lora.LoraLayer, device) -> torch.Tensor:
+    if isinstance(layer, (peft.tuners.lora.Linear8bitLt, peft.tuners.lora.Linear4bit)):
+        adapter = layer.active_adapter
+        return (
+            peft.utils.transpose(
+                layer.lora_B[adapter].weight.detach().to(device)
+                @ layer.lora_A[adapter].weight.detach().to(device),
+                getattr(layer, "fan_in_fan_out", False),
+            )
+            * layer.scaling[adapter]
+        )
+
+    return layer.get_delta_weight().to(device)
+
+
+def find_lora_modules(model: peft.LoraModel) -> Dict[str, peft.tuners.lora.LoraLayer]:
+    modules: Dict[str, peft.tuners.lora.LoraLayer] = {}
+
+    key_list = [key for key, _ in model.model.named_modules() if "lora" not in key]
+    for key in key_list:
+        try:
+            # pylint: disable=protected-access
+            _parent, target, _target_name = peft.utils._get_submodules(model.model, key)
+        except AttributeError:
+            continue
+
+        if isinstance(target, peft.tuners.lora.LoraLayer):
+            modules[key] = target
+
+    return modules
+
+
+def update_weights(
+    target: peft.tuners.lora.LoraLayer, new_weight: torch.Tensor, reinit: bool, device
+):
+    if reinit:
+        for adapter_name in target.lora_A:
+            target.reset_lora_parameters(adapter_name)
+        for adapter_name in target.lora_embedding_A:
+            target.reset_lora_parameters(adapter_name)
+
+    if isinstance(target, peft.tuners.lora.Linear4bit):
+        # This could be faster, but the quantization of Linear4bit weights occurs
+        # when the module is moved from cpu to gpu. Without meddling *too* deeply in
+        # PEFT's innards or maintaining a duplicate of that codepath, this is good
+        # enough for now.
+        target.weight.quant_state = None
+        target.weight.data = new_weight.cpu()
+        target.to(device)
+    elif isinstance(target, peft.tuners.lora.Linear8bitLt):
+        target.weight = bnb.nn.Int8Params(new_weight, requires_grad=False).to(device)
+    else:
+        target.weight.data = new_weight.to(device)
+
+
+def merge_and_save(
+    model: peft.LoraModel,
+    model_src: str,
+    model_dst: str,
+    reinit: bool = False,
+    quantized: bool = False,
+    cpu_offload: bool = False,
+    actually_save: bool = True,
+):
+    modules = find_lora_modules(model)
+
+    if not quantized:
+        for module_name, target in modules.items():
+            update = target.get_delta_weight(target.active_adapter).detach()
+            target.weight.data += update
+
+            if reinit:
+                for adapter_name in target.lora_A:
+                    target.reset_lora_parameters(adapter_name)
+                for adapter_name in target.lora_embedding_A:
+                    target.reset_lora_parameters(adapter_name)
+        return
+
+    os.makedirs(model_dst, exist_ok=True)
+    shard_paths = sharded_paths(model_src, modules.keys())
+    out_shard_paths = {}
+
+    unique_shards = list(set(shard_paths.values()))
+    for shard_path in unique_shards:
+        out_tensors = {}
+        if shard_path.endswith(".safetensors"):
+            in_tensors = st.load_file(str(Path(model_src) / shard_path))
+        else:
+            in_tensors = torch.load(Path(model_src) / shard_path)
+            if "state_dict" in in_tensors:
+                in_tensors = in_tensors["state_dict"]
+
+        for module_name, target in modules.items():
+            key = module_name + ".weight"
+            if key not in shard_paths or shard_paths[key] != shard_path:
+                continue
+
+            orig_weight = in_tensors[key]
+            old_dev = target.weight.device
+            math_dev = "cpu" if cpu_offload else old_dev
+
+            delta_weight = lora_delta_weight(target, math_dev)
+            new_weight = orig_weight.to(math_dev) + delta_weight
+            del delta_weight
+
+            if actually_save:
+                out_tensors[key] = new_weight.half().cpu()
+
+            update_weights(target, new_weight, reinit=reinit, device=old_dev)
+
+        if actually_save:
+            out_shard_name = shard_path
+            if out_shard_name.startswith("pytorch_model"):
+                out_shard_name = (
+                    out_shard_name.replace("pytorch_model", "model").rstrip(".bin")
+                    + ".safetensors"
+                )
+
+            for module_name in in_tensors:
+                if module_name not in out_tensors:
+                    out_tensors[module_name] = in_tensors[module_name].half()
+                out_shard_paths[module_name] = out_shard_name
+
+            shard_fn = str(Path(model_dst) / out_shard_name)
+            LOG.info(f"saving tensors to {shard_fn}")
+            st.save_file(out_tensors, shard_fn, metadata={"format": "pt"})
+
+        del in_tensors
+        del out_tensors
+        torch.cuda.empty_cache()
+
+    if actually_save and len(unique_shards) > 1:
+        with open(
+            str(Path(model_dst, "model.safetensors.index.json")), "w", encoding="utf-8"
+        ) as file:
+            json.dump({"metadata": {}, "weight_map": out_shard_paths}, file)
+
+
+def load_weight_checkpoint(model: peft.LoraModel, checkpoint_path: str):
+    modules = find_lora_modules(model)
+    shard_paths = sharded_paths(checkpoint_path, modules.keys())
+    unique_shards = list(set(shard_paths.values()))
+
+    for shard_path in unique_shards:
+        tensors = st.load_file(os.path.join(checkpoint_path, shard_path))
+
+        for module_name, target in modules.items():
+            key = module_name + ".weight"
+            if key not in shard_paths or shard_paths[key] != shard_path:
+                continue
+
+            new_weight = tensors[key]
+            update_weights(
+                target, new_weight, reinit=False, device=target.weight.device
+            )

src/axolotl/prompt_tokenizers.py

@@ -13,7 +13,7 @@ from axolotl.prompters import IGNORE_TOKEN_ID
 LOG = logging.getLogger("axolotl")
 
 IGNORE_INDEX = -100
-LLAMA_DEFAULT_PAD_TOKEN = "[PAD]"  # nosec
+LLAMA_DEFAULT_PAD_TOKEN = "<pad>"  # nosec
 LLAMA_DEFAULT_EOS_TOKEN = "</s>"  # nosec
 LLAMA_DEFAULT_BOS_TOKEN = "<s>"  # nosec
 LLAMA_DEFAULT_UNK_TOKEN = "<unk>"  # nosec

src/axolotl/prompters.py

@@ -309,10 +309,6 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
         )
 
     def build_prompt(self, source) -> Generator[str, None, None]:
-        # ignore the system prompt if provided
-        if source[0]["from"] == "system":
-            source.pop(0)
-
         if len(source) < 2:
             # If there isn't a back and forth conversation, ignore it
             # also happens on the data splitting leaving empty conversations
@@ -321,6 +317,12 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
             )
 
         conv = self._conversation.copy()
+
+        # Add the conversation system prompt if provided, otherwise use the default one
+        if source[0]["from"] == "system":
+            conv.system = source[0]["value"]
+            source.pop(0)
+
         roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
 
         try:

src/axolotl/train.py

@@ -0,0 +1,139 @@
+"""Prepare and train a model on a dataset. Can also infer from a model or merge lora"""
+
+import logging
+import os
+import signal
+import sys
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional
+
+import torch
+
+# add src to the pythonpath so we don't need to pip install this
+from datasets import Dataset
+from optimum.bettertransformer import BetterTransformer
+
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.logging_config import configure_logging
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.models import load_model, load_tokenizer
+from axolotl.utils.trainer import setup_trainer
+
+project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
+src_dir = os.path.join(project_root, "src")
+sys.path.insert(0, src_dir)
+
+configure_logging()
+LOG = logging.getLogger("axolotl.train")
+
+
+@dataclass
+class TrainDatasetMeta:
+    """
+    dataclass to capture the dataset specific options for training
+    """
+
+    train_dataset: Dataset
+    eval_dataset: Optional[Dataset] = None
+    total_num_steps: Optional[int] = None
+
+
+def train(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
+    dataset_meta: TrainDatasetMeta,
+):
+    # load the tokenizer first
+    LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
+    tokenizer = load_tokenizer(cfg)
+
+    train_dataset = dataset_meta.train_dataset
+    eval_dataset = dataset_meta.eval_dataset
+    total_num_steps = dataset_meta.total_num_steps
+
+    # Load the model and tokenizer
+    LOG.info("loading model and (optionally) peft_config...")
+    model, peft_config = load_model(cfg, tokenizer, inference=cli_args.inference)
+
+    safe_serialization = cfg.save_safetensors is True
+
+    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
+        possible_checkpoints = [
+            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
+        ]
+        if len(possible_checkpoints) > 0:
+            sorted_paths = sorted(
+                possible_checkpoints,
+                key=lambda path: int(path.split("-")[-1]),
+            )
+            cfg.resume_from_checkpoint = sorted_paths[-1]
+            LOG.info(
+                f"Using Auto-resume functionality to start with checkpoint at {cfg.resume_from_checkpoint}"
+            )
+    resume_from_checkpoint = cfg.resume_from_checkpoint
+
+    trainer = setup_trainer(
+        cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
+    )
+
+    model.config.use_cache = False
+
+    if torch.__version__ >= "2" and sys.platform != "win32":
+        LOG.info("Compiling torch model")
+        model = torch.compile(model)
+
+    # go ahead and presave, so we have the adapter config available to inspect
+    if peft_config:
+        LOG.info(f"Pre-saving adapter config to {cfg.output_dir}")
+        peft_config.save_pretrained(cfg.output_dir)
+
+    # In case we want to stop early with ctrl+c, this is a nice to have to save the pretrained model
+    if cfg.local_rank == 0:
+
+        def terminate_handler(_, __, model):
+            if cfg.flash_optimum:
+                model = BetterTransformer.reverse(model)
+            model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
+            sys.exit(0)
+
+        signal.signal(
+            signal.SIGINT, lambda signum, frame: terminate_handler(signum, frame, model)
+        )
+
+    LOG.info("Starting trainer...")
+    if cfg.group_by_length:
+        LOG.info("hang tight... sorting dataset for group_by_length")
+
+    if not Path(cfg.output_dir).is_dir():
+        os.makedirs(cfg.output_dir, exist_ok=True)
+    tokenizer.save_pretrained(cfg.output_dir)
+    if cfg.flash_optimum:
+        with torch.backends.cuda.sdp_kernel(
+            enable_flash=True, enable_math=True, enable_mem_efficient=True
+        ):
+            trainer.train(resume_from_checkpoint=resume_from_checkpoint)
+    else:
+        trainer.train(resume_from_checkpoint=resume_from_checkpoint)
+
+    LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
+
+    if cfg.relora_steps:
+        if cfg.adapter == "lora" and not (cfg.load_in_4bit or cfg.load_in_8bit):
+            model = model.merge_and_unload()
+        else:
+            # final model weights have already been saved by `ReLoRACallback.on_train_end`
+            return model, tokenizer
+
+    # TODO do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
+    # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple processes attempt to write the same file
+    if cfg.fsdp:
+        trainer.save_model(cfg.output_dir)
+    elif cfg.local_rank == 0:
+        if cfg.flash_optimum:
+            model = BetterTransformer.reverse(model)
+
+        model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
+
+    return model, tokenizer

src/axolotl/utils/callbacks.py

@@ -1,9 +1,19 @@
 """Callbacks for Trainer class"""
 
+from __future__ import annotations
+
 import logging
 import os
+from typing import TYPE_CHECKING, Dict, List
 
+import evaluate
+import numpy as np
+import pandas as pd
+import torch
+import torch.distributed as dist
+from datasets import load_dataset
 from optimum.bettertransformer import BetterTransformer
+from tqdm import tqdm
 from transformers import (
     TrainerCallback,
     TrainerControl,
@@ -13,8 +23,20 @@ from transformers import (
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, IntervalStrategy
 
 from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.distributed import (
+    barrier,
+    gather_scalar_from_all_ranks,
+    get_world_size,
+    is_distributed,
+    is_main_process,
+    zero_first,
+)
+
+if TYPE_CHECKING:
+    from axolotl.utils.trainer import AxolotlTrainingArguments
 
 LOG = logging.getLogger("axolotl.callbacks")
+IGNORE_INDEX = -100
 
 
 class SavePeftModelCallback(TrainerCallback):  # pylint: disable=too-few-public-methods
@@ -33,7 +55,9 @@ class SavePeftModelCallback(TrainerCallback):  # pylint: disable=too-few-public-
         )
 
         peft_model_path = os.path.join(checkpoint_folder, "adapter_model")
-        kwargs["model"].save_pretrained(peft_model_path)
+        kwargs["model"].save_pretrained(
+            peft_model_path, save_safetensors=args.save_safetensors
+        )
 
         return control
 
@@ -94,3 +118,202 @@ class GPUStatsCallback(
             log_gpu_memory_usage(LOG, "while training", self.cfg.device)
             self.logged = True
         return control
+
+
+def bench_eval_callback_factory(trainer, tokenizer):
+    accuracy = evaluate.load("accuracy")
+    abcd_idx = [
+        tokenizer("A", add_special_tokens=False).input_ids[0],
+        tokenizer("B", add_special_tokens=False).input_ids[0],
+        tokenizer("C", add_special_tokens=False).input_ids[0],
+        tokenizer("D", add_special_tokens=False).input_ids[0],
+        tokenizer("E", add_special_tokens=False).input_ids[0],
+        tokenizer("F", add_special_tokens=False).input_ids[0],
+        tokenizer("G", add_special_tokens=False).input_ids[0],
+    ]
+    bench_split = "eval"
+
+    def transform_bench_subject(example):
+        # Split on ':' and trim whitespace
+        parts = example["subject"].split(":")
+        first_part = (
+            parts[0].strip().lower().replace("-", "_")
+        )  # Lowercase the first part
+        second_part = (
+            parts[1].strip().replace("-", "_") if len(parts) > 1 else "all"
+        )  # Replace hyphens with underscores
+
+        # Return the transformed values
+        return {"name": first_part, "subject": second_part}
+
+    if trainer.args.bench_dataset == "mmlu-zs":
+        bench_dataset = load_dataset(
+            "openaccess-ai-collective/mmlu-evals",
+            data_files={
+                "eval": "zero_shot_mmlu_val.json",
+                "test": "zero_shot_mmlu_test.json",
+            },
+        )
+        # bench_dataset = bench_dataset.remove_columns("subject")
+    # MMLU Five-shot (Eval/Test only)
+    elif trainer.args.bench_dataset in ["mmlu", "mmlu-fs"]:
+        bench_dataset = load_dataset(
+            "openaccess-ai-collective/mmlu-evals",
+            data_files={
+                "eval": "five_shot_mmlu_val.json",
+                "test": "five_shot_mmlu_test.json",
+            },
+        )
+        # bench_dataset = bench_dataset.remove_columns('subject')
+    elif "/" in trainer.args.bench_dataset:
+        bench_ds = trainer.args.bench_dataset
+        bench_ds_name = "/".join(bench_ds.split("/", 2)[:2])
+        bench_ds_data_file = "/".join(bench_ds.split("/", 2)[2:])
+        bench_dataset = load_dataset(
+            bench_ds_name,
+            data_files={
+                "eval": bench_ds_data_file,
+            },
+        )
+        bench_dataset["eval"] = bench_dataset["eval"].map(transform_bench_subject)
+    else:
+        raise ValueError(
+            f"unhandled value `{trainer.args.bench_dataset}` for bench_dataset training args"
+        )
+    bench_dataset = bench_dataset[trainer.args.bench_split]
+    if trainer.args.max_bench_samples is not None:
+        bench_dataset = bench_dataset.select(range(trainer.args.max_bench_samples))
+
+    def tokenize_evals(example):
+        source = f"{tokenizer.bos_token}{example['input']}"
+        target = f"{example['output']}{tokenizer.eos_token}"
+
+        tokenized_source = tokenizer(
+            source,
+            max_length=2048,
+            truncation=True,
+            add_special_tokens=False,
+        )
+        tokenized_target = tokenizer(
+            target,
+            max_length=2048,
+            truncation=True,
+            add_special_tokens=False,
+        )
+        input_ids = tokenized_source["input_ids"] + tokenized_target["input_ids"]
+        labels = [IGNORE_INDEX] * len(tokenized_source["input_ids"]) + tokenized_target[
+            "input_ids"
+        ]
+
+        return {
+            "input_ids": input_ids,
+            "labels": labels,
+            "subject": example["subject"],
+        }
+
+    with zero_first(is_main_process()):
+        bench_dataset = bench_dataset.map(tokenize_evals)
+        bench_dataset = bench_dataset.filter(lambda x: x["labels"][-2] in abcd_idx)
+
+    class BenchEvalCallback(TrainerCallback):
+        """
+        TrainerCallback that runs the MMLU evals
+        """
+
+        def on_evaluate(
+            self,
+            args: AxolotlTrainingArguments,
+            state: TrainerState,  # pylint: disable=unused-argument
+            control: TrainerControl,  # pylint: disable=unused-argument
+            metrics: Dict[str, float],  # pylint: disable=unused-argument
+            **kwargs,  # pylint: disable=unused-argument
+        ):
+            data_loader = trainer.get_bench_dataloader(
+                bench_dataset.remove_columns(["input", "subject", "output", "name"])
+            )
+            trainer.model.eval()
+            preds, refs = [], []
+            loss_bench = 0
+            for batch in tqdm(data_loader, total=len(data_loader)):
+                (loss, logits, labels) = trainer.prediction_step(
+                    trainer.model,
+                    batch,
+                    prediction_loss_only=False,
+                )
+                # There are two tokens, the output, and eos token.
+                for i, logit in enumerate(logits):
+                    label_non_zero_id = (batch["labels"][i] != IGNORE_INDEX).nonzero()[
+                        0
+                    ][0]
+                    logit_abcd = logit[label_non_zero_id - 1][abcd_idx]
+                    preds.append(torch.argmax(logit_abcd).item())
+                labels = labels[labels != IGNORE_INDEX].view(-1, 2)[:, 0]
+                refs += [
+                    abcd_idx.index(label) if label in abcd_idx else -1
+                    for label in labels.tolist()
+                ]
+                loss_bench += loss.item()
+            # Extract results by subject.
+            bench_name = bench_dataset["name"]
+            bench_names: dict = {s: {"refs": [], "preds": []} for s in set(bench_name)}
+            for s, p, r in zip(bench_name, preds, refs):  # pylint: disable=invalid-name
+                bench_names[s]["preds"].append(p)
+                bench_names[s]["refs"].append(r)
+            barrier()
+            local_bench_names = bench_names
+            gathered_bench_names: List[Dict] = [{} for _ in range(get_world_size())]
+            # Gather results from all GPUs to GPU 0
+
+            loss_bench_ranks = gather_scalar_from_all_ranks(
+                lambda: loss_bench, get_world_size()
+            )
+            len_data_loader_ranks = gather_scalar_from_all_ranks(
+                lambda: len(data_loader), get_world_size()
+            )
+
+            if is_distributed() and not is_main_process():
+                dist.gather_object(local_bench_names, dst=0)
+            else:
+                if is_distributed():
+                    dist.gather_object(local_bench_names, gathered_bench_names, dst=0)
+                else:
+                    gathered_bench_names = [local_bench_names]
+                bench_loss = sum(loss_bench_ranks) / sum(len_data_loader_ranks)
+                results = {f"{bench_split}_bench_loss": bench_loss}
+
+                # Combine results from all GPUs
+                combined_bench_names: Dict[str, Dict[str, List]] = {}
+                for bench_name in gathered_bench_names:
+                    for name, data in bench_name.items():
+                        if name not in combined_bench_names:
+                            combined_bench_names[name] = {"refs": [], "preds": []}
+                        combined_bench_names[name]["refs"].extend(data["refs"])
+                        combined_bench_names[name]["preds"].extend(data["preds"])
+
+                bench_scores = []
+                bench_refs = []
+                bench_preds = []
+                for (
+                    bench_name
+                ) in combined_bench_names:  # pylint: disable=consider-using-dict-items
+                    bench_score = accuracy.compute(
+                        references=combined_bench_names[bench_name]["refs"],
+                        predictions=combined_bench_names[bench_name]["preds"],
+                    )["accuracy"]
+                    bench_refs.extend(combined_bench_names[bench_name]["refs"])
+                    bench_preds.extend(combined_bench_names[bench_name]["preds"])
+                    if not pd.isna(bench_score):
+                        results[
+                            f"{bench_split}_bench_accuracy_{bench_name}"
+                        ] = bench_score
+                        bench_scores.append(bench_score)
+                    else:
+                        results[f"{bench_split}_bench_accuracy_{bench_name}"] = 0.0
+                        bench_scores.append(0.0)
+                results[f"{bench_split}_bench_average_accuracy"] = np.mean(bench_scores)
+                results[f"{bench_split}_bench_total_accuracy"] = accuracy.compute(
+                    references=bench_refs, predictions=bench_preds
+                )["accuracy"]
+                trainer.log(results)
+
+    return BenchEvalCallback

src/axolotl/utils/config.py

@@ -6,6 +6,7 @@ import os
 import torch
 
 from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.models import load_model_config
 
 LOG = logging.getLogger("axolotl")
 
@@ -69,6 +70,16 @@ def normalize_config(cfg):
     else:
         cfg.torch_dtype = torch.float32
 
+    model_config = load_model_config(cfg)
+
+    # figure out if the model is llama
+    cfg.is_llama_derived_model = (
+        (hasattr(model_config, "model_type") and model_config.model_type == "llama")
+        or cfg.is_llama_derived_model
+        or "llama" in cfg.base_model
+        or (cfg.model_type and "llama" in cfg.model_type.lower())
+    )
+
     log_gpu_memory_usage(LOG, "baseline", cfg.device)
 
 
@@ -97,9 +108,7 @@ def validate_config(cfg):
             "To calculate the equivalent gradient_accumulation_steps, divide batch_size / micro_batch_size / number of gpus.",
         )
     if cfg.load_4bit:
-        raise ValueError(
-            "cfg.load_4bit parameter has been deprecated and replaced by cfg.gptq"
-        )
+        raise ValueError("cfg.load_4bit parameter has been deprecated")
 
     if cfg.adapter == "qlora":
         if cfg.merge_lora:
@@ -126,6 +135,19 @@ def validate_config(cfg):
     if not cfg.load_in_8bit and cfg.adapter == "lora":
         LOG.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
 
+    if cfg.relora_steps:
+        if cfg.adapter not in ("lora", "qlora"):
+            raise ValueError("cfg.adapter must be lora or qlora to use ReLoRA")
+
+        if cfg.fsdp:
+            raise ValueError("fsdp not supported with ReLoRA")
+
+        if cfg.deepspeed:
+            raise ValueError("deepspeed not supported with ReLoRA")
+
+        if cfg.lr_scheduler == "one_cycle":
+            raise ValueError("ReLoRA is not compatible with the one_cycle scheduler")
+
     if cfg.trust_remote_code:
         LOG.warning(
             "`trust_remote_code` is set to true. Please make sure that you reviewed the remote code/model."

src/axolotl/utils/data.py

@@ -54,9 +54,10 @@ DEFAULT_DATASET_PREPARED_PATH = "last_run_prepared"
 
 def prepare_dataset(cfg, tokenizer):
     if not cfg.pretraining_dataset:
-        train_dataset, eval_dataset = load_prepare_datasets(
-            tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
-        )
+        with zero_first(is_main_process()):
+            train_dataset, eval_dataset = load_prepare_datasets(
+                tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
+            )
     else:
         train_dataset = load_pretraining_dataset(
             cfg.pretraining_dataset,
@@ -133,8 +134,17 @@ def load_tokenized_prepared_datasets(
             seed = 42
 
         datasets = []
+
+        def for_d_in_datasets(dataset_configs):
+            for dataset in dataset_configs:
+                if dataset.name and isinstance(dataset.name, list):
+                    for name in dataset.name:
+                        yield DictDefault({**dataset, "name": name})
+                else:
+                    yield dataset
+
         # pylint: disable=invalid-name
-        for d in cfg.datasets:
+        for d in for_d_in_datasets(cfg.datasets):
             ds: Union[Dataset, DatasetDict] = None
             ds_from_hub = False
             try:

src/axolotl/utils/dataloader.py

@@ -243,6 +243,18 @@ class MultipackDistributedDataloader:
             len_remaining -= 1
             if not len_remaining:
                 return
+        # yield a no-op for cases where we don't have any data left to pack
+        for i in range(0, len_remaining):
+            yield self.collate_fn(
+                [
+                    {
+                        "input_ids": [0],
+                        "labels": [-100],
+                        "attention_mask": [True],
+                        "position_ids": [0],
+                    }
+                ]
+            )
 
     def _len_est(self):
         lengths_sum = np.sum(self.lengths)

src/axolotl/utils/distributed.py

@@ -1,8 +1,10 @@
 """
 utility helpers for distributed checks
 """
+import os
 from contextlib import contextmanager
 
+import torch
 import torch.distributed as dist
 from accelerate import Accelerator
 
@@ -43,6 +45,10 @@ def is_main_process():
     return dist.get_rank() == 0
 
 
+def get_world_size():
+    return int(os.getenv("WORLD_SIZE", "1"))
+
+
 @contextmanager
 def zero_first(is_main):
     """
@@ -53,3 +59,37 @@ def zero_first(is_main):
     yield
     if is_main:  # then rank 0 waits after it has run the context
         barrier()
+
+
+def gather_scalar_from_all_ranks(fn, world_size=1):  # pylint: disable=invalid-name
+    """
+    Run a callable 'fn' on all ranks and gather the results on the specified rank.
+
+    Args:
+    - fn (callable): A function that computes the value. This should not have any side effects.
+    - rank (int, optional): The rank that gathers the values. Default is 0.
+    - world_size (int, optional): Total number of processes in the current distributed setup.
+
+    Returns:
+    - A list of computed values from all ranks if on the gathering rank, otherwise None.
+    """
+    value_scalar = fn()
+    if not is_distributed():
+        return [value_scalar]
+    value_tensor = torch.tensor(value_scalar, device=dist.get_rank()).float()
+
+    if not is_main_process():
+        dist.gather(value_tensor, dst=0)
+    else:
+        gathered_tensors = [torch.zeros_like(value_tensor) for _ in range(world_size)]
+        dist.gather(value_tensor, gather_list=gathered_tensors, dst=0)
+
+        # Convert tensors back to their original type (int or float)
+        gathered_values = []
+        for tensor in gathered_tensors:
+            if tensor == tensor.int():
+                gathered_values.append(int(tensor.item()))
+            else:
+                gathered_values.append(float(tensor.item()))
+        return gathered_values
+    return None

src/axolotl/utils/models.py

@@ -4,33 +4,37 @@
 import logging
 import math
 import os
-from pathlib import Path
-from typing import TYPE_CHECKING, Optional, Tuple  # noqa: F401
+from typing import Optional, Tuple  # noqa: F401
 
 import bitsandbytes as bnb
 import torch
 import transformers
 from optimum.bettertransformer import BetterTransformer
-from peft.tuners.lora import LoraLayer
+from peft import PeftConfig, prepare_model_for_kbit_training
 from transformers import (  # noqa: F401
     AutoConfig,
     AutoModelForCausalLM,
     AutoTokenizer,
     BitsAndBytesConfig,
+    GPTQConfig,
     LlamaConfig,
     PreTrainedModel,
     PreTrainedTokenizerBase,
 )
 
-from axolotl.prompt_tokenizers import LLAMA_DEFAULT_PAD_TOKEN
+from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
 from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.dict import DictDefault
 
 LOG = logging.getLogger("axolotl")
 
-if TYPE_CHECKING:
-    from peft import PeftConfig  # noqa: F401
 
-    from axolotl.utils.dict import DictDefault  # noqa: F401
+def load_model_config(cfg):
+    model_config_name = cfg.base_model_config or cfg.base_model
+    trust_remote_code: bool = False or cfg.trust_remote_code
+    return AutoConfig.from_pretrained(
+        model_config_name, trust_remote_code=trust_remote_code
+    )
 
 
 def load_tokenizer(cfg):
@@ -55,11 +59,18 @@ def load_tokenizer(cfg):
         **tokenizer_kwargs,
     )
 
-    if tokenizer.__class__.__name__ in [
-        "LlamaTokenizer",
-        "LlamaTokenizerFast",
-    ]:
-        tokenizer.pad_token = LLAMA_DEFAULT_PAD_TOKEN
+    if (
+        tokenizer.__class__.__name__
+        in [
+            "LlamaTokenizer",
+            "LlamaTokenizerFast",
+            "CodeLlamaTokenizer",
+        ]
+        and hasattr(tokenizer, "pad_token")
+        and not tokenizer.pad_token
+    ):
+        # set a pad_token, but use eos_token so we don't add a new token
+        tokenizer.pad_token = LLAMA_DEFAULT_EOS_TOKEN
 
     LOG.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
     LOG.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
@@ -80,8 +91,10 @@ def load_tokenizer(cfg):
 
 
 def load_model(
-    cfg, tokenizer
-):  # type: (DictDefault, PreTrainedTokenizerBase) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
+    cfg: DictDefault,
+    tokenizer: PreTrainedTokenizerBase,
+    inference: bool = False,
+) -> Tuple[PreTrainedModel, Optional[PeftConfig]]:
     """
     Load a model for a given configuration and tokenizer.
     """
@@ -91,14 +104,9 @@ def load_model(
 
     # TODO refactor as a kwarg
     load_in_8bit = cfg.load_in_8bit
-    cfg.is_llama_derived_model = (
-        "llama" in base_model
-        or (cfg.model_type and "llama" in cfg.model_type.lower())
-        or cfg.is_llama_derived_model
-    )
 
     if cfg.is_llama_derived_model and cfg.flash_attention:
-        if cfg.device not in ["mps", "cpu"] and not cfg.inference:
+        if cfg.device not in ["mps", "cpu"] and not inference:
             from axolotl.monkeypatch.llama_attn_hijack_flash import (
                 replace_llama_attn_with_flash_attn,
             )
@@ -140,39 +148,24 @@ def load_model(
     if (
         cfg.is_llama_derived_model
         and (cfg.max_packed_sequence_len or cfg.sample_packing)
-        and not cfg.inference
+        and not inference
     ):
         from axolotl.monkeypatch.llama_expand_mask import hijack_expand_mask
 
         LOG.info("patching _expand_mask")
         hijack_expand_mask()
 
-    try:
-        if cfg.gptq:
-            from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
-                replace_peft_model_with_int4_lora_model,
-            )
-
-            replace_peft_model_with_int4_lora_model()
-    except Exception as err:
-        LOG.exception(err)
-        raise err
-
-    if not cfg.gptq and (
-        (cfg.adapter == "lora" and load_in_8bit)
-        or (cfg.adapter == "qlora" and cfg.load_in_4bit)
-    ):
-        try:
-            from peft import prepare_model_for_kbit_training
-        except ImportError:
-            # For backward compatibility
-            from peft import (
-                prepare_model_for_int8_training as prepare_model_for_kbit_training,
-            )
-
     model_kwargs = {}
     if cfg.model_revision:
         model_kwargs["revision"] = cfg.model_revision
+    if cfg.gptq:
+        model_config = load_model_config(cfg)
+        if hasattr(model_config, "quantization_config"):
+            LOG.warning("model config does not contain quantization_config information")
+        else:
+            model_kwargs["quantization_config"] = GPTQConfig(
+                **model_config.quantization_config
+            )
     if cfg.adapter == "qlora" and cfg.load_in_4bit:
         model_kwargs["quantization_config"] = BitsAndBytesConfig(
             load_in_4bit=True,
@@ -183,45 +176,7 @@ def load_model(
             bnb_4bit_quant_type="nf4",
         )
     try:
-        if cfg.gptq and cfg.is_llama_derived_model:
-            from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
-            from huggingface_hub import snapshot_download
-
-            try:
-                snapshot_download_kwargs = {}
-                if cfg.base_model_ignore_patterns:
-                    snapshot_download_kwargs[
-                        "ignore_patterns"
-                    ] = cfg.base_model_ignore_patterns
-                cache_model_path = Path(
-                    snapshot_download(base_model, **snapshot_download_kwargs)
-                )
-                files = (
-                    list(cache_model_path.glob("*.pt"))
-                    + list(cache_model_path.glob("*.safetensors"))
-                    + list(cache_model_path.glob("*.bin"))
-                )
-                if len(files) > 0:
-                    model_path = str(files[0])
-                else:
-                    LOG.warning(
-                        "unable to find a cached model file, this will likely fail..."
-                    )
-                    model_path = str(cache_model_path)
-            except Exception:  # pylint: disable=broad-exception-caught
-                model_path = cfg.base_model
-            model, _ = load_llama_model_4bit_low_ram(
-                base_model_config if base_model_config else base_model,
-                model_path,
-                device_map=cfg.device_map,
-                half=cfg.fp16,
-                groupsize=cfg.gptq_groupsize if cfg.gptq_groupsize else -1,
-                is_v1_model=cfg.gptq_model_v1
-                if cfg.gptq_model_v1 is not None
-                else True,
-            )
-            load_in_8bit = False
-        elif cfg.is_llama_derived_model and not cfg.trust_remote_code:
+        if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
             from transformers import LlamaForCausalLM
 
             config_kwargs = {}
@@ -267,15 +222,24 @@ def load_model(
         #     )
         #     model.train() # sets to train instead of eval mode
         elif model_type and not cfg.trust_remote_code:
-            model = getattr(transformers, model_type).from_pretrained(
-                base_model,
-                device_map=cfg.device_map,
-                load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
-                load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-                torch_dtype=cfg.torch_dtype,
-                trust_remote_code=cfg.trust_remote_code or False,
-                **model_kwargs,
-            )
+            if cfg.gptq:
+                model = AutoModelForCausalLM.from_pretrained(
+                    base_model,
+                    device_map=cfg.device_map,
+                    torch_dtype=cfg.torch_dtype,
+                    trust_remote_code=cfg.trust_remote_code or False,
+                    **model_kwargs,
+                )
+            else:
+                model = getattr(transformers, model_type).from_pretrained(
+                    base_model,
+                    device_map=cfg.device_map,
+                    load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
+                    load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
+                    torch_dtype=cfg.torch_dtype,
+                    trust_remote_code=cfg.trust_remote_code or False,
+                    **model_kwargs,
+                )
         else:
             config = AutoConfig.from_pretrained(
                 base_model,
@@ -342,36 +306,46 @@ def load_model(
     if model.device.type == "cuda":
         log_gpu_memory_usage(LOG, "after model load", model.device)
 
-    if not cfg.gptq and (
-        (cfg.adapter == "lora" and load_in_8bit)
-        or (cfg.adapter == "qlora" and cfg.load_in_4bit)
+    # make sure these are fp32 per Ramesh et al. (2021)
+    for name, module in model.named_modules():
+        if "norm" in name:
+            module.to(torch.float32)
+        if "lm_head" in name or "embed_tokens" in name:
+            if hasattr(module, "weight"):
+                module.to(torch.float32)
+
+    needs_fa2_dtype = cfg.adapter or cfg.fsdp
+    if (cfg.adapter == "lora" and load_in_8bit) or (
+        cfg.adapter == "qlora" and cfg.load_in_4bit
     ):
         LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
+        if cfg.gradient_checkpointing:
+            model.gradient_checkpointing_enable()
         model = prepare_model_for_kbit_training(
             model, use_gradient_checkpointing=cfg.gradient_checkpointing
         )
+        needs_fa2_dtype = True
+
+    # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
+    # convert them back to fp16/bf16 for flash-attn compatibility.
+    if needs_fa2_dtype or (cfg.flash_attention and cfg.is_llama_derived_model):
+        LOG.info("converting modules to %s for flash attention", cfg.torch_dtype)
+        for name, module in model.named_modules():
+            if "norm" in name:
+                module.to(cfg.torch_dtype)
+            if "lm_head" in name or "embed_tokens" in name:
+                if hasattr(module, "weight"):
+                    module.to(cfg.torch_dtype)
 
     model, lora_config = load_adapter(model, cfg, cfg.adapter)
 
     if cfg.ddp and not load_in_8bit:
         model.to(f"cuda:{cfg.local_rank}")
 
-    if cfg.gptq:
-        # Scales to half
-        LOG.info("Fitting 4bit scales and zeros to half")
-        for _, module in model.named_modules():
-            if "Autograd4bitQuantLinear" in str(type(module)) or "Linear4bitLt" in str(
-                type(module)
-            ):
-                if hasattr(module, "is_v1_model") and module.is_v1_model:
-                    module.zeros = module.zeros.half()
-                module.scales = module.scales.half()
-                module.bias = module.bias.half()
-
     if (
         torch.cuda.device_count() > 1
         and int(os.getenv("WORLD_SIZE", "1")) > 1
-        and (cfg.gptq or cfg.load_in_4bit)
+        and (cfg.load_in_4bit)
     ):
         # llama is PROBABLY model parallelizable, but the default isn't that it is
         # so let's only set it for the 4bit, see
@@ -397,15 +371,15 @@ def load_model(
     return model, lora_config
 
 
-def load_adapter(model, cfg, adapter):
-    # type: (PreTrainedModel, DictDefault, Optional[str]) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
+def load_adapter(model, cfg, adapter, inference=False):
+    # type: (PreTrainedModel, DictDefault, Optional[str], bool) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
 
     if adapter is None:
         return model, None
     if hasattr(model, "enable_input_require_grads"):
         model.enable_input_require_grads()
     if adapter in ["lora", "qlora"]:
-        return load_lora(model, cfg)
+        return load_lora(model, cfg, inference=inference)
     if adapter == "llama-adapter":
         return load_llama_adapter(model, cfg)
 
@@ -437,12 +411,8 @@ def load_llama_adapter(model, cfg):
     return model, peft_config
 
 
-def find_all_linear_names(bits, model):
-    cls = (
-        bnb.nn.Linear4bit
-        if bits == 4
-        else (bnb.nn.Linear8bitLt if bits == 8 else torch.nn.Linear)
-    )
+def find_all_linear_names(model):
+    cls = (bnb.nn.Linear4bit, bnb.nn.Linear8bitLt, torch.nn.Linear)
     lora_module_names = set()
     for name, module in model.named_modules():
         if isinstance(module, cls):
@@ -455,21 +425,15 @@ def find_all_linear_names(bits, model):
     return list(lora_module_names)
 
 
-def load_lora(model, cfg):
-    # type: (PreTrainedModel, DictDefault) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
+def load_lora(model, cfg, inference=False):
+    # type: (PreTrainedModel, DictDefault, bool) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
 
     from peft import LoraConfig, PeftModel, get_peft_model
 
     lora_target_modules = list(cfg.lora_target_modules or [])
 
     if cfg.lora_target_linear:
-        bits = None
-        if cfg.load_in_4bit:
-            bits = 4
-        elif cfg.load_in_8bit:
-            bits = 8
-
-        linear_names = find_all_linear_names(bits, model)
+        linear_names = find_all_linear_names(model)
         LOG.info(f"found linear modules: {repr(linear_names)}")
         lora_target_modules = list(set(lora_target_modules + linear_names))
 
@@ -489,27 +453,11 @@ def load_lora(model, cfg):
         model = PeftModel.from_pretrained(
             model,
             cfg.lora_model_dir,
-            is_trainable=not cfg.inference,
+            is_trainable=(not inference),
         )
     else:
         model = get_peft_model(model, lora_config)
 
-    for name, module in model.named_modules():
-        if isinstance(module, LoraLayer):
-            module = module.to(cfg.torch_dtype)
-        if "norm" in name:
-            module = module.to(torch.float32)
-        if "lm_head" in name or "embed_tokens" in name:
-            if hasattr(module, "weight"):
-                module = module.to(cfg.torch_dtype)
-
-    # LlamaRMSNorm layers are in fp32 after kbit_training, so we need to
-    # convert them back to fp16/bf16 for flash-attn compatibility.
-    if cfg.flash_attention and cfg.is_llama_derived_model:
-        for name, module in model.named_modules():
-            if "norm" in name:
-                module = module.to(cfg.torch_dtype)
-
     model.print_trainable_parameters()
 
     return model, lora_config

src/axolotl/utils/tokenization.py

@@ -8,13 +8,13 @@ from termcolor import colored
 LOG = logging.getLogger("axolotl")
 
 
-def check_dataset_labels(dataset, tokenizer):
+def check_dataset_labels(dataset, tokenizer, num_examples=5, text_only=False):
     # the dataset is already shuffled, so let's just check the first 5 elements
-    for idx in range(5):
-        check_example_labels(dataset[idx], tokenizer)
+    for idx in range(num_examples):
+        check_example_labels(dataset[idx], tokenizer, text_only=text_only)
 
 
-def check_example_labels(example, tokenizer):
+def check_example_labels(example, tokenizer, text_only=False):
     # Get the input_ids, labels, and attention_mask from the dataset
     input_ids = example["input_ids"]
     labels = example["labels"]
@@ -29,8 +29,10 @@ def check_example_labels(example, tokenizer):
         decoded_input_token = tokenizer.decode(input_id)
         # Choose the color based on whether the label has the ignore value or not
         color = "red" if label_id == -100 else ("yellow" if label_id == 0 else "green")
-        colored_token = colored(decoded_input_token, color) + colored(
-            f"({label_id}, {mask}, {input_id})", "white"
+        colored_token = colored(decoded_input_token, color) + (
+            not text_only
+            and colored(f"({label_id}, {mask}, {input_id})", "white")
+            or ""
         )
         colored_tokens.append(colored_token)
 

src/axolotl/utils/trainer.py

@@ -10,31 +10,30 @@ from functools import partial
 from pathlib import Path
 from typing import Optional, Union
 
-import bitsandbytes as bnb
 import numpy as np
 import torch.cuda
 import transformers
 from datasets import Dataset, set_caching_enabled
-from torch import nn
 from torch.optim.lr_scheduler import OneCycleLR
-from torch.utils.data import DataLoader, DistributedSampler, RandomSampler
-from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
-from transformers.trainer_pt_utils import (
-    SequentialDistributedSampler,
-    get_parameter_names,
+from torch.utils.data import (
+    DataLoader,
+    DistributedSampler,
+    RandomSampler,
+    SequentialSampler,
 )
+from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
+from transformers.trainer_pt_utils import SequentialDistributedSampler
 
+from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils.callbacks import (
     GPUStatsCallback,
     SaveBetterTransformerModelCallback,
     SavePeftModelCallback,
+    bench_eval_callback_factory,
 )
 from axolotl.utils.collators import DataCollatorForSeq2Seq
 from axolotl.utils.dataloader import MultipackDistributedDataloader
-from axolotl.utils.schedulers import (
-    InterpolatingLogScheduler,
-    get_cosine_schedule_with_quadratic_warmup,
-)
+from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
 
 LOG = logging.getLogger("axolotl")
 
@@ -127,6 +126,35 @@ class AxolotlTrainingArguments(TrainingArguments):
         default=1,
         metadata={"help": "the multiplier for the max len for packed sequences"},
     )
+    relora_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how often to reset for ReLoRA"},
+    )
+    relora_warmup_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how many warmup steps to take after reset for ReLoRA"},
+    )
+    bench_split: Optional[str] = field(
+        default="eval", metadata={"help": "The benchmark split to run on"}
+    )
+    bench_dataset: Optional[str] = field(
+        default="pharaouk/dharma-1/dharma_1_mini.json",
+        metadata={
+            "help": "Benchmark dataset to use: options are `mmlu-zs`, `mmlu-fs`, or the full path to the dataset file"
+        },
+    )
+    do_bench_eval: Optional[bool] = field(
+        default=False, metadata={"help": "Whether to run the Benchmark evaluation."}
+    )
+    max_bench_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "If set, only evaluates on `max_bench_samples` of the benchmark dataset."
+        },
+    )
+    bench_source_max_len: int = field(
+        default=2048, metadata={"help": "Maximum source sequence length for bench."}
+    )
 
 
 class AxolotlTrainer(Trainer):
@@ -136,6 +164,10 @@ class AxolotlTrainer(Trainer):
 
     args = None  # type: AxolotlTrainingArguments
 
+    def __init__(self, *args, bench_data_collator=None, **kwargs):
+        self.bench_data_collator = bench_data_collator
+        super().__init__(*args, **kwargs)
+
     def create_scheduler(
         self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
     ):
@@ -226,6 +258,31 @@ class AxolotlTrainer(Trainer):
             )
         return super().get_eval_dataloader(eval_dataset)
 
+    def _get_bench_sampler(
+        self, bench_dataset: Dataset
+    ) -> Optional[torch.utils.data.Sampler]:
+        if self.args.world_size <= 1:
+            return SequentialSampler(bench_dataset)
+        return None
+
+    def get_bench_dataloader(
+        self,
+        bench_dataset: Dataset,
+    ) -> Union[DataLoader, MultipackDistributedDataloader]:
+        dataloader_params = {
+            "batch_size": self.args.eval_batch_size,
+            "collate_fn": self.bench_data_collator,
+            "num_workers": self.args.dataloader_num_workers,
+            "pin_memory": self.args.dataloader_pin_memory,
+        }
+
+        if not isinstance(bench_dataset, torch.utils.data.IterableDataset):
+            dataloader_params["sampler"] = self._get_bench_sampler(bench_dataset)
+            dataloader_params["drop_last"] = self.args.dataloader_drop_last
+
+        return DataLoader(bench_dataset, **dataloader_params)
+        # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
+
     def compute_loss(self, model, inputs, return_outputs=False):
         # use one's weighted cross entropy loss calc
         # if self.args.sample_packing:
@@ -265,13 +322,46 @@ class OneCycleLRSchedulerTrainer(AxolotlTrainer):
         return self.lr_scheduler
 
 
+class ReLoRATrainer(AxolotlTrainer):
+    """
+    Trainer subclass that uses the OneCycleLR scheduler
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.lr_scheduler = None
+
+    def create_scheduler(
+        self,
+        num_training_steps: int,
+        optimizer: Optional[torch.optim.Optimizer] = None,
+    ):
+        optimizer = self.optimizer if optimizer is None else optimizer
+        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
+
+        if self.args.relora_steps:
+            warmup_steps = (
+                self.args.relora_warmup_steps if self.args.relora_warmup_steps else 10
+            )
+            self.lr_scheduler = ReLoRAScheduler(
+                optimizer,
+                lr_scheduler,
+                self.args.relora_steps,
+                warmup_steps,
+            )
+        else:
+            self.lr_scheduler = lr_scheduler
+
+        return self.lr_scheduler
+
+
 def add_position_ids(sample):
     sample["position_ids"] = torch.arange(len(sample["input_ids"]))
     return sample
 
 
 def drop_long_seq(sample, sequence_len=2048):
-    return len(sample["input_ids"]) <= sequence_len
+    return len(sample["input_ids"]) <= sequence_len and len(sample["input_ids"]) > 0
 
 
 @contextmanager
@@ -311,6 +401,16 @@ def calculate_total_num_steps(cfg, train_dataset, tokenizer):
             LOG.info(f"📝 UPDATE CONFIG WITH: `total_num_tokens: {total_num_tokens}`")
             cfg.total_num_tokens = total_num_tokens
 
+        if not cfg.total_supervised_tokens:
+            total_supervised_tokens = (
+                train_dataset.data.column("labels")
+                .to_pandas()
+                .apply(lambda x: np.sum(np.array(x) != -100))
+                .sum()
+            )
+            LOG.info(f"`total_supervised_tokens: {total_supervised_tokens}`")
+            cfg.total_supervised_tokens = total_supervised_tokens
+
         if cfg.sample_packing_eff_est:
             total_num_steps = (
                 # match count to len est in dataloader
@@ -414,23 +514,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         training_arguments_kwargs["seed"] = cfg.seed
 
     if cfg.gradient_checkpointing:
-        if cfg.gptq:
-            from alpaca_lora_4bit.gradient_checkpointing import (
-                apply_gradient_checkpointing,
-            )
-
-            gradient_checkpointing_ratio = (
-                cfg.gradient_checkpointing_ratio
-                if cfg.gradient_checkpointing_ratio
-                else 1.0
-            )
-            apply_gradient_checkpointing(
-                model, checkpoint_ratio=gradient_checkpointing_ratio
-            )
-        else:
-            training_arguments_kwargs[
-                "gradient_checkpointing"
-            ] = cfg.gradient_checkpointing
+        training_arguments_kwargs["gradient_checkpointing"] = cfg.gradient_checkpointing
     if cfg.fsdp:
         training_arguments_kwargs["fsdp"] = cfg.fsdp
         if cfg.fsdp_config:
@@ -484,6 +568,20 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
             "steps" if cfg.save_steps else "epoch"
         )
 
+    if cfg.do_bench_eval:
+        training_arguments_kwargs["do_bench_eval"] = cfg.do_bench_eval
+        if cfg.bench_dataset:
+            training_arguments_kwargs["bench_dataset"] = cfg.bench_dataset
+
+    # DDP Config
+    if cfg.ddp_timeout:
+        training_arguments_kwargs["ddp_timeout"] = cfg.ddp_timeout
+    # see https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html
+    if cfg.ddp_bucket_cap_mb:
+        training_arguments_kwargs["ddp_bucket_cap_mb"] = cfg.ddp_bucket_cap_mb
+    if cfg.ddp_broadcast_buffers is not None:
+        training_arguments_kwargs["ddp_broadcast_buffers"] = cfg.ddp_broadcast_buffers
+
     training_args = AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
         max_steps=total_num_steps if cfg.max_steps else -1,
         max_seq_length=cfg.sequence_len,
@@ -517,6 +615,8 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         weight_decay=cfg.weight_decay if cfg.weight_decay is not None else 0.0,
         sample_packing=cfg.sample_packing if cfg.sample_packing else False,
         sample_packing_seq_len_multiplier=cfg.micro_batch_size,
+        relora_steps=cfg.relora_steps,
+        relora_warmup_steps=cfg.relora_warmup_steps,
         **training_arguments_kwargs,
     )
 
@@ -526,69 +626,13 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         if Path(cfg.torchdistx_path).exists():
             sys.path.append(cfg.torchdistx_path)
             importlib.import_module("torchdistx")
-    if (
-        cfg.optimizer == "adamw_bnb_8bit"
-        and not cfg.gptq
-        and "deepspeed" not in training_arguments_kwargs
-        and not cfg.fsdp
-    ):
-        decay_parameters = get_parameter_names(model, [nn.LayerNorm])
-        decay_parameters = [name for name in decay_parameters if "bias" not in name]
-        optimizer_grouped_parameters = [
-            {
-                "params": [
-                    p
-                    for n, p in model.named_parameters()
-                    if (n in decay_parameters and p.requires_grad)
-                ],
-                "weight_decay": training_args.weight_decay,
-            },
-            {
-                "params": [
-                    p
-                    for n, p in model.named_parameters()
-                    if (n not in decay_parameters and p.requires_grad)
-                ],
-                "weight_decay": 0.0,
-            },
-        ]
-
-        optimizer = bnb.optim.Adam8bit(
-            optimizer_grouped_parameters,
-            betas=(training_args.adam_beta1, training_args.adam_beta2),
-            eps=training_args.adam_epsilon,
-            lr=training_args.learning_rate,
-        )
-
-        if cfg.lr_scheduler == "one_cycle":
-            lr_scheduler_kwargs = (
-                cfg.lr_scheduler_kwargs if cfg.lr_scheduler_kwargs else {}
-            )
-            lr_scheduler = OneCycleLR(
-                optimizer,
-                cfg.learning_rate,
-                total_steps=total_num_steps,
-                epochs=cfg.num_epochs,
-                div_factor=cfg.lr_div_factor if cfg.lr_div_factor else 6,
-                **lr_scheduler_kwargs,
-            )
-        elif cfg.lr_scheduler == "log_sweep":
-            lr_scheduler = InterpolatingLogScheduler(
-                optimizer,
-                cfg.warmup_steps,
-                cfg.log_sweep_min_lr if cfg.log_sweep_min_lr else 1e-10,
-                cfg.log_sweep_max_lr if cfg.log_sweep_max_lr else 10,
-            )
-        else:
-            lr_scheduler = transformers.get_cosine_schedule_with_warmup(
-                optimizer,
-                training_args.warmup_steps,
-                total_num_steps,
-            )
-        trainer_kwargs["optimizers"] = (optimizer, lr_scheduler)
 
     callbacks = []
     callbacks.append(GPUStatsCallback(cfg))
+
+    if cfg.relora_steps:
+        callbacks.append(ReLoRACallback(cfg))
+
     # TODO on_save callback to sync checkpoints to GCP/AWS in background
     if cfg.early_stopping_patience:
         early_stop_cb = EarlyStoppingCallback(
@@ -606,10 +650,12 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         callbacks.append(SaveBetterTransformerModelCallback)
 
     data_collator_kwargs = {
-        "padding": True,
+        "padding": True,  # True/"longest" is the default
     }
-    if cfg.collator_pad_to_longest:
-        data_collator_kwargs["padding"] = "longest"
+    if cfg.pad_to_sequence_len:
+        data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
+            cfg.sequence_len / 64
+        )
     else:
         # A100 is best at 64, while others at 8. Let's use the larger so we don't have to check
         # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
@@ -633,11 +679,11 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
                 num_proc=32,
             )
 
-    trainer_cls = (
-        OneCycleLRSchedulerTrainer
-        if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora")
-        else AxolotlTrainer
-    )
+    trainer_cls = AxolotlTrainer
+    if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora"):
+        trainer_cls = OneCycleLRSchedulerTrainer
+    elif cfg.relora_steps:
+        trainer_cls = ReLoRATrainer
     trainer = trainer_cls(
         model=model,
         train_dataset=train_dataset,
@@ -648,8 +694,16 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
             return_tensors="pt",
             **data_collator_kwargs,
         ),
+        bench_data_collator=transformers.DataCollatorForSeq2Seq(
+            tokenizer,
+            return_tensors="pt",
+            **data_collator_kwargs,
+        ),
         callbacks=callbacks,
         **trainer_kwargs,
     )
 
+    if cfg.do_bench_eval:
+        trainer.add_callback(bench_eval_callback_factory(trainer, tokenizer))
+
     return trainer