repalce linear layers for qlora as well as add peft

* fix(config): Set eos/bos to tokenizer if different

* chore: fix lint

README.md

@@ -32,7 +32,6 @@ Features:
   - [How to Use Custom Pretokenized Dataset](#how-to-use-your-custom-pretokenized-dataset)
 - [Config](#config)
   - [Train](#train)
-  - [Training w/ Deepspeed](#training-with-deepspeed)
   - [Inference](#inference)
   - [Merge LORA to Base](#merge-lora-to-base)
 - [Common Errors](#common-errors-)
@@ -115,6 +114,25 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   docker compose up -d
   ```
 
+  <details>
+
+  <summary>Docker advanced</summary>
+
+  A more powerful Docker command to run would be this:
+
+  ```bash
+  docker run --gpus '"all"' --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=volume,src=axolotl,target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-py3.10-cu118-2.0.1
+  ```
+
+  It additionally:
+  * Prevents memory issues when running e.g. deepspeed (e.g. you could hit SIGBUS/signal 7 error) through `--ipc` and `--ulimit` args.
+  * Persists the downloaded HF data (models etc.) and your modifications to axolotl code through `--mount`/`-v` args.
+  * The `--name` argument simply makes it easier to refer to the container in vscode (`Dev Containers: Attach to Running Container...`) or in your terminal.
+
+  [More information on nvidia website](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html#setincshmem)
+
+  </details>
+
 #### Conda/Pip venv
   1. Install python >=**3.9**
 
@@ -356,6 +374,13 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
         - typescript
       type: ... # unimplemented custom format
 
+  # fastchat conversation
+  # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+  datasets:
+    - path: ...
+      type: sharegpt
+      conversation: chatml
+
   # local
   datasets:
     - path: data.jsonl # or json
@@ -394,7 +419,7 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
 
 <details>
 
-<summary>All yaml options</summary>
+<summary>All yaml options (click me)</summary>
 
 ```yaml
 # This is the huggingface model that contains *.pt, *.safetensors, or *.bin files
@@ -461,7 +486,9 @@ datasets:
     data_files: # Optional[str] path to source data files
     shards: # Optional[int] number of shards to split data into
     name: # Optional[str] name of dataset configuration to load
-    conversation:  # Optional[str] fastchat conversation type, only used with type: sharegpt
+
+    # Optional[str] fastchat conversation type, only used with type: sharegpt
+    conversation:  # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
 
   # Custom user prompt
   - path: repo
@@ -591,14 +618,14 @@ gradient_accumulation_steps: 1
 # The number of samples to include in each batch. This is the number of samples sent to each GPU.
 micro_batch_size: 2
 eval_batch_size:
-num_epochs: 3
+num_epochs: 4
 warmup_steps: 100
 learning_rate: 0.00003
 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: # Leave empty to eval at each epoch
+eval_steps: # Leave empty to eval at each epoch, integers for every N steps. decimal for fraction of total steps
 save_total_limit: # Checkpoints saved at a time
 # Maximum number of iterations to train for. It precedes num_epochs which means that
 # if both are set, num_epochs will not be guaranteed.
@@ -815,14 +842,41 @@ Run
 accelerate launch -m axolotl.cli.train your_config.yml
 ```
 
-#### Multi-GPU
+#### Preprocess dataset
+
+You can optionally pre-tokenize dataset with the following before finetuning.
+This is recommended for large datasets.
+
+- Set `push_dataset_to_hub: hf_user/repo` to push it to Huggingface.
+- Use `--debug` to see preprocessed examples.
 
-You can optionally pre-tokenize dataset with the following before finetuning:
 ```bash
-CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train your_config.yml --prepare_ds_only
+python -m axolotl.cli.preprocess your_config.yml
 ```
 
-##### Config
+#### Multi-GPU
+
+Below are the options available in axolotl for training with multiple GPUs. Note that DeepSpeed
+is the recommended multi-GPU option currently because FSDP may experience
+[loss instability](https://github.com/huggingface/transformers/issues/26498).
+
+##### DeepSpeed
+
+Deepspeed is an optimization suite for multi-gpu systems allowing you to train much larger models than you
+might typically be able to fit into your GPU's VRAM. More information about the various optimization types
+for deepspeed is available at https://huggingface.co/docs/accelerate/main/en/usage_guides/deepspeed#what-is-integrated
+
+We provide several default deepspeed JSON configurations for ZeRO stage 1, 2, and 3.
+
+```yaml
+deepspeed: deepspeed/zero1.json
+```
+
+```shell
+accelerate launch -m axolotl.cli.train examples/llama-2/config.py --deepspeed deepspeed/zero1.json
+```
+
+##### FSDP
 
 - llama FSDP
 ```yaml
@@ -847,24 +901,6 @@ wandb_run_id:
 wandb_log_model:
 ```
 
-### Training with Deepspeed
-
-Deepspeed is an optimization suite for multi-gpu systems allowing you to train much larger models than you
-might typically be able to fit into your GPU's VRAM. More information about the various optimization types
-for deepspeed is available at https://huggingface.co/docs/accelerate/main/en/usage_guides/deepspeed#what-is-integrated
-
-We provide several default deepspeed JSON configurations for ZeRO stage 1, 2, and 3.
-
-```shell
-accelerate launch -m axolotl.cli.train examples/llama-2/config.py --deepspeed deepspeed/zero1.json
-```
-
-or
-
-```yaml
-deepspeed: deepspeed/zero1.json
-```
-
 ### Inference
 
 Pass the appropriate flag to the train command:

deepspeed/zero3.json

@@ -1,14 +1,6 @@
 {
   "zero_optimization": {
     "stage": 3,
-    "offload_optimizer": {
-      "device": "cpu",
-      "pin_memory": true
-    },
-    "offload_param": {
-      "device": "cpu",
-      "pin_memory": true
-    },
     "overlap_comm": true,
     "contiguous_gradients": true,
     "sub_group_size": 0,
@@ -41,12 +33,13 @@
     }
   },
   "scheduler": {
-    "type": "WarmupLR",
+    "type": "WarmupDecayLR",
     "params": {
       "warmup_min_lr": "auto",
       "warmup_max_lr": "auto",
       "warmup_num_steps": "auto",
-      "warmup_type": "linear"
+      "warmup_type": "linear",
+      "total_num_steps": "auto"
     }
   },
   "gradient_accumulation_steps": "auto",

docs/faq.md

@@ -12,3 +12,7 @@ This usually happens when you run out of system RAM.
 > Exitcode -7 while using deepspeed
 
 Try upgrading deepspeed w: `pip install -U deepspeed`
+
+> AttributeError: 'DummyOptim' object has no attribute 'step'
+
+You may be using deepspeed with single gpu. Please don't set `deepspeed:` in yaml or cli.

examples/cerebras/qlora.yml

@@ -49,7 +49,7 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 save_steps:
 debug:
 deepspeed:

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

@@ -34,7 +34,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +54,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 save_steps:
 debug:
 deepspeed:

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

@@ -36,7 +36,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,7 +56,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 save_steps:
 debug:
 deepspeed:

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

@@ -34,7 +34,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +54,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 save_steps:
 debug:
 deepspeed:

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

@@ -36,7 +36,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,7 +56,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 save_steps:
 debug:
 deepspeed:

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

@@ -34,7 +34,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +54,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 save_steps:
 debug:
 deepspeed:

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

@@ -36,7 +36,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,7 +56,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 save_steps:
 debug:
 deepspeed:

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

@@ -53,7 +53,7 @@ output_dir: ./qlora-out
 # decrease if OOM, increase for max VRAM utilization
 micro_batch_size: 1
 gradient_accumulation_steps: 2
-num_epochs: 3
+num_epochs: 4
 # Optimizer for QLoRA
 optimizer: paged_adamw_32bit
 torchdistx_path:

examples/gptj/qlora.yml

@@ -46,7 +46,7 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 save_steps:
 debug:
 deepspeed:

examples/jeopardy-bot/config.yml

@@ -24,7 +24,7 @@ wandb_log_model:
 output_dir: ./jeopardy-bot-7b
 gradient_accumulation_steps: 1
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine

examples/llama-2/gptq-lora.yml

@@ -37,7 +37,7 @@ wandb_log_model:
 output_dir: ./model-out
 gradient_accumulation_steps: 1
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_torch
 adam_beta2: 0.95
 adam_eps: 0.00001

examples/llama-2/lora.yml

@@ -34,7 +34,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +54,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 eval_table_size:
 eval_table_max_new_tokens: 128
 save_steps:

examples/llama-2/qlora.yml

@@ -36,7 +36,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,7 +56,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 eval_table_size:
 save_steps:
 debug:

examples/llama-2/relora.yml

@@ -40,7 +40,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 4
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -60,7 +60,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 save_steps: 50
 debug:
 deepspeed:

examples/llama-2/tiny-llama.yml

@@ -34,7 +34,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +54,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 eval_table_size:
 save_steps:
 debug:

examples/mistral/config.yml

@@ -26,7 +26,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.000005
@@ -46,7 +46,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 eval_table_size:
 eval_table_max_new_tokens: 128
 save_steps:

examples/mistral/qlora.yml

@@ -63,7 +63,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+eval_steps: 0.05
 eval_table_size:
 eval_table_max_new_tokens: 128
 save_steps:

examples/mpt-7b/config.yml

@@ -26,7 +26,7 @@ wandb_log_model:
 output_dir: ./mpt-alpaca-7b
 gradient_accumulation_steps: 1
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine

examples/pythia/lora.yml

@@ -23,7 +23,7 @@ wandb_log_model:
 output_dir: ./lora-alpaca-pythia
 gradient_accumulation_steps: 1
 micro_batch_size: 4
-num_epochs: 3
+num_epochs: 4
 learning_rate: 0.00001
 train_on_inputs: false
 group_by_length: false
@@ -33,5 +33,5 @@ early_stopping_patience:
 resume_from_checkpoint:
 local_rank:
 weight_decay: 0.1
-eval_steps: 20
+eval_steps: 0.05
 logging_steps: 1

examples/redpajama/config-3b.yml

@@ -27,7 +27,7 @@ wandb_log_model:
 output_dir: ./redpajama-alpaca-3b
 batch_size: 4
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine

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

@@ -26,7 +26,7 @@ wandb_log_model:
 output_dir: ./lora-replit
 batch_size: 8
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer:
 torchdistx_path:
 lr_scheduler:

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

@@ -51,7 +51,7 @@ output_dir: ./qlora-out
 # decrease if OOM, increase for max VRAM utilization
 micro_batch_size: 1
 gradient_accumulation_steps: 1
-num_epochs: 3
+num_epochs: 4
 # Optimizer for QLoRA
 optimizer: paged_adamw_32bit
 torchdistx_path:

gitbook/README.md

@@ -0,0 +1 @@
+# Page

gitbook/SUMMARY.md

@@ -0,0 +1,4 @@
+# Table of contents
+
+* [Page](README.md)
+* [Small dev details](small-dev-details.md)

gitbook/small-dev-details.md

@@ -0,0 +1,3 @@
+# Small dev details
+
+/

requirements.txt

@@ -31,3 +31,4 @@ scikit-learn==1.2.2
 pynvml
 art
 fschat==0.2.29
+tensor_parallel

scripts/finetune.py

@@ -45,8 +45,6 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
         shard(cfg=parsed_cfg, cli_args=parsed_cli_args)
     else:
         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)
 
 

src/axolotl/cli/__init__.py

@@ -222,7 +222,9 @@ def load_datasets(
 ) -> TrainDatasetMeta:
     tokenizer = load_tokenizer(cfg)
 
-    train_dataset, eval_dataset, total_num_steps = prepare_dataset(cfg, tokenizer)
+    train_dataset, eval_dataset, total_num_steps, prompters = prepare_dataset(
+        cfg, tokenizer
+    )
 
     if cli_args.debug or cfg.debug:
         LOG.info("check_dataset_labels...")
@@ -238,6 +240,10 @@ def load_datasets(
             text_only=cli_args.debug_text_only,
         )
 
+        LOG.info("printing prompters...")
+        for prompter in prompters:
+            LOG.info(prompter)
+
     return TrainDatasetMeta(
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,

src/axolotl/cli/preprocess.py

@@ -0,0 +1,53 @@
+"""
+CLI to run training on a model
+"""
+import logging
+from pathlib import Path
+
+import fire
+import transformers
+from colorama import Fore
+
+from axolotl.cli import (
+    check_accelerate_default_config,
+    check_user_token,
+    load_cfg,
+    load_datasets,
+    print_axolotl_text_art,
+)
+from axolotl.common.cli import PreprocessCliArgs
+from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
+
+LOG = logging.getLogger("axolotl.cli.preprocess")
+
+
+def do_cli(config: Path = Path("examples/"), **kwargs):
+    # pylint: disable=duplicate-code
+    print_axolotl_text_art()
+    parsed_cfg = load_cfg(config, **kwargs)
+    check_accelerate_default_config()
+    check_user_token()
+    parser = transformers.HfArgumentParser((PreprocessCliArgs))
+    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
+        return_remaining_strings=True
+    )
+    if not parsed_cfg.dataset_prepared_path:
+        msg = (
+            Fore.RED
+            + "preprocess CLI called without dataset_prepared_path set, "
+            + f"using default path: {DEFAULT_DATASET_PREPARED_PATH}"
+            + Fore.RESET
+        )
+        LOG.warning(msg)
+        parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
+
+    _ = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    LOG.info(
+        Fore.GREEN
+        + f"Success! Preprocessed data path: `dataset_prepared_path: {parsed_cfg.dataset_prepared_path}`"
+        + Fore.RESET
+    )
+
+
+if __name__ == "__main__":
+    fire.Fire(do_cli)

src/axolotl/cli/train.py

@@ -6,7 +6,6 @@ from pathlib import Path
 
 import fire
 import transformers
-from colorama import Fore
 
 from axolotl.cli import (
     check_accelerate_default_config,
@@ -16,7 +15,6 @@ from axolotl.cli import (
     print_axolotl_text_art,
 )
 from axolotl.common.cli import TrainerCliArgs
-from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.train import train
 
 LOG = logging.getLogger("axolotl.cli.train")
@@ -32,18 +30,7 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
-    if parsed_cli_args.prepare_ds_only and not parsed_cfg.dataset_prepared_path:
-        msg = (
-            Fore.RED
-            + "--prepare_ds_only called without dataset_prepared_path set."
-            + Fore.RESET
-        )
-        LOG.warning(msg)
-        parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
-
     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)
 
 

src/axolotl/common/cli.py

@@ -25,11 +25,22 @@ class TrainerCliArgs:
     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)
 
 
+@dataclass
+class PreprocessCliArgs:
+    """
+    dataclass representing arguments for preprocessing only
+    """
+
+    debug: bool = field(default=False)
+    debug_text_only: bool = field(default=False)
+    debug_num_examples: int = field(default=1)
+    prompter: Optional[str] = field(default=None)
+
+
 def load_model_and_tokenizer(
     *,
     cfg: DictDefault,

src/axolotl/core/trainer_builder.py

@@ -14,6 +14,7 @@ from functools import partial
 from pathlib import Path
 from typing import Optional, Union
 
+import tensor_parallel as tp
 import torch
 import transformers
 from datasets import Dataset
@@ -33,6 +34,7 @@ from axolotl.utils.callbacks import (
 )
 from axolotl.utils.collators import DataCollatorForSeq2Seq
 from axolotl.utils.dataloader import MultipackDistributedDataloader
+from axolotl.utils.distributed import is_distributed
 from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
 
 try:
@@ -102,6 +104,9 @@ class AxolotlTrainingArguments(TrainingArguments):
     bench_source_max_len: int = field(
         default=2048, metadata={"help": "Maximum source sequence length for bench."}
     )
+    tensor_parallel: bool = field(
+        default=False, metadata={"help": "Use tensor parallelism to train"}
+    )
 
 
 class AxolotlTrainer(Trainer):
@@ -111,7 +116,8 @@ class AxolotlTrainer(Trainer):
 
     args = None  # type: AxolotlTrainingArguments
 
-    def __init__(self, *args, bench_data_collator=None, **kwargs):
+    def __init__(self, *args, num_epochs=1, bench_data_collator=None, **kwargs):
+        self.num_epochs = num_epochs
         self.bench_data_collator = bench_data_collator
         super().__init__(*args, **kwargs)
 
@@ -182,6 +188,7 @@ class AxolotlTrainer(Trainer):
                     packing_efficiency_estimate=self.args.sample_packing_efficiency,
                     sample_packing_seq_len_multiplier=self.args.sample_packing_seq_len_multiplier,
                     device_count=int(os.environ.get("WORLD_SIZE", 1)),
+                    num_epochs=self.num_epochs,
                 )
             )
         return super().get_train_dataloader()
@@ -205,6 +212,7 @@ class AxolotlTrainer(Trainer):
                     packing_efficiency_estimate=self.args.sample_packing_efficiency,
                     sample_packing_seq_len_multiplier=self.args.eval_batch_size,
                     device_count=int(os.environ.get("WORLD_SIZE", 1)),
+                    num_epochs=self.num_epochs,
                 )
             )
         return super().get_eval_dataloader(eval_dataset)
@@ -243,6 +251,14 @@ class AxolotlTrainer(Trainer):
         #     return (loss, outputs) if return_outputs else loss
         return super().compute_loss(model, inputs, return_outputs=return_outputs)
 
+    def _wrap_model(self, model, training=True, dataloader=None):
+        if self.args.tensor_parallel:
+            model = tp.tensor_parallel(model, distributed=is_distributed())
+            model.hf_device_map = tp.infer_sharded_device_map(model)
+        else:
+            model = super()._wrap_model(model, training=training, dataloader=dataloader)
+        return model
+
 
 class OneCycleLRSchedulerTrainer(AxolotlTrainer):
     """
@@ -368,7 +384,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         return trainer_kwargs, trainer_cls
 
     def hook_post_create_trainer(self, trainer):
-        # TODO
+        if self.cfg.tensor_parallel:
+            trainer.model = trainer.accelerator.prepare_model(
+                trainer.model, device_placement=True
+            )
         return trainer
 
     def get_callbacks(self):
@@ -612,6 +631,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         ] = self.cfg.micro_batch_size
         training_arguments_kwargs["relora_steps"] = self.cfg.relora_steps
         training_arguments_kwargs["relora_warmup_steps"] = self.cfg.relora_warmup_steps
+        training_arguments_kwargs["tensor_parallel"] = self.cfg.tensor_parallel is True
+
         training_arguments_kwargs = self.hook_pre_create_training_args(
             training_arguments_kwargs
         )
@@ -680,6 +701,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 **data_collator_kwargs,
             ),
             callbacks=self.get_callbacks(),
+            num_epochs=self.cfg.num_epochs,
             **trainer_kwargs,
         )
         trainer = self.hook_post_create_trainer(trainer)

src/axolotl/monkeypatch/llama_embeddings_hijack.py

@@ -1,40 +0,0 @@
-"""
-patch to add noisy embeddings per https://arxiv.org/abs/2310.05914
-"""
-
-import torch
-import transformers.models.llama.modeling_llama
-from transformers.utils import logging
-
-logger = logging.get_logger(__name__)
-
-
-def replace_llama_embeddings_with_uniform_distribution(noise_alpha=5):
-    # pylint: disable=duplicate-code
-    def noised_embed(orig_embed, noise_alpha, model):
-        def new_func(input_ids):
-            # during training, we add noise to the embedding
-            # during generation, we don't add noise to the embedding
-            if model.training:
-                embed_init = orig_embed(input_ids)
-                dims = torch.tensor(embed_init.size(1) * embed_init.size(2))
-                mag_norm = noise_alpha / torch.sqrt(dims)
-                return embed_init + torch.zeros_like(embed_init).uniform_(
-                    -mag_norm, mag_norm
-                )
-            return orig_embed(input_ids)
-
-        return new_func
-
-    def post_init(orig_post_init):
-        def new_func(self):
-            orig_post_init(self)
-            self.embed_tokens.forward = noised_embed(
-                self.embed_tokens.forward, noise_alpha, self
-            )
-
-        return new_func
-
-    transformers.models.llama.modeling_llama.LlamaModel.post_init = post_init(
-        transformers.models.llama.modeling_llama.LlamaModel.post_init
-    )

src/axolotl/monkeypatch/mistral_embeddings_hijack.py

@@ -1,40 +0,0 @@
-"""
-patch to add noisy embeddings per https://arxiv.org/abs/2310.05914
-"""
-
-import torch
-import transformers.models.mistral.modeling_mistral
-from transformers.utils import logging
-
-logger = logging.get_logger(__name__)
-
-
-def replace_mistral_embeddings_with_uniform_distribution(noise_alpha=5):
-    # pylint: disable=duplicate-code
-    def noised_embed(orig_embed, noise_alpha, model):
-        def new_func(input_ids):
-            # during training, we add noise to the embedding
-            # during generation, we don't add noise to the embedding
-            if model.training:
-                embed_init = orig_embed(input_ids)
-                dims = torch.tensor(embed_init.size(1) * embed_init.size(2))
-                mag_norm = noise_alpha / torch.sqrt(dims)
-                return embed_init + torch.zeros_like(embed_init).uniform_(
-                    -mag_norm, mag_norm
-                )
-            return orig_embed(input_ids)
-
-        return new_func
-
-    def post_init(orig_post_init):
-        def new_func(self):
-            orig_post_init(self)
-            self.embed_tokens.forward = noised_embed(
-                self.embed_tokens.forward, noise_alpha, self
-            )
-
-        return new_func
-
-    transformers.models.mistral.modeling_mistral.MistralModel.post_init = post_init(
-        transformers.models.mistral.modeling_mistral.MistralModel.post_init
-    )

src/axolotl/monkeypatch/neft_embeddings.py

@@ -0,0 +1,65 @@
+"""
+patches implemented through the trainer hooks to enable NEFT/noisy embeddings per https://arxiv.org/abs/2310.05914
+"""
+import torch
+from peft import PeftModel
+from transformers import PreTrainedModel
+
+
+def patch_neft(alpha, model):
+    embeddings = None
+    if isinstance(model, PreTrainedModel):
+        embeddings = model.get_input_embeddings()
+    if isinstance(model, PeftModel):
+        embeddings = model.base_model.get_input_embeddings()
+    if not embeddings:
+        raise ValueError(f"unhandled model class for neft: {model.__class__.__name__}")
+    embeddings.noisy_embedding_alpha = alpha
+    old_forward = embeddings.forward
+
+    # This hack seems to be needed to properly use a custom forward pass
+    # all credits to: https://discuss.pytorch.org/t/how-can-i-replace-the-forward-method-of-a-predefined-torchvision-model-with-my-customized-forward-function/54224/11
+    bound_method = neft_forward.__get__(  # pylint: disable=no-value-for-parameter
+        embeddings, embeddings.__class__
+    )
+    setattr(embeddings, "forward", bound_method)
+
+    embeddings._old_forward = old_forward  # pylint: disable=protected-access
+    return model
+
+
+def unpatch_neft(model):
+    embeddings = None
+    if isinstance(model, PreTrainedModel):
+        embeddings = model.get_input_embeddings()
+    if isinstance(model, PeftModel):
+        embeddings = model.base_model.get_input_embeddings()
+    if not embeddings:
+        raise ValueError(f"unhandled model class for neft: {model.__class__.__name__}")
+    if hasattr(embeddings, "_old_forward"):
+        embeddings.forward = embeddings._old_forward  # pylint: disable=protected-access
+        del embeddings._old_forward  # pylint: disable=protected-access
+        del embeddings.noisy_embedding_alpha
+
+
+def neft_forward(self, inputs: torch.Tensor):
+    embeddings = self._old_forward(inputs)  # pylint: disable=protected-access
+
+    if self.training:
+        dims = torch.tensor(embeddings.size(1) * embeddings.size(2))
+        mag_norm = self.noisy_embedding_alpha / torch.sqrt(dims)
+        embeddings = embeddings + torch.zeros_like(embeddings).uniform_(
+            -mag_norm, mag_norm
+        )
+
+    return embeddings
+
+
+def pretrain_hook(cfg, trainer):
+    if cfg.noisy_embedding_alpha:
+        trainer.model = patch_neft(cfg.noisy_embedding_alpha, trainer.model)
+
+
+def post_train_hook(cfg, trainer):
+    if cfg.noisy_embedding_alpha:
+        unpatch_neft(trainer.model)

src/axolotl/prompt_strategies/sharegpt.py

@@ -24,7 +24,7 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
     )
     field_human = ds_cfg["field_human"] if ds_cfg and "field_human" in ds_cfg else None
     field_model = ds_cfg["field_model"] if ds_cfg and "field_model" in ds_cfg else None
-    return SimpleShareGPTPromptTokenizingStrategy(
+    strategy = SimpleShareGPTPromptTokenizingStrategy(
         ShareGPTPrompterV2(
             conversation=conversation,
             role_key_model=field_model,
@@ -34,6 +34,9 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
         cfg.train_on_inputs,
         cfg.sequence_len,
     )
+    if ds_cfg and "strict" in ds_cfg:
+        strategy.strict = ds_cfg["strict"]
+    return strategy
 
 
 def load_role(tokenizer, cfg):
@@ -59,8 +62,26 @@ class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
     basic sharegpt strategy to grab conversations from the sample row
     """
 
+    _strict = True
+
+    @property
+    def strict(self):
+        return self._strict
+
+    @strict.setter
+    def strict(self, strict):
+        self._strict = strict
+
     def get_conversation_thread(self, prompt):
-        return prompt["conversations"]
+        conversations = prompt["conversations"]
+        if self.strict:
+            return conversations
+        # remap roles - allow for assistant turn
+        role_map = {"human": "human", "assistant": "gpt", "gpt": "gpt"}
+        turns = [
+            {"from": role_map[t["from"]], "value": t["value"]} for t in conversations
+        ]
+        return turns
 
 
 class SimpleRoleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):

src/axolotl/prompt_tokenizers.py

@@ -245,6 +245,7 @@ class ReflectionPromptTokenizingStrategy(PromptTokenizingStrategy):
         raise NotImplementedError
 
     def tokenize_prompt(self, prompt):
+        # pylint: disable=duplicate-code
         (
             instruction,
             input,  # pylint: disable=redefined-builtin

src/axolotl/prompters.py

@@ -4,10 +4,12 @@ import logging
 from enum import Enum
 from typing import Generator, Optional, Union
 
+from colorama import Fore
 from fastchat.conversation import Conversation, get_conv_template
 
 LOG = logging.getLogger("axolotl")
 IGNORE_TOKEN_ID = -100
+REPR_TEMPLATE = "\n<start>\n" + Fore.CYAN + "{full_prompt}" + Fore.RESET + "\n<end>\n"
 
 
 class PromptStyle(Enum):
@@ -55,20 +57,15 @@ class AlpacaPrompter:
             )
             self.system_format = "<|im_start|>system\n{system}<|im_end|>\n"
 
-    def build_prompt(
-        self,
-        instruction: str,
-        input: Union[None, str] = None,  # pylint: disable=redefined-builtin
-        output: Union[None, str] = None,
-    ) -> Generator[str, None, None]:
+    def _build_result(self, instruction, input_text, output):
         # returns the full prompt from instruction and optional input
         # if a label (=response, =output) is provided, it's also appended.
-        if input:
+        if input_text:
             res = (
                 self.system_format.format(system=self.system_prompt)
                 if self.system_prompt
                 else ""
-            ) + self.turn_format.format(instruction=instruction, input=input)
+            ) + self.turn_format.format(instruction=instruction, input=input_text)
         else:
             res = (
                 self.system_format.format(system=self.system_no_input_prompt)
@@ -77,7 +74,21 @@ class AlpacaPrompter:
             ) + self.turn_no_input_format.format(instruction=instruction)
         if output:
             res = f"{res}{output}"
-        yield res
+
+        return res
+
+    def build_prompt(
+        self,
+        instruction: str,
+        input: Union[None, str] = None,  # pylint: disable=redefined-builtin
+        output: Union[None, str] = None,
+    ) -> Generator[str, None, None]:
+        yield self._build_result(instruction, input, output)
+
+    def __repr__(self) -> str:
+        return REPR_TEMPLATE.format(
+            full_prompt=self._build_result("{instruction}", "{input}", "{output}")
+        )
 
 
 class UnpromptedPrompter(AlpacaPrompter):
@@ -191,14 +202,14 @@ class ReflectAlpacaPrompter:
             )
             self.response_split = "ASSISTANT:"
 
-    def build_prompt(
+    def _build_result(
         self,
         instruction: str,
         input: Union[None, str] = None,  # pylint: disable=redefined-builtin
         output: Union[None, str] = None,
         reflection: Union[None, str] = None,
         corrected: Union[None, str] = None,
-    ) -> Generator[str, None, None]:
+    ):
         # returns the full prompt from instruction and optional input
         # if a label (=response, =output) is provided, it's also appended.
         if input:
@@ -212,7 +223,30 @@ class ReflectAlpacaPrompter:
                 corrected=corrected,
             )
             res = f"{res}{label}"
-        yield res
+
+        return res
+
+    def build_prompt(
+        self,
+        instruction: str,
+        input: Union[None, str] = None,  # pylint: disable=redefined-builtin
+        output: Union[None, str] = None,
+        reflection: Union[None, str] = None,
+        corrected: Union[None, str] = None,
+    ) -> Generator[str, None, None]:
+        # pylint: disable=duplicate-code
+        yield self._build_result(
+            instruction,
+            input,
+            output,
+            reflection,
+            corrected,
+        )
+
+    def __repr__(self) -> str:
+        return REPR_TEMPLATE.format(
+            full_prompt=self._build_result("{instruction}", "{input}", "{output}")
+        )
 
 
 SHAREGPT_ASSERTION_FAILED_ROLE = (
@@ -247,7 +281,7 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
         if role_key_model:
             self.role_key_model = role_key_model
 
-    def build_prompt(self, source) -> Generator[str, None, None]:
+    def _build_result(self, source):
         if len(source) < 2:
             # If there isn't a back and forth conversation, ignore it
             # also happens on the data splitting leaving empty conversations
@@ -282,11 +316,20 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
                 LOG.warning(f"{SHAREGPT_ASSERTION_FAILED_ROLE}: {sentence}")
             conv.append_message(role, sentence["value"])
 
-        for part in conv.get_turns():
+        return conv.get_turns()
+
+    def build_prompt(self, source) -> Generator[str, None, None]:
+        turns = self._build_result(source)
+
+        for part in turns:
             if part[0] and not part[1]:
                 LOG.warning(f"role with empty message: {part[0]}")
             yield part
 
+    def __repr__(self) -> str:
+        turns = self._build_result([{"from": "{from}", "value": "{value}"}])
+        return "\n".join([REPR_TEMPLATE.format(full_prompt=part) for part in turns])
+
 
 class ShareGPTPrompterV2(ShareGPTPrompter):
     """
@@ -304,3 +347,15 @@ class ShareGPTPrompterV2(ShareGPTPrompter):
             role_key_human=role_key_human,
             role_key_model=role_key_model,
         )
+
+
+class UnsupportedPrompter:
+    """
+    A dummy class for custom prompters
+    """
+
+    def __init__(self) -> None:
+        pass
+
+    def __repr__(self):
+        return "Pre-tokenized or custom dataset types are unsupported for logging"

src/axolotl/train.py

@@ -16,6 +16,7 @@ from transformers.deepspeed import is_deepspeed_zero3_enabled
 
 from axolotl.common.cli import TrainerCliArgs
 from axolotl.logging_config import configure_logging
+from axolotl.monkeypatch import neft_embeddings
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
@@ -107,6 +108,7 @@ def train(
     if cfg.group_by_length:
         LOG.info("hang tight... sorting dataset for group_by_length")
 
+    pretrain_hooks(cfg, trainer)
     if cfg.flash_optimum:
         with torch.backends.cuda.sdp_kernel(
             enable_flash=True, enable_math=True, enable_mem_efficient=True
@@ -114,6 +116,7 @@ def train(
             trainer.train(resume_from_checkpoint=resume_from_checkpoint)
     else:
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
+    post_train_hooks(cfg, trainer)
 
     LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
 
@@ -163,3 +166,23 @@ def train(
         trainer.create_model_card(model_name=cfg.output_dir.lstrip("./"))
 
     return model, tokenizer
+
+
+def pretrain_hooks(cfg, trainer):
+    """
+    Run hooks right before kicking off the training
+    :param cfg:
+    :param trainer:
+    :return:
+    """
+    neft_embeddings.pretrain_hook(cfg, trainer)
+
+
+def post_train_hooks(cfg, trainer):
+    """
+    Run hooks right after training completes
+    :param cfg:
+    :param trainer:
+    :return:
+    """
+    neft_embeddings.post_train_hook(cfg, trainer)

src/axolotl/utils/bench.py

@@ -1,10 +1,13 @@
 """Benchmarking and measurement utilities"""
 import functools
+import logging
 
 import pynvml
 import torch
 from pynvml.nvml import NVMLError
 
+LOG = logging.getLogger("axolotl.utils.bench")
+
 
 def check_cuda_device(default_value):
     """
@@ -62,7 +65,14 @@ def gpu_memory_usage_smi(device=0):
 
 
 def log_gpu_memory_usage(log, msg, device):
-    usage, cache, misc = gpu_memory_usage_all(device)
+    if not torch.cuda.is_available():
+        return (0, 0, 0)
+
+    try:
+        usage, cache, misc = gpu_memory_usage_all(device)
+    except ValueError as exc:
+        LOG.exception(exc)
+        return (0, 0, 0)
     extras = []
     if cache > 0:
         extras.append(f"+{cache:.03f}GB cache")

src/axolotl/utils/config.py

@@ -369,6 +369,10 @@ def validate_config(cfg):
             "If you want to full finetune, please turn off load_in_8bit and load_in_4bit."
         )
 
+    if cfg.tensor_parallel and cfg.gradient_checkpointing:
+        raise ValueError(
+            "TensorParallelPreTrainedModel does not support gradient checkpointing"
+        )
     # TODO
     # MPT 7b
     # https://github.com/facebookresearch/bitsandbytes/issues/25

src/axolotl/utils/data.py

@@ -3,7 +3,7 @@ import functools
 import hashlib
 import logging
 from pathlib import Path
-from typing import Dict, List, Tuple, Union
+from typing import Any, Dict, List, Tuple, Union
 
 import torch
 from datasets import (
@@ -36,6 +36,7 @@ from axolotl.prompters import (
     MultipleChoiceExplainPrompter,
     ReflectAlpacaPrompter,
     SummarizeTLDRPrompter,
+    UnsupportedPrompter,
 )
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process, zero_first
@@ -55,9 +56,10 @@ def md5(to_hash: str, encoding: str = "utf-8") -> str:
 
 
 def prepare_dataset(cfg, tokenizer):
+    prompters = []
     if not cfg.pretraining_dataset:
         with zero_first(is_main_process()):
-            train_dataset, eval_dataset = load_prepare_datasets(
+            train_dataset, eval_dataset, prompters = load_prepare_datasets(
                 tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
             )
     else:
@@ -70,7 +72,7 @@ def prepare_dataset(cfg, tokenizer):
         # https://discuss.huggingface.co/t/how-to-use-huggingface-trainer-streaming-datasets-without-wrapping-it-with-torchdatas-iterablewrapper/25230
         train_dataset = train_dataset.with_format("torch")
         eval_dataset = None
-        return train_dataset, eval_dataset, cfg.max_steps
+        return train_dataset, eval_dataset, cfg.max_steps, prompters
 
     with zero_first(is_main_process()):
         train_dataset, eval_dataset = process_datasets_for_packing(
@@ -83,7 +85,7 @@ def prepare_dataset(cfg, tokenizer):
         LOG.info(f"Maximum number of steps set at {total_num_steps}")
     else:
         total_num_steps = calculate_total_num_steps(cfg, train_dataset, tokenizer)
-    return train_dataset, eval_dataset, total_num_steps
+    return train_dataset, eval_dataset, total_num_steps, prompters
 
 
 def load_tokenized_prepared_datasets(
@@ -109,6 +111,7 @@ def load_tokenized_prepared_datasets(
         else Path(default_dataset_prepared_path) / ds_hash
     )
     dataset = None
+    prompters = []
     use_auth_token = cfg.hf_use_auth_token
     try:
         if cfg.push_dataset_to_hub:
@@ -147,13 +150,13 @@ def load_tokenized_prepared_datasets(
                     yield dataset
 
         # pylint: disable=invalid-name
-        for d in for_d_in_datasets(cfg.datasets):
+        for config_dataset in for_d_in_datasets(cfg.datasets):
             ds: Union[Dataset, DatasetDict] = None
             ds_from_hub = False
             try:
                 load_dataset(
-                    d.path,
-                    name=d.name,
+                    config_dataset.path,
+                    name=config_dataset.name,
                     streaming=True,
                     token=use_auth_token,
                 )
@@ -162,33 +165,33 @@ def load_tokenized_prepared_datasets(
                 pass
 
             # prefer local dataset, even if hub exists
-            local_path = Path(d.path)
+            local_path = Path(config_dataset.path)
             if local_path.exists():
                 if local_path.is_dir():
                     # TODO dirs with arrow or parquet files could be loaded with `load_from_disk`
                     ds = load_dataset(
-                        d.path,
-                        name=d.name,
-                        data_files=d.data_files,
+                        config_dataset.path,
+                        name=config_dataset.name,
+                        data_files=config_dataset.data_files,
                         streaming=False,
                         split=None,
                     )
                 elif local_path.is_file():
                     ds_type = "json"
-                    if d.ds_type:
-                        ds_type = d.ds_type
-                    elif ".parquet" in d.path:
+                    if config_dataset.ds_type:
+                        ds_type = config_dataset.ds_type
+                    elif ".parquet" in config_dataset.path:
                         ds_type = "parquet"
-                    elif ".arrow" in d.path:
+                    elif ".arrow" in config_dataset.path:
                         ds_type = "arrow"
-                    elif ".csv" in d.path:
+                    elif ".csv" in config_dataset.path:
                         ds_type = "csv"
-                    elif ".txt" in d.path:
+                    elif ".txt" in config_dataset.path:
                         ds_type = "text"
                     ds = load_dataset(
                         ds_type,
-                        name=d.name,
-                        data_files=d.path,
+                        name=config_dataset.name,
+                        data_files=config_dataset.path,
                         streaming=False,
                         split=None,
                     )
@@ -198,25 +201,25 @@ def load_tokenized_prepared_datasets(
                     )
             elif ds_from_hub:
                 ds = load_dataset(
-                    d.path,
-                    name=d.name,
+                    config_dataset.path,
+                    name=config_dataset.name,
                     streaming=False,
-                    data_files=d.data_files,
+                    data_files=config_dataset.data_files,
                     token=use_auth_token,
                 )
             else:
-                if isinstance(d.data_files, str):
+                if isinstance(config_dataset.data_files, str):
                     fp = hf_hub_download(
-                        repo_id=d.path,
+                        repo_id=config_dataset.path,
                         repo_type="dataset",
-                        filename=d.data_files,
+                        filename=config_dataset.data_files,
                     )
-                elif isinstance(d.data_files, list):
+                elif isinstance(config_dataset.data_files, list):
                     fp = []
-                    for file in d.data_files:
+                    for file in config_dataset.data_files:
                         fp.append(
                             hf_hub_download(
-                                repo_id=d.path,
+                                repo_id=config_dataset.path,
                                 repo_type="dataset",
                                 filename=file,
                             )
@@ -226,21 +229,27 @@ def load_tokenized_prepared_datasets(
                         "data_files must be either a string or list of strings"
                     )
                 ds = load_dataset(
-                    "json", name=d.name, data_files=fp, streaming=False, split=None
+                    "json",
+                    name=config_dataset.name,
+                    data_files=fp,
+                    streaming=False,
+                    split=None,
                 )
             if not ds:
                 raise ValueError("unhandled dataset load")
             # support for using a subset of the data
-            if d.shards:
+            if config_dataset.shards:
                 if "train" in ds:
                     ds = ds.shuffle(seed=seed)["train"].shard(
-                        num_shards=d.shards, index=0
+                        num_shards=config_dataset.shards, index=0
                     )
                 else:
-                    ds = ds.shuffle(seed=seed).shard(num_shards=d.shards, index=0)
+                    ds = ds.shuffle(seed=seed).shard(
+                        num_shards=config_dataset.shards, index=0
+                    )
 
             d_base_type = d_prompt_style = None
-            d_type = d.type
+            d_type = config_dataset.type
             if isinstance(d_type, str):
                 d_type_split = d_type.split(":")
                 d_base_type = d_type_split[0]
@@ -249,108 +258,26 @@ def load_tokenized_prepared_datasets(
                 ds = ds["train"]
             elif (
                 isinstance(ds, DatasetDict)
-                and d.train_on_split
-                and d.train_on_split in ds
+                and config_dataset.train_on_split
+                and config_dataset.train_on_split in ds
             ):
-                ds = ds[d.train_on_split]
+                ds = ds[config_dataset.train_on_split]
             elif isinstance(ds, DatasetDict):
                 raise ValueError(
-                    f"no train split found for dataset {d.path}, you may specify a split with 'train_on_split: `"
-                )
-            if (
-                "input_ids" in ds.features
-                and "attention_mask" in ds.features
-                and "labels" in ds.features
-            ):
-                # dataset is already tokenized, just drop it straight in
-                datasets.append(ds)
-            elif isinstance(d.type, DictDefault):
-                ds_strategy = load("user_defined", tokenizer, cfg, d.type.to_dict())
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            elif ds_strategy := load(d.type, tokenizer, cfg, d):
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            elif d_base_type == "alpaca":
-                ds_strategy = AlpacaPromptTokenizingStrategy(
-                    AlpacaPrompter(d_prompt_style),
-                    tokenizer,
-                    cfg.train_on_inputs,
-                    cfg.sequence_len,
-                )
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            elif d_base_type == "explainchoice":
-                ds_strategy = AlpacaMultipleChoicePromptTokenizingStrategy(
-                    MultipleChoiceExplainPrompter(d_prompt_style),
-                    tokenizer,
-                    cfg.train_on_inputs,
-                    cfg.sequence_len,
-                )
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            elif d_base_type == "concisechoice":
-                ds_strategy = AlpacaMultipleChoicePromptTokenizingStrategy(
-                    MultipleChoiceConcisePrompter(d_prompt_style),
-                    tokenizer,
-                    cfg.train_on_inputs,
-                    cfg.sequence_len,
-                )
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            elif d_base_type == "summarizetldr":
-                ds_strategy = SummarizeTLDRPromptTokenizingStrategy(
-                    SummarizeTLDRPrompter(d_prompt_style),
-                    tokenizer,
-                    cfg.train_on_inputs,
-                    cfg.sequence_len,
-                )
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            elif d_base_type == "jeopardy":
-                ds_strategy = JeopardyPromptTokenizingStrategy(
-                    JeopardyPrompter(d_prompt_style),
-                    tokenizer,
-                    cfg.train_on_inputs,
-                    cfg.sequence_len,
-                )
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            elif d_base_type == "oasst":
-                ds_strategy = OpenAssistantPromptTokenizingStrategy(
-                    AlpacaPrompter(d_prompt_style),
-                    tokenizer,
-                    cfg.train_on_inputs,
-                    cfg.sequence_len,
-                )
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            elif d_base_type == "gpteacher":
-                ds_strategy = GPTeacherPromptTokenizingStrategy(
-                    GPTeacherPrompter(d_prompt_style),
-                    tokenizer,
-                    cfg.train_on_inputs,
-                    cfg.sequence_len,
-                )
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            elif d_base_type == "reflection":
-                ds_strategy = AlpacaReflectionPTStrategy(
-                    ReflectAlpacaPrompter(d_prompt_style),
-                    tokenizer,
-                    cfg.train_on_inputs,
-                    cfg.sequence_len,
-                )
-                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
-                datasets.append(ds_wrapper)
-            else:
-                suffix = ""
-                if ":load_" in d.type:
-                    suffix = f" Did you mean {d.type.replace(':load_', '.load_')}?"
-                LOG.error(f"unhandled prompt tokenization strategy: {d.type}. {suffix}")
-                raise ValueError(
-                    f"unhandled prompt tokenization strategy: {d.type} {suffix}"
+                    f"no train split found for dataset {config_dataset.path}, you may specify a split with 'train_on_split: `"
                 )
+
+            dataset_wrapper, dataset_prompter = get_dataset_wrapper(
+                config_dataset=config_dataset,
+                dataset=ds,
+                tokenizer=tokenizer,
+                cfg=cfg,
+                d_base_type=d_base_type,
+                d_prompt_style=d_prompt_style,
+            )
+            datasets.append(dataset_wrapper)
+            prompters.append(dataset_prompter)
+
         LOG.info("merging datasets")
         dataset = concatenate_datasets(datasets)
 
@@ -368,14 +295,14 @@ def load_tokenized_prepared_datasets(
                     f"{cfg.push_dataset_to_hub}/{ds_hash}", private=True
                 )
 
-    return dataset
+    return dataset, prompters
 
 
 def load_prepare_datasets(
     tokenizer: PreTrainedTokenizerBase,
     cfg,
     default_dataset_prepared_path,
-) -> Tuple[Dataset, Dataset]:
+) -> Tuple[Dataset, Dataset, List[Any]]:
     max_packed_sequence_len = (
         cfg.max_packed_sequence_len if cfg.max_packed_sequence_len else cfg.sequence_len
     )
@@ -384,6 +311,7 @@ def load_prepare_datasets(
     )  # make sure we don't accidentally set it larger than sequence_len
 
     tokenizer_name = tokenizer.__class__.__name__
+    prompters = []
     if cfg.max_packed_sequence_len is not None:
         # see if we can go ahead and load the stacked dataset
         seed = f"@{str(cfg.seed)}" if cfg.seed else ""
@@ -439,7 +367,7 @@ def load_prepare_datasets(
                     f"{cfg.push_dataset_to_hub}/{ds_hash}", private=True
                 )
         else:
-            dataset = load_tokenized_prepared_datasets(
+            dataset, prompters = load_tokenized_prepared_datasets(
                 tokenizer, cfg, default_dataset_prepared_path
             )
 
@@ -481,7 +409,7 @@ def load_prepare_datasets(
                         private=True,
                     )
     else:
-        dataset = load_tokenized_prepared_datasets(
+        dataset, prompters = load_tokenized_prepared_datasets(
             tokenizer, cfg, default_dataset_prepared_path
         )
 
@@ -532,7 +460,124 @@ def load_prepare_datasets(
         train_dataset = dataset
         eval_dataset = None
 
-    return train_dataset, eval_dataset
+    return train_dataset, eval_dataset, prompters
+
+
+def get_dataset_wrapper(
+    config_dataset, dataset, tokenizer, cfg, d_base_type, d_prompt_style
+):
+    dataset_wrapper = None
+    dataset_prompter = None
+
+    if (
+        "input_ids" in dataset.features
+        and "attention_mask" in dataset.features
+        and "labels" in dataset.features
+    ):
+        # dataset is already tokenized, just drop it straight in
+        dataset_prompter = UnsupportedPrompter()
+        dataset_wrapper = dataset
+    elif isinstance(config_dataset.type, DictDefault):
+        ds_strategy = load(
+            "user_defined", tokenizer, cfg, config_dataset.type.to_dict()
+        )
+        dataset_prompter = UnsupportedPrompter()
+        dataset_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+    elif ds_strategy := load(config_dataset.type, tokenizer, cfg, config_dataset):
+        dataset_prompter = UnsupportedPrompter()
+        dataset_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+    elif d_base_type == "alpaca":
+        dataset_prompter = AlpacaPrompter(d_prompt_style)
+        ds_strategy = AlpacaPromptTokenizingStrategy(
+            dataset_prompter,
+            tokenizer,
+            cfg.train_on_inputs,
+            cfg.sequence_len,
+        )
+        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+        dataset_wrapper = ds_wrapper
+    elif d_base_type == "explainchoice":
+        dataset_prompter = MultipleChoiceExplainPrompter(d_prompt_style)
+        ds_strategy = AlpacaMultipleChoicePromptTokenizingStrategy(
+            dataset_prompter,
+            tokenizer,
+            cfg.train_on_inputs,
+            cfg.sequence_len,
+        )
+        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+        dataset_wrapper = ds_wrapper
+    elif d_base_type == "concisechoice":
+        dataset_prompter = MultipleChoiceConcisePrompter(d_prompt_style)
+        ds_strategy = AlpacaMultipleChoicePromptTokenizingStrategy(
+            dataset_prompter,
+            tokenizer,
+            cfg.train_on_inputs,
+            cfg.sequence_len,
+        )
+        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+        dataset_wrapper = ds_wrapper
+    elif d_base_type == "summarizetldr":
+        dataset_prompter = SummarizeTLDRPrompter(d_prompt_style)
+        ds_strategy = SummarizeTLDRPromptTokenizingStrategy(
+            dataset_prompter,
+            tokenizer,
+            cfg.train_on_inputs,
+            cfg.sequence_len,
+        )
+        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+        dataset_wrapper = ds_wrapper
+    elif d_base_type == "jeopardy":
+        dataset_prompter = JeopardyPrompter(d_prompt_style)
+        ds_strategy = JeopardyPromptTokenizingStrategy(
+            dataset_prompter,
+            tokenizer,
+            cfg.train_on_inputs,
+            cfg.sequence_len,
+        )
+        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+        dataset_wrapper = ds_wrapper
+    elif d_base_type == "oasst":
+        dataset_prompter = AlpacaPrompter(d_prompt_style)
+        ds_strategy = OpenAssistantPromptTokenizingStrategy(
+            dataset_prompter,
+            tokenizer,
+            cfg.train_on_inputs,
+            cfg.sequence_len,
+        )
+        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+        dataset_wrapper = ds_wrapper
+    elif d_base_type == "gpteacher":
+        dataset_prompter = GPTeacherPrompter(d_prompt_style)
+        ds_strategy = GPTeacherPromptTokenizingStrategy(
+            dataset_prompter,
+            tokenizer,
+            cfg.train_on_inputs,
+            cfg.sequence_len,
+        )
+        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+        dataset_wrapper = ds_wrapper
+    elif d_base_type == "reflection":
+        dataset_prompter = ReflectAlpacaPrompter(d_prompt_style)
+        ds_strategy = AlpacaReflectionPTStrategy(
+            dataset_prompter,
+            tokenizer,
+            cfg.train_on_inputs,
+            cfg.sequence_len,
+        )
+        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
+        dataset_wrapper = ds_wrapper
+    else:
+        suffix = ""
+        if ":load_" in config_dataset.type:
+            suffix = f" Did you mean {config_dataset.type.replace(':load_', '.load_')}?"
+        LOG.error(
+            f"unhandled prompt tokenization strategy: {config_dataset.type}. {suffix}"
+        )
+        raise ValueError(
+            f"unhandled prompt tokenization strategy: {config_dataset.type} {suffix}"
+        )
+
+    return dataset_wrapper, dataset_prompter
 
 
 def encode_pretraining(

src/axolotl/utils/dataloader.py

@@ -3,6 +3,9 @@ import hashlib
 import itertools
 import logging
 import math
+import time
+from queue import Queue
+from threading import Thread
 from typing import Any, Callable, List, Union
 
 import numba
@@ -149,6 +152,8 @@ class MultipackDistributedDataloader:
         packing_efficiency_estimate: float = 1.0,
         sample_packing_seq_len_multiplier: int = 1,
         device_count: int = 1,
+        prefetch_max: int = 1000,
+        num_epochs: int = 1,
     ):
         # Dataset
         self.dataset = dataset
@@ -167,6 +172,7 @@ class MultipackDistributedDataloader:
         self.seq_max_length = seq_max_length
         self.batch_max_length = batch_size * seq_max_length
         self.collate_fn = collate_fn
+        self.num_epochs = num_epochs
 
         self.num_replicas = 1
         self.rank = 0
@@ -177,6 +183,44 @@ class MultipackDistributedDataloader:
         self.packing_efficiency_estimate = packing_efficiency_estimate or 1.0
         self.device_count = device_count
 
+        # maxsize is maximum number of samples in queue
+        self.prefetch_max = prefetch_max
+        self.queue: Queue = Queue(maxsize=prefetch_max)
+        self.thread = None
+
+    def _worker(self):
+        LOG.info(
+            f"[WORKER] Epochs: {self.num_epochs}, Samples: {self.len_w_stats()*self.batch_size}"
+        )
+        for epoch in range(self.num_epochs):
+            for sample in self._internal_batch_generator():
+                while True:
+                    if self.queue.full():
+                        time.sleep(1)
+                    else:
+                        break
+                self.queue.put(sample)
+
+            # stop the queue when epoch is done
+            self.queue.put(None)
+
+    def __iter__(self):
+        if hasattr(self.sampler, "set_epoch"):
+            new_epoch = self.sampler.epoch + 1
+            self.sampler.set_epoch(new_epoch)
+            LOG.info(f"calling sampler.set_epoch({new_epoch})")
+
+        if self.thread is None:
+            self.thread = Thread(target=self._worker, daemon=True)
+            self.thread.start()
+
+        while True:
+            item = self.queue.get()
+
+            if item is None:
+                break
+            yield item
+
     def generate_batches(self, set_stats=False):
         LOG.info("generating packed batches")
         if self.sampler:
@@ -206,11 +250,7 @@ class MultipackDistributedDataloader:
 
         return batches, totseqs
 
-    def __iter__(self):
-        if hasattr(self.sampler, "set_epoch"):
-            new_epoch = self.sampler.epoch + 1
-            self.sampler.set_epoch(new_epoch)
-            LOG.info(f"calling sampler.set_epoch({new_epoch})")
+    def _internal_batch_generator(self):
         all_batches, _ = self.generate_batches(set_stats=True)
         features = self.dataset.features.keys()
         len_remaining = self._len_est()

src/axolotl/utils/models.py

@@ -7,6 +7,7 @@ from typing import Optional, Tuple  # noqa: F401
 import bitsandbytes as bnb
 import torch
 import transformers
+import transformers.utils.bitsandbytes
 from optimum.bettertransformer import BetterTransformer
 from peft import PeftConfig, prepare_model_for_kbit_training
 from peft.tuners.lora import QuantLinear
@@ -31,7 +32,7 @@ LOG = logging.getLogger("axolotl")
 
 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
+    trust_remote_code = cfg.trust_remote_code is True
     return AutoConfig.from_pretrained(
         model_config_name, trust_remote_code=trust_remote_code
     )
@@ -72,11 +73,6 @@ def load_tokenizer(cfg):
         # 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}")
-    LOG.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")
-    LOG.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
-
     if tokenizer.__class__.__name__ == "GPTNeoXTokenizerFast":
         tokenizer.add_special_tokens({"pad_token": "[PAD]"})
         os.environ["TOKENIZERS_PARALLELISM"] = "false"
@@ -98,6 +94,11 @@ def load_tokenizer(cfg):
             ]
         )
 
+    LOG.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
+    LOG.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
+    LOG.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")
+    LOG.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
+
     return tokenizer
 
 
@@ -180,26 +181,6 @@ def load_model(
         LOG.info("patching with flash attention")
         replace_mistral_attn_with_flash_attn(packed=cfg.sample_packing)
 
-    if cfg.is_llama_derived_model and cfg.noisy_embedding_alpha:
-        from axolotl.monkeypatch.llama_embeddings_hijack import (
-            replace_llama_embeddings_with_uniform_distribution,
-        )
-
-        LOG.info("patching with noisy embeddings")
-        replace_llama_embeddings_with_uniform_distribution(
-            noise_alpha=cfg.noisy_embedding_alpha
-        )
-
-    if cfg.is_mistral_derived_model and cfg.noisy_embedding_alpha:
-        from axolotl.monkeypatch.mistral_embeddings_hijack import (
-            replace_mistral_embeddings_with_uniform_distribution,
-        )
-
-        LOG.info("patching with noisy embeddings")
-        replace_mistral_embeddings_with_uniform_distribution(
-            noise_alpha=cfg.noisy_embedding_alpha
-        )
-
     if cfg.is_llama_derived_model and cfg.xpos_rope:
         from axolotl.monkeypatch.xpos_rope_llama_monkey_patch import (
             replace_llama_rope_with_xpos_rope,
@@ -241,7 +222,7 @@ def load_model(
             load_in_4bit=True,
             llm_int8_threshold=6.0,
             llm_int8_has_fp16_weight=False,
-            bnb_4bit_compute_dtype=cfg.torch_dtype,
+            bnb_4bit_compute_dtype=torch.float16,
             bnb_4bit_use_double_quant=True,
             bnb_4bit_quant_type="nf4",
         )
@@ -255,7 +236,12 @@ def load_model(
             model_kwargs["use_flash_attention_2"] = True
 
     try:
-        if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
+        if (
+            cfg.is_llama_derived_model
+            and not cfg.trust_remote_code
+            and not cfg.gptq
+            and not cfg.tensor_parallel
+        ):
             from transformers import LlamaForCausalLM
 
             config_kwargs = {}
@@ -321,7 +307,7 @@ def load_model(
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
                 **model_kwargs,
             )
-        elif model_type and not cfg.trust_remote_code:
+        elif model_type and not cfg.trust_remote_code and not cfg.tensor_parallel:
             if cfg.gptq:
                 model = AutoModelForCausalLM.from_pretrained(
                     base_model,
@@ -336,6 +322,17 @@ def load_model(
                     trust_remote_code=cfg.trust_remote_code or False,
                     **model_kwargs,
                 )
+        elif cfg.tensor_parallel:
+            model_kwargs.pop("device_map")
+            model = AutoModelForCausalLM.from_pretrained(
+                base_model,
+                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,
+                low_cpu_mem_usage=True,
+                offload_state_dict=True,
+                trust_remote_code=cfg.trust_remote_code or False,
+                **model_kwargs,
+            )
         else:
             config = AutoConfig.from_pretrained(
                 base_model,
@@ -386,15 +383,18 @@ def load_model(
             **model_kwargs,
         )
 
-    embeddings_len = (
-        math.ceil(len(tokenizer) / 32) * 32
-        if cfg.resize_token_embeddings_to_32x
-        else len(tokenizer)
-    )
-    if model.get_input_embeddings().num_embeddings < embeddings_len:
-        model.resize_token_embeddings(embeddings_len)
-    else:
-        model.tie_weights()
+    try:
+        embeddings_len = (
+            math.ceil(len(tokenizer) / 32) * 32
+            if cfg.resize_token_embeddings_to_32x
+            else len(tokenizer)
+        )
+        if model.get_input_embeddings().num_embeddings < embeddings_len:
+            model.resize_token_embeddings(embeddings_len)
+        else:
+            model.tie_weights()
+    except NotImplementedError:
+        LOG.warning("`resize_token_embeddings` not implemented on model")
 
     if (
         hasattr(model.config, "max_position_embeddings")
@@ -406,6 +406,20 @@ def load_model(
         )
         model.config.max_position_embeddings = cfg.sequence_len
 
+    if (
+        hasattr(model.config, "bos_token_id")
+        and model.config.bos_token_id
+        and model.config.bos_token_id != tokenizer.bos_token_id
+    ):
+        model.config.bos_token_id = tokenizer.bos_token_id
+
+    if (
+        hasattr(model.config, "eos_token_id")
+        and model.config.eos_token_id
+        and model.config.eos_token_id != tokenizer.eos_token_id
+    ):
+        model.config.eos_token_id = tokenizer.eos_token_id
+
     if model.device.type == "cuda":
         log_gpu_memory_usage(LOG, "after model load", model.device)
 
@@ -483,7 +497,12 @@ def load_adapter(model, cfg, adapter, inference=False):
     if adapter is None:
         return model, None
     if hasattr(model, "enable_input_require_grads"):
-        model.enable_input_require_grads()
+        try:
+            model.enable_input_require_grads()
+        except NotImplementedError:
+            LOG.warning("enable_input_require_grads not implemented on model")
+    if adapter == "qlora" and cfg.tensor_parallel:
+        model, _ = load_tp_qlora(model)
     if adapter in ["lora", "qlora"]:
         return load_lora(model, cfg, inference=inference)
     if adapter == "llama-adapter":
@@ -535,6 +554,25 @@ def find_all_linear_names(model):
     return list(lora_module_names)
 
 
+def load_tp_qlora(model):
+    from transformers.utils.bitsandbytes import replace_with_bnb_linear
+
+    model = replace_with_bnb_linear(
+        model,
+        quantization_config=BitsAndBytesConfig(
+            load_in_4bit=True,
+            llm_int8_threshold=6.0,
+            llm_int8_has_fp16_weight=False,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type="nf4",
+        ),
+    )
+    model.is_loaded_in_4bit = True
+
+    return model, None
+
+
 def load_lora(model, cfg, inference=False):
     # type: (PreTrainedModel, DictDefault, bool) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
 

src/axolotl/utils/trainer.py

@@ -216,6 +216,7 @@ def calculate_total_num_steps(cfg, train_dataset, tokenizer):
                 packing_efficiency_estimate=cfg.sample_packing_eff_est,
                 sample_packing_seq_len_multiplier=cfg.micro_batch_size,
                 device_count=int(os.environ.get("WORLD_SIZE", 1)),
+                num_epochs=cfg.num_epochs,
             )
             data_loader_len = data_loader.len_w_stats()
             actual_eff = data_loader.efficiency()