Update README.md

* start at index 0

* add test to check for missing turns

* apply black

* Update test_prompt_tokenizers.py

* Update src/axolotl/monkeypatch/fastchat_conversation_turns.py

Co-authored-by: Motoki Wu <tokestermw@gmail.com>

* fix linting

* apply black

* add more tests for llama/sharegpt

* make logic clearer

---------

Co-authored-by: Motoki Wu <tokestermw@gmail.com>

README.md

@@ -36,7 +36,9 @@ Features:
   - [Train](#train)
   - [Inference](#inference)
   - [Merge LORA to Base](#merge-lora-to-base)
+  - [Special Tokens](#special-tokens)
 - [Common Errors](#common-errors-)
+  - [Tokenization Mismatch b/w Training & Inference](#tokenization-mismatch-bw-inference--training)
 - [Need Help?](#need-help-)
 - [Badge](#badge-)
 - [Community Showcase](#community-showcase)
@@ -100,7 +102,7 @@ pip3 install -e '.[flash-attn,deepspeed]'
 ```
 
 ### Usage
-```bash
+```bashtet
 # finetune lora
 accelerate launch -m axolotl.cli.train examples/openllama-3b/lora.yml
 
@@ -251,6 +253,13 @@ Have dataset(s) in one of the following format (JSONL recommended):
   ```json
   {"conversations": [{"from": "...", "value": "..."}]}
   ```
+- `llama-2`: the json is the same format as `sharegpt` above, with the following config (see the [config section](#config) for more details)
+    ```yml
+    datasets:
+      - path: <your-path>
+        type: sharegpt
+        conversation: llama-2
+    ```
 - `completion`: raw corpus
   ```json
   {"text": "..."}
@@ -691,9 +700,11 @@ warmup_ratio: 0.05  # cannot use with warmup_steps
 learning_rate: 0.00003
 lr_quadratic_warmup:
 logging_steps:
+eval_steps: # Leave empty to eval at each epoch, integers for every N steps. decimal for fraction of total steps
+evals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_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, integers for every N steps. decimal for fraction of total steps
+saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_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.
@@ -772,7 +783,7 @@ max_grad_norm:
 # Augmentation techniques
 # NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings
 # currently only supported on Llama and Mistral
-noisy_embedding_alpha:
+neftune_noise_alpha:
 
 # Whether to bettertransformers
 flash_optimum:
@@ -968,6 +979,22 @@ wandb_name:
 wandb_log_model:
 ```
 
+##### Special Tokens
+
+It is important to have special tokens like delimiters, end-of-sequence, beginning-of-sequence in your tokenizer's vocabulary.  This will help you avoid tokenization issues and help your model train better.  You can do this in axolotl like this:
+
+```yml
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"
+tokens: # these are delimiters
+  - "<|im_start|>"
+  - "<|im_end|>"
+```
+
+When you include these tokens in your axolotl config, axolotl adds these tokens to the tokenizer's vocabulary.
+
 ### Inference
 
 Pass the appropriate flag to the train command:
@@ -1020,6 +1047,10 @@ Please reduce any below
   - `gradient_accumulation_steps`
   - `sequence_len`
 
+If it does not help, try running without deepspeed and without accelerate (replace "accelerate launch" with "python") in the command.
+
+Using adamw_bnb_8bit might also save you some memory.
+
 > `failed (exitcode: -9)`
 
 Usually means your system has run out of system memory.
@@ -1042,6 +1073,20 @@ It's safe to ignore it.
 
 See the [NCCL](docs/nccl.md) guide.
 
+
+### Tokenization Mismatch b/w Inference & Training
+
+For many formats, Axolotl constructs prompts by concatenating token ids _after_ tokenizing strings.  The reason for concatenating token ids rather than operating on strings is to maintain precise accounting for attention masks.
+
+If you decode a prompt constructed by axolotl, you might see spaces between tokens (or lack thereof) that you do not expect, especially around delimiters and special tokens.  When you are starting out with a new format, you should always do the following:
+
+1. Materialize some data using `python -m axolotl.cli.preprocess your_config.yml --debug`, and then decode the first few rows with your model's tokenizer.
+2. During inference, right before you pass a tensor of token ids to your model, decode these tokens back into a string.
+3. Make sure the inference string from #2 looks **exactly** like the data you fine tuned on from #1, including spaces and new lines.  If they aren't the same adjust your inference server accordingly.
+4. As an additional troubleshooting step, you can look look at the token ids between 1 and 2 to make sure they are identical.
+
+Having misalignment between your prompts during training and inference can cause models to perform very poorly, so it is worth checking this.  See [this blog post](https://hamel.dev/notes/llm/05_tokenizer_gotchas.html) for a concrete example.
+
 ## Need help? 🙋♂️
 
 Join our [Discord server](https://discord.gg/HhrNrHJPRb) where we can help you

deepspeed/zero3_bf16.json

@@ -0,0 +1,39 @@
+{
+  "zero_optimization": {
+    "stage": 3,
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "sub_group_size": 0,
+    "reduce_bucket_size": "auto",
+    "stage3_prefetch_bucket_size": "auto",
+    "stage3_param_persistence_threshold": "auto",
+    "stage3_max_live_parameters": 0,
+    "stage3_max_reuse_distance": 0,
+    "stage3_gather_16bit_weights_on_model_save": true
+  },
+  "bf16": {
+    "enabled": true
+  },
+  "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": "auto",
+      "eps": "auto",
+      "weight_decay": "auto"
+    }
+  },
+  "gradient_accumulation_steps": "auto",
+  "train_batch_size": "auto",
+  "train_micro_batch_size_per_gpu": "auto",
+  "wall_clock_breakdown": false
+}

docker/Dockerfile

@@ -10,7 +10,7 @@ ARG PYTORCH_VERSION="2.0.1"
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
 
 RUN apt-get update && \
-    apt-get install -y vim curl
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
 
 WORKDIR /workspace
 

examples/cerebras/btlm-ft.yml

@@ -72,8 +72,8 @@ gptq_groupsize:
 gptq_model_v1:
 
 warmup_steps: 32
-eval_steps:
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 save_total_limit:
 
 debug:

examples/cerebras/qlora.yml

@@ -49,8 +49,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

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

@@ -54,8 +54,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -56,8 +56,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -54,8 +54,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -56,8 +56,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -54,8 +54,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -56,8 +56,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -51,8 +51,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40
-eval_steps: 5
-save_steps: 43
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -80,8 +80,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 5
-save_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.000001

examples/falcon/config-7b.yml

@@ -51,8 +51,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40
-eval_steps: 5
-save_steps: 43
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/gptj/qlora.yml

@@ -46,8 +46,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/jeopardy-bot/config.yml

@@ -42,8 +42,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 110
-save_steps: 660
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/llama-2/fft_optimized.yml

@@ -58,9 +58,9 @@ flash_attn_fuse_qkv: false
 flash_attn_fuse_mlp: true
 
 warmup_steps: 100
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed: #deepspeed/zero2.json # multi-gpu only
 weight_decay: 0.1

examples/llama-2/gptq-lora.yml

@@ -62,8 +62,8 @@ flash_attention:
 sdp_attention:
 flash_optimum:
 warmup_steps: 100
-eval_steps:
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/llama-2/lora.yml

@@ -54,10 +54,10 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/llama-2/qlora.yml

@@ -56,9 +56,9 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/llama-2/relora.yml

@@ -60,8 +60,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps: 50
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/llama-2/tiny-llama.yml

@@ -54,9 +54,9 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mamba/config.yml

@@ -47,10 +47,10 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 10
-eval_steps:
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:  0.25
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mistral/config.yml

@@ -46,10 +46,10 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mistral/mixtral.yml

@@ -1,5 +1,5 @@
-base_model: DiscoResearch/mixtral-7b-8expert
-model_type: MixtralForCausalLM
+base_model: mistralai/Mixtral-8x7B-v0.1
+model_type: AutoModelForCausalLM
 tokenizer_type: LlamaTokenizer
 trust_remote_code: true
 
@@ -14,6 +14,15 @@ dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: ./qlora-out
 
+## You can optionally freeze the entire model and unfreeze a subset of parameters
+unfrozen_parameters:
+#  - lm_head.*
+#  - model.embed_tokens.*
+#  - model.layers.2[0-9]+.block_sparse_moe.gate.*
+#  - model.layers.2[0-9]+.block_sparse_moe.experts.*
+#  - model.layers.3[0-9]+.block_sparse_moe.gate.*
+#  - model.layers.3[0-9]+.block_sparse_moe.experts.*
+
 adapter: qlora
 lora_model_dir:
 
@@ -67,10 +76,10 @@ loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3
 
 warmup_steps: 10
-eval_steps:
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed: deepspeed/zero2.json
 weight_decay: 0.0

examples/mistral/qlora.yml

@@ -66,10 +66,10 @@ loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mpt-7b/config.yml

@@ -44,8 +44,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 110
-save_steps: 660
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0001

examples/openllama-3b/config.yml

@@ -49,8 +49,8 @@ flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/openllama-3b/lora.yml

@@ -54,8 +54,8 @@ flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/openllama-3b/qlora.yml

@@ -48,8 +48,8 @@ flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/phi/phi-ft.yml

@@ -59,8 +59,8 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 100
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/phi/phi-qlora.yml

@@ -59,8 +59,8 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 100
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/pythia/lora.yml

@@ -33,5 +33,5 @@ early_stopping_patience:
 resume_from_checkpoint:
 local_rank:
 weight_decay: 0.1
-eval_steps: 0.05
+evals_per_epoch: 4
 logging_steps: 1

examples/qwen/lora.yml

@@ -56,10 +56,10 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/qwen/qlora.yml

@@ -56,10 +56,10 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/redpajama/config-3b.yml

@@ -45,8 +45,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 110
-save_steps: 660
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0001

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

@@ -45,8 +45,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 50
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0

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

@@ -78,8 +78,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 50
-save_steps: 50
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

requirements.txt

@@ -2,7 +2,7 @@
 auto-gptq==0.5.1
 packaging
 peft==0.6.0
-transformers @ git+https://github.com/huggingface/transformers.git@df5c5c62ae253055336f5bb0828ca8e3e15ab6bd
+transformers @ git+https://github.com/huggingface/transformers.git@ebfdb9ca62205279d5019ef1403877461b3b2da4
 tokenizers==0.15.0
 bitsandbytes>=0.41.1
 accelerate==0.24.1
@@ -29,7 +29,7 @@ scipy
 scikit-learn==1.2.2
 pynvml
 art
-fschat==0.2.29
+fschat==0.2.34
 gradio==3.50.2
 tensorboard
 

src/axolotl/cli/train.py

@@ -22,8 +22,8 @@ LOG = logging.getLogger("axolotl.cli.train")
 
 def do_cli(config: Path = Path("examples/"), **kwargs):
     # pylint: disable=duplicate-code
-    print_axolotl_text_art()
     parsed_cfg = load_cfg(config, **kwargs)
+    print_axolotl_text_art()
     check_accelerate_default_config()
     check_user_token()
     parser = transformers.HfArgumentParser((TrainerCliArgs))

src/axolotl/core/trainer_builder.py

@@ -692,6 +692,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             and self.cfg.lr_scheduler not in ("one_cycle", "log_sweep")
             else "cosine"
         )
+        training_arguments_kwargs["lr_scheduler_kwargs"] = (
+            self.cfg.lr_scheduler_kwargs if self.cfg.lr_scheduler_kwargs else {}
+        )
         training_arguments_kwargs["weight_decay"] = (
             self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
         )
@@ -712,6 +715,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs
         )
         training_arguments_kwargs["model_type"] = self.cfg.model_config_type
+
+        if self.cfg.neftune_noise_alpha is not None:
+            training_arguments_kwargs[
+                "neftune_noise_alpha"
+            ] = self.cfg.neftune_noise_alpha
+
         training_args = (
             AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
                 **training_arguments_kwargs,

src/axolotl/models/mixtral/__init__.py

@@ -1,6 +0,0 @@
-"""
-Custom modeling code for mixtral
-"""
-
-from .configuration_moe_mistral import MixtralConfig  # noqa
-from .modeling_moe_mistral import MixtralForCausalLM  # noqa

src/axolotl/models/mixtral/configuration_moe_mistral.py

@@ -1,154 +0,0 @@
-# coding=utf-8
-# Copyright 2023 Mistral AI and the HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Mistral model configuration"""
-
-from transformers.configuration_utils import PretrainedConfig
-from transformers.utils import logging
-
-logger = logging.get_logger(__name__)
-
-MISTRAL_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "mistralai/Mistral-7B-v0.1": "https://huggingface.co/mistralai/Mistral-7B-v0.1/resolve/main/config.json",
-    "mistralai/Mistral-7B-Instruct-v0.1": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1/resolve/main/config.json",
-}
-
-
-class MixtralConfig(PretrainedConfig):
-    r"""
-    This is the configuration class to store the configuration of a [`MistralModel`]. It is used to instantiate an
-    Mistral model according to the specified arguments, defining the model architecture. Instantiating a configuration
-    with the defaults will yield a similar configuration to that of the Mistral-7B-v0.1 or Mistral-7B-Instruct-v0.1.
-
-    [mistralai/Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
-    [mistralai/Mistral-7B-Instruct-v0.1](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1)
-
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
-
-
-    Args:
-        vocab_size (`int`, *optional*, defaults to 32000):
-            Vocabulary size of the Mistral model. Defines the number of different tokens that can be represented by the
-            `inputs_ids` passed when calling [`MistralModel`]
-        hidden_size (`int`, *optional*, defaults to 4096):
-            Dimension of the hidden representations.
-        intermediate_size (`int`, *optional*, defaults to 14336):
-            Dimension of the MLP representations.
-        num_hidden_layers (`int`, *optional*, defaults to 32):
-            Number of hidden layers in the Transformer encoder.
-        num_attention_heads (`int`, *optional*, defaults to 32):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        num_key_value_heads (`int`, *optional*, defaults to 8):
-            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
-            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
-            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
-            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
-            by meanpooling all the original heads within that group. For more details checkout [this
-            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `8`.
-        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
-            The non-linear activation function (function or string) in the decoder.
-        max_position_embeddings (`int`, *optional*, defaults to `4096*32`):
-            The maximum sequence length that this model might ever be used with. Mistral's sliding window attention
-            allows sequence of up to 4096*32 tokens.
-        initializer_range (`float`, *optional*, defaults to 0.02):
-            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
-            The epsilon used by the rms normalization layers.
-        use_cache (`bool`, *optional*, defaults to `True`):
-            Whether or not the model should return the last key/values attentions (not used by all models). Only
-            relevant if `config.is_decoder=True`.
-        pad_token_id (`int`, *optional*):
-            The id of the padding token.
-        bos_token_id (`int`, *optional*, defaults to 1):
-            The id of the "beginning-of-sequence" token.
-        eos_token_id (`int`, *optional*, defaults to 2):
-            The id of the "end-of-sequence" token.
-        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
-            Whether the model's input and output word embeddings should be tied.
-        rope_theta (`float`, *optional*, defaults to 10000.0):
-            The base period of the RoPE embeddings.
-        sliding_window (`int`, *optional*, defaults to 4096):
-            Sliding window attention window size. If not specified, will default to `4096`.
-        attention_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout ratio for the attention probabilities.
-
-    ```python
-    >>> from transformers import MistralModel, MistralConfig
-
-    >>> # Initializing a Mistral 7B style configuration
-    >>> configuration = MixtralConfig()
-
-    >>> # Initializing a model from the Mistral 7B style configuration
-    >>> model = MixtralModel(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-    ```"""
-
-    model_type = "mistral"
-    keys_to_ignore_at_inference = ["past_key_values"]
-
-    def __init__(
-        self,
-        vocab_size=32000,
-        hidden_size=4096,
-        intermediate_size=14336,
-        num_hidden_layers=32,
-        num_attention_heads=32,
-        num_key_value_heads=8,
-        hidden_act="silu",
-        max_position_embeddings=4096 * 32,
-        initializer_range=0.02,
-        rms_norm_eps=1e-6,
-        use_cache=True,
-        pad_token_id=None,
-        bos_token_id=1,
-        eos_token_id=2,
-        tie_word_embeddings=False,
-        rope_theta=10000.0,
-        attention_dropout=0.0,
-        num_experts_per_token=2,
-        num_experts=8,
-        **kwargs,
-    ):
-        self.vocab_size = vocab_size
-        self.max_position_embeddings = max_position_embeddings
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-
-        # for backward compatibility
-        if num_key_value_heads is None:
-            num_key_value_heads = num_attention_heads
-
-        self.num_key_value_heads = num_key_value_heads
-        self.hidden_act = hidden_act
-        self.initializer_range = initializer_range
-        self.rms_norm_eps = rms_norm_eps
-        self.use_cache = use_cache
-        self.rope_theta = rope_theta
-        self.attention_dropout = attention_dropout
-        self.num_experts = num_experts
-        self.num_experts_per_token = num_experts_per_token
-
-        # pylint: disable=duplicate-code
-        super().__init__(
-            pad_token_id=pad_token_id,
-            bos_token_id=bos_token_id,
-            eos_token_id=eos_token_id,
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )

src/axolotl/monkeypatch/fastchat_conversation_turns.py

@@ -83,14 +83,21 @@ def get_turns(  # pylint: disable=too-many-return-statements
                 yield role + ":", ""
         return
     if self.sep_style == SeparatorStyle.LLAMA2:
-        seps = [self.sep, self.sep2]
         if self.system_message:
+            if self.messages:
+                # For llama, the system message is incorporated into the first human instruction
+                first_role, first_msg = self.messages[0]
+                if first_role == self.roles[0]:
+                    system_prompt += first_msg
+                    self.messages.pop(0)
             yield "", system_prompt
-        else:
-            yield "", "[INST] "
-        for i, (role, message) in enumerate(self.messages[1:]):
+        for i, (role, message) in enumerate(self.messages):
             if message:
-                yield role + " ", message + seps[i % 2]
+                if (i % 2 == 0 and not self.system_message) or (
+                    i % 2 != 0 and self.system_message
+                ):
+                    role = "<s> " + role
+                yield role + " ", message
             else:
                 yield role, ""
         return

src/axolotl/monkeypatch/mixtral/__init__.py

@@ -0,0 +1,22 @@
+"""
+Patches to support multipack for mixtral
+"""
+import transformers
+
+
+def replace_mixtral_attn_with_multipack_flash_attn():
+    from .modeling_mixtral import (
+        MixtralMultipackFlashAttention2,
+        mixtral_decoder_layer_forward,
+        mixtral_model_forward,
+    )
+
+    transformers.models.mixtral.modeling_mixtral.MixtralDecoderLayer.forward = (
+        mixtral_decoder_layer_forward
+    )
+    transformers.models.mixtral.modeling_mixtral.MixtralModel.forward = (
+        mixtral_model_forward
+    )
+    transformers.models.mixtral.modeling_mixtral.MISTRAL_ATTENTION_CLASSES[
+        "flash_attention_2"
+    ] = MixtralMultipackFlashAttention2

src/axolotl/monkeypatch/mixtral/modeling_mixtral.py

@@ -0,0 +1,379 @@
+"""
+Mixtral modeling for multipack
+"""
+# pylint: disable=missing-module-docstring,unused-argument,protected-access,pointless-string-statement,duplicate-code
+import logging
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from einops import rearrange
+from flash_attn import flash_attn_varlen_qkvpacked_func
+from transformers import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.models.mixtral.modeling_mixtral import (
+    MixtralFlashAttention2,
+    apply_rotary_pos_emb,
+    repeat_kv,
+)
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+
+LOG = logging.getLogger("axolotl.monkeypatch.mixtral")
+
+
+class MixtralMultipackFlashAttention2(MixtralFlashAttention2):
+    """
+    Custom multipack implementation w flash attention 2
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._flash_attn_uses_top_left_mask = True
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(
+            bsz, q_len, self.num_heads, self.head_dim
+        ).transpose(1, 2)
+        key_states = key_states.view(
+            bsz, q_len, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+        value_states = value_states.view(
+            bsz, q_len, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin, position_ids
+        )
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        if cu_seqlens is not None and max_seqlen is not None and cu_seqlens.dim() == 1:
+            # special handling using sample packing
+            qkv = torch.stack(
+                [query_states, key_states, value_states], dim=2
+            )  # [bsz, nh, 3, q_len, hd]
+            qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
+            qkv = rearrange(qkv, "b s ... -> (b s) ...")
+
+            attn_output = flash_attn_varlen_qkvpacked_func(
+                qkv,
+                cu_seqlens,
+                max_seqlen,
+                dropout_p=self.attention_dropout,
+                softmax_scale=None,
+                causal=True,
+            )
+            attn_output = rearrange(attn_output, "(b s) ... -> b s ...", b=bsz)
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+def mixtral_decoder_layer_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: Optional[bool] = False,
+    output_router_logits: Optional[bool] = False,
+    use_cache: Optional[bool] = False,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[torch.Tensor] = None,
+    **kwargs,
+) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+    if "padding_mask" in kwargs:
+        warnings.warn(
+            "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+        )
+    """
+    Args:
+        hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+        attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+            `(batch, sequence_length)` where padding elements are indicated by 0.
+        past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+            returned tensors for more detail.
+        output_router_logits (`bool`, *optional*):
+            Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
+            should not be returned during inference.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+            (see `past_key_values`).
+    """
+
+    residual = hidden_states
+
+    hidden_states = self.input_layernorm(hidden_states)
+
+    # Self Attention
+    hidden_states, self_attn_weights, present_key_value = self.self_attn(
+        hidden_states=hidden_states,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_value=past_key_value,
+        output_attentions=output_attentions,
+        use_cache=use_cache,
+        cu_seqlens=cu_seqlens,
+        max_seqlen=max_seqlen,
+    )
+    hidden_states = residual + hidden_states
+
+    # Fully Connected
+    residual = hidden_states
+    hidden_states = self.post_attention_layernorm(hidden_states)
+    hidden_states, router_logits = self.block_sparse_moe(hidden_states)
+    hidden_states = residual + hidden_states
+
+    outputs = (hidden_states,)
+
+    if output_attentions:
+        outputs += (self_attn_weights,)
+
+    if use_cache:
+        outputs += (present_key_value,)
+
+    if output_router_logits:
+        outputs += (router_logits,)
+
+    return outputs
+
+
+def mixtral_model_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+) -> Union[Tuple, MoeModelOutputWithPast]:
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_router_logits = (
+        output_router_logits
+        if output_router_logits is not None
+        else self.config.output_router_logits
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # retrieve input_ids and inputs_embeds
+    if input_ids is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
+        )
+    if input_ids is not None:
+        batch_size, seq_length = input_ids.shape
+    elif inputs_embeds is not None:
+        batch_size, seq_length, _ = inputs_embeds.shape
+    else:
+        raise ValueError(
+            "You have to specify either decoder_input_ids or decoder_inputs_embeds"
+        )
+
+    past_key_values_length = 0
+
+    if use_cache:
+        use_legacy_cache = not isinstance(past_key_values, Cache)
+        if use_legacy_cache:
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+        past_key_values_length = past_key_values.get_usable_length(seq_length)
+
+    cu_seqlens = None
+    max_seqlen = None
+    if position_ids is None:
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+        position_ids = torch.arange(
+            past_key_values_length,
+            seq_length + past_key_values_length,
+            dtype=torch.long,
+            device=device,
+        )
+        position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+    else:
+        position_ids = position_ids.view(-1, seq_length).long()
+        cu_seqlens, max_seqlen = get_cu_seqlens_from_pos_ids(position_ids)
+        cu_seqlens = cu_seqlens.squeeze()
+
+    if inputs_embeds is None:
+        inputs_embeds = self.embed_tokens(input_ids)
+
+    if attention_mask is not None and self._use_flash_attention_2 and use_cache:
+        is_padding_right = attention_mask[:, -1].sum().item() != batch_size
+        if is_padding_right:
+            raise ValueError(
+                "You are attempting to perform batched generation with padding_side='right'"
+                " this may lead to unexpected behaviour for Flash Attention version of Mixtral. Make sure to "
+                " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+            )
+
+    if self._use_flash_attention_2:
+        # 2d mask is passed through the layers
+        attention_mask = (
+            attention_mask
+            if (attention_mask is not None and 0 in attention_mask)
+            else None
+        )
+    else:
+        # 4d mask is passed through the layers
+        attention_mask = _prepare_4d_causal_attention_mask(
+            attention_mask,
+            (batch_size, seq_length),
+            inputs_embeds,
+            past_key_values_length,
+            sliding_window=self.config.sliding_window,
+        )
+
+    hidden_states = inputs_embeds
+
+    if self.gradient_checkpointing and self.training:
+        if use_cache:
+            LOG.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+            )
+            use_cache = False
+
+    # decoder layers
+    all_hidden_states = () if output_hidden_states else None
+    all_self_attns = () if output_attentions else None
+    all_router_logits = () if output_router_logits else None
+    next_decoder_cache = None
+
+    for decoder_layer in self.layers:
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if self.gradient_checkpointing and self.training:
+            layer_outputs = self._gradient_checkpointing_func(
+                decoder_layer.__call__,
+                hidden_states,
+                attention_mask,
+                position_ids,
+                past_key_values,
+                output_attentions,
+                output_router_logits,
+                use_cache,
+                cu_seqlens,
+                max_seqlen,
+            )
+        else:
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                output_attentions=output_attentions,
+                output_router_logits=output_router_logits,
+                use_cache=use_cache,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+
+        hidden_states = layer_outputs[0]
+
+        if use_cache:
+            next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+        if output_attentions:
+            all_self_attns += (layer_outputs[1],)
+
+        if output_router_logits:
+            all_router_logits += (layer_outputs[-1],)
+
+    hidden_states = self.norm(hidden_states)
+
+    # add hidden states from the last decoder layer
+    if output_hidden_states:
+        all_hidden_states += (hidden_states,)
+
+    next_cache = None
+    if use_cache:
+        next_cache = (
+            next_decoder_cache.to_legacy_cache()
+            if use_legacy_cache
+            else next_decoder_cache
+        )
+
+    if not return_dict:
+        return tuple(
+            v
+            for v in [
+                hidden_states,
+                next_cache,
+                all_hidden_states,
+                all_self_attns,
+                all_router_logits,
+            ]
+            if v is not None
+        )
+
+    return MoeModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+        router_logits=all_router_logits,
+    )

src/axolotl/monkeypatch/neft_embeddings.py

@@ -1,65 +0,0 @@
-"""
-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/llama2_chat.py

@@ -81,8 +81,9 @@ class LLama2ChatTokenizingStrategy(PromptTokenizingStrategy):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.sequence_len = 4096
-        self.tokenizer.add_special_tokens({"pad_token": "<pad>"})
+        self.tokenizer.add_special_tokens(
+            {"pad_token": getattr(self.tokenizer, "pad_token", "<pad>")}
+        )
         # https://huggingface.co/meta-llama/Llama-2-7b-chat-hf/blob/main/added_tokens.json
 
     def tokenize_prompt(self, prompt):

src/axolotl/prompters.py

@@ -33,8 +33,8 @@ class AlpacaPrompter(Prompter):
     Base class for alpaca prompters
     """
 
-    system_prompt = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n"
-    system_no_input_prompt = "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n"
+    system_prompt = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request."
+    system_no_input_prompt = "Below is an instruction that describes a task. Write a response that appropriately completes the request."
     system_format: str = "{system}"
     turn_format: str
     turn_no_input_format: str

src/axolotl/train.py

@@ -16,8 +16,8 @@ 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.freeze import freeze_parameters_except
 from axolotl.utils.models import load_model, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
 
@@ -78,6 +78,9 @@ def train(
             )
     resume_from_checkpoint = cfg.resume_from_checkpoint
 
+    if cfg.unfrozen_parameters:
+        freeze_parameters_except(model, cfg.unfrozen_parameters)
+
     trainer = setup_trainer(
         cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
     )
@@ -176,21 +179,19 @@ def train(
     return model, tokenizer
 
 
-def pretrain_hooks(cfg, trainer):
+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):
+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/config.py

@@ -77,6 +77,15 @@ def normalize_config(cfg):
     else:
         cfg.torch_dtype = torch.float32
 
+    if cfg.saves_per_epoch:
+        save_steps = 1.0 / (cfg.saves_per_epoch * cfg.num_epochs)
+        if save_steps < 1.0:  # prevent saves on every step
+            cfg.save_steps = save_steps
+    if cfg.evals_per_epoch:
+        eval_steps = 1.0 / (cfg.evals_per_epoch * cfg.num_epochs)
+        if eval_steps < 1.0:  # prevent evals on every step
+            cfg.eval_steps = eval_steps
+
     cfg.dataset_processes = cfg.dataset_processes or os.cpu_count()
 
     if not cfg.base_model_config:
@@ -352,6 +361,27 @@ def validate_config(cfg):
                 cfg.datasets[idx].type = cfg.datasets[idx].type.replace(
                     "sharegpt_simple", "sharegpt"
                 )
+
+    if cfg.saves_per_epoch and cfg.save_steps:
+        raise ValueError(
+            "save_steps and saves_per_epoch are mutually exclusive and cannot be used together."
+        )
+    if cfg.saves_per_epoch and cfg.save_strategy and cfg.save_strategy != "steps":
+        raise ValueError(
+            "save_strategy must be empty or set to `steps` when used with saves_per_epoch."
+        )
+    if cfg.evals_per_epoch and cfg.eval_steps:
+        raise ValueError(
+            "eval_steps and evals_per_epoch are mutually exclusive and cannot be used together."
+        )
+    if (
+        cfg.evals_per_epoch
+        and cfg.evaluation_strategy
+        and cfg.evaluation_strategy != "steps"
+    ):
+        raise ValueError(
+            "evaluation_strategy must be empty or set to `steps` when used with evals_per_epoch."
+        )
     if cfg.save_strategy and cfg.save_steps and cfg.save_strategy != "steps":
         raise ValueError(
             "save_strategy and save_steps mismatch. Please set save_strategy to 'steps' or remove save_steps."
@@ -404,6 +434,20 @@ def validate_config(cfg):
             "wandb_run_id sets the ID of the run. If you would like to set the name, please use wandb_name instead."
         )
 
+    if cfg.noisy_embedding_alpha is not None:
+        # Deprecated, use neftune_noise_alpha
+        LOG.warning("noisy_embedding_alpha is deprecated, use neftune_noise_alpha")
+        if cfg.neftune_noise_alpha is None:
+            cfg.neftune_noise_alpha = cfg.noisy_embedding_alpha
+        else:
+            # User is providing both; bail and have them sort out their settings
+            raise ValueError(
+                "noisy_embedding_alpha is deprecated, use neftune_noise_alpha; both are set, please remove the deprecated noisy_embedding_alpha setting"
+            )
+
+    if cfg.neftune_noise_alpha is not None and cfg.neftune_noise_alpha <= 0.0:
+        raise ValueError("neftune_noise_alpha must be > 0.0")
+
     # TODO
     # MPT 7b
     # https://github.com/facebookresearch/bitsandbytes/issues/25

src/axolotl/utils/freeze.py

@@ -0,0 +1,38 @@
+"""
+module to freeze/unfreeze parameters by name
+"""
+import logging
+import re
+
+from axolotl.utils.distributed import is_main_process
+
+LOG = logging.getLogger("axolotl.utils.freeze")
+
+
+def freeze_parameters_except(model, regex_patterns):
+    """
+    Freezes all layers of the given model except for the layers that match given regex patterns.
+    Periods in the patterns are treated as literal periods, not as wildcard characters.
+
+    Parameters:
+    - model (nn.Module): The PyTorch model to be modified.
+    - regex_patterns (list of str): List of regex patterns to match layer names to keep unfrozen.
+
+    Returns:
+    None; the model is modified in place.
+    """
+    # Escape periods and compile the regex patterns
+    compiled_patterns = [
+        re.compile(pattern.replace(".", "\\.")) for pattern in regex_patterns
+    ]
+
+    # First, freeze all parameters in the model
+    for param in model.parameters():
+        param.requires_grad = False
+
+    # Unfreeze layers that match the regex patterns
+    for name, param in model.named_parameters():
+        if any(pattern.match(name) for pattern in compiled_patterns):
+            if is_main_process():
+                LOG.debug(f"unfreezing {name}")
+            param.requires_grad = True

src/axolotl/utils/models.py

@@ -21,6 +21,7 @@ from transformers import (  # noqa: F401
     PreTrainedModel,
     PreTrainedTokenizerBase,
 )
+from transformers.deepspeed import is_deepspeed_zero3_enabled
 
 from axolotl.models.mamba import fix_mamba_attn_for_loss
 from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
@@ -54,25 +55,19 @@ def check_model_config(cfg: DictDefault, model_config: AutoConfig):
 def load_model_config(cfg):
     model_config_name = cfg.base_model_config or cfg.base_model
     trust_remote_code = cfg.trust_remote_code is True
-    model_type = cfg.model_type
 
-    if model_type == "MixtralForCausalLM":
-        from axolotl.models.mixtral.configuration_moe_mistral import MixtralConfig
-
-        model_config = MixtralConfig.from_pretrained(model_config_name)
-    else:
-        try:
-            model_config = AutoConfig.from_pretrained(
-                model_config_name, trust_remote_code=trust_remote_code
+    try:
+        model_config = AutoConfig.from_pretrained(
+            model_config_name, trust_remote_code=trust_remote_code
+        )
+    except ValueError as err:
+        if "mamba" in model_config_name:
+            return addict.Dict(
+                {
+                    "model_type": "mamba",
+                }
             )
-        except ValueError as err:
-            if "mamba" in model_config_name:
-                return addict.Dict(
-                    {
-                        "model_type": "mamba",
-                    }
-                )
-            raise err
+        raise err
 
     if cfg.model_config:
         for key, val in cfg.model_config.items():
@@ -255,6 +250,18 @@ def load_model(
         LOG.info("patching with flash attention")
         replace_mistral_attn_with_flash_attn(packed=cfg.sample_packing)
 
+    if (
+        cfg.model_config_type == "mixtral"
+        and cfg.flash_attention
+        and cfg.sample_packing
+    ):
+        from axolotl.monkeypatch.mixtral import (
+            replace_mixtral_attn_with_multipack_flash_attn,
+        )
+
+        LOG.info("patching with flash attention")
+        replace_mixtral_attn_with_multipack_flash_attn()
+
     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,
@@ -279,6 +286,9 @@ def load_model(
     model_kwargs["max_memory"] = cfg.max_memory
     model_kwargs["torch_dtype"] = cfg.torch_dtype
 
+    if is_deepspeed_zero3_enabled():
+        del model_kwargs["device_map"]
+
     if cfg.model_revision:
         model_kwargs["revision"] = cfg.model_revision
     if cfg.gptq:
@@ -302,15 +312,26 @@ def load_model(
             bnb_4bit_quant_type="nf4",
         )
     # sample packing uses custom FA2 patch
-    if cfg.flash_attention and not cfg.sample_packing:
-        if (
-            cfg.is_llama_derived_model
-            or cfg.is_falcon_derived_model
-            or cfg.is_mistral_derived_model
-        ):
-            # TODO enable once properly supported in transformers
-            # model_kwargs["attn_implementation"] = "flash_attention_2"
-            model_kwargs["use_flash_attention_2"] = True  # legacy, to be deprecated
+    if cfg.flash_attention:
+        if not cfg.sample_packing:
+            if (
+                cfg.is_llama_derived_model
+                or cfg.is_falcon_derived_model
+                or cfg.is_mistral_derived_model
+                or model_config.model_type == "mixtral"
+            ):
+                model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "flash_attention_2"
+                )
+        else:
+            if model_config.model_type == "mixtral":
+                model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "flash_attention_2"
+                )
+            else:
+                model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "eager"
+                )
 
     try:
         if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
@@ -372,15 +393,6 @@ def load_model(
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
                 **model_kwargs,
             )
-        elif model_type == "MixtralForCausalLM":
-            from axolotl.models.mixtral import MixtralForCausalLM
-
-            model = MixtralForCausalLM.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,
-                **model_kwargs,
-            )
         elif model_type == "MambaLMHeadModel":
             # FIXME this is janky at best and hacked together to make it work
             MambaLMHeadModel = fix_mamba_attn_for_loss()  # pylint: disable=invalid-name

src/axolotl/utils/trainer.py

@@ -276,6 +276,7 @@ def prepare_optim_env(cfg):
         setup_fsdp_envs(cfg)
     elif cfg.deepspeed:
         os.environ["ACCELERATE_USE_DEEPSPEED"] = "true"
+        os.environ["ACCELERATE_DEEPSPEED_CONFIG_FILE"] = cfg.deepspeed
 
 
 def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps):

tests/test_prompt_tokenizers.py

@@ -114,6 +114,76 @@ class TestPromptTokenizationStrategies(unittest.TestCase):
                 in self._caplog.records[0].message
             )
 
+    def test_sharegpt_llama(self):
+        "Make sure the sharegpt/llama is tokenized and formatted correctly."
+        prompter = ShareGPTPrompterV2(conversation="llama-2")
+        strat = ShareGPTPromptTokenizingStrategy(
+            prompter,
+            self.tokenizer,
+            False,
+            2048,
+        )
+
+        def tokenize(conv):
+            return strat.tokenize_prompt(conv)["input_ids"]
+
+        def decode(ids):
+            return strat.tokenizer.decode(ids)
+
+        # Multi-turn conversations
+        multi_turn_conv = {
+            "conversations": [
+                {"from": "system", "value": "lorem"},
+                {"from": "human", "value": "abc"},
+                {"from": "gpt", "value": "ipsum"},
+                {"from": "human", "value": "123"},
+                {"from": "gpt", "value": "sit"},
+            ]
+        }
+        # fmt: off
+        mt_ids = tokenize(multi_turn_conv)
+        assert decode(mt_ids) == '<s> [INST] <<SYS>>\nlorem\n<</SYS>>\n\nabc [/INST] ipsum</s><s> [INST] 123 [/INST] sit</s>'
+        assert mt_ids == [1, 518, 25580, 29962, 3532, 14816, 29903, 6778, 13, 29880, 3668, 13, 29966, 829, 14816, 29903, 6778, 13, 13, 10736, 518, 29914, 25580, 29962, 23421, 2, 1, 518, 25580, 29962, 29871, 29896, 29906, 29941, 518, 29914, 25580, 29962, 7845, 2]
+
+        # Single-turn conversations
+        single_turn_conv = {
+            "conversations": [
+                {"from": "system", "value": "lorem"},
+                {"from": "human", "value": "abc"},
+                {"from": "gpt", "value": "ipsum"},
+            ]
+        }
+
+        st_ids = tokenize(single_turn_conv)
+        assert decode(st_ids) == '<s> [INST] <<SYS>>\nlorem\n<</SYS>>\n\nabc [/INST] ipsum</s>'
+        assert st_ids == [1, 518, 25580, 29962, 3532, 14816, 29903, 6778, 13, 29880, 3668, 13, 29966, 829, 14816, 29903, 6778, 13, 13, 10736, 518, 29914, 25580, 29962, 23421, 2]
+
+        # No system message, single-turn
+        no_sys_conv = {
+            "conversations": [
+                {"from": "human", "value": "abc"},
+                {"from": "gpt", "value": "ipsum"},
+            ]
+        }
+
+        ns_ids = tokenize(no_sys_conv)
+        assert decode(ns_ids) == '<s> [INST] abc [/INST] ipsum</s>'
+        assert ns_ids == [1, 518, 25580, 29962, 25638, 518, 29914, 25580, 29962, 23421, 2]
+
+        # No system message, multi-turn
+        no_sys_mt_conv = {
+            "conversations": [
+                {"from": "human", "value": "abc"},
+                {"from": "gpt", "value": "ipsum"},
+                {"from": "human", "value": "123"},
+                {"from": "gpt", "value": "sit"},
+            ]
+        }
+        ns_mt_ids = tokenize(no_sys_mt_conv)
+        assert decode(ns_mt_ids) == '<s> [INST] abc [/INST] ipsum</s><s> [INST] 123 [/INST] sit</s>'
+        assert ns_mt_ids == [1, 518, 25580, 29962, 25638, 518, 29914, 25580, 29962, 23421, 2, 1, 518, 25580, 29962, 29871, 29896, 29906, 29941, 518, 29914, 25580, 29962, 7845, 2]
+        # fmt: on
+
     def test_sharegpt_changes_roles(self):
         conversation = {
             "roles": ["USER", "CHARACTER"],