Merge branch 'main' into flx_attn_support

fixes silly mistake

.github/workflows/multi-gpu-e2e.yml

@@ -4,10 +4,6 @@ on:
   pull_request:
     paths:
       - 'tests/e2e/multigpu/*.py'
-      - 'requirements.txt'
-      - 'setup.py'
-      - 'pyproject.toml'
-      - '.github/workflows/multi-gpu-e2e.yml'
   workflow_dispatch:
   schedule:
     - cron: '0 0 * * 1,4'  # Runs at 00:00 UTC every monday & thursday

README.md

@@ -50,14 +50,13 @@ Features:
 ## 🚀 Quick Start
 
 **Requirements**:
-
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
 - PyTorch ≥2.4.1
 
 ### Installation
 
-```bash
+```shell
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs
@@ -69,7 +68,7 @@ Other installation approaches are described [here](https://axolotl-ai-cloud.gith
 
 ### Your First Fine-tune
 
-```bash
+```shell
 # Fetch axolotl examples
 axolotl fetch examples
 

_quarto.yml

@@ -3,12 +3,10 @@ project:
 
 website:
   title: "Axolotl"
-  description: "We make fine-tuning accessible, scalable, and fun"
+  description: "Fine-tuning"
   favicon: favicon.jpg
-
   navbar:
-    logo: image/axolotl_logo_digital_white.svg
-    title: false
+    title: Axolotl
     background: dark
     pinned: false
     collapse: false
@@ -27,58 +25,33 @@ website:
       contents:
         - text: Home
           href: index.qmd
-
-        - section: "Getting Started"
+        - section: "How-To Guides"
           contents:
+          # TODO Edit folder structure after we have more docs.
             - docs/getting-started.qmd
             - docs/installation.qmd
-            - docs/cli.qmd
+            - docs/debugging.qmd
             - docs/inference.qmd
-
-        - section: "Dataset Formats"
-          contents: docs/dataset-formats/*
-
-        - section: "Deployments"
-          contents:
+            - docs/multipack.qmd
+            - docs/fsdp_qlora.qmd
+            - docs/input_output.qmd
+            - docs/rlhf.qmd
+            - docs/nccl.qmd
+            - docs/mac.qmd
             - docs/multi-gpu.qmd
             - docs/multi-node.qmd
-            - docs/ray-integration.qmd
-            - docs/amd_hpc.qmd
-            - docs/mac.qmd
-
-        - section: "How To Guides"
-          contents:
-            - docs/multimodal.qmd
-            - docs/rlhf.qmd
-            - docs/reward_modelling.qmd
-            - docs/lr_groups.qmd
-            - docs/lora_optims.qmd
-
-        - section: "Core Concepts"
-          contents:
-            - docs/batch_vs_grad.qmd
-            - docs/dataset_preprocessing.qmd
-            - docs/multipack.qmd
-
-        - section: "Advanced Features"
-          contents:
-            - docs/fsdp_qlora.qmd
             - docs/unsloth.qmd
-            - docs/torchao.qmd
-            - docs/custom_integrations.qmd
-
-        - section: "Troubleshooting"
-          contents:
-            - docs/faq.qmd
-            - docs/debugging.qmd
-            - docs/nccl.qmd
-
+            - docs/amd_hpc.qmd
+            - docs/ray-integration.qmd
+        - section: "Dataset Formats"
+          contents: docs/dataset-formats/*
         - section: "Reference"
           contents:
             - docs/config.qmd
+        - docs/faq.qmd
 
 format:
   html:
-    theme: darkly
+    theme: materia
     css: styles.css
     toc: true

cicd/multigpu.py

@@ -37,11 +37,15 @@ temp_dir = tempfile.mkdtemp()
 with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
     f.write(dockerfile_contents)
 
-cicd_image = Image.from_dockerfile(
-    pathlib.Path(temp_dir) / "Dockerfile",
-    force_build=True,
-    gpu="A10G",
-).env(df_args)
+cicd_image = (
+    Image.from_dockerfile(
+        pathlib.Path(temp_dir) / "Dockerfile",
+        force_build=True,
+        gpu="A10G",
+    )
+    .env(df_args)
+    .pip_install("fastapi==0.110.0", "pydantic==2.6.3")
+)
 
 app = App("Axolotl CI/CD", secrets=[])
 

docs/amd_hpc.qmd

@@ -1,5 +1,5 @@
 ---
-title: AMD GPUs on HPC Systems
+title: Training with AMD GPUs on HPC Systems
 description: A comprehensive guide for using Axolotl on distributed systems with AMD GPUs
 ---
 

docs/cli.qmd

@@ -1,19 +1,28 @@
----
-title: "CLI Reference"
-format:
-  html:
-    toc: true
-    toc-expand: 2
-    toc-depth: 3
-execute:
-  enabled: false
----
+# Axolotl CLI Documentation
 
 The Axolotl CLI provides a streamlined interface for training and fine-tuning large language models. This guide covers
 the CLI commands, their usage, and common examples.
 
+### Table of Contents
 
-## Basic Commands
+- Basic Commands
+- Command Reference
+  - fetch
+  - preprocess
+  - train
+  - inference
+  - merge-lora
+  - merge-sharded-fsdp-weights
+  - evaluate
+  - lm-eval
+- Legacy CLI Usage
+- Remote Compute with Modal Cloud
+  - Cloud Configuration
+  - Running on Modal Cloud
+  - Cloud Configuration Options
+
+
+### Basic Commands
 
 All Axolotl commands follow this general structure:
 
@@ -23,9 +32,9 @@ axolotl <command> [config.yml] [options]
 
 The config file can be local or a URL to a raw YAML file.
 
-## Command Reference
+### Command Reference
 
-### fetch
+#### fetch
 
 Downloads example configurations and deepspeed configs to your local machine.
 
@@ -40,7 +49,7 @@ axolotl fetch deepspeed_configs
 axolotl fetch examples --dest path/to/folder
 ```
 
-### preprocess
+#### preprocess
 
 Preprocesses and tokenizes your dataset before training. This is recommended for large datasets.
 
@@ -65,7 +74,7 @@ dataset_prepared_path: Local folder for saving preprocessed data
 push_dataset_to_hub: HuggingFace repo to push preprocessed data (optional)
 ```
 
-### train
+#### train
 
 Trains or fine-tunes a model using the configuration specified in your YAML file.
 
@@ -86,38 +95,7 @@ axolotl train config.yml --no-accelerate
 axolotl train config.yml --resume-from-checkpoint path/to/checkpoint
 ```
 
-It is possible to run sweeps over multiple hyperparameters by passing in a sweeps config.
-
-```bash
-# Basic training with sweeps
-axolotl train config.yml --sweep path/to/sweep.yaml
-```
-
-Example sweep config:
-```yaml
-_:
-  # This section is for dependent variables we need to fix
-  - load_in_8bit: false
-    load_in_4bit: false
-    adapter: lora
-  - load_in_8bit: true
-    load_in_4bit: false
-    adapter: lora
-
-# These are independent variables
-learning_rate: [0.0003, 0.0006]
-lora_r:
-  - 16
-  - 32
-lora_alpha:
-  - 16
-  - 32
-  - 64
-```
-
-
-
-### inference
+#### inference
 
 Runs inference using your trained model in either CLI or Gradio interface mode.
 
@@ -137,7 +115,7 @@ cat prompt.txt | axolotl inference config.yml \
     --base-model="./completed-model"
 ```
 
-### merge-lora
+#### merge-lora
 
 Merges trained LoRA adapters into the base model.
 
@@ -159,7 +137,7 @@ gpu_memory_limit: Limit GPU memory usage
 lora_on_cpu: Load LoRA weights on CPU
 ```
 
-### merge-sharded-fsdp-weights
+#### merge-sharded-fsdp-weights
 
 Merges sharded FSDP model checkpoints into a single combined checkpoint.
 
@@ -168,7 +146,7 @@ Merges sharded FSDP model checkpoints into a single combined checkpoint.
 axolotl merge-sharded-fsdp-weights config.yml
 ```
 
-### evaluate
+#### evaluate
 
 Evaluates a model's performance using metrics specified in the config.
 
@@ -177,27 +155,27 @@ Evaluates a model's performance using metrics specified in the config.
 axolotl evaluate config.yml
 ```
 
-### lm-eval
+#### lm-eval
 
 Runs LM Evaluation Harness on your model.
 
 ```bash
 # Basic evaluation
 axolotl lm-eval config.yml
+
+# Evaluate specific tasks
+axolotl lm-eval config.yml --tasks arc_challenge,hellaswag
 ```
 
 Configuration options:
 
 ```yaml
-# List of tasks to evaluate
-lm_eval_tasks:
-  - arc_challenge
-  - hellaswag
-lm_eval_batch_size: # Batch size for evaluation
-output_dir: # Directory to save evaluation results
+lm_eval_tasks: List of tasks to evaluate
+lm_eval_batch_size: Batch size for evaluation
+output_dir: Directory to save evaluation results
 ```
 
-## Legacy CLI Usage
+### Legacy CLI Usage
 
 While the new Click-based CLI is preferred, Axolotl still supports the legacy module-based CLI:
 
@@ -217,18 +195,12 @@ accelerate launch -m axolotl.cli.inference config.yml \
     --lora_model_dir="./outputs/lora-out" --gradio
 ```
 
-::: {.callout-important}
-When overriding CLI parameters in the legacy CLI, use same notation as in yaml file (e.g., `--lora_model_dir`).
-
-**Note:** This differs from the new Click-based CLI, which uses dash notation (e.g., `--lora-model-dir`). Keep this in mind if you're referencing newer documentation or switching between CLI versions.
-:::
-
-## Remote Compute with Modal Cloud
+### Remote Compute with Modal Cloud
 
 Axolotl supports running training and inference workloads on Modal cloud infrastructure. This is configured using a
 cloud YAML file alongside your regular Axolotl config.
 
-### Cloud Configuration
+#### Cloud Configuration
 
 Create a cloud config YAML with your Modal settings:
 
@@ -243,17 +215,13 @@ branch: main    # Git branch to use (optional)
 volumes:        # Persistent storage volumes
   - name: axolotl-cache
     mount: /workspace/cache
-  - name: axolotl-data
-    mount: /workspace/data
-  - name: axolotl-artifacts
-    mount: /workspace/artifacts
 
 env:            # Environment variables
   - WANDB_API_KEY
   - HF_TOKEN
 ```
 
-### Running on Modal Cloud
+#### Running on Modal Cloud
 
 Commands that support the --cloud flag:
 
@@ -271,18 +239,18 @@ axolotl train config.yml --cloud cloud_config.yml --no-accelerate
 axolotl lm-eval config.yml --cloud cloud_config.yml
 ```
 
-### Cloud Configuration Options
+#### Cloud Configuration Options
 
 ```yaml
-provider: # compute provider, currently only `modal` is supported
-gpu: # GPU type to use
-gpu_count: # Number of GPUs (default: 1)
-memory: # RAM in GB (default: 128)
-timeout: # Maximum runtime in seconds
-timeout_preprocess: # Preprocessing timeout
-branch: # Git branch to use
-docker_tag: # Custom Docker image tag
-volumes: # List of persistent storage volumes
-env: # Environment variables to pass
-secrets: # Secrets to inject
+provider: compute provider, currently only `modal` is supported
+gpu: GPU type to use
+gpu_count: Number of GPUs (default: 1)
+memory: RAM in GB (default: 128)
+timeout: Maximum runtime in seconds
+timeout_preprocess: Preprocessing timeout
+branch: Git branch to use
+docker_tag: Custom Docker image tag
+volumes: List of persistent storage volumes
+env: Environment variables to pass
+secrets: Secrets to inject
 ```

docs/config.qmd

@@ -166,7 +166,7 @@ datasets:
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
     # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train
     # See examples at `docs/dataset-formats/conversation.qmd`
-    # Note: If the below 4 fields are set to empty, defaults to training only on the last message.
+    # Note: If the below 4 fields are empty, defaults to training only on the last message.
 
     # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.
     roles_to_train: ["assistant"]  # default
@@ -174,7 +174,6 @@ datasets:
     # - all: train on all EOS tokens
     # - turn (default): train on the EOS token at the end of each trainable turn
     # - last: train on the last EOS token in the conversation
-    # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.
     train_on_eos: last
     # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.
     message_field_training: training

docs/custom_integrations.qmd

@@ -1,57 +0,0 @@
----
-title: Custom Integrations
-toc: true
-toc-depth: 3
----
-
-```{python}
-#| echo: false
-
-import re
-
-def process_readme(integration_name):
-    try:
-        path = f'../src/axolotl/integrations/{integration_name}/README.md'
-        with open(path, 'r') as f:
-            txt = f.read()
-            # Remove h1 headings
-            txt = re.sub(r'^# .*\n?', '', txt, flags=re.MULTILINE)
-            # Convert h2 to h3
-            txt = re.sub(r'^## ', '### ', txt, flags=re.MULTILINE)
-            return txt
-    except FileNotFoundError:
-        return None
-
-def print_section(name, folder_name):
-    output = f"\n## {name}\n"
-    content = process_readme(folder_name)
-    if content:
-        output += content
-    output += f"\nPlease see reference [here](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/{folder_name})\n"
-    return output
-```
-
-```{python}
-#| output: asis
-#| echo: false
-
-# Introduction text
-print("""
-Axolotl adds custom features through `integrations`. They are located within the `src/axolotl/integrations` directory.
-
-To enable them, please check the respective documentations.
-""")
-
-# Sections
-sections = [
-    ("Cut Cross Entropy", "cut_cross_entropy"),
-    ("Grokfast", "grokfast"),
-    ("Knowledge Distillation (KD)", "kd"),
-    ("Liger Kernels", "liger"),
-    ("Language Model Evaluation Harness (LM Eval)", "lm_eval"),
-    ("Spectrum", "spectrum")
-]
-
-for section_name, folder_name in sections:
-    print(print_section(section_name, folder_name))
-```

docs/dataset-formats/conversation.qmd

@@ -6,9 +6,7 @@ order: 3
 
 ## sharegpt
 
-::: {.callout-important}
-ShareGPT is deprecated!. Please see [chat_template](#chat_template) section below.
-:::
+IMPORTANT: ShareGPT is deprecated!. Please see [chat_template](#chat_template) section below.
 
 ## pygmalion
 
@@ -104,10 +102,6 @@ datasets:
     type: chat_template
 ```
 
-::: {.callout-important}
-Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.
-:::
-
 5. (Advanced) Using fine-grained control over tokens and turns to train in a conversation
 
 For a data sample that looks like:
@@ -155,6 +149,4 @@ datasets:
     message_field_training_detail: train_detail
 ```
 
-::: {.callout-tip}
-It is not necessary to set both `message_field_training` and `message_field_training_detail` at once.
-:::
+Tip: It is not necessary to use both `message_field_training` and `message_field_training_detail` at a time.

docs/dataset-formats/index.qmd

@@ -13,7 +13,7 @@ As there are a lot of available options in Axolotl, this guide aims to provide a
 
 Axolotl supports 3 kinds of training methods: pre-training, supervised fine-tuning, and preference-based post-training (e.g. DPO, ORPO, PRMs). Each method has their own dataset format which are described below.
 
-## Pre-training
+## [Pre-training](pretraining.qmd)
 
 When aiming to train on large corpora of text datasets, pre-training is your go-to choice. Due to the size of these datasets, downloading the entire-datasets before beginning training would be prohibitively time-consuming. Axolotl supports [streaming](https://huggingface.co/docs/datasets/en/stream) to only load batches into memory at a time.
 
@@ -96,10 +96,6 @@ One step is equal to `sequence_len * micro_batch_size * gradient_accumulation_st
 
 It is recommended to leave this off if downloading from Hugging Face hub as it would download the entire dataset which can be very large.
 
-### Reference
-
-Please see docs [here](pretraining.qmd).
-
 ## Supervised fine-tuning (SFT)
 
 Supervised fine-tuning is the process of training models to respond to an instruction or chat input.
@@ -124,7 +120,7 @@ If you went through the flow chart and did not find one that matches, it is reco
 You can mix and match within each approach or across approaches to train a model on a variety of datasets.
 :::
 
-### Pre-Tokenized Dataset
+### [Pre-Tokenized Dataset](tokenized.qmd)
 
 We suggest this approach when you want to bring your own tokenized dataset.
 
@@ -149,9 +145,7 @@ datasets:
 `type: ` is empty!
 :::
 
-Reference: [Pre-Tokenized Dataset Documentation](tokenized.qmd).
-
-### Template Free Dataset
+### [Template Free Dataset](template_free.qmd)
 
 We reccomend this approach when you want granular control over the prompt formatting, special tokens, and masking, whilst letting Axolotl handle the tokenization. This is very useful if your dataset has unique prompts that differ across samples and where one single general template wouldn't suffice.
 
@@ -188,9 +182,7 @@ datasets:
     type: input_output
 ```
 
-Reference: [Template Free Documentation](template_free.qmd).
-
-### Conversation Dataset
+### [Conversation Dataset](conversation.qmd)
 
 `conversation` messages are a list of messages which usually contain a `role` and `content` key.
 
@@ -266,7 +258,7 @@ Newer conversation datasets usually follow the OpenAI format.
 
 Axolotl supports both as well as allowing customization of any kind of key.
 
-#### Chat Template Usage
+#### [Chat Template Usage](conversation.qmd#chat_template)
 
 To properly use this method, it is important to identify three things:
 
@@ -348,19 +340,9 @@ datasets:
       narrator: ["narrator"]
 ```
 
-::: {.callout-tip}
-As chat_templates may use hardcoded EOS/EOT tokens that are different from the tokenizer's EOS, it is highly recommended to set them. For example, `ChatML` uses `<|im_end|>` to end turns.
+#### Applying `chat_template`
 
-```yaml
-special_tokens:
-  eos_token: <|im_end|>
-```
-
-:::
-
-##### Applying `chat_template`
-
-Once all the above steps are completed, you could combine all these configs together to form a bespoke configuration for your custom dataset.
+Once all the above steps are completed, you could combine all these configs together to form a bespoke configuration for your  custom dataset. The final step would be to correctly set the EOS token in your config:
 
 ```yaml
 datasets:
@@ -409,17 +391,7 @@ If this config were to be applied to the sample dataset above, the output would
 
 The first number refers to the label, the second refers to the `token_id`. For example, `-100` labels appear on non-assistant portions, meaning that they are masked during. For assistant portions, the label is the same as the `token_id`.
 
-::: {.callout-note}
-
-If during `preprocess`, there are a lot of warnings of `Could not find content __ boundary`, please check the FAQ section for [chat_templates](../faq.qmd#chat-templates).
-
-:::
-
-#### Reference
-
-Please see docs [here](conversation.qmd).
-
-### Instruction Dataset
+### [Instruction Dataset](inst_tune.qmd)
 
 Instruction datasets are used to train instruction-following models and comprise a prompt, containing an instruction, and a single response. In contrast to chat datasets which may be multi-turn, instruct datasets are typically single-turn.
 
@@ -451,9 +423,6 @@ datasets:
 
 Axolotl supports many kinds of instruction dataset. All of them can be found here (https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html) with their respective type and sample row format.
 
-
-Reference: [Instruction Dataset Documentation](inst_tune.qmd).
-
 #### Custom Instruct Prompt Format
 
 Due to the myriad possibilities of instruction formats, Axolotl allows customizing your own instruction format without having to dive into the code directly.
@@ -484,8 +453,6 @@ datasets:
 
 The config sets that the `field_instruction` is actually named `input`, and the `field_input` is empty as we don't have an `input` in this sample. Generally, `instruction` can be thought as the question to the model, and `input` as the additional information with `output` being the response. It is not necessary to have an `input` nor `system`. In the end, the most important part is to understand what format you want it to look like and how you can customize this to your use case.
 
-Reference: [Custom Instruct Prompt Format Documentation](inst_tune.qmd#how-to-add-custom-prompt-format).
-
 ## Reinforcement Learning from Human Feedback (RLHF)
 
-As there are multiple RLHF methods with their own dataset requirements. Please see [RLHF documentation](../rlhf.qmd) for more detail.
+As there are multiple RLHF methods with their own dataset requirements. Please see [RLHF datasets](../rlhf.qmd) documentation for more detail.

docs/dataset-formats/pretraining.qmd

@@ -27,6 +27,7 @@ pretraining_dataset:
     type: pretrain
     trust_remote_code:
     skip: # number of rows of data to skip over from the beginning
+...
 ```
 
 :::

docs/dataset-formats/template_free.qmd

@@ -1,239 +1,7 @@
 ---
 title: Template-Free
 description: Construct prompts without a template.
-toc: true
-toc-depth: 3
 order: 4
 ---
 
-## Background {#sec-background}
-
-### Masking Inputs {#masking-inputs}
-
-One of the most popular features of
-[axolotl](https://github.com/axolotl-ai-cloud/axolotl) is
-setting the following configuration value:
-
-
-```yaml
-train_on_inputs: false
-```
-
-If you declare a [dataset formats](https://github.com/axolotl-ai-cloud/axolotl?tab=readme-ov-file#dataset)
-such as `alpaca` or `chatml`, axolotl knows what is an input
-(i.e. human) vs. an output (i.e. the assistant) and masks the input
-labels so that your model can focus on predicting the outputs only.
-
-### You may not want prompt templates {#sec-you-may-not-want-prompt-templates}
-
-However, there are many situations where you don't want to use one of
-these formats or templates. This is because they can:
-
--   Add unnecessary boilerplate to your prompts.
--   Create artifacts like special delimiters `<|im_start|>` that can
-    quickly become footguns if you don't include them correctly at
-    inference time.
--   Enforce a *chat* interface when you do not want one. Sometimes you
-    just want to fine-tune a model to a very specific task and do NOT
-    want multi-turn conversations, roles, etc.
--   Limit you to only certain roles that the template allows.
-
-### The `input_output` format {#sec-the-inputoutput-format}
-
-You can construct your prompts without a template by using the
-`input_output` format, by setting `type: input_output` in your
-configuration file like this:
-
-**config.yml**
-
-```yaml
-train_on_inputs: false # Mask segments of your data
-datasets:
-  - path: output.jsonl
-    type: input_output  # use template free prompt construction
-```
-
-Unlike `type: completion`, which is also template-free,
-`type: input_output` allows you to mask segments of your text. More
-details on how this works are described below.
-
-## Usage {#sec-usage}
-
-This is how you can use the `input_output` format:
-
-### 1. Prepare Data {#sec-1-prepare-data}
-
-To use the `input_output` format, collect your data in the following
-format into a jsonl file (below is the first row from the file
-`output`.jsonl` pretty printed):
-
-```bash
-$ head -n1 output.jsonl | python -m json.tool
-```
-
-:::{.cell-output .cell-output-stdout}
-    {
-        "segments": [
-            {
-                "label": true,
-                "text": "<s>Hello\n"
-            },
-            {
-                "label": true,
-                "text": "hi there!. "
-            },
-            {
-                "label": false,
-                "text": "goodbye "
-            },
-            {
-                "label": true,
-                "text": "farewell</s>"
-            }
-        ]
-    }
-:::
-
-Set `label:false` when you want to mask a segment of text so that the
-model isn't trained on it. Some things to keep in mind:
-
-> [!IMPORTANT]
-> 1.  **EOS, BOS, spaces, newlines etc. are entirely up to you. Axolotl
-    concatenates all the segments as-is.** The tokenizer doesn't add
-    anything additional. Notice how I added spaces, newlines, `<s>`
-    (BOS), and `</s>` (EOS) myself.
-> 2.  Make sure you check the materialized output to validate that the
-    prompt is getting assembled how you like.
-
-### 2. Use `type: input_output` {#sec-2-use-type-inputoutput}
-
-Let's materialize data with our `output.jsonl` file by setting
-`type: input_output` in our axolotl config:
-
-```yaml
-# training_config.yaml
-base_model: mistralai/Mistral-7B-v0.1
-data_seed: 49
-seed: 49
-
-datasets:
-  - path: output.jsonl
-    type: input_output
-val_set_size: 0.1
-
-sequence_len: 896
-sample_packing: false
-
-micro_batch_size: 2
-gradient_accumulation_steps: 3
-eval_batch_size: 2
-num_epochs: 1
-learning_rate: 0.0002
-
-train_on_inputs: false
-special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"
-```
-
-You can use the following command to materialize your data. The
-`--debug` flag will print the tokens, along with the labels so you can
-verify that the correct items are being ignored:
-
-```bash
-axolotl preprocess training_config.yaml --debug
-
-...
-[2024-03-05 23:36:46,969] [INFO] [axolotl.check_example_labels:35] [PID:607731] [RANK:0] <s>(1, 1) Hello(22557, 22557)
-(13, 13) hi(12014, 12014) there(736, 736) !(28808, 28808) .(28723, 28723) (28705, 28705) good(-100, 1179) bye(-100, 17664) (-100, 28705) fare(19111, 19111) well(5458, 5458) </s>(2, 2)
-
-```
-
-The format is `decoded_token`(`label`, `token_id`), for example,
-`<s>(1, 1)` means that the token is `<s>`, the label is `1` and the
-token_id is `1`. When the label is `-100` then that token is ignored for
-training.
-
-### 3. Check the prompts {#sec-3-check-the-prompts}
-
-Here is another way to check the materialized output:
-
-```python
-from transformers import AutoTokenizer
-from datasets import load_from_disk
-import yaml
-
-directory = !ls last_run_prepared/
-with open('training_config.yaml', 'r') as f:
-    cfg = yaml.safe_load(f)
-model_id = cfg['base_model']
-tok = AutoTokenizer.from_pretrained(model_id)
-ds = load_from_disk(f'last_run_prepared/{directory[0]}/')
-```
-
-```python
->>> row = ds[0]
->>> print(tok.decode(row['input_ids']))
-<s> Hello
-    hi there!.  goodbye  farewell</s>
-```
-
-We can check that the right tokens are ignored by comparing the labels
-to each token:
-
-```python
-import pandas as pd
-pd.DataFrame([{'token': tok.decode(i), 'label': l, 'id':i} for i,l in
-              zip(row['input_ids'], row['labels'])])
-```
-
-| token | label | id    |
-|-------|-------|-------|
-| 0     | \<s\> | 1     |
-| 1     | Hello | 22557 |
-| 2     | \\n   | 13    |
-| 3     | hi    | 12014 |
-| 4     | there | 736   |
-| 5     | !     | 28808 |
-| 6     | .     | 28723 |
-| 7     |       | 28705 |
-| 8     | good  | -100  |
-| 9     | bye   | -100  |
-| 10    |       | -100  |
-| 11    | fare  | 19111 |
-| 12    | well  | 5458  |
-| 13    | \</s\>| 2     |
-
-
-
-If we look at the input data, the above table seems correct! (The jsonl
-version is repeated below for reference):
-
-
-```bash
-$ head -n1 output.jsonl | python -m json.tool
-```
-
-:::{.cell-output .cell-output-stdout}
-    {
-        "segments": [
-            {
-                "label": true,
-                "text": "<s>Hello\n"
-            },
-            {
-                "label": true,
-                "text": "hi there!. "
-            },
-            {
-                "label": false,
-                "text": "goodbye "
-            },
-            {
-                "label": true,
-                "text": "farewell</s>"
-            }
-        ]
-    }
-:::
+See [these docs](../input_output.qmd).

docs/dataset_preprocessing.qmd

@@ -3,11 +3,8 @@ title: Dataset Preprocessing
 description: How datasets are processed
 ---
 
-## Overview
-
 Dataset pre-processing is the step where Axolotl takes each dataset you've configured alongside
-the [dataset format](docs/dataset-formats) and prompt strategies to:
-
+the (dataset format)[../dataset-formats/] and prompt strategies to:
  - parse the dataset based on the *dataset format*
  - transform the dataset to how you would interact with the model based on the *prompt strategy*
  - tokenize the dataset based on the configured model & tokenizer
@@ -15,12 +12,10 @@ the [dataset format](docs/dataset-formats) and prompt strategies to:
 
 The processing of the datasets can happen one of two ways:
 
-1. Before kicking off training by calling `axolotl preprocess config.yaml --debug`
+1. Before kicking off training by calling `python -m axolotl.cli.preprocess /path/to/your.yaml --debug`
 2. When training is started
 
-### What are the benefits of pre-processing?
-
-When training interactively or for sweeps
+What are the benefits of pre-processing? When training interactively or for sweeps
 (e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly
 slow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent
 training parameters so that it will intelligently pull from its cache when possible.
@@ -33,12 +28,8 @@ default path of `./last_run_prepared/`, but will ignore anything already cached
 setting `dataset_prepared_path: ./last_run_prepared`, the trainer will use whatever pre-processed
 data is in the cache.
 
-### What are the edge cases?
-
-Let's say you are writing a custom prompt strategy or using a user-defined
+What are the edge cases? Let's say you are writing a custom prompt strategy or using a user-defined
 prompt template. Because the trainer cannot readily detect these changes, we cannot change the
-calculated hash value for the pre-processed dataset.
-
-If you have `dataset_prepared_path: ...` set
+calculated hash value for the pre-processed dataset. If you have `dataset_prepared_path: ...` set
 and change your prompt templating logic, it may not pick up the changes you made and you will be
 training over the old prompt.

docs/debugging.qmd

@@ -31,13 +31,11 @@ While debugging it's helpful to simplify your test scenario as much as possible.
     - Set `CUDA_VISIBLE_DEVICES` to a single GPU, ex: `export CUDA_VISIBLE_DEVICES=0`.
     - Set `dataset_processes: 1` in your axolotl config or run the training command with `--dataset_processes=1`.
 2. **Use a small dataset**: Construct or use a small dataset from HF Hub. When using a small dataset, you will often have to make sure `sample_packing: False` and `eval_sample_packing: False` to avoid errors.  If you are in a pinch and don't have time to construct a small dataset but want to use from the HF Hub, you can shard the data (this will still tokenize the entire dataset, but will only use a fraction of the data for training.  For example, to shard the dataset into 20 pieces, add the following to your axolotl config):
-
     ```yaml
-    datasets:
+    dataset:
         ...
         shards: 20
     ```
-
 3. **Use a small model**: A good example of a small model is [TinyLlama/TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0).
 4. **Minimize iteration time**: Make sure the training loop finishes as fast as possible, with these settings.
     - `micro_batch_size: 1`
@@ -87,7 +85,7 @@ The easiest way to get started is to modify the [.vscode/launch.json](../.vscode
 
 For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
 
-```json
+```jsonc
 // .vscode/launch.json
 {
     "version": "0.2.0",
@@ -134,7 +132,7 @@ For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 acceler
 
 Below is the [./vscode/tasks.json](../.vscode/tasks.json) file that defines the `cleanup-for-dataprep` task.  This task is run before each debugging session when you use the above configuration.  Note how there are two tasks that delete the two folders mentioned above.  The third task `cleanup-for-dataprep` is a composite task that combines the two tasks.  A composite task is necessary because VSCode does not allow you to specify multiple tasks in the `preLaunchTask` argument of the `launch.json` file.
 
-```json
+```jsonc
 // .vscode/tasks.json
 // this file is used by launch.json
 {

docs/faq.qmd

@@ -3,7 +3,6 @@ title: FAQ
 description: Frequently asked questions
 ---
 
-### General
 
 **Q: The trainer stopped and hasn't progressed in several minutes.**
 
@@ -25,28 +24,6 @@ description: Frequently asked questions
 
 > A: This is usually an issue with the GPU. This can be resolved through setting the os environment variable `CUDA_VISIBLE_DEVICES=0`. If you are on runpod, this is usually a pod issue. Starting a new pod should take care of it.
 
-### Chat templates
-
 **Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**
 
 > A: This means that the property mapping for the stated attribute does not exist when building `chat_template` prompt. For example, if `no attribute 'content'`, please check you have added the correct mapping for `content` under `message_property_mappings`.
-
-**Q: `Empty template generated for turn ___`**
-
-> A: The `content` is empty for that turn.
-
-**Q: `Could not find content start/end boundary for turn __`**
-
-> A: The specific turn's start/end could not be detected. Please ensure you have set the `eos_token` following your `chat_template`. Otherwise, this could be a `chat_template` which doesn't use proper boundaries for each turn (like system). On the rare occurrence, make sure your content is not `[[dummy_message]]`. Please let us know about this.
-
-**Q: `Content end boundary is before start boundary for turn ___`**
-
-> A: This is an edge case which should not occur. Please create an Issue if this happens.
-
-**Q: `Content end boundary is the same as start boundary for turn ___. This is likely an empty turn.`**
-
-> A: This is likely an empty turn.
-
-**Q: The EOS/EOT token is incorrectly being masked or not being masked.**
-
-> A: This is because of the mismatch between `tokenizer.eos_token` and EOS/EOT token in template. Please make sure to set `eos_token` under `special_tokens` to the same EOS/EOT token as in template.

docs/getting-started.qmd

@@ -1,5 +1,5 @@
 ---
-title: "Quickstart"
+title: "Getting Started with Axolotl"
 format:
   html:
     toc: true
@@ -17,12 +17,12 @@ Let's start by fine-tuning a small language model using LoRA. This example uses
 Assuming `axolotl` is installed (if not, see our [Installation Guide](installation.qmd))
 
 1. Download example configs:
-```bash
+```shell
 axolotl fetch examples
 ```
 
 2. Run the training:
-```bash
+```shell
 axolotl train examples/llama-3/lora-1b.yml
 ```
 
@@ -108,7 +108,7 @@ Please consult the supported [Dataset Formats](dataset-formats/) for more detail
 
 3. Run the training:
 
-```bash
+```shell
 axolotl train my_training.yml
 ```
 
@@ -118,7 +118,7 @@ axolotl train my_training.yml
 
 After training, test your model:
 
-```bash
+```shell
 axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out"
 ```
 
@@ -126,7 +126,7 @@ axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out"
 
 For large datasets, preprocess first:
 
-```bash
+```shell
 axolotl preprocess my_training.yml
 ```
 
@@ -134,7 +134,7 @@ axolotl preprocess my_training.yml
 
 Launch a Gradio interface:
 
-```bash
+```shell
 axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
 ```
 

docs/inference.qmd

@@ -1,10 +1,11 @@
 ---
-title: "Inference"
+title: "Inference Guide"
 format:
   html:
     toc: true
     toc-depth: 3
     number-sections: true
+    code-tools: true
 execute:
   enabled: false
 ---

docs/input_output.qmd

@@ -3,4 +3,263 @@ title: Template-free prompt construction
 description: "Template-free prompt construction with the `input_output` format"
 ---
 
-The documentation moved to [here](dataset-formats/template_free.qmd).
+<!-- TOC -->
+
+- [Background](#background)
+    - [Masking Inputs](#masking-inputs)
+    - [You may not want prompt templates](#you-may-not-want-prompt-templates)
+    - [The `input_output` format](#the-input_output-format)
+- [Usage](#usage)
+    - [1. Prepare Data](#1-prepare-data)
+    - [2. Use `type: input_output`](#2-use-type-input_output)
+    - [3. Check the prompts](#3-check-the-prompts)
+
+<!-- /TOC -->
+
+<a id="markdown-background" name="background"></a>
+
+## Background
+
+<a id="markdown-masking-inputs" name="masking-inputs"></a>
+
+### Masking Inputs
+
+One of the most popular features of
+[axolotl](https://github.com/axolotl-ai-cloud/axolotl) is
+setting the following configuration value:
+
+
+```yaml
+train_on_inputs: false
+```
+
+If you declare a [dataset formats](https://github.com/axolotl-ai-cloud/axolotl?tab=readme-ov-file#dataset)
+such as `alpaca` or `chatml`, axolotl knows what is an input
+(i.e. human) vs. an output (i.e. the assistant) and masks the input
+labels so that your model can focus on predicting the outputs only.
+
+<a id="markdown-you-may-not-want-prompt-templates" name="you-may-not-want-prompt-templates"></a>
+
+### You may not want prompt templates
+
+However, there are many situations where you don't want to use one of
+these formats or templates. This is because they can:
+
+-   Add unnecessary boilerplate to your prompts.
+-   Create artifacts like special delimiters `<|im_start|>` that can
+    quickly become footguns if you don't include them correctly at
+    inference time.
+-   Enforce a *chat* interface when you do not want one. Sometimes you
+    just want to fine-tune a model to a very specific task and do NOT
+    want multi-turn conversations, roles, etc.
+-   Limit you to only certain roles that the template allows.
+
+<a id="markdown-the-inputoutput-format" name="the-inputoutput-format"></a>
+
+### The `input_output` format
+
+You can construct your prompts without a template by using the
+`input_output` format, by setting `type: input_output` in your
+configuration file like this:
+
+**config.yml**
+
+```yaml
+train_on_inputs: false # Mask segments of your data
+datasets:
+  - path: output.jsonl
+    type: input_output  # use template free prompt construction
+```
+
+Unlike `type: completion`, which is also template-free,
+`type: input_output` allows you to mask segments of your text. More
+details on how this works are described below.
+
+<a id="markdown-usage" name="usage"></a>
+
+## Usage
+
+This is how you can use the `input_output` format:
+
+<a id="markdown-1-prepare-data" name="1-prepare-data"></a>
+
+### 1. Prepare Data
+
+To use the `input_output` format, collect your data in the following
+format into a jsonl file (below is the first row from the file
+`output`.jsonl` pretty printed):
+
+```bash
+$ head -n1 output.jsonl | python -m json.tool
+```
+
+:::{.cell-output .cell-output-stdout}
+    {
+        "segments": [
+            {
+                "label": true,
+                "text": "<s>Hello\n"
+            },
+            {
+                "label": true,
+                "text": "hi there!. "
+            },
+            {
+                "label": false,
+                "text": "goodbye "
+            },
+            {
+                "label": true,
+                "text": "farewell</s>"
+            }
+        ]
+    }
+:::
+
+Set `label:false` when you want to mask a segment of text so that the
+model isn't trained on it. Some things to keep in mind:
+
+> [!IMPORTANT]
+> 1.  **EOS, BOS, spaces, newlines etc. are entirely up to you. Axolotl
+    concatenates all the segments as-is.** The tokenizer doesn't add
+    anything additional. Notice how I added spaces, newlines, `<s>`
+    (BOS), and `</s>` (EOS) myself.
+> 2.  Make sure you check the materialized output to validate that the
+    prompt is getting assembled how you like.
+
+<a id="markdown-2-use-type-inputoutput" name="2-use-type-inputoutput"></a>
+
+### 2. Use `type: input_output`
+
+Let's materialize data with our `output.jsonl` file by setting
+`type: input_output` in our axolotl config:
+
+```yaml
+# training_config.yaml
+base_model: mistralai/Mistral-7B-v0.1
+data_seed: 49
+seed: 49
+
+datasets:
+  - path: output.jsonl
+    type: input_output
+val_set_size: 0.1
+
+sequence_len: 896
+sample_packing: false
+
+micro_batch_size: 2
+gradient_accumulation_steps: 3
+eval_batch_size: 2
+num_epochs: 1
+learning_rate: 0.0002
+
+train_on_inputs: false
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"
+```
+
+You can use the following command to materialize your data. The
+`--debug` flag will print the tokens, along with the labels so you can
+verify that the correct items are being ignored:
+
+```bash
+$ python -m axolotl.cli.preprocess training_config.yaml --debug
+
+...
+[2024-03-05 23:36:46,969] [INFO] [axolotl.check_example_labels:35] [PID:607731] [RANK:0] <s>(1, 1) Hello(22557, 22557)
+(13, 13) hi(12014, 12014) there(736, 736) !(28808, 28808) .(28723, 28723) (28705, 28705) good(-100, 1179) bye(-100, 17664) (-100, 28705) fare(19111, 19111) well(5458, 5458) </s>(2, 2)
+
+```
+
+The format is `decoded_token`(`label`, `token_id`), for example,
+`<s>(1, 1)` means that the token is `<s>`, the label is `1` and the
+token_id is `1`. When the label is `-100` then that token is ignored for
+training.
+
+<a id="markdown-3-check-the-prompts" name="3-check-the-prompts"></a>
+
+### 3. Check the prompts
+
+Here is another way to check the materialized output:
+
+```python
+from transformers import AutoTokenizer
+from datasets import load_from_disk
+import yaml
+
+directory = !ls last_run_prepared/
+with open('training_config.yaml', 'r') as f:
+    cfg = yaml.safe_load(f)
+model_id = cfg['base_model']
+tok = AutoTokenizer.from_pretrained(model_id)
+ds = load_from_disk(f'last_run_prepared/{directory[0]}/')
+```
+
+```python
+>>> row = ds[0]
+>>> print(tok.decode(row['input_ids']))
+<s> Hello
+    hi there!.  goodbye  farewell</s>
+```
+
+We can check that the right tokens are ignored by comparing the labels
+to each token:
+
+```python
+import pandas as pd
+pd.DataFrame([{'token': tok.decode(i), 'label': l, 'id':i} for i,l in
+              zip(row['input_ids'], row['labels'])])
+```
+
+| token | label | id    |
+|-------|-------|-------|
+| 0     | \<s\> | 1     |
+| 1     | Hello | 22557 |
+| 2     | \\n   | 13    |
+| 3     | hi    | 12014 |
+| 4     | there | 736   |
+| 5     | !     | 28808 |
+| 6     | .     | 28723 |
+| 7     |       | 28705 |
+| 8     | good  | -100  |
+| 9     | bye   | -100  |
+| 10    |       | -100  |
+| 11    | fare  | 19111 |
+| 12    | well  | 5458  |
+| 13    | \</s\>| 2     |
+
+
+
+If we look at the input data, the above table seems correct! (The jsonl
+version is repeated below for reference):
+
+
+```bash
+$ head -n1 output.jsonl | python -m json.tool
+```
+
+:::{.cell-output .cell-output-stdout}
+    {
+        "segments": [
+            {
+                "label": true,
+                "text": "<s>Hello\n"
+            },
+            {
+                "label": true,
+                "text": "hi there!. "
+            },
+            {
+                "label": false,
+                "text": "goodbye "
+            },
+            {
+                "label": true,
+                "text": "farewell</s>"
+            }
+        ]
+    }
+:::

docs/installation.qmd

@@ -1,10 +1,11 @@
 ---
-title: "Installation"
+title: "Installation Guide"
 format:
   html:
     toc: true
     toc-depth: 3
     number-sections: true
+    code-tools: true
 execute:
   enabled: false
 ---

docs/lora_optims.qmd

@@ -1,6 +1,7 @@
 ---
 title: "LoRA Optimizations"
-description: "Custom autograd functions and Triton kernels in Axolotl for optimized LoRA fine-tuning"
+description: "Custom autograd functions and Triton kernels in Axolotl for optimized
+LoRA fine-tuning"
 ---
 
 Inspired by [Unsloth](https://github.com/unslothai/unsloth), we've implemented two

docs/mac.qmd

@@ -19,5 +19,4 @@ Current support:
 - [ ] DeepSpeed
 
 Untested:
-
 - FSDP

docs/multi-gpu.qmd

@@ -1,5 +1,5 @@
 ---
-title: "Multi-GPU"
+title: "Multi-GPU Training Guide"
 format:
   html:
     toc: true
@@ -35,11 +35,7 @@ deepspeed: deepspeed_configs/zero1.json
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
-# Passing arg via config
-axolotl train config.yml
-
-# Passing arg via cli
-axolotl train config.yml --deepspeed deepspeed_configs/zero1.json
+accelerate launch -m axolotl.cli.train examples/llama-2/config.yml --deepspeed deepspeed_configs/zero1.json
 ```
 
 ### ZeRO Stages {#sec-zero-stages}
@@ -74,7 +70,25 @@ For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).
 
 ### Liger Kernel Integration {#sec-liger}
 
-Please see [docs](custom_integrations.qmd#liger) for more info.
+::: {.callout-note}
+Liger Kernel provides efficient Triton kernels for LLM training, offering:
+
+- 20% increase in multi-GPU training throughput
+- 60% reduction in memory usage
+- Compatibility with both FSDP and DeepSpeed
+:::
+
+Configuration:
+
+```{.yaml}
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_layer_norm: true
+liger_fused_linear_cross_entropy: true
+```
 
 ## Troubleshooting {#sec-troubleshooting}
 

docs/multi-node.qmd

@@ -13,7 +13,7 @@ You will also need to have the same configuration file for your model on each ma
 Make sure the main machine is reachable by other machines.
 :::
 
-## Accelerate
+# Accelerate
 
 You will need to create a configuration for accelerate, either by using `accelerate config` and follow the instructions or you can use one of the preset below:
 
@@ -51,17 +51,17 @@ fsdp_config:
 
 All you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.
 
-## Raytrain
+# Raytrain
 
 Please see ray train doc [here](ray-integration.qmd).
 
-## Torchrun
+# Torchrun
 
 If you are using Infiniband, we recommend torchrun to utilize the full bandwidth.
 
 Set the following env (change buffersize/socketname depending on your system):
 
-```bash
+```yaml
 export NCCL_IB_DISABLE=0
 export NCCL_SOCKET_IFNAME="eth0,en,eth,em,bond"
 export NCCL_BUFFSIZE=2097152

docs/nccl.qmd

@@ -13,13 +13,13 @@ Often, this timeout will happen after 30 minutes (the default setting) and is ac
 
 Forcing cross-GPU communication via [NVLink](https://en.wikipedia.org/wiki/NVLink) may help without increasing timeouts. To verify that your configuration is leveraging NVLink run the following command:
 
-```bash
+```shell
 nvidia-smi nvlink --status
 ```
 
 To force NCCL to use NVLink, simply set this in the environment:
 
-```bash
+```shell
 export NCCL_P2P_LEVEL=NVL
 ```
 
@@ -33,13 +33,13 @@ If NVLink is not available in your environment there are other options for ``NCC
 
 To validate that acceptable data transfer speeds exist for your training job, running [NCCL Tests](https://github.com/NVIDIA/nccl-tests/blob/master/README.md) can help pinpoint bottlenecks, for example:
 
-```bash
+```shell
 ./build/all_reduce_perf -b 8 -e 128M -f 2 -g 3
 ```
 
 It can be useful when debugging NCCL communication timeouts to activate additional logging in both PyTorch and NCCL:
 
-```bash
+```shell
 export NCCL_DEBUG=INFO
 export NCCL_DEBUG_SUBSYS=ALL
 export TORCH_DISTRIBUTED_DEBUG=INFO

docs/ray-integration.qmd

@@ -1,5 +1,5 @@
 ---
-title: Ray Train
+title: Ray Train integration
 description: How to use Axolotl with Ray Train
 ---
 
@@ -9,7 +9,7 @@ With the `--use-ray` CLI flag, Axolotl will use Ray Train's [`TorchTrainer`](htt
 
 ## Ray cluster setup
 
-A prerequisite using the Ray Train integration is to setup a Ray cluster on your desired node(s). For a detailed guide on how you can get started with ray clusters, check the official Ray docs [here](https://docs.ray.io/en/latest/cluster/getting-started.html).
+A prerequisite using the Ray Train integration is to setup a Ray cluster on your desired node(s). For a detailed guide on how you can get started with ray clusters, check the official Ray docs here: https://docs.ray.io/en/latest/cluster/getting-started.html
 
 Every Ray cluster has one _head_ node and a set of worker nodes. The head node is just like any other worker node, but it also runs certain special processes related to scheduling and orchestration. Ray-enabled scripts are run on the head node and depending on the resources (number of CPUs, GPUs, etc) they request, will be scheduled to run certain tasks on the worker nodes. For more on key concepts behind a Ray cluster, you can refer this [doc](https://docs.ray.io/en/latest/cluster/key-concepts.html#cluster-key-concepts).
 
@@ -58,11 +58,13 @@ You can find an example configuration at `configs/llama-3/lora-1b-ray.yaml`.
 The key parameters to note here are:
 
 ```yaml
+...
 use_ray: true
 ray_num_workers: 4
 # optional
 resources_per_worker:
     GPU: 1
+...
 ```
 
 - `use_ray`: This is the flag that enables the Ray Train integration. You can either use the corresponding `--use-ray` flag in the CLI or set `use_ray` in the config file.

docs/rlhf.qmd

@@ -3,22 +3,22 @@ title: "RLHF (Beta)"
 description: "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback."
 back-to-top-navigation: true
 toc: true
-toc-depth: 4
+toc-depth: 3
 ---
 
-## Overview
+# Overview
 
 Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human
 feedback. Various methods include, but not limited to:
 
+- Proximal Policy Optimization (PPO) (not yet supported in axolotl)
 - [Direct Preference Optimization (DPO)](#dpo)
 - [Identity Preference Optimization (IPO)](#ipo)
 - [Kahneman-Tversky Optimization (KTO)](#kto)
 - [Odds Ratio Preference Optimization (ORPO)](#orpo)
-- Proximal Policy Optimization (PPO) (not yet supported in axolotl)
 
 
-## RLHF using Axolotl
+# RLHF using Axolotl
 
 ::: {.callout-important}
 This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.
@@ -30,7 +30,7 @@ We rely on the [TRL](https://github.com/huggingface/trl) library for implementat
 You can find what each method supports by going into `src/axolotl/prompt_strategies/{method}` where `{method}` is one of our supported methods. The `type: ` can be retrieved from `{method}.{function_name}`.
 :::
 
-### DPO
+## DPO
 
 Example config:
 
@@ -47,7 +47,7 @@ datasets:
 
 DPO supports the following types with the following dataset format:
 
-#### chatml.argilla
+### chatml.argilla
 
 ```json
 {
@@ -58,7 +58,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### chatml.argilla_chat
+### chatml.argilla_chat
 
 ```json
 {
@@ -73,7 +73,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### chatml.icr
+### chatml.icr
 
 ```json
 {
@@ -84,7 +84,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### chatml.intel
+### chatml.intel
 
 ```json
 {
@@ -95,7 +95,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### chatml.prompt_pairs
+### chatml.prompt_pairs
 
 ```json
 {
@@ -106,7 +106,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### chatml.ultra
+### chatml.ultra
 
 ```json
 {
@@ -123,7 +123,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.argilla
+### llama3.argilla
 
 ```json
 {
@@ -134,7 +134,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.argilla_chat
+### llama3.argilla_chat
 
 ```json
 {
@@ -149,7 +149,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.icr
+### llama3.icr
 
 ```json
 {
@@ -160,7 +160,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.intel
+### llama3.intel
 
 ```json
 {
@@ -171,7 +171,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.prompt_pairs
+### llama3.prompt_pairs
 
 ```json
 {
@@ -182,7 +182,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.ultra
+### llama3.ultra
 
 ```json
 {
@@ -199,7 +199,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### zephyr.nectar
+### zephyr.nectar
 
 ```json
 {
@@ -218,7 +218,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-#### chat_template.default
+### chat_template.default
 
 ```yaml
 rl: dpo
@@ -264,7 +264,7 @@ Sample input format:
 }
 ```
 
-#### user_defined.default
+### user_defined.default
 
 For custom behaviors,
 
@@ -295,7 +295,7 @@ The input format is a simple JSON input with customizable fields based on the ab
 }
 ```
 
-### IPO
+## IPO
 
 As IPO is just DPO with a different loss function, all supported options for DPO works here.
 
@@ -303,7 +303,7 @@ As IPO is just DPO with a different loss function, all supported options for DPO
 rl: ipo
 ```
 
-### ORPO
+## ORPO
 
 Paper: https://arxiv.org/abs/2403.07691
 
@@ -320,7 +320,7 @@ datasets:
 
 ORPO supports the following types with the following dataset format:
 
-#### chat_template.argilla
+### chat_template.argilla
 
 ```json
 {
@@ -339,7 +339,7 @@ ORPO supports the following types with the following dataset format:
 }
 ```
 
-### KTO
+## KTO
 
 ```yaml
 rl: kto
@@ -360,7 +360,7 @@ gradient_checkpointing_kwargs:
 
 KTO supports the following types with the following dataset format:
 
-#### chatml.argilla
+### chatml.argilla
 
 ```json
 {
@@ -370,7 +370,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### chatml.argilla_chat
+### chatml.argilla_chat
 
 ```json
 {
@@ -383,7 +383,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### chatml.intel
+### chatml.intel
 
 ```json
 {
@@ -393,7 +393,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### chatml.prompt_pairs
+### chatml.prompt_pairs
 
 ```json
 {
@@ -403,7 +403,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### chatml.ultra
+### chatml.ultra
 
 ```json
 {
@@ -413,7 +413,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.argilla
+### llama3.argilla
 
 ```json
 {
@@ -423,7 +423,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.argilla_chat
+### llama3.argilla_chat
 
 ```json
 {
@@ -434,7 +434,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.intel
+### llama3.intel
 
 ```json
 {
@@ -444,7 +444,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.prompt_pairs
+### llama3.prompt_pairs
 
 ```json
 {
@@ -454,7 +454,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### llama3.ultra
+### llama3.ultra
 
 ```json
 {
@@ -464,7 +464,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-#### user_defined.default
+### user_defined.default
 
 For custom behaviors,
 
@@ -494,49 +494,7 @@ The input format is a simple JSON input with customizable fields based on the ab
 }
 ```
 
-### GRPO
-
-GRPO uses custom reward functions and transformations. Please have them ready locally.
-
-For ex, to load OpenAI's GSM8K and use a random reward for completions:
-
-```python
-# rewards.py
-import random
-
-def rand_reward_func(completions, **kwargs) -> list[float]:
-    return [random.uniform(0, 1) for _ in completions]
-
-def oai_gsm8k_transform(cfg, *args, **kwargs):
-    def transform_fn(example, tokenizer=None):
-        label = example["answer"].split("####")[-1].strip().replace(",", "")
-        return {
-            "prompt": [{"role": "user", "content": example["question"]},],
-            "answer": label,
-        }
-    return transform_fn, {"remove_columns": ["question"]}
-```
-
-```yaml
-rl: grpo
-
-trl:
-    beta: 0.001
-    max_completion_length: 256
-    use_vllm: True
-    vllm_device: auto
-    vllm_gpu_memory_utilization: 0.15
-    num_generations: 4
-    reward_funcs: ["rewards.rand_reward_func"]    # format: '{file_name}.{fn_name}'
-datasets:
-  - path: openai/gsm8k
-    name: main
-    type: rewards.oai_gsm8k_transform  # format: '{file_name}.{fn_name}'
-```
-
-To see other examples of custom reward functions, please see [TRL GRPO Docs](https://github.com/huggingface/trl/blob/main/docs/source/grpo_trainer.md#using-a-custom-reward-function).
-
-### Using local dataset files
+## Using local dataset files
 
 ```yaml
 datasets:
@@ -547,7 +505,7 @@ datasets:
     type: chatml.intel
 ```
 
-### TRL auto-unwrapping for PEFT
+## TRL auto-unwrapping for PEFT
 
 TRL supports auto-unwrapping PEFT models for RL training paradigms which rely on a reference model. This significantly reduces memory pressure as an additional refreference model does not need to be loaded, and reference model log-probabilities can be obtained by disabling PEFT adapters. This is enabled by default. To turn it off, pass the following config:
 

docs/torchao.qmd

@@ -3,12 +3,6 @@ title: "PyTorch ao"
 description: "Custom data types and layouts for training and inference"
 ---
 
-To use experimental optimizers (`AdamWFp8`, `AdamW4bit`, `AdamW8bit`) from Pytorch Ao, please install the package as shown below.
-
-::: {.callout-tip}
-Some experimental optimizers are already present in regular Pytorch, so please re-check if you actually need this package!
-:::
-
 ### Installation
 
 Stable Release from the PyTorch index

docs/unsloth.qmd

@@ -8,12 +8,6 @@ description: "Hyper-optimized QLoRA finetuning for single GPUs"
 Unsloth provides hand-written optimized kernels for LLM finetuning that slightly improve speed and VRAM over
 standard industry baselines.
 
-::: {.callout-important}
-Due to breaking changes in transformers `v4.48.0`, users will need to downgrade to `<=v4.47.1` to use this patch.
-
-This will later be deprecated in favor of [LoRA Optimizations](lora_optims.qmd).
-:::
-
 
 ### Installation
 
@@ -23,7 +17,7 @@ The following will install the correct unsloth and extras from source.
 python scripts/unsloth_install.py | sh
 ```
 
-### Usage
+### Using unsloth w Axolotl
 
 Axolotl exposes a few configuration options to try out unsloth and get most of the performance gains.
 

index.qmd

@@ -1,7 +1,7 @@
 ---
-# toc-location: right-body
-# toc-title: Table Of Contents
-# toc-expand: 2
+toc-location: right-body
+toc-title: Table Of Contents
+toc-expand: 2
 ---
 
 ```{python}

requirements.txt

@@ -7,7 +7,7 @@ mamba-ssm==1.2.0.post1
 flash-attn==2.7.4.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.3
+liger-kernel==0.5.2
 # END section
 
 packaging==23.2

src/axolotl/cli/cloud/modal_.py

@@ -123,6 +123,8 @@ class ModalCloud(Cloud):
         if env := self.get_env():
             image = image.env(env)
 
+        image = image.pip_install("fastapi==0.110.0", "pydantic==2.6.3")
+
         return image
 
     def get_secrets(self):
@@ -258,21 +260,25 @@ class ModalCloud(Cloud):
 
 
 def _preprocess(config_yaml: str, volumes=None):
-    Path("/workspace/mounts").mkdir(parents=True, exist_ok=True)
-    with open("/workspace/mounts/config.yaml", "w", encoding="utf-8") as f_out:
+    Path("/workspace/artifacts/axolotl").mkdir(parents=True, exist_ok=True)
+    with open(
+        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
+    ) as f_out:
         f_out.write(config_yaml)
-    run_folder = "/workspace/mounts"
+    run_folder = "/workspace/artifacts/axolotl"
     run_cmd(
-        "axolotl preprocess /workspace/mounts/config.yaml --dataset-processes=8",
+        "axolotl preprocess /workspace/artifacts/axolotl/config.yaml --dataset-processes=8",
         run_folder,
         volumes,
     )
 
 
 def _train(config_yaml: str, accelerate: bool = True, volumes=None, **kwargs):
-    with open("/workspace/mounts/config.yaml", "w", encoding="utf-8") as f_out:
+    with open(
+        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
+    ) as f_out:
         f_out.write(config_yaml)
-    run_folder = "/workspace/mounts"
+    run_folder = "/workspace/artifacts/axolotl"
     if accelerate:
         accelerate_args = "--accelerate"
     else:
@@ -281,18 +287,20 @@ def _train(config_yaml: str, accelerate: bool = True, volumes=None, **kwargs):
     if num_processes := kwargs.pop("num_processes", None):
         num_processes_args = f"--num-processes {num_processes}"
     run_cmd(
-        f"axolotl train {accelerate_args} {num_processes_args} /workspace/mounts/config.yaml",
+        f"axolotl train {accelerate_args} {num_processes_args} /workspace/artifacts/axolotl/config.yaml",
         run_folder,
         volumes,
     )
 
 
 def _lm_eval(config_yaml: str, volumes=None):
-    with open("/workspace/mounts/config.yaml", "w", encoding="utf-8") as f_out:
+    with open(
+        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
+    ) as f_out:
         f_out.write(config_yaml)
-    run_folder = "/workspace/mounts"
+    run_folder = "/workspace/artifacts/axolotl"
     run_cmd(
-        "axolotl lm-eval /workspace/mounts/config.yaml",
+        "axolotl lm-eval /workspace/artifacts/axolotl/config.yaml",
         run_folder,
         volumes,
     )

src/axolotl/core/trainer_builder.py

@@ -831,7 +831,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             if "max_length" in kwargs:
                 kwargs.pop("max_length")
         elif use_batch_sampler_collator:
-            if self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
+            if self.cfg.flex_attention is True:
+                collator = V2BatchSamplerDataCollatorForSeq2Seq
+            elif self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
                 collator = V2BatchSamplerDataCollatorForSeq2Seq
             elif (
                 self.cfg.model_config_type in ["llama"]

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -1,10 +1,6 @@
 # Cut Cross Entropy
 
-Cut Cross Entropy reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.
-
-See https://github.com/apple/ml-cross-entropy
-
-## Usage
+### Usage
 
 ```yaml
 plugins:
@@ -12,19 +8,3 @@ plugins:
 
 cut_cross_entropy: true
 ```
-
-## Citation
-
-```bib
-@article{wijmans2024cut,
-  author       = {Erik Wijmans and
-                  Brody Huval and
-                  Alexander Hertzberg and
-                  Vladlen Koltun and
-                  Philipp Kr\"ahenb\"uhl},
-  title        = {Cut Your Losses in Large-Vocabulary Language Models},
-  journal      = {arXiv},
-  year         = {2024},
-  url          = {https://arxiv.org/abs/2411.09009},
-}
-```

src/axolotl/integrations/grokfast/README.md

@@ -2,7 +2,7 @@
 
 See https://github.com/ironjr/grokfast
 
-## Usage
+### Usage
 
 ```yaml
 plugins:
@@ -11,14 +11,3 @@ plugins:
 grokfast_alpha: 2.0
 grokfast_lamb: 0.98
 ```
-
-## Citation
-
-```bib
-@article{lee2024grokfast,
-    title={{Grokfast}: Accelerated Grokking by Amplifying Slow Gradients},
-    author={Lee, Jaerin and Kang, Bong Gyun and Kim, Kihoon and Lee, Kyoung Mu},
-    journal={arXiv preprint arXiv:2405.20233},
-    year={2024}
-}
-```

src/axolotl/integrations/kd/README.md

@@ -1,23 +0,0 @@
-# Knowledge Distillation
-
-## Usage
-
-```yaml
-plugins:
-  - "axolotl.integrations.kd.KDPlugin"
-
-kd_trainer: True
-kd_ce_alpha: 0.1
-kd_alpha: 0.9
-kd_temperature: 1.0
-
-torch_compile: True  # torch>=2.5.1, recommended to reduce vram
-
-datasets:
-  - path: ...
-    type: "axolotl.integrations.kd.chat_template"
-    field_messages: "messages_combined"
-    logprobs_field: "llm_text_generation_vllm_logprobs"  # for kd only, field of logprobs
-```
-
-An example dataset can be found at [`axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample`](https://huggingface.co/datasets/axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample)

src/axolotl/integrations/kd/topk_logprob/LICENSE.md

@@ -0,0 +1,58 @@
+### AXOLOTL COMMUNITY LICENSE AGREEMENT
+
+This Axolotl Community License Agreement (“Agreement”) is entered into by and between Axolotl AI Corp. (“Axolotl”) and
+any individual or entity (“Licensee”) who wishes to use the Software (as defined below) in accordance with the terms
+and conditions set forth in this Agreement.
+
+1.  Definitions
+    1.1 “Licensee” refers to any individual or entity who has obtained a copy of the Software under this Agreement.
+    1.2 “Plugin Integration” means independent integration software modules which may or may not be offered by Axolotl,
+        which may be licensed separately by their respective  authors and/or licensors.
+    1.3 “Software” refers to the specific sub-directory of the Axolotl, Inc. software located at
+        https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations and its subdirectories which
+        permits Plugin Integrations to integrate with the Axolotl service.
+2.  Grant of License
+    2.1	Axolotl hereby grants Licensee a worldwide, non-exclusive, royalty-free, license to use, copy, modify, merge,
+        publish, distribute, sublicense, and/or otherwise exploit the Software, subject to the following conditions:
+        - Licensee must comply with all the terms and conditions of this Agreement.
+        - Licensee must include the original copyright notice and disclaimer of warranty in all copies or substantial
+          portions of the Software.
+    2.2 Licensee may use the Software for any lawful purpose, except as restricted in Section 3.
+3.  Restrictions
+    3.1 Licensee shall not use the Software for any activity that constitutes a commercial activity of offering for
+        free or for sale any services, platform, or equivalent  to third parties for the purposes of allowing such
+        third parties to fine-tune artificial intelligence models.
+    3.2 Licensee shall not:
+        - Use the Software for any illegal or unauthorized purpose.
+        - Reverse engineer, decompile, or disassemble the Software.
+        - Remove or modify any copyright, trademark, or other proprietary notices contained in the Software.
+        - Use the Software in a way that could damage, disable, overburden, or impair the functionality of the
+          Software or interfere with any third-party use of the Software.
+    3.3 Axolotl reserves the right to restrict certain Plugin Integrations for use with the Software. To the extent Licensee integrates a permitted, applicable Plugin Integration with the Software, Licensee shall comply with any additional terms and conditions imposed by the licensors of such Plugin Integration for use of such Plugin Integrations. Licensee shall contact Axolotl if it has questions about whether its use of the Software falls beyond the scope of this Agreement.
+4.  Intellectual Property Rights
+    4.1 Axolotl and its contributors retain all intellectual property rights in and to the Software. Licensee
+        acknowledges that this Agreement does not transfer any ownership rights or intellectual property rights to
+        Licensee.
+5.  Disclaimer of Warranty
+    5.1 THE SOFTWARE IS PROVIDED “AS IS,” WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED
+        TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NON-INFRINGEMENT. IN NO EVENT SHALL
+        THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY, WHETHER IN AN ACTION OF
+        CONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+        DEALINGS IN THE SOFTWARE.
+6.  Termination
+    6.1 Axolotl may terminate this Agreement at any time if Licensee fails to comply with any of the terms and
+        conditions set forth herein. Upon termination, Licensee shall cease all use of the Software and destroy any
+        copies in its possession.
+7.  Governing Law
+    7.1 This Agreement shall be governed by and construed in accordance with the laws of the State of California,
+        without regards to conflicts of laws provisions thereof.
+8.  Entire Agreement
+    8.1 This Agreement constitutes the entire agreement between Axolotl and Licensee with respect to the subject matter
+        hereof and supersedes all prior or contemporaneous understandings or agreements between the parties concerning
+        the Software, whether written or oral. Axolotl may update the terms of this Agreement from time to time, and
+        Licensee’s continued use of the Software after any such updates shall constitute acceptance of updated terms
+        on a go-forward basis.  Axolotl will use commercially reasonable efforts to provide Licensee notice of any
+        material updates. By using the Software, Licensee acknowledges that it has read, understood, and agrees to be
+        bound by the terms and conditions of this Agreement.
+
+This Agreement was last updated on August 23, 2024.

src/axolotl/integrations/kd/topk_logprob/forward_kl.py

@@ -1,16 +1,14 @@
 # Copyright 2024 Axolotl AI. All rights reserved.
 #
-# Licensed under the Apache License, Version 2.0 (the "License");
+# This software may be used and distributed according to
+# the terms of the Axolotl Community License Agreement (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.
+# 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.
 
 """
 loss for top_k KL divergence

src/axolotl/integrations/liger/README.md

@@ -1,36 +0,0 @@
-# Liger Kernel Integration
-
-Liger Kernel provides efficient Triton kernels for LLM training, offering:
-
-- 20% increase in multi-GPU training throughput
-- 60% reduction in memory usage
-- Compatibility with both FSDP and DeepSpeed
-
-See https://github.com/linkedin/Liger-Kernel
-
-## Usage
-
-```yaml
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-```
-
-## Citation
-
-```bib
-@article{hsu2024ligerkernelefficienttriton,
-      title={Liger Kernel: Efficient Triton Kernels for LLM Training},
-      author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},
-      year={2024},
-      eprint={2410.10989},
-      archivePrefix={arXiv},
-      primaryClass={cs.LG},
-      url={https://arxiv.org/abs/2410.10989},
-      journal={arXiv preprint arXiv:2410.10989},
-}
-```

src/axolotl/integrations/lm_eval/README.md

@@ -1,10 +1,6 @@
 # LM Eval Harness
 
-Run evaluation on model using the popular lm-evaluation-harness library.
-
-See https://github.com/EleutherAI/lm-evaluation-harness
-
-## Usage
+### Usage
 
 ```yaml
 plugins:
@@ -14,22 +10,4 @@ lm_eval_tasks:
   - gsm8k
   - hellaswag
   - arc_easy
-
-lm_eval_batch_size: # Batch size for evaluation
-output_dir: # Directory to save evaluation results
-```
-
-## Citation
-
-```bib
-@misc{eval-harness,
-  author       = {Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and Reynolds, Laria and Schoelkopf, Hailey and Skowron, Aviya and Sutawika, Lintang and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy},
-  title        = {A framework for few-shot language model evaluation},
-  month        = 07,
-  year         = 2024,
-  publisher    = {Zenodo},
-  version      = {v0.4.3},
-  doi          = {10.5281/zenodo.12608602},
-  url          = {https://zenodo.org/records/12608602}
-}
 ```

src/axolotl/integrations/spectrum/README.md

@@ -1,17 +1,15 @@
-# Spectrum: Targeted Training on Signal to Noise Ratio
+## Spectrum: Targeted Training on Signal to Noise Ratio
 
 by Eric Hartford, Lucas Atkins, Fernando Fernandes, David Golchinfar
 
 This plugin contains code to freeze the bottom fraction of modules in a model, based on the Signal-to-Noise Ratio (SNR).
 
-See https://github.com/cognitivecomputations/spectrum
-
-## Overview
+### Overview
 
 Spectrum is a tool for scanning and evaluating the Signal-to-Noise Ratio (SNR) of layers in large language models.
 By identifying the top n% of layers with the highest SNR, you can optimize training efficiency.
 
-## Usage
+### Usage
 
 ```yaml
 plugins:
@@ -21,17 +19,3 @@ spectrum_top_fraction: 0.5
 # Optional if using a pre-scanned model as your base_model. Useful if using a model mirror
 spectrum_model_name: meta-llama/Meta-Llama-3.1-8B
 ```
-
-## Citation
-
-```bib
-@misc{hartford2024spectrumtargetedtrainingsignal,
-      title={Spectrum: Targeted Training on Signal to Noise Ratio},
-      author={Eric Hartford and Lucas Atkins and Fernando Fernandes Neto and David Golchinfar},
-      year={2024},
-      eprint={2406.06623},
-      archivePrefix={arXiv},
-      primaryClass={cs.LG},
-      url={https://arxiv.org/abs/2406.06623},
-}
-```

src/axolotl/monkeypatch/multipack.py

@@ -25,7 +25,6 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "gemmoe",
     "starcoder2",
     "deepseek_v2",
-    "deepseek_v3",
 ]
 
 

src/axolotl/monkeypatch/utils.py

@@ -95,6 +95,103 @@ def get_cu_seqlens(attn_mask):
     return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)
 
 
+def get_packed_mask_from_pos_ids(position_ids):
+    if len(position_ids.shape) == 1:
+        position_ids = position_ids.unsqueeze(0)
+
+    device = position_ids.device
+    results = []
+
+    for i, row in enumerate(position_ids):
+        # Count the number of consecutive zeros from the right side
+        padding_length = (row == 0).int().flip(dims=[0]).cumprod(dim=0).sum().item()
+
+        # Adjust the row to exclude padding
+        adjusted_row = row[:-padding_length] if padding_length else row.clone()
+
+        # Find where the position resets to 0 (indicating a new sequence)
+        seq_starts = torch.cat(
+            [
+                torch.tensor([True], dtype=torch.bool, device=device),
+                adjusted_row[1:] == 0,
+            ]
+        )
+        # Get the indices where the sequence starts
+        start_indices = torch.cat(
+            [
+                torch.nonzero(seq_starts).unbind(dim=1)[0],
+                torch.tensor([len(adjusted_row)], dtype=torch.int32, device=device),
+            ]
+        )
+        # Calculate the sequence lengths
+        seq_lengths = start_indices[1:] - start_indices[:-1]
+        # Append the padding length to the sequence lengths
+        doc_mask = torch.ones(len(row), dtype=torch.int32, device=device)
+        for i, seq_len in enumerate(seq_lengths):
+            start_id = start_indices[i]
+            doc_mask[start_id : start_id + seq_len] = (
+                (i+1) * doc_mask[start_id : start_id + seq_len]
+            )
+        if padding_length:
+            doc_mask[len(adjusted_row) :] = 0 * doc_mask[len(adjusted_row) :]
+
+        results.append(doc_mask)
+
+    return torch.stack(results)
+
+
+def get_seqlens_from_pos_ids(position_ids):
+    """generate a sequence length set using pos ids for doc mask creation in flex attention"""
+    if len(position_ids.shape) == 1:
+        position_ids = position_ids.unsqueeze(0)
+    max_seq_len = position_ids.shape[1]
+
+    device = position_ids.device
+    results = []
+    totalseqlens = []
+
+    for row in position_ids:
+        # Count the number of consecutive zeros from the right side
+        padding_length = (row == 0).int().flip(dims=[0]).cumprod(dim=0).sum().item()
+
+        # Adjust the row to exclude padding
+        adjusted_row = row[:-padding_length] if padding_length else row.clone()
+
+        # Find where the position resets to 0 (indicating a new sequence)
+        seq_starts = torch.cat(
+            [
+                torch.tensor([True], dtype=torch.bool, device=device),
+                adjusted_row[1:] == 0,
+            ]
+        )
+        # Get the indices where the sequence starts
+        start_indices = torch.cat(
+            [
+                torch.nonzero(seq_starts).unbind(dim=1)[0],
+                torch.tensor([len(adjusted_row)], dtype=torch.int32, device=device),
+            ]
+        )
+        # Calculate the sequence lengths
+        seq_lengths = start_indices[1:] - start_indices[:-1]
+        # Append the padding length to the sequence lengths
+        if padding_length:
+            seq_lengths = torch.cat(
+                [
+                    seq_lengths,
+                    torch.tensor(
+                        [len(row) - torch.sum(seq_lengths)],
+                        dtype=torch.int32,
+                        device=device,
+                    ),
+                ]
+            )
+
+        results.append(seq_lengths)
+        totalseqlens.append(len(adjusted_row))
+
+    return results, torch.tensor(totalseqlens, dtype=torch.int32, device=device)
+
+
 def get_cu_seqlens_from_pos_ids(position_ids):
     """generate a cumulative sequence length mask for flash attention using pos ids"""
     if len(position_ids.shape) == 1:
@@ -176,7 +273,10 @@ def mask_2d_to_4d(
     when they attend to each other within that sequence.
     This expansion transforms the mask to lower triangular form to prevent future peeking.
     """
-    bsz, src_len = mask.size()
+
+    if len(mask.size()) == 4:
+        return mask
+    bsz, src_len = int(mask.size()[0]), int(mask.size()[1])
     tgt_len = tgt_len if tgt_len is not None else src_len
 
     mask = mask.unsqueeze(1).unsqueeze(2)

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -55,7 +55,6 @@ class ChatTemplate(str, Enum):
     phi_3 = "phi_3"  # pylint: disable=invalid-name
     phi_35 = "phi_35"  # pylint: disable=invalid-name
     deepseek_v2 = "deepseek_v2"  # pylint: disable=invalid-name
-    deepseek_v3 = "deepseek_v3"  # pylint: disable=invalid-name
     jamba = "jamba"  # pylint: disable=invalid-name
     jinja = "jinja"  # pylint: disable=invalid-name
     qwen_25 = "qwen_25"  # pylint: disable=invalid-name
@@ -824,6 +823,7 @@ class AxolotlInputConfig(
     xformers_attention: Optional[bool] = None
     sdp_attention: Optional[bool] = None
     s2_attention: Optional[bool] = None
+    flex_attention: Optional[bool] = None
     flash_attention: Optional[bool] = None
     flash_attn_cross_entropy: Optional[bool] = None
     flash_attn_rms_norm: Optional[bool] = None
@@ -1790,6 +1790,26 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
                 )
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_flex_torch_version(cls, data):
+        if (data.get("flex_attention") is not None) and (
+            data.get("flex_attention") is True
+        ):
+            env_capabilities = data.get("env_capabilities", {})
+            torch_version = env_capabilities.get("torch_version")
+
+            if torch_version is None:
+                import torch
+
+                torch_version = str(torch.__version__).split("+", maxsplit=1)[0]
+
+            if version.parse(torch_version) < version.parse("2.5.1"):
+                raise ValueError(
+                    "Flex attention is not supported on torch version < 2.5.1"
+                )
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_torch_compile_auto(cls, data):

src/axolotl/utils/models.py

@@ -403,7 +403,7 @@ class ModelLoader:
 
         if (
             self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES
-            and self.cfg.flash_attention
+            and (self.cfg.flash_attention or self.cfg.flex_attention)
             and self.cfg.sample_packing
         ):
             if "auto_map" in self.model_config:
@@ -707,7 +707,13 @@ class ModelLoader:
         """
         sample packing uses custom FA2 patch
         """
-        if self.cfg.flash_attention:
+
+        if self.cfg.flex_attention:
+            self.model_kwargs["attn_implementation"] = "flex_attention"
+            self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                "flex_attention"
+            )
+        elif self.cfg.flash_attention:
             if not self.cfg.sample_packing and self.cfg.s2_attention:
                 pass
             self.model_kwargs["attn_implementation"] = "flash_attention_2"
@@ -1113,7 +1119,7 @@ class ModelLoader:
         should_convert = (
             # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
             # convert them back to fp16/bf16 for flash-attn compatibility.
-            ((needs_fa2_dtype or self.cfg.flash_attention) and not qlora_fsdp)
+            ((needs_fa2_dtype or self.cfg.flash_attention or self.cfg.flex_attention) and not qlora_fsdp)
             or self.cfg.cut_cross_entropy  # Cut cross entropy requires embedding layers to be in fp16/bf16 for backward pass
         )
 

styles.css

@@ -1,193 +1,5 @@
-/* TYPOGRAPHY SECTION */
+/* css styles */
 
-/* Import fonts */
-@import url('https://fonts.googleapis.com/css2?family=Be+Vietnam+Pro:wght@400;500&display=swap');
-@import url('https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@400&display=swap');
-
-/* Typography hierarchy */
-:root {
-    --font-title: 'Be Vietnam Pro', sans-serif;
-    --font-body: 'JetBrains Mono', monospace;
-}
-
-/* Title (h1) */
-h1 {
-    font-family: var(--font-title);
-    font-weight: 400;
-    font-size: 6rem;
-    line-height: 1.1;
-    letter-spacing: -0.05em;
-    font-feature-settings: "ss01" on;
-}
-
-/* Heading (h2) */
-h2 {
-    font-family: var(--font-title);
-    font-weight: 500;
-    font-size: 2rem;
-    line-height: 1.2;
-    letter-spacing: -0.03em;
-    font-feature-settings: "ss01" on;
-}
-
-/* Subtitle/Preamble */
-h3,
-h4 {
-    font-family: var(--font-body);
-    font-weight: 400;
-    font-size: 1.5rem;
-    line-height: 1.5;
-    letter-spacing: -0.02em;
-}
-
-/* Body text */
-body {
-    font-family: var(--font-body);
-    font-weight: 400;
-    font-size: 1rem;
-    line-height: 1.5;
-    letter-spacing: -0.02em;
-}
-
-/* Links */
-a {
-    font-family: var(--font-body);
-    font-weight: 400;
-    font-size: 0.875rem;
-    line-height: 1;
-    letter-spacing: -0.02em;
-}
-
-/* NAV BAR SECTION */
-
-/* Navbar logo styling */
-.navbar-brand img {
-    height: 32px;
-    margin-right: 10px;
-}
-
-/* COLORS SECTION */
-
-/* Brand colors */
-:root {
-    --white: #ffffff;
-    --greige-300: #EEEEE7;
-    --greige-600: #CCCAC0;
-    --black: #141310;
-    --lime: #E3F8A8;
-    --cyan: #A0F4EA;
-    --purple: #C8D0F8;
-}
-
-/* Base styles */
-body {
-    background-color: var(--black);
-    color: var(--greige-300);
-}
-
-/* Navigation */
-.navbar {
-    background-color: var(--black) !important;
-}
-
-.navbar-dark .navbar-nav .nav-link {
-    color: var(--greige-300);
-}
-
-.navbar-dark .navbar-nav .nav-link:hover {
-    color: var(--lime);
-}
-
-/* Sidebar */
-.sidebar-navigation {
-    background-color: var(--black);
-    border-right: 1px solid var(--greige-600);
-}
-
-.sidebar nav[role="doc-toc"] ul>li>a {
-    color: var(--greige-300);
-}
-
-.sidebar nav[role="doc-toc"] ul>li>a:hover {
-    color: var(--lime);
-}
-
-/* Links */
-a {
-    color: var(--lime);
-}
-
-a:hover {
-    color: var(--cyan);
-}
-
-/* Headers */
-h1,
-h2,
-h3,
-h4,
-h5,
-h6 {
-    color: var(--white);
-}
-
-/* Code blocks */
-pre {
-    background-color: #1a1a1a !important;
-    border: 1px solid var(--greige-600);
-}
-
-/* Tables */
-.table {
-    color: var(--greige-300);
-}
-
-/* TOC */
-#toc-title {
-    color: var(--white);
-}
-
-.toc-active {
-    color: var(--lime) !important;
-}
-
-/* Buttons */
-.btn-primary {
-    background-color: var(--lime);
-    color: var(--black);
-    border: none;
-}
-
-.btn-primary:hover {
-    background-color: var(--cyan);
-    color: var(--black);
-}
-
-/* For inline code (single backtick) */
-code {
-    background-color: #1a1a1a !important;
-    color: var(--lime) !important;
-    padding: 2px 4px;
-    border-radius: 4px;
-}
-
-/* For inline code that is also a link */
-a code {
-    color: var(--cyan) !important;
-}
-
-/* For code blocks (triple backtick) */
-pre.sourceCode {
-    background-color: #1a1a1a !important;
-}
-
-/* Make comments in bash/shell scripts green */
-code span.co {
-    color: #5cb85c !important;
-}
-
-/* Remove underlines from JSON comments and make them green */
-code span.er {
-    color: #5cb85c !important;
-    text-decoration: none !important;
+img[alt="Axolotl"] {
+    content: url("https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/887513285d98132142bf5db2a74eb5e0928787f1/image/axolotl_logo_digital_black.svg") !important;
 }

tests/e2e/test_deepseekv3.py

@@ -1,130 +0,0 @@
-"""
-E2E tests for lora llama
-"""
-
-import logging
-import os
-from pathlib import Path
-
-import pytest
-
-from axolotl.cli.args import TrainerCliArgs
-from axolotl.common.datasets import load_datasets
-from axolotl.train import train
-from axolotl.utils.config import normalize_config, validate_config
-from axolotl.utils.dict import DictDefault
-
-LOG = logging.getLogger("axolotl.tests.e2e")
-os.environ["WANDB_DISABLED"] = "true"
-
-
-class TestDeepseekV3:
-    """
-    Test case for DeepseekV3 models
-    """
-
-    @pytest.mark.parametrize(
-        "sample_packing",
-        [True, False],
-    )
-    def test_lora_deepseekv3(self, temp_dir, sample_packing):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "axolotl-ai-co/DeepSeek-V3-11M",
-                "trust_remote_code": True,
-                "sample_packing": sample_packing,
-                "flash_attention": True,
-                "sequence_len": 2048,
-                "adapter": "lora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0,
-                "datasets": [
-                    {
-                        "path": "mlabonne/FineTome-100k",
-                        "type": "chat_template",
-                        "field_messages": "conversations",
-                        "message_property_mappings": {
-                            "role": "from",
-                            "content": "value",
-                        },
-                        "drop_system_message": True,
-                        "split": "train[:1%]",
-                    },
-                ],
-                "special_tokens": {
-                    "bos_token": "<｜begin▁of▁sentence｜>",
-                    "eos_token": "<｜end▁of▁sentence｜>",
-                },
-                "chat_template": "deepseek_v3",
-                "num_epochs": 1,
-                "micro_batch_size": 1,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_bnb_8bit",
-                "lr_scheduler": "cosine",
-                "max_steps": 5,
-                "save_safetensors": True,
-                "bf16": True,
-            }
-        )
-        cfg = validate_config(cfg)
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.safetensors").exists()
-
-    @pytest.mark.parametrize(
-        "sample_packing",
-        [True, False],
-    )
-    def test_fft_deepseekv3(self, temp_dir, sample_packing):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "axolotl-ai-co/DeepSeek-V3-11M",
-                "trust_remote_code": True,
-                "sample_packing": sample_packing,
-                "flash_attention": True,
-                "sequence_len": 2048,
-                "val_set_size": 0,
-                "datasets": [
-                    {
-                        "path": "mlabonne/FineTome-100k",
-                        "type": "chat_template",
-                        "field_messages": "conversations",
-                        "message_field_role": "from",
-                        "message_field_content": "value",
-                        "split": "train[:1%]",
-                    },
-                ],
-                "chat_template": "deepseek_v3",
-                "special_tokens": {
-                    "bos_token": "<｜begin▁of▁sentence｜>",
-                    "eos_token": "<｜end▁of▁sentence｜>",
-                },
-                "num_epochs": 1,
-                "micro_batch_size": 1,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_bnb_8bit",
-                "lr_scheduler": "cosine",
-                "max_steps": 5,
-                "save_safetensors": True,
-                "bf16": True,
-            }
-        )
-        cfg = validate_config(cfg)
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "model.safetensors").exists()