keep some softmax layers

more fixes

.gitignore

@@ -1,6 +1,7 @@
 **/axolotl.egg-info
 configs
 last_run_prepared/
+outputs
 .vscode
 _site/
 
@@ -185,3 +186,6 @@ out/
 
 # vim
 *.swp
+
+# symlinked to axolotl-artifacts in docker containers
+outputs

cicd/cicd.sh

@@ -4,7 +4,6 @@ set -e
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
 pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ /workspace/axolotl/tests/
-# pytest -v --durations=10 -n8 --dist loadfile /workspace/axolotl/tests/patched/
-pytest -v --durations=10 -n1 --dist loadfile /workspace/axolotl/tests/e2e/patched/
-pytest -v --durations=10 -n1 --dist loadfile /workspace/axolotl/tests/e2e/integrations/
+pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/
+pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
 pytest -v --durations=10 --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/

cicd/multigpu.py

@@ -1,6 +1,6 @@
 """
- modal application to run axolotl gpu tests in Modal
- """
+modal application to run axolotl gpu tests in Modal
+"""
 # pylint: disable=duplicate-code
 
 import os

docs/config.qmd

@@ -127,34 +127,40 @@ datasets:
     # - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to if the tokenizer does not have a chat template else default to tokenizer. E.g. tokenizer_default_fallback_chatml.
     # - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.
     chat_template: tokenizer_default
-    # Custom jinja template for chat template. This will be only used if `chat_template` is set to `jinja` or empty (in which case chat_template is automatically set to `jinja`).
+
+    # Custom jinja chat template. Used only if `chat_template: jinja` or empty.
     chat_template_jinja:
-    # The key in the data example that contains the messages. Default is "messages".
+
+    # Key containing the messages (default: "messages")
     field_messages: messages
-    # The key in the message turn that contains the role. Default is "role".
+    # Key for role in each message (default: "role")
     message_field_role: role
-    # The key in the message turn that contains the content. Default is "content".
+    # Key for content in each message (default:  "content")
     message_field_content: content
-    # Optional[Dict[str, List]]. Roles mapping for the messages.
+
+    # Optional[Dict[str, List]]. Roles mapping in the messages. The default is:
     roles:
       user: ["human", "user"]
-      assistant: ["gpt", "assistant", "ai"]
+      assistant: ["gpt", "assistant"]
       system: ["system"]
+      tool: ["tool"]
 
-    ## NOTE: Leaving the below empty will default to using the simple legacy tokenization strategy where only last message is trained on.
+    # 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 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: ["gpt", "assistant"]
+    roles_to_train: ["assistant"]  # default
     # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:
     # - all: train on all EOS tokens
-    # - turn: train on the EOS token at the end of each trainable turn
+    # - turn (default): train on the EOS token at the end of each trainable turn
     # - last: train on the last EOS token in the conversation
     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
     # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.
     # The value of the key is a List[Dict] containing `begin_offset` (start character index in content), `end_offset` (end character index in content), and `train` (boolean whether to train).
-    # See example at `docs/dataset-formats/conversation.qmd`
     message_field_training_detail: train_detail
 
 
@@ -239,6 +245,9 @@ sample_packing_group_size: 100000
 # The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.
 sample_packing_bin_size: 200
 
+# Use batch flattening for speedups when not using sample_packing
+batch_flattening:
+
 # Passed through to transformers when loading the model when launched without accelerate
 # Use `sequential` when training w/ model parallelism to limit memory
 device_map:
@@ -331,7 +340,8 @@ comet_experiment_config: # Dictionary for additional configuration settings, see
 output_dir: ./completed-model
 
 # Whether to use torch.compile and which backend to use
-torch_compile:  # bool
+# setting to `auto` will enable torch compile when torch>=2.5.1
+torch_compile:  # Optional[Union[Literal["auto"], bool]]
 torch_compile_backend:  # Optional[str]
 
 # Training hyperparameters
@@ -363,6 +373,10 @@ eval_table_size: # Approximate number of predictions sent to wandb depending on
 eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
 eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"]
 
+profiler_steps: # enable the pytorch profiler to capture the first N steps of training to the output_dir.
+                # see https://pytorch.org/blog/understanding-gpu-memory-1/ for more information
+                # snapshots can be visualized @ https://pytorch.org/memory_viz
+
 loss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)
 loss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)
 

docs/dataset-formats/conversation.qmd

@@ -68,6 +68,8 @@ We recommend checking the below examples for other usecases.
 datasets:
   - path: ...
     type: chat_template
+    roles_to_train:
+    train_on_eos:
 ```
 
 2. Using the `gemma` chat template to override the tokenizer_config.json's chat template on OpenAI messages format, training on all assistant messages.
@@ -77,7 +79,7 @@ chat_template: gemma # this overwrites the tokenizer's chat_template
 datasets:
   - path: ...
     type: chat_template
-    roles_to_train: ["assistant"]
+    roles_to_train: ["assistant"]  # default value
 ```
 
 3. Using the tokenizer_config.json's chat template or `chatml` as fallback if the former's chat template does not exist, on OpenAI messages format, training on all assistant messages.
@@ -87,7 +89,6 @@ chat_template: tokenizer_default_fallback_chatml # this overwrites the tokenizer
 datasets:
   - path: ...
     type: chat_template
-    roles_to_train: ["assistant"]
 ```
 
 4. Using a custom jinja template on OpenAI messages format, training on all assistant messages.
@@ -99,7 +100,6 @@ chat_template_jinja: "{{ bos_token }}{% for message in messages %}{% if (message
 datasets:
   - path: ...
     type: chat_template
-    roles_to_train: ["assistant"]
 ```
 
 5. (Advanced) Using fine-grained control over tokens and turns to train in a conversation

examples/cerebras/btlm-ft.yml

@@ -1,6 +1,10 @@
 base_model: cerebras/btlm-3b-8k-base
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: GPT2Tokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 tokenizer_use_fast: true
 tokenizer_legacy: true

examples/cerebras/qlora.yml

@@ -1,4 +1,7 @@
 base_model: cerebras/Cerebras-GPT-1.3B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 load_in_8bit: false
 load_in_4bit: true
 strict: false

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

@@ -1,6 +1,9 @@
 base_model: codellama/CodeLlama-13b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

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

@@ -1,6 +1,9 @@
 base_model: codellama/CodeLlama-13b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

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

@@ -1,6 +1,9 @@
 base_model: codellama/CodeLlama-34b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

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

@@ -1,6 +1,9 @@
 base_model: codellama/CodeLlama-34b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

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

@@ -1,6 +1,9 @@
 base_model: codellama/CodeLlama-7b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

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

@@ -1,6 +1,9 @@
 base_model: codellama/CodeLlama-7b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/dbrx/16bit-lora.yaml

@@ -1,4 +1,7 @@
 base_model: LnL-AI/dbrx-base-converted-v2
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/dbrx/8bit-lora.yaml

@@ -1,4 +1,7 @@
 base_model: LnL-AI/dbrx-base-converted-v2
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: true

examples/dbrx/fft-ds-zero3.yaml

@@ -1,4 +1,7 @@
 base_model: LnL-AI/dbrx-base-converted-v2
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/deepseek-v2/fft-fsdp-16b.yaml

@@ -1,4 +1,6 @@
 base_model: deepseek-ai/DeepSeek-V2-Lite
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 trust_remote_code: true
 
 load_in_8bit: false

examples/deepseek-v2/qlora-fsdp-2_5.yaml

@@ -1,4 +1,7 @@
 base_model: axolotl-quants/DeepSeek-V2.5-bnb-nf4-bf16
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

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

@@ -1,7 +1,12 @@
 base_model: tiiuae/falcon-7b
-trust_remote_code: true
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
+trust_remote_code: true
 
 load_in_8bit: true
 load_in_4bit: false

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

@@ -1,10 +1,15 @@
 # 1b: tiiuae/falcon-rw-1b
 # 40b: tiiuae/falcon-40b
 base_model: tiiuae/falcon-7b
-# required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
-trust_remote_code: true
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
+trust_remote_code: true
+
 
 load_in_8bit: false
 # enable 4bit for QLoRA

examples/falcon/config-7b.yml

@@ -1,7 +1,12 @@
 base_model: tiiuae/falcon-7b
-trust_remote_code: true
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
+trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: false

examples/gemma/qlora.yml

@@ -1,7 +1,10 @@
 # use google/gemma-7b if you have access
 base_model: mhenrichsen/gemma-7b
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/gemma2/qlora.yml

@@ -1,6 +1,9 @@
 base_model: google/gemma-2-9b
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/gemma2/reward-model.yaml

@@ -1,6 +1,9 @@
 base_model: google/gemma-2-2b
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForSequenceClassification
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/gptj/qlora.yml

@@ -1,4 +1,7 @@
 base_model: EleutherAI/gpt-j-6b
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 load_in_8bit: false
 load_in_4bit: true
 strict: false

examples/jamba/qlora.yaml

@@ -1,4 +1,7 @@
 base_model: ai21labs/Jamba-v0.1
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/jamba/qlora_deepspeed.yaml

@@ -1,4 +1,6 @@
 base_model: ai21labs/Jamba-v0.1
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 trust_remote_code: true
 
 load_in_8bit: false

examples/jamba/qlora_fsdp_large.yaml

@@ -1,5 +1,8 @@
 base_model: ai21labs/AI21-Jamba-1.5-Large
+# optionally might have model_type or tokenizer_type
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_4bit: true
 strict: false

examples/jeopardy-bot/config.yml

@@ -1,6 +1,10 @@
 base_model: huggyllama/llama-7b
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 load_in_8bit: false
 datasets:
   - path: openaccess-ai-collective/jeopardy

examples/llama-2/fft_optimized.yml

@@ -1,6 +1,9 @@
 base_model: NousResearch/Llama-2-7b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/llama-2/gptq-lora.yml

@@ -1,8 +1,13 @@
 base_model: TheBloke/Llama-2-7B-GPTQ
-gptq: true
-gptq_disable_exllama: true
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+gptq: true
+gptq_disable_exllama: true
+
 tokenizer_use_fast: true
 tokenizer_legacy: true
 load_in_8bit: false

examples/llama-2/lisa.yml

@@ -1,6 +1,9 @@
 base_model: NousResearch/Llama-2-7b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/llama-2/loftq.yml

@@ -1,6 +1,9 @@
 base_model: NousResearch/Llama-2-7b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/llama-2/lora.yml

@@ -1,6 +1,9 @@
 base_model: NousResearch/Llama-2-7b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/llama-2/qlora-fsdp.yml

@@ -1,6 +1,9 @@
 base_model: NousResearch/Llama-2-7b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/llama-2/qlora.yml

@@ -1,6 +1,9 @@
 base_model: NousResearch/Llama-2-7b-hf
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/llama-3-vision/lora-11b.yaml

@@ -1,5 +1,9 @@
 base_model: alpindale/Llama-3.2-11B-Vision-Instruct
+# optionally might have model_type or tokenizer_type or processor_type
 processor_type: AutoProcessor
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 strict: false
 
 # these 3 lines are needed for now to handle vision chat templates w images

examples/llama-3/fft-8b-liger-fsdp.yaml

@@ -1,4 +1,6 @@
 base_model: NousResearch/Meta-Llama-3.1-8B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 plugins:
   - axolotl.integrations.liger.LigerPlugin

examples/llama-3/fft-8b.yaml

@@ -1,4 +1,6 @@
 base_model: NousResearch/Meta-Llama-3.1-8B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/llama-3/instruct-dpo-lora-8b.yml

@@ -1,6 +1,9 @@
 base_model: meta-llama/Meta-Llama-3-8B-Instruct
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/llama-3/instruct-lora-8b.yml

@@ -1,6 +1,9 @@
 base_model: NousResearch/Meta-Llama-3-8B-Instruct
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/llama-3/lora-1b-deduplicate-dpo.yml

@@ -1,6 +1,9 @@
 base_model: meta-llama/Llama-3.2-1B
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/llama-3/lora-1b-deduplicate-sft.yml

@@ -1,6 +1,9 @@
 base_model: meta-llama/Llama-3.2-1B
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/llama-3/lora-1b.yml

@@ -1,4 +1,6 @@
 base_model: NousResearch/Llama-3.2-1B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/llama-3/lora-8b.yml

@@ -1,6 +1,9 @@
 base_model: NousResearch/Meta-Llama-3-8B
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/llama-3/qlora-1b-kto.yaml

@@ -1,4 +1,6 @@
 base_model: meta-llama/Llama-3.2-1B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/llama-3/qlora-1b.yml

@@ -1,4 +1,6 @@
 base_model: NousResearch/Llama-3.2-1B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/llama-3/qlora-fsdp-405b.yaml

@@ -1,5 +1,8 @@
 base_model: hugging-quants/Meta-Llama-3.1-405B-BNB-NF4-BF16
+# optionally might have model_type or tokenizer_type
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_4bit: true
 strict: false

examples/llama-3/qlora-fsdp-70b.yaml

@@ -1,6 +1,9 @@
 base_model: casperhansen/llama-3-70b-fp16
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: AutoTokenizer  # PreTrainedTokenizerFast
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/llama-3/qlora.yml

@@ -1,6 +1,9 @@
 base_model: NousResearch/Meta-Llama-3-8B
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/mamba/config.yml

@@ -1,7 +1,10 @@
 base_model: state-spaces/mamba-2.8b
+# optionally might have model_type or tokenizer_type or tokenizer_config
 model_type: MambaLMHeadModel
 tokenizer_type: AutoTokenizer
 tokenizer_config: EleutherAI/gpt-neox-20b
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/mistral/bigstral-ds-zero3.yaml

@@ -1,6 +1,10 @@
 base_model: mistral-community/Mixtral-8x22B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/mistral/config.yml

@@ -1,6 +1,9 @@
 base_model: mistralai/Mistral-7B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/mistral/lora-mps.yml

@@ -1,6 +1,9 @@
 base_model: mistralai/Mistral-7B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/mistral/lora.yml

@@ -1,6 +1,9 @@
 base_model: mistralai/Mistral-7B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/mistral/mistral-dpo-qlora.yml

@@ -4,8 +4,11 @@
 #face problems with the special tokens.
 
 base_model: mistralai/Mistral-7B-Instruct-v0.2
+# optionally might have model_type or tokenizer_type
 model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/mistral/mistral-qlora-fsdp.yml

@@ -1,6 +1,10 @@
 base_model: mistralai/Mixtral-8x7B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/mistral/mistral-qlora-orpo.yml

@@ -1,6 +1,9 @@
 base_model: mistralai/Mistral-7B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/mistral/mixtral-8x22b-qlora-fsdp.yml

@@ -1,6 +1,9 @@
 base_model: mistral-community/Mixtral-8x22B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/mistral/mixtral-qlora-fsdp.yml

@@ -1,6 +1,10 @@
 base_model: mistralai/Mixtral-8x7B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/mistral/mixtral.yml

@@ -1,6 +1,10 @@
 base_model: mistralai/Mixtral-8x7B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/mistral/mixtral_22.yml

@@ -1,6 +1,10 @@
 base_model: mistral-community/Mixtral-8x22B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/mistral/qlora.yml

@@ -1,6 +1,9 @@
 base_model: mistralai/Mistral-7B-v0.1
+# optionally might have model_type or tokenizer_type
 model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/mpt-7b/config.yml

@@ -1,5 +1,9 @@
 base_model: mosaicml/mpt-7b
+# optionally might have model_type or tokenizer_type
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true  # required for mpt as their model class is not merged into transformers yet
 load_in_8bit: false
 datasets:

examples/openllama-3b/config.yml

@@ -1,6 +1,10 @@
 base_model: openlm-research/open_llama_3b_v2
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 load_in_8bit: false
 load_in_4bit: false
 strict: false

examples/openllama-3b/lora.yml

@@ -1,6 +1,10 @@
 base_model: openlm-research/open_llama_3b_v2
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 load_in_8bit: true
 load_in_4bit: false
 strict: false

examples/openllama-3b/qlora.yml

@@ -1,6 +1,10 @@
 base_model: openlm-research/open_llama_3b_v2
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 load_in_8bit: false
 load_in_4bit: true
 strict: false

examples/phi/lora-3.5.yaml

@@ -1,6 +1,9 @@
 base_model: microsoft/Phi-3.5-mini-instruct
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/phi/phi-ft.yml

@@ -1,6 +1,9 @@
 base_model: microsoft/phi-1_5
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/phi/phi-qlora.yml

@@ -1,6 +1,9 @@
 base_model: microsoft/phi-1_5
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/phi/phi2-ft.yml

@@ -1,6 +1,9 @@
 base_model: microsoft/phi-2
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/phi/phi3-ft-fsdp.yml

@@ -1,6 +1,9 @@
 base_model: microsoft/Phi-3-mini-4k-instruct
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/phi/phi3-ft.yml

@@ -1,7 +1,11 @@
 base_model: microsoft/Phi-3-mini-4k-instruct
+# optionally might have model_type or tokenizer_type
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 chat_template: phi_3
 
 load_in_8bit: false

examples/pythia-12b/config.yml

@@ -1,7 +1,11 @@
 base_model: EleutherAI/pythia-12b-deduped
 base_model_ignore_patterns: pytorch*  # prefer safetensors
+# optionally might have model_type or tokenizer_type
 model_type: GPTNeoXForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 load_in_8bit: false
 load_in_4bit: false
 gptq: false

examples/pythia/lora.yml

@@ -1,4 +1,7 @@
 base_model: EleutherAI/pythia-1.4b-deduped
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 load_in_8bit: true
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

examples/qwen/lora.yml

@@ -1,6 +1,9 @@
 base_model: Qwen/Qwen-7B
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 trust_remote_code: true
 

examples/qwen/qlora.yml

@@ -1,6 +1,9 @@
 base_model: Qwen/Qwen-7B
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 trust_remote_code: true
 

examples/qwen/qwen2-moe-lora.yaml

@@ -1,4 +1,7 @@
 base_model: Qwen/Qwen1.5-MoE-A2.7B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/qwen/qwen2-moe-qlora.yaml

@@ -1,4 +1,7 @@
 base_model: Qwen/Qwen1.5-MoE-A2.7B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/qwen2/dpo.yaml

@@ -1,4 +1,6 @@
 base_model: Qwen/Qwen2.5-0.5B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 strict: false
 

examples/qwen2/qlora-fsdp.yaml

@@ -1,4 +1,7 @@
 base_model: Qwen/Qwen2-7B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

examples/redpajama/config-3b.yml

@@ -1,6 +1,10 @@
 base_model: togethercomputer/RedPajama-INCITE-Chat-3B-v1
+# optionally might have model_type or tokenizer_type
 model_type: GPTNeoXForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code:
 load_in_8bit: false
 datasets:

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

@@ -1,4 +1,7 @@
 base_model: replit/replit-code-v1-3b
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 load_in_8bit: false
 datasets:

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

@@ -1,6 +1,10 @@
 base_model: stabilityai/stablelm-2-1_6b
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: false

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

@@ -1,6 +1,10 @@
 base_model: stabilityai/stablelm-2-1_6b
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
 trust_remote_code: true
 
 load_in_8bit: true

examples/starcoder2/qlora.yml

@@ -1,4 +1,6 @@
 base_model: bigcode/starcoder2-3b
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

examples/tiny-llama/lora-mps.yml

@@ -1,6 +1,9 @@
 base_model: TinyLlama/TinyLlama_v1.1
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/tiny-llama/lora.yml

@@ -1,5 +1,8 @@
 base_model: TinyLlama/TinyLlama_v1.1
+# optionally might have model_type or tokenizer_type
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: true
 load_in_4bit: false

examples/tiny-llama/pretrain.yml

@@ -1,7 +1,9 @@
 base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
-
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: false

examples/tiny-llama/qlora.yml

@@ -1,6 +1,9 @@
 base_model: TinyLlama/TinyLlama_v1.1
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

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

@@ -1,9 +1,14 @@
 # An example finetuning Saleforce's XGen-7b model with 8k context using qlora
 # on Tim Dettmer's Guanaco dataset.
 base_model: Salesforce/xgen-7b-8k-base
-trust_remote_code: true
+# optionally might have model_type or tokenizer_type
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
 load_in_8bit: false
 # enable 4bit for QLoRA
 load_in_4bit: true

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

@@ -1,6 +1,9 @@
 base_model: 01-ai/Yi-34B-Chat
+# optionally might have model_type or tokenizer_type
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
 
 load_in_8bit: false
 load_in_4bit: true

requirements.txt

@@ -7,26 +7,31 @@ mamba-ssm==1.2.0.post1
 flash-attn==2.7.0.post2
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.4.2
+liger-kernel==0.5.2
 # END section
 
 packaging==23.2
+
 peft==0.14.0
-transformers>=4.46.3
+transformers==4.47.1
 tokenizers>=0.20.1
-accelerate==1.2.0
+accelerate==1.2.1
 datasets==3.1.0
 deepspeed==0.16.1
+trl==0.12.1
+
+optimum==1.16.2
+hf_transfer
+sentencepiece
+gradio==3.50.2
+
 pydantic==2.6.3
 addict
 fire
 PyYAML>=6.0
 requests
-sentencepiece
 wandb
 einops
-optimum==1.16.2
-hf_transfer
 colorama
 numba
 numpy>=1.24.4,<=2.0.1
@@ -36,7 +41,6 @@ scipy
 scikit-learn==1.4.2
 nvidia-ml-py==12.560.30
 art
-gradio==3.50.2
 tensorboard
 python-dotenv==1.0.1
 
@@ -45,7 +49,6 @@ s3fs>=2024.5.0
 gcsfs>=2024.5.0
 # adlfs
 
-trl==0.12.1
 zstandard==0.22.0
 fastcore
 
@@ -55,5 +58,7 @@ langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
-torchao==0.5.0
+torchao==0.7.0
 schedulefree==1.3.0
+
+axolotl-contribs-lgpl==0.0.1b2

scripts/unsloth_install.py

@@ -32,5 +32,5 @@ else:
     raise RuntimeError(f"Torch = {v} too new!")
 x = x.format(cuda.replace(".", ""), "-ampere" if is_ampere else "")
 print(
-    f'pip install unsloth-zoo==2024.11.7 && pip install --no-deps "unsloth[{x}]==2024.11.9"'
+    f'pip install unsloth-zoo==2024.12.1 && pip install --no-deps "unsloth[{x}]==2024.12.4"'
 )

src/axolotl/__init__.py

@@ -1,3 +1,7 @@
 """Axolotl - Train and fine-tune large language models"""
 
+import pkgutil
+
+__path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
+
 __version__ = "0.6.0"

src/axolotl/cli/evaluate.py

@@ -0,0 +1,52 @@
+"""
+CLI to run training on a model
+"""
+import logging
+from pathlib import Path
+from typing import Dict, Union
+
+import fire
+from dotenv import load_dotenv
+from transformers.hf_argparser import HfArgumentParser
+
+from axolotl.cli import (
+    check_accelerate_default_config,
+    check_user_token,
+    load_cfg,
+    load_datasets,
+    load_rl_datasets,
+    print_axolotl_text_art,
+)
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.evaluate import evaluate
+
+LOG = logging.getLogger("axolotl.cli.evaluate")
+
+
+def do_evaluate(cfg, cli_args) -> Dict[str, float]:
+    # pylint: disable=duplicate-code
+    print_axolotl_text_art()
+    check_accelerate_default_config()
+    check_user_token()
+
+    if cfg.rl:  # and cfg.rl != "orpo":
+        dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
+    else:
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+    return evaluate(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+
+
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
+    # pylint: disable=duplicate-code
+    parsed_cfg = load_cfg(config, **kwargs)
+    parser = HfArgumentParser(TrainerCliArgs)
+    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
+        return_remaining_strings=True
+    )
+    do_evaluate(parsed_cfg, parsed_cli_args)
+
+
+if __name__ == "__main__":
+    load_dotenv()
+    fire.Fire(do_cli)

src/axolotl/cli/integrations/convert_diff_transformer.py

@@ -0,0 +1,207 @@
+"""CLI to convert a transformers model's attns to diff attns."""
+import logging
+import warnings
+from pathlib import Path
+from time import time
+from typing import Union
+
+import fire
+import torch
+import yaml
+from colorama import Fore
+from dotenv import load_dotenv
+from transformers import HfArgumentParser
+
+from axolotl.cli import load_cfg, print_axolotl_text_art
+from axolotl.common.cli import ConvertDiffTransformerCliArgs, load_model_and_tokenizer
+from axolotl.integrations.diff_transformer.convert import convert_to_diff_attn
+from axolotl.utils.yaml import dump_yaml_preserved_order
+
+LOG = logging.getLogger(__name__)
+
+
+def test_inference(model, tokenizer, prompt="The quick brown fox"):
+    """Run test inference and return generation time"""
+    try:
+        inputs = tokenizer(prompt, return_tensors="pt")
+        inputs = {
+            k: v.to(device=model.device, dtype=torch.long) for k, v in inputs.items()
+        }
+
+        start = time()
+        with torch.no_grad():
+            outputs = model.generate(
+                **inputs,
+                max_new_tokens=20,
+                num_beams=1,
+                do_sample=False,
+                pad_token_id=tokenizer.pad_token_id,
+                use_cache=False,
+            )
+        elapsed = time() - start
+
+        generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
+        LOG.info("Prompt: %s", prompt)
+        LOG.info("Generated: %s", generated_text)
+        LOG.info("Generation time: %.2fs", elapsed)
+
+        return elapsed, generated_text
+
+    except Exception as exc:
+        LOG.error("Inference failed: %s", str(exc))
+        raise
+
+
+def convert_diff_transformer(cfg, cli_args, config_path):
+    debug_info = {}
+
+    # Load model and tokenizer
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
+        model.to(cfg.device, dtype=cfg.torch_dtype)
+
+    # Log original model info
+    LOG.info(
+        "Original model config:\n\t- Hidden size: %d\n\t- Num attention heads: %d",
+        model.config.hidden_size,
+        model.config.num_attention_heads,
+    )
+
+    # Test original model
+    if cli_args.debug:
+        LOG.info("Testing original model...")
+        debug_info["orig_time"], debug_info["orig_text"] = test_inference(
+            model, tokenizer
+        )
+
+    # Convert attention
+    LOG.info("Converting to differential attention...")
+    if cli_args.split_heads and cli_args.zero_init:
+        LOG.warning(
+            Fore.YELLOW
+            + "Warning: Using split_heads with zero_init is not recommended; "
+            + "split_heads will preclude the effects of zero_init"
+            + Fore.RESET
+        )
+    try:
+        model = convert_to_diff_attn(
+            model=model,
+            zero_init=cli_args.zero_init,
+            sublayer_norm=cli_args.sublayer_norm,
+            split_heads=cli_args.split_heads,
+        )
+        model.to(cfg.device, dtype=cfg.torch_dtype)
+    except Exception as exc:
+        LOG.error(Fore.RED + "Conversion failed: %s" + Fore.RESET, str(exc))
+        raise
+
+    # Test converted model
+    if cli_args.debug:
+        LOG.info("Testing converted model...")
+        debug_info["conv_time"], debug_info["conv_text"] = test_inference(
+            model, tokenizer
+        )
+
+    # Save if requested
+    if cfg.output_dir:
+        # Save model and tokenizer
+        LOG.info("Saving converted model to %s", cfg.output_dir)
+        model.save_pretrained(cfg.output_dir)
+        tokenizer.save_pretrained(cfg.output_dir)
+
+        # Modify config to reflect new path / differential attention
+        output_config_path = Path(cfg.output_dir) / "axolotl_config.yml"
+        LOG.info("Saving updated config to %s", output_config_path)
+
+        with open(config_path, "r", encoding="utf-8") as file:
+            modified_cfg = yaml.safe_load(file) or {}
+
+        modified_cfg["base_model"] = cfg.output_dir
+        modified_cfg["diff_attention"] = True
+        plugin_class = (
+            "axolotl.integrations.diff_transformer.DifferentialTransformerPlugin"
+        )
+        if "plugins" in modified_cfg:
+            modified_cfg["plugins"].append(plugin_class)
+        else:
+            modified_cfg["plugins"] = [plugin_class]
+
+        dump_yaml_preserved_order(
+            data=modified_cfg,
+            reference_yaml_path=config_path,
+            output_path=output_config_path,
+        )
+    else:
+        LOG.info("Not saving converted model to disk")
+        LOG.info("Pass --output-dir path/to/save to save model")
+
+    if cli_args.debug:
+        LOG.info(
+            Fore.GREEN
+            + "Conversion successful!\n"
+            + f"Original generation time: {debug_info['orig_time']:.2f}s\n"
+            + f"Converted generation time: {debug_info['conv_time']:.2f}s"
+            + Fore.RESET
+        )
+
+        if debug_info["orig_text"] == debug_info["conv_text"]:
+            LOG.info(
+                Fore.GREEN
+                + "Generations match!\n"
+                + "Model generation:\n"
+                + "*" * 50
+                + "\n"
+                + f"{debug_info['orig_text']}\n"
+                + "*" * 50
+                + "\n"
+                + Fore.RESET
+            )
+            debug_info["generations_match"] = True
+        else:
+            message = (
+                "Generations do not match.\n"
+                + "Original generation:\n"
+                + "*" * 50
+                + "\n"
+                + f"{debug_info['orig_text']}\n"
+                + "*" * 50
+                + "\n"
+                + "Converted generation:\n"
+                + "*" * 50
+                + "\n"
+                + f"{debug_info['conv_text']}\n"
+                + "*" * 50
+                + "\n"
+            )
+            debug_info["generations_match"] = False
+
+            if cli_args.zero_init and not cli_args.sublayer_norm:
+                LOG.info(Fore.RED + message + Fore.RESET)
+                debug_info["match_expected"] = True
+            else:
+                LOG.info(
+                    Fore.YELLOW
+                    + message
+                    + "However, this is expected since --zero-init"
+                    + " and --no-sublayer-norm were not passed."
+                    + Fore.RESET
+                )
+                debug_info["match_expected"] = False
+
+    return model, debug_info
+
+
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
+    print_axolotl_text_art()
+
+    cfg = load_cfg(config, **kwargs)
+    parser = HfArgumentParser(ConvertDiffTransformerCliArgs)
+    cli_args, _ = parser.parse_args_into_dataclasses(return_remaining_strings=True)
+
+    convert_diff_transformer(cfg, cli_args, config)
+
+
+if __name__ == "__main__":
+    load_dotenv()
+    fire.Fire(do_cli)

src/axolotl/cli/integrations/convert_rala.py

@@ -0,0 +1,197 @@
+"""CLI to convert a transformers model's attns to rala attns."""
+import logging
+import warnings
+from pathlib import Path
+from time import time
+from typing import Union
+
+import fire
+import torch
+import yaml
+from colorama import Fore
+from dotenv import load_dotenv
+from transformers import HfArgumentParser
+
+from axolotl.cli import load_cfg, print_axolotl_text_art
+from axolotl.common.cli import ConvertDiffTransformerCliArgs, load_model_and_tokenizer
+from axolotl.integrations.rala.convert import convert_to_rala
+from axolotl.utils.yaml import dump_yaml_preserved_order
+
+LOG = logging.getLogger(__name__)
+
+
+def test_inference(model, tokenizer, prompt="The quick brown fox"):
+    """Run test inference and return generation time"""
+    try:
+        inputs = tokenizer(prompt, return_tensors="pt")
+        inputs = {
+            k: v.to(device=model.device, dtype=torch.long) for k, v in inputs.items()
+        }
+
+        start = time()
+        with torch.no_grad():
+            outputs = model.generate(
+                **inputs,
+                max_new_tokens=20,
+                num_beams=1,
+                do_sample=False,
+                pad_token_id=tokenizer.pad_token_id,
+                use_cache=False,
+            )
+        elapsed = time() - start
+
+        generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
+        LOG.info("Prompt: %s", prompt)
+        LOG.info("Generated: %s", generated_text)
+        LOG.info("Generation time: %.2fs", elapsed)
+
+        return elapsed, generated_text
+
+    except Exception as exc:
+        LOG.error("Inference failed: %s", str(exc))
+        raise
+
+
+def convert_rala(cfg, cli_args, config_path):
+    debug_info = {}
+
+    # Load model and tokenizer
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
+        model.to(cfg.device, dtype=cfg.torch_dtype)
+
+    # Log original model info
+    LOG.info(
+        "Original model config:\n\t- Hidden size: %d\n\t- Num attention heads: %d",
+        model.config.hidden_size,
+        model.config.num_attention_heads,
+    )
+
+    # Test original model
+    if cli_args.debug:
+        LOG.info("attention layers to RALA attention")
+        debug_info["orig_time"], debug_info["orig_text"] = test_inference(
+            model, tokenizer
+        )
+
+    # Convert attention
+    try:
+        model = convert_to_rala(
+            model=model,
+            zero_init=cli_args.zero_init,
+        )
+        model.to(cfg.device, dtype=cfg.torch_dtype)
+    except Exception as exc:
+        LOG.error(Fore.RED + "Conversion failed: %s" + Fore.RESET, str(exc))
+        raise
+
+    # Test converted model
+    if cli_args.debug:
+        LOG.info("Testing converted model...")
+        debug_info["conv_time"], debug_info["conv_text"] = test_inference(
+            model, tokenizer
+        )
+
+    # Save if requested
+    if cfg.output_dir:
+        # Save model and tokenizer
+        LOG.info("Saving converted model to %s", cfg.output_dir)
+        model.save_pretrained(cfg.output_dir)
+        tokenizer.save_pretrained(cfg.output_dir)
+
+        # Modify config to reflect new path / differential attention
+        output_config_path = Path(cfg.output_dir) / "axolotl_config.yml"
+        LOG.info("Saving updated config to %s", output_config_path)
+
+        with open(config_path, "r", encoding="utf-8") as file:
+            modified_cfg = yaml.safe_load(file) or {}
+
+        modified_cfg["base_model"] = cfg.output_dir
+        modified_cfg["rala_attention"] = True
+        plugin_class = "axolotl.integrations.rala.RalaPlugin"
+        if "plugins" in modified_cfg:
+            modified_cfg["plugins"].append(plugin_class)
+        else:
+            modified_cfg["plugins"] = [plugin_class]
+
+        dump_yaml_preserved_order(
+            data=modified_cfg,
+            reference_yaml_path=config_path,
+            output_path=output_config_path,
+        )
+    else:
+        LOG.info("Not saving converted model to disk")
+        LOG.info("Pass --output-dir path/to/save to save model")
+
+    if cli_args.debug:
+        LOG.info(
+            Fore.GREEN
+            + "Conversion successful!\n"
+            + f"Original generation time: {debug_info['orig_time']:.2f}s\n"
+            + f"Converted generation time: {debug_info['conv_time']:.2f}s"
+            + Fore.RESET
+        )
+
+        if debug_info["orig_text"] == debug_info["conv_text"]:
+            LOG.info(
+                Fore.GREEN
+                + "Generations match!\n"
+                + "Model generation:\n"
+                + "*" * 50
+                + "\n"
+                + f"{debug_info['orig_text']}\n"
+                + "*" * 50
+                + "\n"
+                + Fore.RESET
+            )
+            debug_info["generations_match"] = True
+        else:
+            message = (
+                "Generations do not match.\n"
+                + "Original generation:\n"
+                + "*" * 50
+                + "\n"
+                + f"{debug_info['orig_text']}\n"
+                + "*" * 50
+                + "\n"
+                + "Converted generation:\n"
+                + "*" * 50
+                + "\n"
+                + f"{debug_info['conv_text']}\n"
+                + "*" * 50
+                + "\n"
+            )
+            debug_info["generations_match"] = False
+
+            if cli_args.zero_init and not cli_args.sublayer_norm:
+                LOG.info(Fore.RED + message + Fore.RESET)
+                debug_info["match_expected"] = True
+            else:
+                LOG.info(
+                    Fore.YELLOW
+                    + message
+                    + "However, this is expected since --zero-init"
+                    + " and --no-sublayer-norm were not passed."
+                    + Fore.RESET
+                )
+                debug_info["match_expected"] = False
+
+    return model, debug_info
+
+
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
+    print_axolotl_text_art()
+
+    cfg = load_cfg(config, **kwargs)
+    if cfg.rala_attention:
+        cfg.rala_attention = False
+    parser = HfArgumentParser(ConvertDiffTransformerCliArgs)
+    cli_args, _ = parser.parse_args_into_dataclasses(return_remaining_strings=True)
+
+    convert_rala(cfg, cli_args, config)
+
+
+if __name__ == "__main__":
+    load_dotenv()
+    fire.Fire(do_cli)

src/axolotl/cli/main.py

@@ -12,7 +12,13 @@ from axolotl.cli.utils import (
     build_command,
     fetch_from_github,
 )
-from axolotl.common.cli import PreprocessCliArgs, TrainerCliArgs
+from axolotl.common.cli import (
+    ConvertDiffTransformerCliArgs,
+    EvaluateCliArgs,
+    PreprocessCliArgs,
+    TrainerCliArgs,
+)
+from axolotl.utils import set_pytorch_cuda_alloc_conf
 from axolotl.utils.config.models.input.v0_4_1 import AxolotlInputConfig
 
 
@@ -48,6 +54,9 @@ def train(config: str, accelerate: bool, **kwargs):
     """Train or fine-tune a model."""
     kwargs = {k: v for k, v in kwargs.items() if v is not None}
 
+    # Enable expandable segments for cuda allocation to improve VRAM usage
+    set_pytorch_cuda_alloc_conf()
+
     if accelerate:
         base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.train"]
         if config:
@@ -60,6 +69,34 @@ def train(config: str, accelerate: bool, **kwargs):
         do_cli(config=config, **kwargs)
 
 
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@click.option(
+    "--accelerate/--no-accelerate",
+    default=True,
+    help="Use accelerate launch for multi-GPU training",
+)
+@add_options_from_dataclass(EvaluateCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+def evaluate(config: str, accelerate: bool, **kwargs):
+    """Evaluate a model."""
+    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+    # Enable expandable segments for cuda allocation to improve VRAM usage
+    set_pytorch_cuda_alloc_conf()
+
+    if accelerate:
+        base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.evaluate"]
+        if config:
+            base_cmd.append(config)
+        cmd = build_command(base_cmd, kwargs)
+        subprocess.run(cmd, check=True)  # nosec B603
+    else:
+        from axolotl.cli.evaluate import do_cli
+
+        do_cli(config=config, **kwargs)
+
+
 @cli.command()
 @click.argument("config", type=click.Path(exists=True, path_type=str))
 @click.option(
@@ -211,6 +248,32 @@ def merge_lora(
     do_cli(config=config, **kwargs)
 
 
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@add_options_from_dataclass(ConvertDiffTransformerCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+def convert_diff_transformer(config: str, **kwargs):
+    """Convert model attention layers to differential attention layers."""
+    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+    from axolotl.cli.integrations.convert_diff_transformer import do_cli
+
+    do_cli(config=config, **kwargs)
+
+
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@add_options_from_dataclass(ConvertDiffTransformerCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+def convert_rala(config: str, **kwargs):
+    """Convert model attention layers to RALA attention layers."""
+    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+    from axolotl.cli.integrations.convert_rala import do_cli
+
+    do_cli(config=config, **kwargs)
+
+
 @cli.command()
 @click.argument("directory", type=click.Choice(["examples", "deepspeed_configs"]))
 @click.option("--dest", help="Destination directory")