Feat: add gemma3n support (#2852)

* feat: add gemma3n cce

* feat: add sample config

* feat: add gemma3n multimodal mode

* feat: add audio example

* feat: support audio and return pixel values in collator

* feat: support unmask only assistant region (gemma3n for now)

* feat(doc): add notes for audio loading

* feat: add audio support for gemma3n

* feat: update examples

* feat: add gemma3n to the docs

* fix: add link at top

* feat(doc): clarify additional requirements

* fix: mllama missing aspect ratio

* fix: mllama need attention fixes for fa2

* Partially Revert "fix: mllama need attention fixes for fa2"

This reverts commit a0bfdd1777.

* fix: disable FA2 for mllama in vision mode

* feat: update configs to use proper attention

* fix: support other vision features

* feat(doc): clarify requirements for gemma3n

examples/gemma3n/README.md

@@ -0,0 +1,19 @@
+# Gemma-3n
+
+## Requirements
+
+In addition to Axolotl's requirements, Gemma-3n requires
+
+```
+pip3 install timm
+```
+
+If you will load audio datasets, please also install
+
+```
+pip3 install librosa
+```
+
+## Usage
+
+See example configs and the [multimodal doc](https://docs.axolotl.ai/docs/multimodal.html).

examples/gemma3n/gemma-3n-e2b-qlora.yml

@@ -0,0 +1,74 @@
+base_model: google/gemma-3n-E2B-it
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+cut_cross_entropy: true
+
+load_in_8bit: false
+load_in_4bit: true
+
+# for use with fft to only train on language model layers
+# unfrozen_parameters:
+  # - model.language_model.*
+  # - lm_head
+  # - embed_tokens
+
+
+chat_template: gemma3n
+eot_tokens:
+  - <end_of_turn>
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    split: train[:1%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+# lora_target_linear: # Does not work with gemma3n currently
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 4
+optimizer: muon
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+# flash_attention: true  # Any attention impl does not work with gemma3n now
+
+warmup_ratio: 0.1
+evals_per_epoch:
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/gemma3n/gemma-3n-e2b-vision-audio-qlora.yml

@@ -0,0 +1,80 @@
+base_model: google/gemma-3n-E2B-it
+processor_type: AutoProcessor
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+cut_cross_entropy: true
+
+# for use with fft to only train on language model layers
+# unfrozen_parameters:
+  # - model.language_model.*
+  # - lm_head
+  # - embed_tokens
+
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+chat_template: gemma3n
+eot_tokens:
+  - <end_of_turn>
+
+# sample dataset below requires downloading audio/image in advance
+# wget https://huggingface.co/datasets/Nanobit/text-vision-audio-2k-test/resolve/main/African_elephant.jpg
+# wget https://huggingface.co/datasets/Nanobit/text-vision-audio-2k-test/resolve/main/En-us-African_elephant.oga
+datasets:
+  - path: Nanobit/text-vision-audio-2k-test
+    type: chat_template
+    data_files:
+      - dataset.jsonl
+dataset_prepared_path:
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+pad_to_sequence_len: false
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: muon
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+# flash_attention: true  # Any attention impl does not work with gemma3n now
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

examples/gemma3n/gemma-3n-e2b-vision-qlora.yml

@@ -0,0 +1,75 @@
+base_model: google/gemma-3n-E2B-it
+processor_type: AutoProcessor
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+cut_cross_entropy: true
+
+# for use with fft to only train on language model layers
+# unfrozen_parameters:
+  # - model.language_model.*
+  # - lm_head
+  # - embed_tokens
+
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+chat_template: gemma3n
+eot_tokens:
+  - <end_of_turn>
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+dataset_prepared_path:
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+pad_to_sequence_len: false
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: muon
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+# flash_attention: true  # Any attention impl does not work with gemma3n now
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

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

@@ -15,8 +15,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.0
 output_dir: ./outputs/out
 
@@ -40,7 +39,7 @@ wandb_log_model:
 gradient_accumulation_steps: 4
 micro_batch_size: 1
 num_epochs: 1
-optimizer: adamw_bnb_8bit
+optimizer: muon
 lr_scheduler: cosine
 learning_rate: 0.0002
 
@@ -50,8 +49,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-flash_attention: true
-eager_attention:
+# flash_attention: true  # use for text-only mode
+sdp_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/llava/lora-7b.yaml

@@ -11,8 +11,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.0
 output_dir: ./outputs/out
 
@@ -36,7 +35,7 @@ wandb_log_model:
 gradient_accumulation_steps: 4
 micro_batch_size: 1
 num_epochs: 1
-optimizer: adamw_bnb_8bit
+optimizer: muon
 lr_scheduler: cosine
 learning_rate: 0.0002
 

examples/mistral/mistral-small-3.1-24B-lora.yml

@@ -48,8 +48,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet.
-eager_attention:
+# flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet.
+sdp_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/pixtral/lora-12b.yml

@@ -11,8 +11,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.0
 output_dir: ./outputs/out
 
@@ -36,7 +35,7 @@ wandb_log_model:
 gradient_accumulation_steps: 4
 micro_batch_size: 1
 num_epochs: 1
-optimizer: adamw_bnb_8bit
+optimizer: muon
 lr_scheduler: cosine
 learning_rate: 0.0002
 
@@ -46,8 +45,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet
-eager_attention:
+# flash_attention:  # PixtralVisionModel does not support Flash Attention 2.0 yet
+sdp_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch: 1