feat: add gemma7b support (#971)

Co-authored-by: Evan Smothers <ebs@fb.com>

README.md

@@ -48,7 +48,7 @@ torchtune currently supports the following models.
 | [Llama2](https://llama.meta.com/llama2/)   | 7B, 13B, 70B [[models](torchtune/models/llama2/_model_builders.py), [configs](recipes/configs/llama2/)]        |
 | [Code-Llama2](https://huggingface.co/codellama)   | 7B, 13B, 70B [[model](torchtune/models/code_llama2/_model_builders.py), [configs](recipes/configs/code_llama2/)] |
 | [Mistral](https://huggingface.co/mistralai)   | 7B [[model](torchtune/models/mistral/_model_builders.py), [configs](recipes/configs/mistral/)] |
-| [Gemma](https://huggingface.co/collections/google/gemma-release-65d5efbccdbb8c4202ec078b)   | 2B [[model](torchtune/models/gemma/_model_builders.py), [configs](recipes/configs/gemma/)] |
+| [Gemma](https://huggingface.co/collections/google/gemma-release-65d5efbccdbb8c4202ec078b)   | 2B, 7B [[model](torchtune/models/gemma/_model_builders.py), [configs](recipes/configs/gemma/)] |
 | [Microsoft Phi3](https://huggingface.co/collections/microsoft/phi-3-6626e15e9585a200d2d761e3) | Mini [[model](torchtune/models/phi3/), [configs](recipes/configs/phi3/)]
 
 We'll be adding a number of new models in the coming weeks, including support for 70B versions and MoEs.

docs/source/api_ref_models.rst

@@ -91,7 +91,7 @@ Pre-trained models can be downloaded from the Hugging Face Hub with the followin
 gemma
 -----
 
-All models from the `Gemma family <https://blog.google/technology/developers/gemma-open-models/>`_.
+Models of size 2B and 7B from the `Gemma family <https://blog.google/technology/developers/gemma-open-models/>`_.
 
 Pre-trained models can be downloaded from the Hugging Face Hub with the following command:
 
@@ -104,3 +104,4 @@ Pre-trained models can be downloaded from the Hugging Face Hub with the followin
     :nosignatures:
 
     gemma.gemma_2b
+    gemma.gemma_7b

recipes/configs/gemma/2B_qlora_single_device.yaml

@@ -5,7 +5,7 @@
 # this run:
 #   tune download google/gemma-2b --hf-token <HF_TOKEN> --output-dir /tmp/gemma --ignore-patterns ""
 #
-# To launch on 4 devices, run the following command from root:
+# To launch on a single device, run the following command from root:
 #   tune run lora_finetune_single_device --config gemma/2B_qlora_single_device
 #
 # You can add specific overrides through the command line. For example

recipes/configs/gemma/7B_full.yaml

@@ -0,0 +1,76 @@
+# Config for multi-device full finetuning in full_finetune_distributed.py
+# using a gemma 7B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download google/gemma-7b --hf-token <HF_TOKEN> --output-dir /tmp/gemma-7b --ignore-patterns "gemma-7b.gguf"
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 4 full_finetune_distributed --config gemma/7B_full
+#
+# You can add specific overrides through the command line. For example
+# to override the checkpointer directory while launching training
+# you can run:
+#   tune run --nnodes 1 --nproc_per_node 4 full_finetune_distributed --config gemma/7B_full checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only when the model is being fine-tuned on 2+ GPUs.
+
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.gemma.gemma_tokenizer
+  path: /tmp/gemma-7b/tokenizer.model
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_dataset
+  train_on_input: True
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.gemma.gemma_7b
+
+checkpointer:
+  _component_: torchtune.utils.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/gemma-7b/
+  checkpoint_files: [
+    model-00001-of-00004.safetensors,
+    model-00002-of-00004.safetensors,
+    model-00003-of-00004.safetensors,
+    model-00004-of-00004.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/gemma
+  model_type: GEMMA
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 1
+epochs: 1
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 2e-5
+loss:
+  _component_: torch.nn.CrossEntropyLoss
+max_steps_per_epoch: null
+gradient_accumulation_steps: 1
+
+# Training env
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True
+memory_efficient_fsdp_wrap: False
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  _component_: torchtune.utils.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+output_dir: /tmp/alpaca-gemma-finetune
+log_every_n_steps: 1
+log_peak_memory_stats: False

recipes/configs/gemma/7B_lora.yaml

@@ -0,0 +1,85 @@
+# Config for multi-device LoRA finetuning in lora_finetune_distributed.py
+# using a gemma 7B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download google/gemma-7b --hf-token <HF_TOKEN> --output-dir /tmp/gemma-7b --ignore-patterns "gemma-7b.gguf"
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 4 lora_finetune_distributed --config gemma/7B_lora
+#
+# You can add specific overrides through the command line. For example
+# to override the checkpointer directory while launching training
+# you can run:
+#   tune run --nnodes 1 --nproc_per_node 4 lora_finetune_distributed --config gemma/7B_lora checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only when the model is being fine-tuned on 2+ GPUs.
+
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.gemma.gemma_tokenizer
+  path: /tmp/gemma-7b/tokenizer.model
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_dataset
+  train_on_input: True
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.gemma.lora_gemma_7b
+  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
+  apply_lora_to_mlp: True
+  lora_rank: 64
+  lora_alpha: 16
+
+checkpointer:
+  _component_: torchtune.utils.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/gemma-7b/
+  checkpoint_files: [
+    model-00001-of-00004.safetensors,
+    model-00002-of-00004.safetensors,
+    model-00003-of-00004.safetensors,
+    model-00004-of-00004.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/gemma-7b/
+  model_type: GEMMA
+resume_from_checkpoint: False
+
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 2e-5
+
+lr_scheduler:
+  _component_: torchtune.modules.get_cosine_schedule_with_warmup
+  num_warmup_steps: 100
+
+loss:
+  _component_: torch.nn.CrossEntropyLoss
+
+# Fine-tuning arguments
+batch_size: 4
+epochs: 3
+max_steps_per_epoch: null
+gradient_accumulation_steps: 1
+
+# Training env
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  _component_: torchtune.utils.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+output_dir: /tmp/alpaca-gemma-lora
+log_every_n_steps: 1
+log_peak_memory_stats: False

recipes/configs/gemma/7B_lora_single_device.yaml

@@ -0,0 +1,92 @@
+# Config for multi-device LoRA finetuning in lora_finetune_single_device.py
+# using a gemma 7B model
+#
+# This config assumes that you've run the following command before launching
+# this run (torchtune does not use gguf so you can ignore it to save time and space):
+#   tune download google/gemma-7b --hf-token <HF_TOKEN> --output-dir /tmp/gemma-7b --ignore-patterns "gemma-7b.gguf"
+#
+# To launch on a single device, run the following command from root:
+#   tune run lora_finetune_single_device --config gemma/7B_lora_single_device
+#
+# You can add specific overrides through the command line. For example
+# to override the checkpointer directory while launching training
+# you can run:
+#   tune run lora_finetune_single_device --config gemma/7B_lora_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.gemma.gemma_tokenizer
+  path: /tmp/gemma-7b/tokenizer.model
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_dataset
+  train_on_input: True
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.gemma.lora_gemma_7b
+  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
+  apply_lora_to_mlp: True
+  lora_rank: 8
+  lora_alpha: 16
+
+checkpointer:
+  _component_: torchtune.utils.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/gemma-7b/
+  checkpoint_files: [
+    model-00001-of-00004.safetensors,
+    model-00002-of-00004.safetensors,
+    model-00003-of-00004.safetensors,
+    model-00004-of-00004.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/gemma-7b/
+  model_type: GEMMA
+resume_from_checkpoint: False
+
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 5e-5
+
+lr_scheduler:
+  _component_: torchtune.modules.get_cosine_schedule_with_warmup
+  num_warmup_steps: 100
+
+loss:
+  _component_: torch.nn.CrossEntropyLoss
+
+# Fine-tuning arguments
+batch_size: 8
+epochs: 1
+max_steps_per_epoch: null
+gradient_accumulation_steps: 2
+compile: False
+
+# Training env
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  _component_: torchtune.utils.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+output_dir: /tmp/alpaca-gemma-lora
+log_every_n_steps: 1
+log_peak_memory_stats: False
+
+# Show case the usage of pytorch profiler
+# Set enabled to False as it's only needed for debugging training
+profiler:
+  _component_: torchtune.utils.profiler
+  enabled: False
+  output_dir: /tmp/alpaca-gemma-finetune/torchtune_perf_tracing.json

recipes/configs/gemma/7B_qlora_single_device.yaml

@@ -0,0 +1,92 @@
+# Config for multi-device QLoRA finetuning in lora_finetune_single_device.py
+# using a gemma 7B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download google/gemma-7b --hf-token <HF_TOKEN> --output-dir /tmp/gemma-7b --ignore-patterns "gemma-7b.gguf"
+#
+# To launch on a single device, run the following command from root:
+#   tune run lora_finetune_single_device --config gemma/7B_qlora_single_device
+#
+# You can add specific overrides through the command line. For example
+# to override the checkpointer directory while launching training
+# you can run:
+#   tune run lora_finetune_single_device --config gemma/7B_qlora_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.gemma.gemma_tokenizer
+  path: /tmp/gemma-7b/tokenizer.model
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_dataset
+  train_on_input: True
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.gemma.qlora_gemma_7b
+  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
+  apply_lora_to_mlp: True
+  lora_rank: 64
+  lora_alpha: 16
+
+checkpointer:
+  _component_: torchtune.utils.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/gemma-7b/
+  checkpoint_files: [
+    model-00001-of-00004.safetensors,
+    model-00002-of-00004.safetensors,
+    model-00003-of-00004.safetensors,
+    model-00004-of-00004.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/gemma-7b/
+  model_type: GEMMA
+resume_from_checkpoint: False
+
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 2e-5
+
+lr_scheduler:
+  _component_: torchtune.modules.get_cosine_schedule_with_warmup
+  num_warmup_steps: 100
+
+loss:
+  _component_: torch.nn.CrossEntropyLoss
+
+# Fine-tuning arguments
+batch_size: 4
+epochs: 3
+max_steps_per_epoch: null
+gradient_accumulation_steps: 4
+compile: False
+
+# Training env
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  _component_: torchtune.utils.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+output_dir: /tmp/alpaca-gemma-lora
+log_every_n_steps: 1
+log_peak_memory_stats: False
+
+# Show case the usage of pytorch profiler
+# Set enabled to False as it's only needed for debugging training
+profiler:
+  _component_: torchtune.utils.profiler
+  enabled: False
+  output_dir: /tmp/alpaca-gemma-finetune/torchtune_perf_tracing.json