feat: add gemma2b variants

README.md

@@ -44,6 +44,7 @@ torchtune currently supports the following models.
 | [Code-Llama2](https://ai.meta.com/blog/code-llama-large-language-model-coding/)   | 7B, 13B, 70B [[models](torchtune/models/code_llama2/_model_builders.py), [configs](recipes/configs/code_llama2/)] |
 | [Mistral](https://huggingface.co/mistralai)   | 7B [[models](torchtune/models/mistral/_model_builders.py), [configs](recipes/configs/mistral/)] |
 | [Gemma](https://huggingface.co/collections/google/gemma-release-65d5efbccdbb8c4202ec078b)   | 2B, 7B [[models](torchtune/models/gemma/_model_builders.py), [configs](recipes/configs/gemma/)] |
+| [Gemma2](https://huggingface.co/docs/transformers/main/en/model_doc/gemma2)   | 2B, 9B, 27B [[models](torchtune/models/gemma2/_model_builders.py), [configs](recipes/configs/gemma2/)] |
 | [Microsoft Phi3](https://huggingface.co/collections/microsoft/phi-3-6626e15e9585a200d2d761e3) | Mini [[models](torchtune/models/phi3/), [configs](recipes/configs/phi3/)]
 | [Qwen2](https://qwenlm.github.io/blog/qwen2/) | 0.5B, 1.5B, 7B [[models](torchtune/models/qwen2/), [configs](recipes/configs/qwen2/)]
 

docs/source/api_ref_models.rst

@@ -320,6 +320,37 @@ To download the Gemma 7B model:
     gemma.gemma_tokenizer
 
 
+gemma2 : # TODO
+-----
+
+Models of size 2B, 9B, 27B from the `Gemma family <https://blog.google/technology/developers/gemma-open-models/>`_.
+
+Important: You need to request access on `Hugging Face <https://huggingface.co/google/gemma-2-2b>`__ to use this model.
+
+To download the Gemma2 2B, 9B, 27B models :
+
+.. code-block:: bash
+
+    tune download google/gemma-2-<MODEL_SIZE>b --ignore-patterns "gemma-2-<MODEL_SIZE>b.gguf"  --hf-token <HF_TOKEN>
+
+
+.. autosummary::
+    :toctree: generated/
+    :nosignatures:
+
+    gemma2.gemma2
+    gemma2.lora_gemma
+    gemma2.gemma2_2b
+    gemma2.lora_gemma2_2b
+    gemma2.qlora_gemma2_2b
+    gemma2.gemma2_9b
+    gemma2.lora_gemma2_9b
+    gemma2.qlora_gemma2_9b
+    gemma2.gemma2_27b
+    gemma2.lora_gemma2_27b
+    gemma2.qlora_gemma2_27b
+    gemma.gemma_tokenizer
+
 clip
 -----
 

recipes/configs/gemma2/27B_full.yaml

@@ -0,0 +1,95 @@
+# Config for multi-device full finetuning in full_finetune_distributed.py
+# using a gemma2 27B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download google/gemma-2-27b --ignore-patterns "gemma-2-27b.gguf"  --hf-token <HF_TOKEN>
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 4 full_finetune_distributed --config gemma2/27B_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 gemma2/27B_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/gemma2-27b/tokenizer.model
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_dataset
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.gemma2.gemma_27b
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/gemma2-27b/
+  checkpoint_files: [
+    model-00001-of-00024.safetensors,
+    model-00002-of-00024.safetensors,
+    model-00003-of-00024.safetensors,
+    model-00004-of-00024.safetensors,
+    model-00005-of-00024.safetensors,
+    model-00006-of-00024.safetensors,
+    model-00007-of-00024.safetensors,
+    model-00008-of-00024.safetensors,
+    model-00009-of-00024.safetensors,
+    model-00010-of-00024.safetensors,
+    model-00011-of-00024.safetensors,
+    model-00012-of-00024.safetensors,
+    model-00013-of-00024.safetensors,
+    model-00014-of-00024.safetensors,
+    model-00015-of-00024.safetensors,
+    model-00016-of-00024.safetensors,
+    model-00017-of-00024.safetensors,
+    model-00018-of-00024.safetensors,
+    model-00019-of-00024.safetensors,
+    model-00020-of-00024.safetensors,
+    model-00021-of-00024.safetensors,
+    model-00022-of-00024.safetensors,
+    model-00023-of-00024.safetensors,
+    model-00024-of-00024.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/gemma2-27b
+  model_type: GEMMA2
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 1
+epochs: 1
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  lr: 2e-5
+loss:
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
+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.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+output_dir: /tmp/alpaca-gemma2-27b-finetune
+log_every_n_steps: 1
+log_peak_memory_stats: False

recipes/configs/gemma2/27B_lora.yaml

@@ -0,0 +1,107 @@
+# Config for multi-device LoRA finetuning in lora_finetune_distributed.py
+# using a gemma2 27B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download google/gemma-2-27b --ignore-patterns "gemma-2-27b.gguf"  --hf-token <HF_TOKEN>
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 4 lora_finetune_distributed --config gemma2/27B_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 gemma2/27B_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/gemma2-27b/tokenizer.model
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_dataset
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.gemma2.lora_gemma2_27b
+  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
+  apply_lora_to_mlp: True
+  lora_rank: 64
+  lora_alpha: 128
+  lora_dropout: 0.0
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/gemma2-27b/
+  checkpoint_files: [
+    model-00001-of-00024.safetensors,
+    model-00002-of-00024.safetensors,
+    model-00003-of-00024.safetensors,
+    model-00004-of-00024.safetensors,
+    model-00005-of-00024.safetensors,
+    model-00006-of-00024.safetensors,
+    model-00007-of-00024.safetensors,
+    model-00008-of-00024.safetensors,
+    model-00009-of-00024.safetensors,
+    model-00010-of-00024.safetensors,
+    model-00011-of-00024.safetensors,
+    model-00012-of-00024.safetensors,
+    model-00013-of-00024.safetensors,
+    model-00014-of-00024.safetensors,
+    model-00015-of-00024.safetensors,
+    model-00016-of-00024.safetensors,
+    model-00017-of-00024.safetensors,
+    model-00018-of-00024.safetensors,
+    model-00019-of-00024.safetensors,
+    model-00020-of-00024.safetensors,
+    model-00021-of-00024.safetensors,
+    model-00022-of-00024.safetensors,
+    model-00023-of-00024.safetensors,
+    model-00024-of-00024.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/gemma2-27b/
+  model_type: GEMMA2
+resume_from_checkpoint: False
+save_adapter_weights_only: False
+
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  lr: 2e-5
+
+lr_scheduler:
+  _component_: torchtune.modules.get_cosine_schedule_with_warmup
+  num_warmup_steps: 10
+
+loss:
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
+
+# 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.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+output_dir: /tmp/alpaca-gemma2-27b-lora
+log_every_n_steps: 1
+log_peak_memory_stats: False

recipes/configs/gemma2/27B_lora_single_device.yaml

@@ -0,0 +1,134 @@
+# Config for multi-device LoRA finetuning in lora_finetune_single_device.py
+# using a gemma2 27B 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-2-27b --ignore-patterns "gemma-2-27b.gguf"  --hf-token <HF_TOKEN>
+#
+# To launch on a single device, run the following command from root:
+#   tune run lora_finetune_single_device --config gemma2/27B_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 gemma2/27B_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/gemma2-27b/tokenizer.model
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_dataset
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.gemma2.lora_gemma2_27b
+  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
+  apply_lora_to_mlp: True
+  lora_rank: 8
+  lora_alpha: 16
+  lora_dropout: 0.0
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/gemma2-27b/
+  checkpoint_files: [
+    model-00001-of-00024.safetensors,
+    model-00002-of-00024.safetensors,
+    model-00003-of-00024.safetensors,
+    model-00004-of-00024.safetensors,
+    model-00005-of-00024.safetensors,
+    model-00006-of-00024.safetensors,
+    model-00007-of-00024.safetensors,
+    model-00008-of-00024.safetensors,
+    model-00009-of-00024.safetensors,
+    model-00010-of-00024.safetensors,
+    model-00011-of-00024.safetensors,
+    model-00012-of-00024.safetensors,
+    model-00013-of-00024.safetensors,
+    model-00014-of-00024.safetensors,
+    model-00015-of-00024.safetensors,
+    model-00016-of-00024.safetensors,
+    model-00017-of-00024.safetensors,
+    model-00018-of-00024.safetensors,
+    model-00019-of-00024.safetensors,
+    model-00020-of-00024.safetensors,
+    model-00021-of-00024.safetensors,
+    model-00022-of-00024.safetensors,
+    model-00023-of-00024.safetensors,
+    model-00024-of-00024.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/gemma2-27b/
+  model_type: GEMMA2
+resume_from_checkpoint: False
+save_adapter_weights_only: False
+
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  lr: 5e-5
+
+lr_scheduler:
+  _component_: torchtune.modules.get_cosine_schedule_with_warmup
+  num_warmup_steps: 10
+
+loss:
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
+
+# 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
+enable_activation_offloading: False
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+output_dir: /tmp/alpaca-gemma2-27b-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.training.setup_torch_profiler
+  enabled: False
+
+  #Output directory of trace artifacts
+  output_dir: ${output_dir}/profiling_outputs
+
+  #`torch.profiler.ProfilerActivity` types to trace
+  cpu: True
+  cuda: True
+
+  #trace options passed to `torch.profiler.profile`
+  profile_memory: False
+  with_stack: False
+  record_shapes: True
+  with_flops: False
+
+  # `torch.profiler.schedule` options:
+  # wait_steps -> wait, warmup_steps -> warmup, active_steps -> active, num_cycles -> repeat
+  wait_steps: 5
+  warmup_steps: 5
+  active_steps: 2
+  num_cycles: 1