update configs to match parallel PR

recipes/configs/gemma2/27B_full.yaml

@@ -23,6 +23,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -53,6 +54,7 @@ loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -69,4 +71,4 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-27b-finetune
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True

recipes/configs/gemma2/27B_lora.yaml

@@ -23,6 +23,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -65,6 +66,7 @@ batch_size: 4
 epochs: 3
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -81,4 +83,4 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-27b-lora
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True

recipes/configs/gemma2/27B_lora_single_device.yaml

@@ -22,6 +22,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -64,7 +65,7 @@ batch_size: 2
 epochs: 1
 max_steps_per_epoch: null
 gradient_accumulation_steps: 8
-compile: False
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -82,7 +83,7 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-27b-lora
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True
 
 # Show case the usage of pytorch profiler
 # Set enabled to False as it's only needed for debugging training

recipes/configs/gemma2/27B_qlora_single_device.yaml

@@ -22,6 +22,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -64,7 +65,7 @@ batch_size: 4
 epochs: 3
 max_steps_per_epoch: null
 gradient_accumulation_steps: 4
-compile: False
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -82,7 +83,7 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-27b-lora
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True
 
 # Show case the usage of pytorch profiler
 # Set enabled to False as it's only needed for debugging training

recipes/configs/gemma2/2B_full.yaml

@@ -23,6 +23,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -55,6 +56,7 @@ loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -71,4 +73,4 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-finetune
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True

recipes/configs/gemma2/2B_lora.yaml

@@ -22,6 +22,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -67,6 +68,7 @@ batch_size: 4
 epochs: 3
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -83,4 +85,4 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-lora
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True

recipes/configs/gemma2/2B_lora_single_device.yaml

@@ -22,6 +22,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -66,7 +67,7 @@ batch_size: 8
 epochs: 3
 max_steps_per_epoch: null
 gradient_accumulation_steps: 2
-compile: False
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -84,7 +85,7 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-lora
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True
 
 # Show case the usage of pytorch profiler
 # Set enabled to False as it's only needed for debugging training

recipes/configs/gemma2/2B_qlora_single_device.yaml

@@ -22,6 +22,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -66,7 +67,7 @@ batch_size: 4
 epochs: 3
 max_steps_per_epoch: null
 gradient_accumulation_steps: 4
-compile: False
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -84,7 +85,7 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-lora
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True
 
 # Show case the usage of pytorch profiler
 # Set enabled to False as it's only needed for debugging training

recipes/configs/gemma2/9B_full.yaml

@@ -23,6 +23,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -53,6 +54,7 @@ loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -69,4 +71,4 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-9b-finetune
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True

recipes/configs/gemma2/9B_lora.yaml

@@ -23,6 +23,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -65,6 +66,7 @@ batch_size: 4
 epochs: 3
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -81,4 +83,4 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-9b-lora
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True

recipes/configs/gemma2/9B_lora_single_device.yaml

@@ -22,6 +22,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -64,7 +65,7 @@ batch_size: 8
 epochs: 1
 max_steps_per_epoch: null
 gradient_accumulation_steps: 2
-compile: False
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -82,7 +83,7 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-9b-lora
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True
 
 # Show case the usage of pytorch profiler
 # Set enabled to False as it's only needed for debugging training

recipes/configs/gemma2/9B_qlora_single_device.yaml

@@ -22,6 +22,7 @@ tokenizer:
 
 # Dataset
 dataset:
+  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
 seed: null
 shuffle: True
@@ -64,7 +65,7 @@ batch_size: 4
 epochs: 3
 max_steps_per_epoch: null
 gradient_accumulation_steps: 4
-compile: False
+compile: False  # pytorch compile, set to true for perf/memory improvement
 
 # Training env
 device: cuda
@@ -82,7 +83,7 @@ metric_logger:
   log_dir: ${output_dir}
 output_dir: /tmp/alpaca-gemma2-9b-lora
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True
 
 # Show case the usage of pytorch profiler
 # Set enabled to False as it's only needed for debugging training