update configs (#1954)

Co-authored-by: Felipe Mello <felipemello@fb.com>
Co-authored-by: ebsmothers <ebs@meta.com>

Author: 3 people

recipes/configs/code_llama2/7B_full_low_memory.yaml

@@ -45,8 +45,8 @@ resume_from_checkpoint: False
 
 # Dataset
 dataset:
-  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
+  packed: False  # True increases speed
 
 seed: null
 shuffle: True
@@ -55,20 +55,20 @@ shuffle: True
 epochs: 1
 max_steps_per_epoch: null
 batch_size: 2
-gradient_accumulation_steps: 1
+gradient_accumulation_steps: 1  # Use to increase virtual batch size
 optimizer:
   _component_: bitsandbytes.optim.PagedAdamW
   lr: 2e-5
-optimizer_in_bwd: True
+optimizer_in_bwd: True  # True saves memory. Requires gradient_accumulation_steps=1
 loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
-compile: False
+compile: False  # pytorch compile, set to true for better perf/memory
 
 # Training env
 device: cuda
 
 # Memory management
-enable_activation_checkpointing: True
+enable_activation_checkpointing: True  # True reduces memory
 enable_activation_offloading: True  # True reduces memory
 dtype: bf16
 
@@ -79,3 +79,28 @@ metric_logger:
   log_dir: /tmp/CodeLlama-7b-hf/logs
 log_every_n_steps: 1
 log_peak_memory_stats: True
+
+# Profiler (disabled)
+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: 3
+  active_steps: 2
+  num_cycles: 1

recipes/configs/code_llama2/7B_lora_single_device.yaml

@@ -18,11 +18,11 @@
 # Model Arguments
 model:
   _component_: torchtune.models.code_llama2.lora_code_llama2_7b
-  lora_attn_modules: ['q_proj', 'v_proj']
-  apply_lora_to_mlp: False
+  lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
+  apply_lora_to_mlp: True
   apply_lora_to_output: False
-  lora_rank: 8
-  lora_alpha: 16
+  lora_rank: 8  # higher increases accuracy and memory
+  lora_alpha: 16  # usually alpha=2*rank
   lora_dropout: 0.0
 
 # Tokenizer
@@ -49,8 +49,8 @@ save_adapter_weights_only: False
 
 # Dataset
 dataset:
-  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_cleaned_dataset
+  packed: False  # True increases speed
 
 seed: null
 shuffle: True
@@ -59,7 +59,7 @@ shuffle: True
 epochs: 1
 max_steps_per_epoch: null
 batch_size: 2
-gradient_accumulation_steps: 16
+gradient_accumulation_steps: 8  # Use to increase virtual batch size
 optimizer:
   _component_: torch.optim.AdamW
   fused: True
@@ -70,13 +70,13 @@ lr_scheduler:
   num_warmup_steps: 100
 loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
-compile: False
+compile: False  # pytorch compile, set to true for better perf/memory
 
 # Training env
 device: cuda
 
 # Memory management
-enable_activation_checkpointing: True
+enable_activation_checkpointing: True  # True reduces memory
 enable_activation_offloading: False  # True reduces memory
 dtype: bf16
 

recipes/configs/code_llama2/7B_qlora_single_device.yaml

@@ -18,11 +18,11 @@
 # Model Arguments
 model:
   _component_: torchtune.models.code_llama2.qlora_code_llama2_7b
-  lora_attn_modules: ['q_proj', 'v_proj', 'k_proj', 'output_proj']
+  lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
   apply_lora_to_mlp: True
   apply_lora_to_output: False
-  lora_rank: 8
-  lora_alpha: 16
+  lora_rank: 8  # higher increases accuracy and memory
+  lora_alpha: 16  # usually alpha=2*rank
   lora_dropout: 0.0
 
 # Tokenizer
@@ -49,16 +49,16 @@ save_adapter_weights_only: False
 
 # Dataset
 dataset:
-  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_cleaned_dataset
+  packed: False  # True increases speed
 seed: null
 shuffle: True
 
 # Fine-tuning arguments
 epochs: 1
 max_steps_per_epoch: null
 batch_size: 2
-gradient_accumulation_steps: 16
+gradient_accumulation_steps: 8  # Use to increase virtual batch size
 optimizer:
   _component_: torch.optim.AdamW
   fused: True
@@ -69,13 +69,13 @@ lr_scheduler:
   num_warmup_steps: 100
 loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
-compile: False
+compile: False  # pytorch compile, set to true for better perf/memory
 
 # Training env
 device: cuda
 
 # Memory management
-enable_activation_checkpointing: True
+enable_activation_checkpointing: True  # True reduces memory
 enable_activation_offloading: False  # True reduces memory
 dtype: bf16
 

recipes/configs/dev/8B_full_experimental.yaml

@@ -26,8 +26,8 @@ tokenizer:
 
 # Dataset
 dataset:
-  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
+  packed: False  # True increases speed
 seed: null
 shuffle: True
 
@@ -57,14 +57,14 @@ optimizer:
 loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
-gradient_accumulation_steps: 1
-compile: False
+gradient_accumulation_steps: 1  # Use to increase virtual batch size
+compile: False  # pytorch compile, set to true for better perf/memory
 
 # Training env
 device: cuda
 
 # Memory management
-enable_activation_checkpointing: False
+enable_activation_checkpointing: False  # True reduces memory
 enable_activation_offloading: False  # True reduces memory
 ac_mode: 'selective'  # ['selective', 'full']
 ac_option: 2 # [int] = ac every positive int layer
@@ -81,3 +81,28 @@ metric_logger:
 output_dir: /tmp/alpaca-llama3-finetune
 log_every_n_steps: null
 log_peak_memory_stats: True
+
+# Profiler (disabled)
+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: 3
+  active_steps: 2
+  num_cycles: 1

recipes/configs/gemma/2B_full.yaml

@@ -23,8 +23,8 @@ tokenizer:
 
 # Dataset
 dataset:
-  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
+  packed: False  # True increases speed
 seed: null
 shuffle: True
 
@@ -54,14 +54,15 @@ optimizer:
 loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
-gradient_accumulation_steps: 1
-compile: False
+gradient_accumulation_steps: 1  # Use to increase virtual batch size
+compile: False  # pytorch compile, set to true for better perf/memory
+optimizer_in_bwd: False  # True saves memory. Requires gradient_accumulation_steps=1
 
 # Training env
 device: cuda
 
 # Memory management
-enable_activation_checkpointing: True
+enable_activation_checkpointing: True  # True reduces memory
 enable_activation_offloading: False  # True reduces memory
 
 # Reduced precision
@@ -74,3 +75,28 @@ metric_logger:
 output_dir: /tmp/alpaca-gemma-finetune
 log_every_n_steps: 1
 log_peak_memory_stats: True
+
+# Profiler (disabled)
+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: 3
+  active_steps: 2
+  num_cycles: 1

recipes/configs/gemma/2B_lora.yaml

@@ -22,18 +22,18 @@ tokenizer:
 
 # Dataset
 dataset:
-  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
+  packed: False  # True increases speed
 seed: null
 shuffle: True
 
 # Model Arguments
 model:
   _component_: torchtune.models.gemma.lora_gemma_2b
-  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
+  lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
   apply_lora_to_mlp: True
-  lora_rank: 64
-  lora_alpha: 128
+  lora_rank: 64  # higher increases accuracy and memory
+  lora_alpha: 128  # usually alpha=2*rank
   lora_dropout: 0.0
 
 checkpointer:
@@ -66,14 +66,14 @@ loss:
 batch_size: 4
 epochs: 3
 max_steps_per_epoch: null
-gradient_accumulation_steps: 1
-compile: False
+gradient_accumulation_steps: 1  # Use to increase virtual batch size
+compile: False  # pytorch compile, set to true for better perf/memory
 
 # Training env
 device: cuda
 
 # Memory management
-enable_activation_checkpointing: True
+enable_activation_checkpointing: True  # True reduces memory
 enable_activation_offloading: False  # True reduces memory
 
 # Reduced precision
@@ -86,3 +86,28 @@ metric_logger:
 output_dir: /tmp/alpaca-gemma-lora
 log_every_n_steps: 1
 log_peak_memory_stats: True
+
+# Profiler (disabled)
+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: 3
+  active_steps: 2
+  num_cycles: 1

recipes/configs/gemma/2B_lora_single_device.yaml

@@ -22,18 +22,18 @@ tokenizer:
 
 # Dataset
 dataset:
-  packed: False # Set to true for great speed ups
   _component_: torchtune.datasets.alpaca_dataset
+  packed: False  # True increases speed
 seed: null
 shuffle: True
 
 # Model Arguments
 model:
   _component_: torchtune.models.gemma.lora_gemma_2b
-  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
+  lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
   apply_lora_to_mlp: True
-  lora_rank: 64
-  lora_alpha: 128
+  lora_rank: 64  # higher increases accuracy and memory
+  lora_alpha: 128  # usually alpha=2*rank
   lora_dropout: 0.0
 
 checkpointer:
@@ -65,14 +65,14 @@ loss:
 batch_size: 4
 epochs: 3
 max_steps_per_epoch: null
-gradient_accumulation_steps: 4
-compile: False
+gradient_accumulation_steps: 8  # Use to increase virtual batch size
+compile: False  # pytorch compile, set to true for better perf/memory
 
 # Training env
 device: cuda
 
 # Memory management
-enable_activation_checkpointing: True
+enable_activation_checkpointing: True  # True reduces memory
 enable_activation_offloading: False  # True reduces memory
 
 # Reduced precision