Update Qwen2.5 configs (#1999)

pytorch · Nov 13, 2024 · 1eb7785 · 1eb7785
1 parent 18d97f0
commit 1eb7785
Show file tree

Hide file tree

Showing 9 changed files with 268 additions and 334 deletions.
diff --git a/recipes/configs/qwen2_5/0_5B_full.yaml → recipes/configs/qwen2_5/0.5B_full.yaml b/recipes/configs/qwen2_5/0_5B_full.yaml → recipes/configs/qwen2_5/0.5B_full.yaml
@@ -1,82 +1,75 @@
 # Config for multi-device full finetuning in full_finetune_distributed.py
 # using a Qwen2.5 0.5B model
 #
-# This config assumes that you've run the following command before launching
-# this run:
-#   tune download Qwen/Qwen2.5-0.5B-Instruct --output-dir /tmp/Qwen2_5-0_5B-Instruct --ignore-patterns None
+# This config assumes that you've run the following command before launching:
+#   tune download Qwen/Qwen2.5-0.5B-Instruct --ignore-patterns None
 #
 # To launch on 2 devices, run the following command from root:
-#   tune run --nnodes 1 --nproc_per_node 2 full_finetune_distributed --config qwen2_5/0_5B_full
+#   tune run --nproc_per_node 2 full_finetune_distributed --config qwen2_5/0.5B_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 2 full_finetune_distributed --config qwen2_5/0_5B_full checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+# to override the checkpointer directory while launching training:
+#   tune run --nproc_per_node 2 full_finetune_distributed --config qwen2_5/0.5B_full checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
 #
-# This config works best when the model is being fine-tuned on 2+ GPUs.
-# Single device full finetuning requires more memory optimizations. It's
-# best to use 0_5B_full_single_device.yaml for those cases
+# This config is for fine-tuning on 2+ GPUs.
+
+# Model arguments
+model:
+  _component_: torchtune.models.qwen2_5.qwen2_5_0_5b
 
 # Tokenizer
 tokenizer:
   _component_: torchtune.models.qwen2_5.qwen2_5_tokenizer
-  path: /tmp/Qwen2_5-0_5B-Instruct/vocab.json
-  merges_file: /tmp/Qwen2_5-0_5B-Instruct/merges.txt
+  path: /tmp/Qwen2.5-0.5B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2.5-0.5B-Instruct/merges.txt
   max_seq_len: null
 
+# Checkpointer
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2.5-0.5B-Instruct
+  checkpoint_files: [model.safetensors]
+  recipe_checkpoint: null
+  output_dir: /tmp/Qwen2.5-0.5B-Instruct-finetune
+  model_type: QWEN2
+resume_from_checkpoint: False
+
 # Dataset
 dataset:
   _component_: torchtune.datasets.alpaca_cleaned_dataset
   packed: False  # True increases speed
 seed: null
 shuffle: True
 
-# Model Arguments
-model:
-  _component_: torchtune.models.qwen2_5.qwen2_5_0_5b
-
-checkpointer:
-  _component_: torchtune.training.FullModelHFCheckpointer
-  checkpoint_dir: /tmp/Qwen2_5-0_5B-Instruct
-  checkpoint_files: [
-    model.safetensors
-  ]
-  recipe_checkpoint: null
-  output_dir: /tmp/Qwen2_5-0_5B-Instruct-finetune
-  model_type: QWEN2
-resume_from_checkpoint: False
-
 # Fine-tuning arguments
-batch_size: 2
 epochs: 1
+max_steps_per_epoch: null
+batch_size: 2
+gradient_accumulation_steps: 8  # Use to increase virtual batch size
 optimizer:
   _component_: torch.optim.AdamW
   fused: True
   lr: 2e-5
+optimizer_in_bwd: False  # True saves memory. Requires gradient_accumulation_steps=1
 loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
-max_steps_per_epoch: null
-gradient_accumulation_steps: 8  # 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  # True reduces memory
+# Memory management / performance
+enable_activation_checkpointing: False  # True reduces memory
 enable_activation_offloading: False  # True reduces memory
-
-# Reduced precision
 dtype: bf16
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
 
 # Logging
+output_dir: /tmp/Qwen2.5-0.5B-Instruct-finetune
 metric_logger:
   _component_: torchtune.training.metric_logging.DiskLogger
-  log_dir: ${output_dir}
-output_dir: /tmp/Qwen2_5-0_5B-Instruct-finetune
+  log_dir: ${output_dir}/logs
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True
 
 # Profiler (disabled)
 profiler:

diff --git a/...figs/qwen2_5/0_5B_full_single_device.yaml → ...figs/qwen2_5/0.5B_full_single_device.yaml b/...figs/qwen2_5/0_5B_full_single_device.yaml → ...figs/qwen2_5/0.5B_full_single_device.yaml
@@ -1,86 +1,75 @@
 # Config for single device full finetuning in full_finetune_single_device.py
 # using a Qwen2.5 0.5B
 #
-# This config assumes that you've run the following command before launching
-# this run:
-#   tune download Qwen/Qwen2.5-0.5B-Instruct --output-dir /tmp/Qwen2_5-0_5B-Instruct --ignore-patterns None
-#
-# The default config uses an optimizer from bitsandbytes. If you do not have it installed,
-# you can install it with
-#   pip install bitsandbytes
+# This config assumes that you've run the following command before launching:
+#   tune download Qwen/Qwen2.5-0.5B-Instruct --ignore-patterns None
 #
 # To launch on a single device, run the following command from root:
-#   tune run full_finetune_single_device --config qwen2_5/0_5B_full_single_device
+#   tune run full_finetune_single_device --config qwen2_5/0.5B_full_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 full_finetune_single_device --config qwen2_5/0_5B_full_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+# to override the checkpointer directory while launching training:
+#   tune run full_finetune_single_device --config qwen2_5/0.5B_full_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
 #
 # This config works only for training on single device.
 
+# Model arguments
+model:
+  _component_: torchtune.models.qwen2_5.qwen2_5_0_5b
+
 # Tokenizer
 tokenizer:
   _component_: torchtune.models.qwen2_5.qwen2_5_tokenizer
-  path: /tmp/Qwen2_5-0_5B-Instruct/vocab.json
-  merges_file: /tmp/Qwen2_5-0_5B-Instruct/merges.txt
+  path: /tmp/Qwen2.5-0.5B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2.5-0.5B-Instruct/merges.txt
   max_seq_len: null
 
+# Checkpointer
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2.5-0.5B-Instruct
+  checkpoint_files: [model.safetensors]
+  recipe_checkpoint: null
+  output_dir: /tmp/Qwen2.5-0.5B-Instruct-finetune
+  model_type: QWEN2
+resume_from_checkpoint: False
+
 # Dataset
 dataset:
   _component_: torchtune.datasets.alpaca_cleaned_dataset
   packed: False  # True increases speed
 seed: null
 shuffle: True
 
-# Model Arguments
-model:
-  _component_: torchtune.models.qwen2_5.qwen2_5_0_5b
-
-checkpointer:
-  _component_: torchtune.training.FullModelHFCheckpointer
-  checkpoint_dir: /tmp/Qwen2_5-0_5B-Instruct
-  checkpoint_files: [
-    model.safetensors
-  ]
-  recipe_checkpoint: null
-  output_dir: /tmp/Qwen2_5-0_5B-Instruct-finetune
-  model_type: QWEN2
-resume_from_checkpoint: False
-
 # Fine-tuning arguments
-batch_size: 2
 epochs: 1
+max_steps_per_epoch: null
+batch_size: 2
+gradient_accumulation_steps: 8  # Use to increase virtual batch size
 optimizer:
   _component_: torch.optim.AdamW
   fused: True
   lr: 2e-5
-
+optimizer_in_bwd: False  # True saves memory. Requires gradient_accumulation_steps=1
 loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
-optimizer_in_bwd: False  # True saves memory. Requires gradient_accumulation_steps=1
 
-max_steps_per_epoch: null
-gradient_accumulation_steps: 8  # Use to increase virtual batch size
-compile: False  # pytorch compile, set to true for better perf/memory
-
-# Training environment
+# Training env
 device: cuda
 
-# Memory management
-enable_activation_checkpointing: True  # True reduces memory
+# Memory management / performance
+enable_activation_checkpointing: False  # True reduces memory
 enable_activation_offloading: False  # True reduces memory
-
-# Reduced precision
 dtype: bf16
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
 
 # Logging
+output_dir: /tmp/Qwen2.5-0.5B-Instruct-finetune
 metric_logger:
   _component_: torchtune.training.metric_logging.DiskLogger
-  log_dir: ${output_dir}
-output_dir: /tmp/Qwen2_5-0_5B-Instruct-finetune
+  log_dir: ${output_dir}/logs
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True
 
 # Profiler (disabled)
 profiler:

diff --git a/recipes/configs/qwen2_5/0_5B_lora.yaml → recipes/configs/qwen2_5/0.5B_lora.yaml b/recipes/configs/qwen2_5/0_5B_lora.yaml → recipes/configs/qwen2_5/0.5B_lora.yaml
@@ -1,23 +1,19 @@
 # Config for multi-device LoRA finetuning in lora_finetune_distributed.py
 # using a Qwen2.5 0.5B model
 #
-# This config assumes that you've run the following command before launching
-# this run:
-#   tune download Qwen/Qwen2.5-0.5B-Instruct --output-dir /tmp/Qwen2_5-0_5B-Instruct --ignore-patterns None
+# This config assumes that you've run the following command before launching:
+#   tune download Qwen/Qwen2.5-0.5B-Instruct --ignore-patterns None
 #
 # To launch on 2 devices, run the following command from root:
-#   tune run --nnodes 1 --nproc_per_node 2 lora_finetune_distributed --config qwen2_5/0_5B_lora
+#   tune run --nproc_per_node 2 lora_finetune_distributed --config qwen2_5/0.5B_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 2 lora_finetune_distributed --config qwen2_5/0_5B_lora checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+# to override the checkpointer directory while launching training:
+#   tune run --nproc_per_node 2 lora_finetune_distributed --config qwen2_5/0.5B_lora checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
 #
-# This config works best when the model is being fine-tuned on 2+ GPUs.
-# For single device LoRA finetuning please use 0_5B_lora_single_device.yaml
+# This config is for fine-tuning on 2+ GPUs.
 
-
-# Model Arguments
+# Model arguments
 model:
   _component_: torchtune.models.qwen2_5.lora_qwen2_5_0_5b
   lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
@@ -26,71 +22,66 @@ model:
   lora_alpha: 64  # usually alpha=2*rank
   lora_dropout: 0.0
 
+# Tokenizer
 tokenizer:
   _component_: torchtune.models.qwen2_5.qwen2_5_tokenizer
-  path: /tmp/Qwen2_5-0_5B-Instruct/vocab.json
-  merges_file: /tmp/Qwen2_5-0_5B-Instruct/merges.txt
+  path: /tmp/Qwen2.5-0.5B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2.5-0.5B-Instruct/merges.txt
   max_seq_len: null
 
+# Checkpointer
 checkpointer:
   _component_: torchtune.training.FullModelHFCheckpointer
-  checkpoint_dir: /tmp/Qwen2_5-0_5B-Instruct
-  checkpoint_files: [
-    model.safetensors
-  ]
+  checkpoint_dir: /tmp/Qwen2.5-0.5B-Instruct
+  checkpoint_files: [model.safetensors]
   recipe_checkpoint: null
-  output_dir: /tmp/Qwen2_5-0_5B-Instruct-lora-finetune
+  output_dir: /tmp/Qwen2.5-0.5B-Instruct-lora-finetune
   model_type: QWEN2
 resume_from_checkpoint: False
 
-# Dataset and Sampler
+# Dataset
 dataset:
   _component_: torchtune.datasets.alpaca_cleaned_dataset
   packed: False  # True increases speed
-
 seed: null
 shuffle: True
-batch_size: 4
 
-# Optimizer and Scheduler
+# Fine-tuning arguments
+epochs: 1
+max_steps_per_epoch: null
+batch_size: 2
+gradient_accumulation_steps: 8  # Use to increase virtual batch size
 optimizer:
   _component_: torch.optim.AdamW
   fused: True
   weight_decay: 0.01
   lr: 2e-3
-
 lr_scheduler:
   _component_: torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
   num_warmup_steps: 100
-
 loss:
   _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
-# Training
-epochs: 1
-max_steps_per_epoch: null
-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 / performance
+enable_activation_checkpointing: False  # True reduces memory
+enable_activation_offloading: False  # True reduces memory
+dtype: bf16
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
 
 # Logging
-output_dir: /tmp/Qwen2_5-0_5B-Instruct-lora-finetune
+output_dir: /tmp/Qwen2.5-0.5B-Instruct-lora-finetune
 metric_logger:
   _component_: torchtune.training.metric_logging.DiskLogger
-  log_dir: ${output_dir}
+  log_dir: ${output_dir}/logs
 log_every_n_steps: 1
-log_peak_memory_stats: False
+log_peak_memory_stats: True
 
-# Environment
-device: cuda
-dtype: bf16
-enable_activation_checkpointing: True  # True reduces memory
-enable_activation_offloading: False  # True reduces memory
-
-# Show case the usage of pytorch profiler
-# Set enabled to False as it's only needed for debugging training
+# Profiler (disabled)
 profiler:
   _component_: torchtune.training.setup_torch_profiler
-
   enabled: False
 
   #Output directory of trace artifacts
@@ -109,6 +100,6 @@ profiler:
   # `torch.profiler.schedule` options:
   # wait_steps -> wait, warmup_steps -> warmup, active_steps -> active, num_cycles -> repeat
   wait_steps: 5
-  warmup_steps: 5
+  warmup_steps: 3
   active_steps: 2
   num_cycles: 1