Add support to Qwen2-0.5B and Qwen2-1.5B. (#1247)

recipes/configs/qwen2/0.5B_full.yaml

@@ -0,0 +1,75 @@
+# Config for multi-device full finetuning in full_finetune_distributed.py
+# using a Qwen2 0.5B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download Qwen/Qwen2-0.5B-Instruct --output-dir /tmp/Qwen2-0.5B-Instruct --ignore-patterns ""
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 4 full_finetune_distributed --config qwen2/0.5B_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 4 full_finetune_distributed --config qwen2/0.5B_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.yaml for those cases
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen2.qwen2_tokenizer
+  path: /tmp/Qwen2-0.5B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2-0.5B-Instruct/merges.txt
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2.qwen2_0_5b
+
+checkpointer:
+  _component_: torchtune.utils.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2-0.5B-Instruct
+  checkpoint_files: [
+    model.safetensors
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/Qwen2-0.5B-Instruct-finetune
+  model_type: QWEN2
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 2
+epochs: 1
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 5e-6
+loss:
+  _component_: torch.nn.CrossEntropyLoss
+max_steps_per_epoch: null
+gradient_accumulation_steps: 16
+
+
+# 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/Qwen2-0.5B-Instruct-finetune
+log_every_n_steps: 1
+log_peak_memory_stats: False

recipes/configs/qwen2/0.5B_full_single_device.yaml

@@ -0,0 +1,77 @@
+# Config for single device full finetuning in full_finetune_single_device.py
+# using a Qwen2 0.5B
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download Qwen/Qwen2-0.5B-Instruct --output-dir /tmp/Qwen2-0.5B-Instruct --ignore-patterns ""
+#
+# The default config uses an optimizer from bitsandbytes. If you do not have it installed,
+# you can install it with
+#   pip install bitsandbytes
+#
+# To launch on a single device, run the following command from root:
+#   tune run full_finetune_single_device --config qwen2/0.5B_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 qwen2/0.5B_full_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen2.qwen2_tokenizer
+  path: /tmp/Qwen2-0.5B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2-0.5B-Instruct/merges.txt
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+seed: null
+shuffle: True
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2.qwen2_0_5b
+
+checkpointer:
+  _component_: torchtune.utils.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2-0.5B-Instruct
+  checkpoint_files: [
+    model.safetensors
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/Qwen2-0.5B-Instruct-finetune
+  model_type: QWEN2
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 2
+epochs: 1
+optimizer:
+  _component_: bitsandbytes.optim.PagedAdamW
+  lr: 5e-6
+optimizer_in_bwd: True
+loss:
+  _component_: torch.nn.CrossEntropyLoss
+max_steps_per_epoch: null
+gradient_accumulation_steps: 16
+compile: False
+
+# Training environment
+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/Qwen2-0.5B-Instruct-finetune
+log_every_n_steps: 1
+log_peak_memory_stats: False

recipes/configs/qwen2/0.5B_lora.yaml

@@ -0,0 +1,108 @@
+# Config for multi-device LoRA finetuning in lora_finetune_distributed.py
+# using a Qwen2 0.5B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download Qwen/Qwen2-0.5B-Instruct --output-dir /tmp/Qwen2-0.5B-Instruct --ignore-patterns ""
+#
+# To launch on 2 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 2 lora_finetune_distributed --config qwen2/0.5B_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 qwen2/0.5B_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
+# or 0.5B_qlora_single_device.yaml
+
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2.lora_qwen2_0_5b
+  lora_attn_modules: ['q_proj', 'v_proj']
+  apply_lora_to_mlp: False
+  apply_lora_to_output: False
+  lora_rank: 8
+  lora_alpha: 16
+
+tokenizer:
+  _component_: torchtune.models.qwen2.qwen2_tokenizer
+  path: /tmp/Qwen2-0.5B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2-0.5B-Instruct/merges.txt
+
+checkpointer:
+  _component_: torchtune.utils.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2-0.5B-Instruct
+  checkpoint_files: [
+    model.safetensors
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/Qwen2-0.5B-Instruct-lora-finetune
+  model_type: QWEN2
+resume_from_checkpoint: False
+
+# Dataset and Sampler
+dataset:
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+seed: null
+shuffle: True
+batch_size: 2
+
+# Optimizer and Scheduler
+optimizer:
+  _component_: torch.optim.AdamW
+  weight_decay: 0.01
+  lr: 3e-4
+lr_scheduler:
+  _component_: torchtune.modules.get_cosine_schedule_with_warmup
+  num_warmup_steps: 100
+
+loss:
+  _component_: torch.nn.CrossEntropyLoss
+
+# Training
+epochs: 1
+max_steps_per_epoch: null
+gradient_accumulation_steps: 32
+
+# Logging
+output_dir: /tmp/Qwen2-0.5B-Instruct-lora-finetune
+metric_logger:
+  _component_: torchtune.utils.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+log_every_n_steps: 1
+log_peak_memory_stats: False
+
+# Environment
+device: cuda
+dtype: bf16
+enable_activation_checkpointing: False
+
+# Show case the usage of pytorch profiler
+# Set enabled to False as it's only needed for debugging training
+profiler:
+  _component_: torchtune.utils.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

recipes/configs/qwen2/0.5B_lora_single_device.yaml

@@ -0,0 +1,106 @@
+# Config for single device LoRA finetuning in lora_finetune_single_device.py
+# using a Qwen2 0.5B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download Qwen/Qwen2-0.5B-Instruct --output-dir /tmp/Qwen2-0.5B-Instruct --ignore-patterns ""
+#
+# To launch on a single device, run the following command from root:
+#   tune run lora_finetune_single_device --config qwen2/0.5B_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 qwen2/0.5B_lora_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2.lora_qwen2_0_5b
+  lora_attn_modules: ['q_proj', 'v_proj']
+  apply_lora_to_mlp: False
+  apply_lora_to_output: False
+  lora_rank: 8
+  lora_alpha: 16
+
+tokenizer:
+  _component_: torchtune.models.qwen2.qwen2_tokenizer
+  path: /tmp/Qwen2-0.5B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2-0.5B-Instruct/merges.txt
+
+checkpointer:
+  _component_: torchtune.utils.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2-0.5B-Instruct
+  checkpoint_files: [
+    model.safetensors
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/Qwen2-0.5B-Instruct-lora-finetune
+  model_type: QWEN2
+resume_from_checkpoint: False
+
+# Dataset and Sampler
+dataset:
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+seed: null
+shuffle: True
+batch_size: 2
+
+# Optimizer and Scheduler
+optimizer:
+  _component_: torch.optim.AdamW
+  weight_decay: 0.01
+  lr: 3e-4
+lr_scheduler:
+  _component_: torchtune.modules.get_cosine_schedule_with_warmup
+  num_warmup_steps: 100
+
+loss:
+  _component_: torch.nn.CrossEntropyLoss
+
+# Training
+epochs: 1
+max_steps_per_epoch: null
+gradient_accumulation_steps: 64
+compile: False
+
+# Logging
+output_dir: /tmp/Qwen2-0.5B-Instruct-lora-finetune
+metric_logger:
+  _component_: torchtune.utils.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+log_every_n_steps: 1
+log_peak_memory_stats: False
+
+# Environment
+device: cuda
+dtype: bf16
+enable_activation_checkpointing: True
+
+# Show case the usage of pytorch profiler
+# Set enabled to False as it's only needed for debugging training
+profiler:
+  _component_: torchtune.utils.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