Qwen 2.5 VL

recipes/configs/qwen2_5_vision/32B_full.yaml

@@ -0,0 +1,109 @@
+# Config for single device full finetuning in full_finetune_distributed.py
+# using a Qwen2.5 32B
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download Qwen/Qwen2.5-32B-Instruct --output-dir /tmp/Qwen2.5-32B-Instruct
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 4 full_finetune_distributed --config qwen2_5/32B_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_5/32B_full checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config was only tested on a 4xH100 machine.
+
+output_dir: /tmp/torchtune/qwen2_5_32B/full # /tmp may be deleted by your system. Change it to your preference.
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen2_5_vision.Qwen2_5_VLTransform
+  path: /tmp/Qwen2.5-VL-3B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2.5-VL-3B-Instruct/merges.txt
+  max_seq_len: null
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.multimodal.the_cauldron_dataset
+  packed: False  # True increases speed
+  subset: ocrvqa
+seed: null
+shuffle: True
+collate_fn: torchtune.models.qwen2_5_vision.qwen2_5_vl_padded_collate_images
+
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2_5_vision.qwen2_5_vl_32b
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2.5-VL-32B-Instruct
+  checkpoint_files:
+    filename_format: model-{}-of-{}.safetensors
+    max_filename: "00018"
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN2_5_VL
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 2
+epochs: 1
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 5e-6
+optimizer_in_bwd: True  # True saves memory. Requires gradient_accumulation_steps=1
+loss:
+  _component_: torchtune.modules.loss.LinearCrossEntropyLoss
+max_steps_per_epoch: 100
+gradient_accumulation_steps: 1  # Use to increase effective batch size
+clip_grad_norm: null
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
+
+# Training environment
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True  # True reduces memory
+enable_activation_offloading: False  # True reduces memory
+custom_sharded_layers: ['decoder.tok_embeddings', 'decoder.output']
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}/logs
+log_every_n_steps: 1
+log_peak_memory_stats: False
+log_level: INFO  # DEBUG, WARN, etc.
+
+
+# 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/qwen2_5_vision/3B_full_single_device.yaml

@@ -0,0 +1,115 @@
+# Config for single device full finetuning in full_finetune_single_device.py
+# using a Qwen2.5 VL 3B
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download Qwen/Qwen2.5-VL-3B-Instruct --output-dir /tmp/Qwen2.5-VL-3B-Instruct
+#
+# 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_5_vision/3B_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_vision/3B_full_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+output_dir: /tmp/torchtune/qwen2_5_3B_VL/full_single_device # /tmp may be deleted by your system. Change it to your preference.
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen2_5_vision.Qwen2_5_VLTransform
+  path: /tmp/Qwen2.5-VL-3B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2.5-VL-3B-Instruct/merges.txt
+  max_seq_len: null
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.multimodal.the_cauldron_dataset
+  packed: False  # True increases speed
+  subset: ocrvqa
+seed: null
+shuffle: True
+collate_fn: torchtune.models.qwen2_5_vision.qwen2_5_vl_padded_collate_images
+
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2_5_vision.qwen2_5_vl_3b
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2.5-VL-3B-Instruct
+  checkpoint_files: [
+    model-00001-of-00004.safetensors,
+    model-00002-of-00004.safetensors,
+    model-00003-of-00004.safetensors,
+    model-00004-of-00004.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN2_5_VL
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 1
+epochs: 1
+optimizer:
+  _component_: bitsandbytes.optim.PagedAdamW
+  lr: 5e-6
+optimizer_in_bwd: True  # True saves memory. Requires gradient_accumulation_steps=1
+loss:
+  _component_: torchtune.modules.loss.LinearCrossEntropyLoss
+max_steps_per_epoch: null
+gradient_accumulation_steps: 1  # Use to increase effective batch size
+clip_grad_norm: null
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
+
+# Training environment
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True  # True reduces memory
+enable_activation_offloading: False  # True reduces memory
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}/logs
+log_every_n_steps: 1
+log_peak_memory_stats: False
+log_level: INFO  # DEBUG, WARN, etc.
+
+
+# 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/qwen2_5_vision/72B_full.yaml

@@ -0,0 +1,109 @@
+# Config for single device full finetuning in full_finetune_distributed.py
+# using a Qwen2.5 72B
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download Qwen/Qwen2.5-72B-Instruct --output-dir /tmp/Qwen2.5-72B-Instruct
+#
+# To launch on 4 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 8 full_finetune_distributed --config qwen2_5/72B_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 8 full_finetune_distributed --config qwen2_5/72B_full checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config was only tested on a 8xH100 machine.
+
+output_dir: /tmp/torchtune/qwen2_5_72B/full # /tmp may be deleted by your system. Change it to your preference.
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen2_5_vision.Qwen2_5_VLTransform
+  path: /tmp/Qwen2.5-VL-3B-Instruct/vocab.json
+  merges_file: /tmp/Qwen2.5-VL-3B-Instruct/merges.txt
+  max_seq_len: null
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.multimodal.the_cauldron_dataset
+  packed: False  # True increases speed
+  subset: ocrvqa
+seed: null
+shuffle: True
+collate_fn: torchtune.models.qwen2_5_vision.qwen2_5_vl_padded_collate_images
+
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2_5_vision.qwen2_5_vl_72b
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2.5-VL-72B-Instruct
+  checkpoint_files:
+    filename_format: model-{}-of-{}.safetensors
+    max_filename: "00018"
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN2_5_VL
+resume_from_checkpoint: False
+
+# Fine-tuning arguments
+batch_size: 2
+epochs: 1
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 5e-6
+optimizer_in_bwd: True  # True saves memory. Requires gradient_accumulation_steps=1
+loss:
+  _component_: torchtune.modules.loss.LinearCrossEntropyLoss
+max_steps_per_epoch: 100
+gradient_accumulation_steps: 1  # Use to increase effective batch size
+clip_grad_norm: null
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
+
+# Training environment
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True  # True reduces memory
+enable_activation_offloading: False  # True reduces memory
+custom_sharded_layers: ['decoder.tok_embeddings', 'decoder.output']
+
+# Reduced precision
+dtype: bf16
+
+# Logging
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}/logs
+log_every_n_steps: 1
+log_peak_memory_stats: False
+log_level: INFO  # DEBUG, WARN, etc.
+
+
+# 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