[WIP] Federated learning

pyproject.toml

@@ -62,6 +62,13 @@ dev = [
     "wandb",
     "expecttest",
 ]
+federated = [
+    "fastapi",
+    "uvicorn",
+    "aiofiles",
+    "requests",
+    "python-multipart"
+]
 
 [tool.setuptools.dynamic]
 version = {attr = "torchtune.__version__"}

recipes/configs/federation_diloco.yaml

@@ -0,0 +1,37 @@
+# Config for FederationRecipe in federation.py
+#
+# To launch, run the following command from root torchtune directory:
+#    tune run federation --config federation_dilico.yaml
+
+output_dir: /tmp/torchtune/qwen2_5_0_5B/federated # /tmp may be deleted by your system. Change it to your preference.
+
+#
+# Model arguments
+model:
+  _component_: torchtune.models.qwen2_5.qwen2_5_0_5b
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /Users/maxime/.cache/huggingface/hub/models--Qwen--Qwen2.5-0.5B-Instruct/snapshots/7ae557604adf67be50417f59c2c2f167def9a775
+  checkpoint_files: [model.safetensors]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN2
+resume_from_checkpoint: False
+
+device: mps
+dtype: bf16
+seed: 1234
+
+federator:
+  _component_: torchtune.training.federation.DiLoCoFederator
+  h: 50
+  optimizer:
+    _component_: torch.optim.SGD
+    lr: 1
+    nesterov: True
+    momentum: 0.9
+
+  participants:
+    - uioGMzl7zqtkydxP66mJh9dZCaNnSWOb
+    - etbefHUtU0XViIPrqo8nN2TlpvIUTdi7

recipes/configs/qwen2_5/0.5B_full_single_device_diloco_1.yaml

@@ -0,0 +1,114 @@
+# 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
+#
+#
+# To launch on a single device, run the following command from root:
+#   tune run full_finetune_single_device --config qwen2/1.5B_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 qwen2/1.5B_full_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+output_dir: /tmp/torchtune/qwen2_5_0_5B/full_single_device # /tmp may be deleted by your system. Change it to your preference.
+
+# Model arguments
+model:
+  _component_: torchtune.models.qwen2_5.qwen2_5_0_5b
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen2_5.qwen2_5_tokenizer
+  path: /Users/maxime/.cache/huggingface/hub/models--Qwen--Qwen2.5-0.5B-Instruct/snapshots/7ae557604adf67be50417f59c2c2f167def9a775/vocab.json
+  merges_file: /Users/maxime/.cache/huggingface/hub/models--Qwen--Qwen2.5-0.5B-Instruct/snapshots/7ae557604adf67be50417f59c2c2f167def9a775/merges.txt
+  max_seq_len: 4096
+
+# Checkpointer
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /Users/maxime/.cache/huggingface/hub/models--Qwen--Qwen2.5-0.5B-Instruct/snapshots/7ae557604adf67be50417f59c2c2f167def9a775/
+  checkpoint_files: [model.safetensors]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN2
+resume_from_checkpoint: False
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+  packed: False  # True increases speed
+
+seed: 4321
+shuffle: True
+
+# Fine-tuning arguments
+batch_size: 4
+epochs: 1
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 2e-5
+
+optimizer_in_bwd: True  # True saves memory. Requires gradient_accumulation_steps=1
+
+loss:
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
+
+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: mps
+
+# 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: True
+
+# Federation
+federation:
+  _component_: torchtune.training.federation.TuneParticipant
+  endpoint: http://127.0.0.1:8000
+  token: uioGMzl7zqtkydxP66mJh9dZCaNnSWOb
+  temporary_dir: ${output_dir}/federation
+
+# 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: False
+
+  #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_5/0.5B_full_single_device_diloco_2.yaml

@@ -0,0 +1,113 @@
+# 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
+#
+#
+# To launch on a single device, run the following command from root:
+#   tune run full_finetune_single_device --config qwen2/1.5B_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 qwen2/1.5B_full_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+output_dir: /tmp/torchtune/qwen2_5_0_5B/full_single_device # /tmp may be deleted by your system. Change it to your preference.
+
+# Model arguments
+model:
+  _component_: torchtune.models.qwen2_5.qwen2_5_0_5b
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen2_5.qwen2_5_tokenizer
+  path: /Users/maxime/.cache/huggingface/hub/models--Qwen--Qwen2.5-0.5B-Instruct/snapshots/7ae557604adf67be50417f59c2c2f167def9a775/vocab.json
+  merges_file: /Users/maxime/.cache/huggingface/hub/models--Qwen--Qwen2.5-0.5B-Instruct/snapshots/7ae557604adf67be50417f59c2c2f167def9a775/merges.txt
+  max_seq_len: 4096
+
+# Checkpointer
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /Users/maxime/.cache/huggingface/hub/models--Qwen--Qwen2.5-0.5B-Instruct/snapshots/7ae557604adf67be50417f59c2c2f167def9a775/
+  checkpoint_files: [model.safetensors]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN2
+resume_from_checkpoint: False
+
+# Dataset
+dataset:
+  _component_: torchtune.datasets.cnn_dailymail_articles_dataset
+
+seed: 1234
+shuffle: True
+
+# Fine-tuning arguments
+batch_size: 1
+epochs: 1
+optimizer:
+  _component_: torch.optim.AdamW
+  lr: 2e-5
+
+optimizer_in_bwd: True  # True saves memory. Requires gradient_accumulation_steps=1
+
+loss:
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
+
+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: True
+
+# Federation
+federation:
+  _component_: torchtune.training.federation.TuneParticipant
+  endpoint: http://127.0.0.1:8000
+  token: etbefHUtU0XViIPrqo8nN2TlpvIUTdi7
+  temporary_dir: ${output_dir}/federation
+
+# 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: False
+
+  #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