Add Phi4

recipes/configs/phi3/evaluation.yaml

@@ -12,7 +12,7 @@ model:
 # Checkpointer
 checkpointer:
   _component_: torchtune.training.FullModelHFCheckpointer
-  checkpoint_dir: /tmp/Phi-3-mini-4k-instruct
+  checkpoint_dir: /tmp/phi-3
   checkpoint_files: [
     model-00001-of-00002.safetensors,
     model-00002-of-00002.safetensors
@@ -25,7 +25,7 @@ resume_from_checkpoint: False
 # Tokenizer
 tokenizer:
   _component_: torchtune.models.phi3.phi3_mini_tokenizer
-  path: /tmp/Phi-3-mini-4k-instruct/tokenizer.model
+  path: /tmp/phi-3/tokenizer.model
   max_seq_len: null
 
 # Environment

recipes/configs/phi4/evaluation.yaml

@@ -0,0 +1,45 @@
+# Config for EleutherEvalRecipe in eleuther_eval.py
+#
+# To launch, run the following command:
+#    tune run eleuther_eval --config phi4/evaluation
+
+output_dir: ./ # Not needed
+
+# Model Arguments
+model:
+  _component_: torchtune.models.phi4.phi4_14b
+
+# Checkpointer
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/phi-4
+  checkpoint_files: [
+    model-00001-of-00002.safetensors,
+    model-00002-of-00002.safetensors
+  ]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: PHI3_MINI
+resume_from_checkpoint: False
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.phi4.phi4_14b_tokenizer
+  vocab_path: /tmp/phi-4/vocab.json
+  merges_path: /tmp/phi-4/merges.txt
+  max_seq_len: null
+
+# Environment
+device: cuda
+dtype: bf16
+seed: 1234 # It is not recommended to change this seed, b/c it matches EleutherAI's default seed
+
+# EleutherAI specific eval args
+tasks: ["truthfulqa_mc2"]
+limit: null
+max_seq_length: 4096
+batch_size: 8
+enable_kv_cache: True
+
+# Quantization specific args
+quantizer: null

recipes/configs/phi4/full.yaml

@@ -0,0 +1,110 @@
+# Config for multi-device full finetuning in full_finetune_distributed.py
+# using a Phi4 16K Instruct
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download microsoft/phi-4 --output-dir /tmp/phi-4 --hf-token <HF_TOKEN>
+#
+# Run this config on 4 GPUs using the following:
+#  tune run --nproc_per_node 4 full_finetune_distributed --config phi4/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 --nproc_per_node 4 full_finetune_distributed --config phi4/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 mini_low_memory.yaml for those cases
+
+output_dir: /tmp/torchtune/phi-4/full # /tmp may be deleted by your system. Change it to your preference.
+
+# Model arguments
+model:
+  _component_: torchtune.models.phi4.phi4_14b
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.phi4.phi4_14b_tokenizer
+  vocab_path: /tmp/phi-4/vocab.json
+  merges_path: /tmp/phi-4/merges.txt
+  max_seq_len: null
+
+# Checkpointer
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/phi-4
+  checkpoint_files: [
+    model-00001-of-00006.safetensors,
+    model-00002-of-00006.safetensors,
+    model-00003-of-00006.safetensors,
+    model-00004-of-00006.safetensors,
+    model-00005-of-00006.safetensors,
+    model-00006-of-00006.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: PHI3_MINI
+resume_from_checkpoint: False
+
+# Dataset
+dataset:
+  _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: 8  # Use to increase effective batch size
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  lr: 5e-6
+loss:
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
+compile: False  # torch.compile the model + loss, True increases speed + decreases 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
+enable_activation_offloading: False  # True reduces memory
+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
+
+
+# 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/phi4/full_low_memory.yaml

@@ -0,0 +1,111 @@
+# Config for single device full finetuning in full_finetune_single_device.py
+# using a Phi4 16K Instruct
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download microsoft/phi-4 --output-dir /tmp/phi-4 --hf-token <HF_TOKEN>
+#
+# 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 phi4/full_low_memory
+#
+# 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 phi4/full_low_memory checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+output_dir: /tmp/torchtune/phi-4/full_low_memory # /tmp may be deleted by your system. Change it to your preference.
+
+# Model arguments
+model:
+  _component_: torchtune.models.phi4.phi4_14b
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.phi4.phi4_14b_tokenizer
+  vocab_path: /tmp/phi-4/vocab.json
+  merges_path: /tmp/phi-4/merges.txt
+  max_seq_len: null
+
+# Checkpointer
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/phi-4
+  checkpoint_files: [
+    model-00001-of-00006.safetensors,
+    model-00002-of-00006.safetensors,
+    model-00003-of-00006.safetensors,
+    model-00004-of-00006.safetensors,
+    model-00005-of-00006.safetensors,
+    model-00006-of-00006.safetensors,
+  ]
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: PHI3_MINI
+resume_from_checkpoint: False
+
+# Dataset
+dataset:
+  _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: 1  # Use to increase effective batch size
+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.CEWithChunkedOutputLoss
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
+
+# Training env
+device: cuda
+
+# Memory management
+enable_activation_checkpointing: True  # True reduces memory
+enable_activation_offloading: True  # True reduces memory
+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
+
+
+# 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