Add support for QAT + LoRA

TODO: write this

eval_it.sh

@@ -0,0 +1,37 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+LOG_DIR="${LOG_DIR:-/home/andrewor/local/logs/tune/qat_lora}"
+CHECKPOINT_FILE="${CHECKPOINT_FILE:-meta_model_0.pt}"
+GROUP_SIZE="${GROUP_SIZE:=32}"
+export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-1}"
+QUANTIZED_CHECKPOINT_FILE="$(echo "$CHECKPOINT_FILE" | sed 's/\.pt/-8da4w.pt/g')"
+
+tune run quantize --config quantization \
+    model._component_=torchtune.models.llama3.llama3_8b \
+    checkpointer._component_=torchtune.training.FullModelMetaCheckpointer \
+    checkpointer.checkpoint_dir="$LOG_DIR" \
+    checkpointer.output_dir="$LOG_DIR" \
+    checkpointer.checkpoint_files=["$CHECKPOINT_FILE"] \
+    checkpointer.model_type=LLAMA3 \
+    quantizer._component_=torchtune.training.quantization.Int8DynActInt4WeightQuantizer \
+    quantizer.groupsize="$GROUP_SIZE" \
+    > "$LOG_DIR"/quantize.log 2>&1
+
+tune run eleuther_eval --config eleuther_evaluation \
+    batch_size=1 \
+    model._component_=torchtune.models.llama3.llama3_8b \
+    checkpointer._component_=torchtune.training.FullModelTorchTuneCheckpointer \
+    checkpointer.checkpoint_dir="$LOG_DIR" \
+    checkpointer.output_dir="$LOG_DIR" \
+    checkpointer.checkpoint_files=["$QUANTIZED_CHECKPOINT_FILE"] \
+    checkpointer.model_type=LLAMA3 \
+    tokenizer._component_=torchtune.models.llama3.llama3_tokenizer \
+    tokenizer.path=/tmp/Meta-Llama-3-8B-Instruct/original/tokenizer.model \
+    tasks=[wikitext] \
+    quantizer._component_=torchtune.training.quantization.Int8DynActInt4WeightQuantizer \
+    quantizer.groupsize="$GROUP_SIZE" \
+    > "$LOG_DIR"/eval.log 2>&1

recipes/configs/llama2/7B_qat_lora.yaml

@@ -0,0 +1,116 @@
+# Config for multi-device QAT + LoRA finetuning in qat_lora_finetune_distributed.py
+# using a Llama2 7B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download meta-llama/Llama-2-7b-hf --output-dir /tmp/Llama-2-7b-hf --hf-token <HF_TOKEN>
+#
+# To launch on 2 devices, run the following command from root:
+#   tune run --nnodes 1 --nproc_per_node 2 qat_lora_finetune_distributed --config llama2/7B_qat_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 qat_lora_finetune_distributed --config llama2/7B_qat_lora checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+
+
+# Model Arguments
+model:
+  _component_: torchtune.models.llama2.lora_llama2_7b
+  lora_attn_modules: ['q_proj', 'v_proj']
+  apply_lora_to_mlp: False
+  apply_lora_to_output: False
+  lora_rank: 8
+  lora_alpha: 16
+  lora_dropout: 0.0
+
+tokenizer:
+  _component_: torchtune.models.llama2.llama2_tokenizer
+  path: /tmp/Llama-2-7b-hf/tokenizer.model
+  max_seq_len: null
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Llama-2-7b-hf
+  checkpoint_files: [
+    pytorch_model-00001-of-00002.bin,
+    pytorch_model-00002-of-00002.bin
+  ]
+  adapter_checkpoint: null
+  recipe_checkpoint: null
+  output_dir: /tmp/Llama-2-7b-hf
+  model_type: LLAMA2
+resume_from_checkpoint: False
+save_adapter_weights_only: False
+
+# Dataset and Sampler
+dataset:
+  packed: False # Set to true for great speed ups
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+seed: null
+shuffle: True
+batch_size: 2
+
+# Optimizer and Scheduler
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  weight_decay: 0.01
+  lr: 3e-4
+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: 32
+
+# Logging
+output_dir: /tmp/lora_finetune_output
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+log_every_n_steps: 1
+log_peak_memory_stats: True
+
+# Environment
+device: cuda
+dtype: bf16
+enable_activation_checkpointing: False
+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:
+  _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: 5
+  active_steps: 2
+  num_cycles: 1
+
+# QAT arguments
+quantizer:
+  _component_: torchtune.training.quantization.Int8DynActInt4WeightQATQuantizer
+  groupsize: 256

recipes/configs/llama3/8B_qat_lora.yaml

@@ -0,0 +1,89 @@
+# Config for multi-device QAT + LoRA finetuning in qat_lora_finetune_distributed.py
+# using a Llama3 8B Instruct model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download meta-llama/Meta-Llama-3-8B-Instruct --output-dir /tmp/Meta-Llama-3-8B-Instruct --hf-token <HF_TOKEN>
+#
+# To launch on 2 devices, run the following command from root:
+#   tune run --nproc_per_node 2 qat_lora_finetune_distributed --config llama3/8B_qat_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 --nproc_per_node 2 qat_lora_finetune_distributed --config llama3/8B_qat_lora checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.llama3.llama3_tokenizer
+  path: /tmp/Meta-Llama-3-8B-Instruct/original/tokenizer.model
+  max_seq_len: null
+
+# Model Arguments
+model:
+  _component_: torchtune.models.llama3.lora_llama3_8b
+  lora_attn_modules: ['q_proj', 'v_proj']
+  apply_lora_to_mlp: False
+  apply_lora_to_output: False
+  lora_rank: 8
+  lora_alpha: 16
+  lora_dropout: 0.0
+
+checkpointer:
+  _component_: torchtune.training.FullModelMetaCheckpointer
+  checkpoint_dir: /tmp/Meta-Llama-3-8B-Instruct/original/
+  checkpoint_files: [
+    consolidated.00.pth
+  ]
+  recipe_checkpoint: null
+  output_dir: /tmp/Meta-Llama-3-8B-Instruct/
+  model_type: LLAMA3
+resume_from_checkpoint: False
+save_adapter_weights_only: False
+
+# Dataset and Sampler
+dataset:
+  packed: False # Set to true for great speed ups
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+seed: null
+shuffle: True
+batch_size: 2
+
+# Optimizer and Scheduler
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  weight_decay: 0.01
+  lr: 3e-4
+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: 32
+compile: False
+
+# Logging
+output_dir: /tmp/lora_finetune_output
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}
+log_every_n_steps: 1
+log_peak_memory_stats: True
+
+# Environment
+device: cuda
+dtype: bf16
+enable_activation_checkpointing: False
+enable_activation_offloading: False  # True reduces memory
+
+# QAT arguments
+quantizer:
+  _component_: torchtune.training.quantization.Int8DynActInt4WeightQATQuantizer
+  groupsize: 256