chore: wip torchtune, dataset merger

merger.py

@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+import json
+import argparse
+import random
+from collections.abc import Sequence
+from pathlib import Path
+from typing import TypedDict
+
+
+class AlpacaEntry(TypedDict):
+    input: str
+    output: str
+    instruction: str | None
+
+
+class Args(argparse.Namespace):
+    def __init__(self) -> None:
+        super().__init__()
+        self.file1: str = ""
+        self.file2: str = ""
+        self.output: str = ""
+        self.shuffle: bool = False
+        self.seed: int | None = None
+        self.omit_instruction: bool = False
+
+
+def load_file(file_path: str) -> Sequence[AlpacaEntry]:
+    """Load and validate a file in JSON or JSONL format."""
+    path = Path(file_path)
+    is_jsonl = path.suffix.lower() in (".jsonl", ".ljson")
+
+    with open(file_path, "r", encoding="utf-8") as f:
+        if is_jsonl:
+            data: list[AlpacaEntry] = []
+            for line in f:
+                line = line.strip()
+                if line:  # Skip empty lines
+                    entry: AlpacaEntry = json.loads(line)
+                    data.append(entry)
+        else:
+            data = json.load(f)
+
+    # Validate required fields
+    for item in data:
+        if not all(key in item for key in ["input", "output"]):
+            raise ValueError(
+                f"Missing required fields in {file_path}. Each item must have 'input' and 'output' fields."
+            )
+
+    return data
+
+
+def write_output(data: Sequence[AlpacaEntry], output_path: str) -> None:
+    """Write output in JSON or JSONL format based on file extension."""
+    path = Path(output_path)
+    is_jsonl = path.suffix.lower() in (".jsonl", ".ljson")
+    print(path, is_jsonl)
+
+    with open(output_path, "w", encoding="utf-8") as f:
+        if is_jsonl:
+            for item in data:
+                _ = f.write(json.dumps(item, ensure_ascii=False) + "\n")
+        else:
+            json.dump(data, f, ensure_ascii=False, indent=2)
+
+
+def merge_datasets(
+    file1: str, file2: str, shuffle: bool = False, omit_instruction: bool = False
+) -> Sequence[AlpacaEntry]:
+    """Merge two files in JSON/JSONL format."""
+    # Load both files
+    data1 = load_file(file1)
+    data2 = load_file(file2)
+    merged_data = list(data1) + list(data2)
+
+    if omit_instruction:
+        for item in merged_data:
+            _ = item.pop("instruction", None)
+
+    if shuffle:
+        random.shuffle(merged_data)
+
+    return merged_data
+
+
+def parse_args() -> Args:
+    """Parse and validate command line arguments."""
+    parser = argparse.ArgumentParser(description="Merge two Alpaca-format JSON files")
+    _ = parser.add_argument("file1", type=str, help="Path to first JSON file")
+    _ = parser.add_argument("file2", type=str, help="Path to second JSON file")
+    _ = parser.add_argument("output", type=str, help="Path to output merged JSON file")
+    _ = parser.add_argument(
+        "--shuffle", action="store_true", help="Shuffle the merged dataset"
+    )
+    _ = parser.add_argument("--seed", type=int, help="Random seed for shuffling")
+    _ = parser.add_argument(
+        "--omit-instruction",
+        action="store_true",
+        help="Omit instruction field from output",
+    )
+
+    return parser.parse_args(namespace=Args())
+
+
+def main() -> None:
+    args = parse_args()
+
+    try:
+        if args.seed is not None:
+            random.seed(args.seed)
+
+        merged_data = merge_datasets(
+            args.file1, args.file2, args.shuffle, args.omit_instruction
+        )
+
+        write_output(merged_data, args.output)
+        print(f"Successfully merged files. Total entries: {len(merged_data)}")
+
+    except Exception as e:
+        print(f"Error: {str(e)}")
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()

train_torchtune.yaml

@@ -0,0 +1,120 @@
+# Config for single device LoRA finetuning in lora_finetune_single_device.py
+# using a Qwen2.5 7B model
+#
+# This config assumes that you've run the following command before launching
+# this run:
+#   tune download Qwen/Qwen2.5-7B-Instruct --output-dir /tmp/Qwen2.5-7B-Instruct
+#
+# To launch on a single device, run the following command from root:
+#   tune run lora_finetune_single_device --config qwen2_5/7B_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_5/7B_lora_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#
+# This config works only for training on single device.
+
+output_dir: /home/mira/models/qwen2_5_7B_tune/lora_single_device # /tmp may be deleted by your system. Change it to your preference.
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2_5.lora_qwen2_5_7b_base
+  lora_attn_modules: ["q_proj", "k_proj", "v_proj", "output_proj"]
+  apply_lora_to_mlp: True
+  apply_lora_to_output: True
+  lora_rank: 32 # higher increases accuracy and memory
+  lora_alpha: 64 # usually alpha=2*rank
+  lora_dropout: 0.05
+  quantize_base: True
+
+tokenizer:
+  _component_: torchtune.models.qwen2_5.qwen2_5_tokenizer
+  path: /home/mira/models/Qwen2.5-7B-Base/vocab.json
+  merges_file: /home/mira/models/Qwen2.5-7B-Base/merges.txt
+  max_seq_len: 16384
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /home/mira/models/Qwen2.5-7B-Base
+  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
+save_every_n_steps: 100
+resume_from_checkpoint: False
+
+# Dataset and Sampler
+dataset:
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+  packed: True # True increases speed
+seed: 42
+shuffle: False
+batch_size: 1
+
+# Optimizer and Scheduler
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  weight_decay: 0.01
+  lr: 1e-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 # Use to increase effective batch size
+clip_grad_norm: null
+compile: True # torch.compile the model + loss, True increases speed + decreases memory
+
+# Logging
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}/logs
+log_every_n_steps: 5
+log_peak_memory_stats: True
+
+# Environment
+device: cuda
+dtype: fp16
+
+# Activations Offloading
+enable_activation_checkpointing: True # True reduces memory
+enable_activation_offloading: True # 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