add bunch of cleanups and design principle section (#71)

Author: wanchaol

README.md

@@ -4,6 +4,14 @@ Note: This repository is currently under heavy development.
 
 torchtrain contains PyTorch native parallelisms, tools and utilities to train large models.
 
+## Design Principles
+
+TorchTrain is a native PyTorch library with various training techniques. While it utilizes the PyTorch ecosystem for things like data loading (i.e. HuggingFace datasets), the core functionality is written in PyTorch.
+
+* Designed to be easy to understand, use and extend for different training purposes.
+* Minimal changes to the model code, when applying 1D/2D or 3D Parallelisms.
+* Modular components instead of monolithic codebase
+
 # Installation
 
 Install PyTorch from source or install the latest pytorch nightly, then install requirements by

torchtrain/datasets/__init__.py

@@ -2,7 +2,6 @@
 # This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
 
 from torchtrain.datasets.alpaca import build_alpaca_data_loader
-from torchtrain.datasets.pad_batch_sequence import pad_batch_to_longest_seq
 from torchtrain.datasets.tokenizer import create_tokenizer
 
 __all__ = ["build_alpaca_data_loader", "create_tokenizer", "pad_batch_to_longest_seq"]

torchtrain/datasets/pad_batch_sequence.py

@@ -130,12 +130,6 @@ def parallelize_llama(model, world_mesh, parallel_dims, args):
                 "feed_forward.w2": RowwiseParallel(output_layouts=Shard(0)),
                 "feed_forward.w3": ColwiseParallel(),
             }
-            # if layer_id == 0:
-            #     # in first transformer block we need to shard the input
-            #     layer_plan[""] = PrepareModuleInput(
-            #         input_layouts=(Replicate(), None),
-            #         desired_input_layouts=(Shard(0), None),
-            #     )
 
             # adjust num_heads in attention layer to local heads
             attn_layer = transformer_block.attention