Introduction Creating a VS Code extension for Mini-WDL was an exciting journey filled with learning, debugging, and problem-solving. This blog details how I implemented syntax highlighting and language configuration for Mini-WDL, along with the challenges I faced—particularly around marking comments correctly—and how I overcame them. Let’s dive in! 1. Setting Up the Project Step 1: Followed the documentation VsCode To start, I used Yeoman’s generator to create the basic structure of the VS Code extension:

What is WDL? The Workflow Description Language (WDL) is an open-source language designed to simplify complex computational workflows, particularly in genomics and bioinformatics. Developed by the Broad Institute and maintained by the OpenWDL community, WDL allows scientists to define analysis pipelines in a human-readable format. Why It Matters: Standardizes workflow definitions across platforms Enables reproducibility in scientific research Simplifies scaling from laptops to cloud environments WDL Syntax: A Step-by-Step Breakdown Let’s dissect a simple WDL workflow to understand its structure.

1. Expression Node Mapping ASTx BinaryOp → Python ast.BinOp ASTx Structure: BinaryOp( lhs: Expr, rhs: Expr, op: BinaryOpKind ) Python AST Equivalent: ast.BinOp( left: ast.expr, op: operator, right: ast.expr ) Conversion Logic: # Operator mapping table OP_MAP = { BinaryOpKind.add: ast.Add(), BinaryOpKind.sub: ast.Sub(), BinaryOpKind.mul: ast.Mult() } @dispatch def visit(self, node: BinaryOp) -> ast.BinOp: return ast.BinOp( left=self.visit(node.lhs), op=OP_MAP[node.op], right=self.visit(node.rhs) ) 2. Statement Node Mapping ASTx DeleteStmt → Python ast.Delete ASTx Structure:

Before diving into the implementation details, I want to summarize our approach: we’ll be creating JAX/Flax implementations of popular open-source LLM architectures, documenting everything thoroughly, and providing clear notebooks to demonstrate their usage. Project Overview JAX, combined with Flax, provides a powerful framework for implementing high-performance neural networks with benefits like JIT compilation, automatic differentiation, and excellent hardware acceleration support. Our goal is to create clean, well-documented implementations of open-source LLM architectures that can serve as reference material and starting points for further research.

Google Collab provides free GPU/TPU resources perfect for fine-tuning AI models. Here’s a complete guide to fine-tuning models in Colab, from setup to saving your trained model. Setting Up Your Colab Environment # 1. Connect to a GPU runtime # Go to Runtime > Change runtime type > GPU # 2. Verify GPU is available !nvidia-smi # 3. Install necessary libraries !pip install -q transformers datasets accelerate peft bitsandbytes trl tensorboard Method 1: QLoRA Fine-Tuning for Large Models This approach is ideal for 7B+ parameter models on Colab’s limited GPU: