AI / ML

This paper addresses the challenges of deploying Large Language Models on specialized hardware by introducing an MLIR-based compilation pipeline. It utilizes two dialects: TopOp for high-level, framework-agnostic model semantics, and TpuOp for target hardware-specific operations. The method aims to improve the import of trained models into compiler-friendly representations and optimize scheduling for autoregressive inference under memory constraints.

• Introduces TopOp dialect for framework-independent model representation
• Uses TpuOp dialect to map operations to specific AI accelerator hardware
• Addresses memory limitations during autoregressive inference scheduling
• Provides a structured path from trained models to hardware execution

This paper proposes replacing complex multimodal fusion architectures with a single approach: converting all electronic health record components—including free-text narratives, lab values, and vitals—into natural language sequences. A pretrained large language model is then fine-tuned end-to-end for tasks like in-hospital mortality and graft failure prediction without modifying the architecture for specific modalities. The method aims to simplify pipeline design by eliminating the need for dedicated encoders and re-engineering combination mechanisms for each new clinical task.

• Eliminates need for custom multimodal fusion layers in clinical NLP pipelines
• Enables unified training for mortality, graft failure, and other EHR tasks
• Reduces engineering overhead by treating structured data as text tokens
• Challenges prevailing assumption that specialized encoders are required for labs/vitals

This open-source curriculum offers 503 lessons across 20 phases, totaling approximately 320 hours of instruction in Python, TypeScript, Rust, and Julia. It focuses on end-to-end, hands-on construction of AI artifacts like prompts, skills, agents, and MCP servers rather than scattered theoretical concepts. The project aims to bridge the gap between casual AI usage and professional readiness by having learners build functional components from scratch.

• Covers Python, TypeScript, Rust, and Julia for comprehensive AI engineering.
• Each lesson produces a reusable artifact like an agent or MCP server.
• Structured 20-phase curriculum totaling ~320 hours of practical work.
• MIT licensed and free, targeting professional skill development.

An IBM Research blog post on Hugging Face examines why model routing, which appears straightforward in theory, becomes intricate in practice. The article highlights the operational and architectural challenges that emerge when managing dynamic inference traffic across multiple models. It serves as a technical review of the pitfalls involved in scaling routing logic beyond simple use cases.

• Routing logic scales poorly when handling diverse model types and latency requirements.
• Operational complexity increases significantly with dynamic traffic distribution.
• Architectural decisions must account for edge cases in real-world inference loads.
• Simple theoretical models often fail under production-grade traffic patterns.