Tag
This post explains MLIR as a compiler infrastructure framework that provides a flexible dialect stack, describing how it is used across modern ML compilers like XLA, Triton, and Mojo to progressively lower tensor operations to machine code.
This paper introduces benchmarks for natural language to MLIR code generation across multiple dialects and a schema-derived constrained decoding stack that allows a small language model to match or exceed large code LMs on structural verifier tasks, without retraining.
This paper presents an MLIR-based compilation method for large language models, using two custom dialects (TopOp and TpuOp) to lower models from framework-agnostic semantics to hardware-specific instructions. It also introduces a three-stage static compilation for autoregressive inference stages: prefill, prefill_kv, and decode.