Candidate Profile — Candidate #3
Position: Mythic — Compiler Engineer – MLIR / PyTorch Infrastructure (Remote)
Candidate Location: Not specified
Experience: 15.9 years experience
Relevant Experience & Education Highlights
Demonstrates deep expertise in MLIR dialect design, PyTorch-to-MLIR integration via Torch-MLIR, and compiler stacks for heterogeneous accelerators, directly aligning with advancing Mythic's MLIR ecosystem and hardware-aware dialects.
1. Direct, hands-on experience with MLIR, including dialect design, compiler passes, or lowering pipelines
Led development of MLIR-based compilation stacks to simplify model deployment on diverse accelerators.
- Built MLIR-enabled compilation stack targeting commodity and in-house ML accelerators at an autonomous vehicle technology company.
- Served as tech lead for ML compiler frontend, enabling seamless connectivity from PyTorch models to MLIR ecosystem and middle-end optimizations.
- Co-created and maintained Tensor Compute Primitives (mlir-tcp), an open-source mid-level MLIR dialect for ML programs.
- Co-maintained Torch-MLIR, providing first-class compiler support from PyTorch to MLIR.
2. Familiarity with PyTorch compiler technologies, especially Torch-MLIR and integration paths with PyTorch 2.0
Architected tools bridging PyTorch workflows to compiler infrastructures for efficient inference.
- Created GraFX, a source-to-source compiler framework accelerating PyTorch model development, optimization, and deployment lifecycles.
- Integrated state-of-the-art PyTorch models with MLIR for heterogeneous compute platforms in latency-critical applications.
- Developed connectivity from PyTorch ecosystem to MLIR, supporting optimizations for autonomous vehicle inference.
3. Background in heterogeneous or specialized accelerators (e.g., analog compute, NPUs, GPUs, DSPs)
Advanced algorithms and architectures for ML acceleration across FPGAs, AI Engines, and custom accelerators.
- Developed TQT, a quantization technique using trained thresholds for accurate fixed-point inference on hardware-constrained schemes at a leading adaptive computing company.
- Explored novel architectures and algorithms for efficient ML workloads on FPGAs and AI Engines for cloud and edge use cases.
- Optimized deep learning deployments onto variety of commodity and in-house ML accelerators.
Requirements & Candidate Alignment
| Mythic Requirement | Candidate Qualification |
|---|---|
| 3+ Years of experience in compiler or high-performance systems development. | 15.9 years total experience, including multiple staff-level roles in ML compilers and high-performance systems. |
| Proficiency in modern C++ (C++14/17/20) and Python. | Python proficiency via Jupyter Notebook, PyTorch workflows, and compiler frontend development. |
| Direct, hands-on experience with MLIR, including dialect design, compiler passes, or lowering pipelines. | Extensive MLIR hands-on work building compilation stacks, leading frontends, co-creating mlir-tcp dialect, and maintaining Torch-MLIR. |
| Strong understanding of compiler IRs and transformations, with the ability to reason about lowering from high-level ops to hardware-aware representations. | Proven compiler IR expertise through PyTorch-to-MLIR lowering pipelines and optimizations for heterogeneous accelerators. |
| Experience architecting complete MLIR flows: from frontend dialects down to hardware-aware dialects, including conversion to and from existing IRs. | Architected full MLIR flows from PyTorch frontends to middle-end and accelerator targets. |
| Familiarity with PyTorch compiler technologies, especially Torch-MLIR and integration paths with PyTorch 2.0 (TorchDynamo, TorchInductor). | Co-maintainer of Torch-MLIR with direct PyTorch integration for MLIR ecosystems. |
| Background in heterogeneous or specialized accelerators (e.g., analog compute, NPUs, GPUs, DSPs). | Hands-on with FPGAs, AI Engines, and custom ML accelerators for edge and cloud inference. |
If you would like to discuss this candidate or other critical roles, here is a link to my calendar to schedule a call:
Schedule a CallContact:
Finding the signal in the noise since 2005