Skip to main content
TALENT247 // SPECIALIST AI EXECUTIVE SEARCH & TECHNICAL RECRUITMENT
INF · AI Infrastructure

Senior ML Compiler Engineer, Long-Context Inference

Seattle, WA · Full-time · Hybrid

AI Specialism

AI Infrastructure

Workplace

Hybrid

Location

Seattle, WA

Compensation

$245k – $340k

annual package

Role Dossier & Specification

REF: SENIOR-ML-COMPILER

THE CLIENT & MANDATE

Our client is a frontier AI research and deployment company operating a multi-accelerator compute fleet across NVIDIA GPUs and custom cloud AI ASICs/TPUs, with a dedicated compiler and performance engineering hub in Seattle, WA. Serving flagship models with 200k–1M+ token context windows across multi-hour coding and research sessions places extreme pressure on compiler graph lowering, HBM memory scheduling, and cross-chip collective fusion.

They have retained Talent247 to hire a Senior ML Compiler Engineer for Long-Context Inference in Seattle, WA. The successful candidate will build multi-target XLA, MLIR, and custom kernel-generation passes that maximize throughput and slash time-to-first-token across heterogeneous frontier accelerator fleets.

KEY RESPONSIBILITIES

  • Design and implement XLA, MLIR, and StableHLO compiler passes optimizing ultra-long-context (200k–1M+ token) prefill and decode graphs across GPUs and custom systolic-array AI ASICs/TPUs.
  • Architect compiler-driven compute-communication overlap transformations (async collective matmuls, Ring/Ulysses context-parallel attention fusion, and pipelined all-to-all MoE dispatch).
  • Optimize HBM-to-SRAM/shared-memory tiling, KV-cache quantization (FP8/INT4), prefix-tree paging layouts, and rematerialization schedules to fit massive batch sizes within tight memory envelopes.
  • Author custom Pallas, Triton, and CUDA/C++ escape-hatch kernels where general compiler lowering leaves roof-line performance on the table, then upstream clean lowering patterns back into the compiler.
  • Build automated cross-backend numerical parity checkers and compiler-regression performance gates across dozens of active production model variants.
  • Partner with model architecture scientists on hardware-co-designed attention mechanisms (MLA, GQA, hybrid sliding-window/global attention) optimized for multi-accelerator compilation.

TECHNICAL ENVIRONMENT & OFFERING

Stack: C++20, MLIR, OpenXLA / StableHLO, LLVM, Python, JAX, PyTorch/TorchXLA, Pallas, Triton, CUDA, multi-pod GPU & cloud TPU/ASIC clusters. Workplace: Hybrid in Seattle, WA. Compensation: $245,000 – $340,000 base salary plus substantial frontier-lab equity, comprehensive healthcare, and 401(k).

Core Requirements

ESSENTIAL QUALIFICATIONS

  • 5+ years of ML compiler, GPU/ASIC toolchain, or high-performance deep learning runtime engineering experience in modern C++ and Python.
  • Deep internals knowledge of OpenXLA (HLO passes, layout assignment, memory space scheduling), MLIR, TVM, or TorchInductor.
  • Strong command of accelerator hardware bottlenecks (systolic arrays, tensor cores, HBM bandwidth, ICI/NVLink interconnect topologies) and roof-line analysis.
  • Practical familiarity with modern long-context LLM serving mechanics—paged KV caches, chunked prefill, context parallelism, and speculative decoding.
  • Proven ability to debug subtle compiler miscompilations, NaN/precision divergence, and multi-host collective deadlocks.

PREFERRED / BONUS EXPERIENCE

  • Hands-on experience writing Pallas (TPU/GPU) or Triton/CUTLASS kernels for custom attention or quantized matrix multiplication.
  • Experience targeting multiple distinct AI hardware backends (e.g., NVIDIA Hopper/Blackwell + Google Cloud TPU or AWS Trainium) from a unified model codebase.
  • Upstream commit history in OpenXLA, MLIR, LLVM, JAX, or Triton.

Position Telemetry

Location Seattle, WA (Hybrid)
Department AI Infrastructure
Employment type Full-time
Workplace type Hybrid
Salary range
$245k – $340k annual package