Quantization Engineer

Project

Work done for a product or program.

• Implement post-training quantization flows including calibration, per-channel and per-group scaling, and outlier handling.
• Diagnose per-layer accuracy loss through sensitivity analysis and assign mixed-precision bit allocations accordingly.
• Implement and validate weight-only schemes such as GPTQ, AWQ and SmoothQuant against the hardware's supported formats.
• Define the quantization operator semantics the compiler and kernel teams must implement bit-exactly.
• Support KV-cache quantization and activation microscaling for long-context LLM inference.
• Produce customer-facing quantization guides, default recipes and accuracy-versus-throughput trade-off curves.
• Coordinate with hardware architects on which numeric formats and scaling granularities the next silicon generation should support.
• Maintain reproducible calibration datasets and versioned quantized model checkpoints.

Function

Work that keeps the discipline running: methods, tools, suppliers and people.

• Develop quantization-aware training recipes with fake-quant insertion, straight-through estimators and learned step size.
• Build accuracy evaluation harnesses that score quantized models on task benchmarks such as MMLU, ImageNet top-1 or WER.

Project

1. Produce and hand over the role's key deliverables
   • Quantization toolkit with PTQ and QAT flows for the target hardware
   • Accuracy-versus-throughput trade-off report per reference model
   • Per-layer sensitivity and mixed-precision bit-allocation policies
   • Bit-exact quantization operator specification for compiler and kernel teams
   • Customer quantization playbook with validated default recipes
2. Develop the softwareGeneral
   • Design to the hardware specification
   • Write and review the code
   • Test on models, emulation or silicon
3. Release and supportGeneral
   • Package and document releases
   • Fix customer-reported defects

Function

1. Maintain the platformGeneral
   • Keep build, test and release tools working
   • Track toolchain and operating system changes
2. Raise qualityGeneral
   • Run code reviews
   • Extend automated tests