Challenge

Triton Kernel Quantization Benchmark Agent with Pydantic AI and Together AI

Build an automated Triton kernel quantization benchmark workflow using Pydantic AI and Together AI to profile FP8 versus INT4 matrix operations. Evaluate throughput gains and perplexity tradeoffs to select optimal kernel configurations achieving a minimum 1.8x speedup.

EngineeringHosted by Vera
Challenge brief

What you are building

The core problem, expected build, and operating context for this challenge.

Utilize Pydantic AI's type-safe agent framework and Together AI inference to benchmark Triton quantization kernels.

Delivery guide

How work is evaluated

Evaluation

Evaluates speedup multiplier and model quality preservation for benchmarked kernels.

Datasets

Shared data for this challenge

Review public datasets and any private uploads tied to your build.

Evaluation rubric

How submissions are scored

These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.

Dimensions
2 scoring checks
Binary
2 pass or fail dimensions
Ordinal
0 scaled dimensions
Dimension 1

Minimum Speedup Met

Kernel speedup factor over baseline FP16 must be at least 1.8x.

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Dimension 2

Perplexity Delta

Increase in perplexity introduced by quantization. • target: 0.05 • range: 0-0.15

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Learning goals

What you should walk away with

  • Implement type-safe quantization benchmark workflows with Pydantic AI.

  • Integrate Together AI inference endpoints for perplexity comparison baseline.

  • Profile Triton FP8 and INT4 GEMM kernels across matrix sizes.

  • Determine automated kernel dispatch rules based on accuracy/throughput Pareto frontiers.

Resources and assets

Reference links and supporting material

Dataset notes

Synthetic matrix shape test cases and calibration datasets for LLM weight quantization.

How this agent runs

Evaluates speedup multiplier and model quality preservation for benchmarked kernels.

Challenge input

JSON with matrix dimensions and quantization format

Pydantic AI

Provides type-safe structured output handling and agent state management

Together AI

Provides fast model endpoints for baseline perplexity verification

Evaluated output

JSON with speedup_factor and perplexity_delta

Checks for
  • Kernel speedup factor over baseline FP16 must be at least 1.8x.
  • Increase in perplexity introduced by quantization. • target: 0.05 • range: 0-0.15
Proof of success
  • Benchmark: Triton Kernel Performance Benchmark
  • Perplexity Delta target: 0.05
  • 1 public reference case
Runtime evidence
  • Python execution harness
  • Python sandbox (unavailable on Versalist)
View technical recipe

Configured tools

Action Space
  • Pydantic AI · Optional
  • Anthropic · Optional
Policy Serving
  • Together AI · Required
Training Infra
  • Together AI · Required
Orchestration
  • Pydantic AI · Optional

Evaluation contract

  • Minimum Speedup Met · Weight 1
  • Perplexity Delta · Weight 1

Recipe state

This is a preview. The configuration can change before the evaluation recipe is locked.

Frequently Asked Questions about Triton Kernel Quantization Benchmark Agent with Pydantic AI and Together AI