Skip to content

requirement_definition_template.md

Latest commit

 

History

History
209 lines (153 loc) · 6.15 KB

requirement_definition_template.md

File metadata and controls

209 lines (153 loc) · 6.15 KB

Requirement Definition

Samples: DGEMM 8-CLI (EN) | CFD 6-agent (JP) | DGEMM 12-agent (EN)

Project Information

  • Project name: [project name]
  • Date: [YYYY-MM-DD]

Optimization Target

Code Source

  • Local files: placed in BaseCode/
  • GitHub repository: [URL]
  • Other: [description]

Target Files

  • Main file: [filename]
  • Dependencies: [file list]

Optimization Goals

Performance Target

[Specific target. e.g., 2x current performance, 50% of theoretical peak]

Reporting

  • Report % of theoretical peak alongside absolute GFLOPS (PM calculates peak from hardware-info skill)

Priority

  • Minimize execution time
  • Maximize throughput
  • Minimize memory usage
  • Improve scalability
  • Maximize energy efficiency
  • Other: [description]

Overview

Application Overview

[Brief description of the application]

Optimization Approach

[Preferred approach. If unspecified, PM decides automatically]

Constraints

Hardware

  • System name: [target environment — supercomputer name, cloud instance, etc.]
  • Available nodes: [node types, GPU models, etc.]

Remote Execution Environment

  • Working directory: [path on remote system; place details in _remote_info/]

Job Resources

Staged Scale-Up

[Node count usage policy]

Resource Constraints

  • Max execution time: [limit]
  • Other constraints: [description]

Job Execution Mode

  • Batch job (recommended)
  • Interactive job
  • Login node execution (not recommended)

If unspecified, batch jobs are used. Most supercomputers prohibit computation on login nodes.

Middleware (Compiler / Parallelization Modules)

Compiler

  • [Select available compilers; refer to _remote_info/]

Parallelization Libraries

  • MPI
  • OpenMP
  • CUDA
  • OpenACC
  • AVX/AVX2/AVX512 (SIMD)
  • Other: [description]

Numerical Libraries

  • Do not use (self-implemented only)
  • Use for performance comparison (MKL, cuBLAS, etc.)
  • Use actively

Parallelization Strategy

[Implementation order and scope. If unspecified, PM decides automatically (evolutionary exploration)]

Accuracy Requirements

  • Same accuracy as existing tests
  • Error tolerance: [description]
  • Other: [description]

Budget (Jobs)

  • Floor: [minimum points to spend before stopping]
  • Target: [expected budget for normal optimization]
  • Ceiling: [hard upper limit, must not exceed]

Point Consumption Rate (Reference)

[System-specific rate. e.g., 0.007 points per elapsed second per GPU]

Time Limit

  • Floor: [minimum working time. e.g., 1 hour]
  • Target: [standard working time]
  • Ceiling: [maximum working time. e.g., 3 hours]

Agent Configuration

Mode

  • SOLO mode (single agent handles all roles)
  • Multi-agent (PM, SE, PG, CD deployed individually)

Operation Mode

  • Autonomous (PM proceeds through all phases without asking for confirmation — recommended for unattended optimization)
  • Vibe Coding (PM pauses at each phase for user feedback — interactive, incremental development)

Worker Count

Corresponds to -w in python3 -m vibecodehpc setup --name <name> -w <N>.

Workers SE PG CD Notes
2 1 1 0 Minimum
4 1 3 0 Small
8 2 5 1 Stable (SE >= 2)
12 2 9 1 Recommended
  • SE count: [specify; 2+ enables cross-monitoring]
  • PG count: [specify]

Agent Configuration Table

Role Agent ID CLI Model Specialization CLI Args Notes
PM PM claude opus -
SE SE1 codex gpt-5.4 -
SE SE2 claude opus -
PG PG1.1 cline sonnet-4.6 OpenMP
PG PG1.2 vibe-local qwen3.5:35b OpenMP --context-window 65536 Ollama
PG PG1.3 gemini Gemini-3 OpenMP CLI default
PG PG2.1 qwen qwen3.5-plus OpenMP Native auth
PG PG2.2 opencode qwen3-next-80b OpenMP
PG PG2.3 kimi kimi-code OpenMP Native auth
  • List all agents explicitly. No default model.
  • CLI Args: Extra flags passed at launch (e.g. --context-window 65536)
  • See CLI Support Matrix for auth and compatibility

CLI Fallback Policy

  • Allow PM to substitute unavailable CLIs with alternatives (flexible — optimization is the goal)
  • Strict: fail and report if specified CLI is not available (for multi-CLI validation experiments)

PG Specialization

Specialization Count Notes
OpenMP [N]
CUDA [N]
MPI [N]
[other] [N]

Specialization is used as the directory name and defines each PG's parallelization scope.

CD (GitHub Integration)

  • Enable
  • Disable
  • Gradual introduction

When using CD:

  • Repository: [URL or org/repo]
  • Branch policy: [description. e.g., create a branch named after the project]

Security Requirements

  • Anonymize user/project information before GitHub push
  • Relativize budget information (hide total budget)
  • Avoid absolute paths (ensure portability)
  • Other: [description]

Additional Instructions

All Agents

[Instructions for all agents]

SE-specific

[SE-specific instructions. e.g., periodically verify all results are plotted in SOTA graphs]

PG-specific

[PG-specific instructions. e.g., record both execution time and throughput]

CD-specific

[CD-specific instructions. e.g., .gitignore policy, push frequency]

Auto-Generated Information (filled by PM)

  • Missing items: [PM fills automatically]
  • Recommended configuration: [PM fills automatically]
  • Initial agent placement: [PM fills automatically]

Sample Requirement Definitions