How amd-agi/geak works
GEAK optimizes AMD GPU workloads through two user-facing workflows: end-to-end model serving and single-kernel tuning. The e2e workflow profiles serving, prioritizes hot kernels, explores configuration and kernel options, and validates overlays; it can delegate kernel work to the kernel workflow. Supporting layers include the standalone workflow runtime, launch-configuration and handoff handling, persistent optimization knowledge, and reference optimization skills. README statements establish the main workflow relationship; source-index references support the shown module connections. Unsampled workflow internals and optional integrations are not expanded.
Main entry points: e2e_store.py E2E knowledge adapter · experience_store.py Kernel experience adapter · e2e_workflow.js Serving optimizer · run_workflow.mjs Workflow executor
Components 32
Optimization workflows
- Serving optimizer
e2e_workflow/e2e_workflow.js - Kernel optimizer
kernel_workflow/kernel_workflow.js - Kernel tuning lane
kernel_workflow/kernel_lane.js - Serving analysis tools
e2e_workflow/scripts - Roofline analysis
e2e_workflow/knowledge/analysis_skills/roofline/roofline_tools.py - FlyDSL overlay
e2e_workflow/knowledge/templates/flydsl_overlay_sitecustomize.py
Workflow runtime
- Workflow executor
interface/runtime/engine/run_workflow.mjs - Agent configuration
interface/runtime/engine/config.mjs - Backend adapter
interface/runtime/engine/backends/base.mjs - Generic backend
interface/runtime/engine/backends/generic.mjs - Workflow schema
interface/runtime/engine/schema.mjs - Experiment support
interface/runtime/engine/experiment.mjs
Optimization knowledge
- E2E knowledge adapter
e2e_workflow/scripts/e2e_store.py - Kernel experience adapter
kernel_workflow/scripts/experience_store.py - Knowledge identity
kb/identity.py - Local knowledge store
kb/store_local.py - Remote store client
kb/store_client.py - Record attestation
kb/attest.py - Record retraction
kb/retract.py - Knowledge curation
kb/curate.py - Fallback ladder
kb/ladder.py - Store plane
kb/plane.py - Remote store
kb/store_remote.py
Expert skills
- Tuning reference skills
perf_knowledge/expert_skills - Gluon authoring tools
perf_knowledge/expert_skills/skills/gluon_authoring/scripts - Tuning tools
perf_knowledge/expert_skills/tuning/tools - Tuning claim checks
perf_knowledge/expert_skills/tuning/validate/claims.py - FlyDSL MoE integration
perf_knowledge/expert_skills/skills/apply_flydsl_moe_to_vllm/flydsl_moe_shim.py
Handoff integration
- E2E handoff runner
interface/run_e2e.py - Launch config
interface/effective_config.py
Other components
- Engineer
- External orchestrator
Connections 40
- Engineer → Serving optimizer: optimizes serving
README.md - Engineer → Kernel optimizer: optimizes kernel
README.md - Serving optimizer → Serving analysis tools: profiles and analyzes
README.md - Serving optimizer → Roofline analysis: uses analysis
- Serving optimizer → FlyDSL overlay: applies overlay
- Serving optimizer → Kernel optimizer: recursively calls
README.md - Kernel optimizer → Kernel tuning lane: orchestrates
README.md - Engineer → Workflow executor: runs workflow
README.md - Workflow executor → Agent configuration: loads selection
interface/runtime/engine/run_workflow.mjs - Workflow executor → Backend adapter: dispatches agent
interface/runtime/engine/run_workflow.mjs - Workflow executor → Generic backend: selects fallback
interface/runtime/engine/run_workflow.mjs - Generic backend → Backend adapter: uses contract
interface/runtime/engine/backends/generic.mjs - Workflow executor → Workflow schema: validates workflow
interface/runtime/engine/run_workflow.mjs - Workflow executor → Experiment support: supports experiments
- External orchestrator → E2E handoff runner: submits handoff
interface/run_e2e.py - E2E handoff runner → Launch config: uses config
interface/run_e2e.py - E2E knowledge adapter → Knowledge identity: uses identity
e2e_workflow/scripts/e2e_store.py - E2E knowledge adapter → Record attestation: uses attestation
e2e_workflow/scripts/e2e_store.py - E2E knowledge adapter → Knowledge curation: uses curation
e2e_workflow/scripts/e2e_store.py - E2E knowledge adapter → Fallback ladder: uses fallback ladder
e2e_workflow/scripts/e2e_store.py - E2E knowledge adapter → Store plane: uses store plane
e2e_workflow/scripts/e2e_store.py - E2E knowledge adapter → Record retraction: uses retraction
e2e_workflow/scripts/e2e_store.py - E2E knowledge adapter → Local knowledge store: uses local store
e2e_workflow/scripts/e2e_store.py - Kernel experience adapter → Knowledge identity: uses identity
kernel_workflow/scripts/experience_store.py - Kernel experience adapter → Record attestation: uses attestation
kernel_workflow/scripts/experience_store.py - Kernel experience adapter → Knowledge curation: uses curation
kernel_workflow/scripts/experience_store.py - Kernel experience adapter → Fallback ladder: uses fallback ladder
kernel_workflow/scripts/experience_store.py - Kernel experience adapter → Store plane: uses store plane
kernel_workflow/scripts/experience_store.py - Kernel experience adapter → Record retraction: uses retraction
kernel_workflow/scripts/experience_store.py - Kernel experience adapter → Local knowledge store: uses local store
kernel_workflow/scripts/experience_store.py - Remote store → Local knowledge store: mirrors store
kb/store_remote.py - Remote store → Remote store client: uses client
kb/store_remote.py - Record attestation → Record retraction: uses retraction
kb/attest.py - Record retraction → Local knowledge store: rewrites records
kb/retract.py - Serving optimizer → Tuning reference skills: draws on knowledge
- Kernel optimizer → Tuning reference skills: draws on knowledge
- Tuning reference skills → Gluon authoring tools: includes tooling
- Tuning reference skills → Tuning tools: includes tooling
- Tuning reference skills → Tuning claim checks: includes validation
- Tuning reference skills → FlyDSL MoE integration: includes integration