data/optimization-roadmap.json

{
  "$schema": "../schemas/optimization-roadmap.schema.json",
  "model": "moonshotai/Kimi-K3",
  "items": [
    {
      "id": "paged-hybrid-cache",
      "priority": 1,
      "scope": "inference",
      "change": "Replace repeated MLA cache concatenation with a paged hybrid KDA-MLA cache.",
      "expectedImpact": "very-high",
      "risk": "medium",
      "verification": [
        "token-logit parity against reference",
        "peak-memory benchmark",
        "prefill and decode throughput",
        "prefix-cache hit correctness"
      ]
    },
    {
      "id": "fused-moe",
      "priority": 1,
      "scope": "inference",
      "change": "Fuse latent projection, routing, GPU dispatch, grouped GEMM and combination.",
      "expectedImpact": "very-high",
      "risk": "high",
      "verification": [
        "expert-selection parity",
        "MXFP4 numerical tolerance",
        "all-to-all traffic",
        "tokens per second by batch size"
      ]
    },
    {
      "id": "reference-test-fixture",
      "priority": 1,
      "scope": "model",
      "change": "Publish a tiny deterministic configuration, fixture tensors and expected logits for the reference implementation.",
      "expectedImpact": "high",
      "risk": "low",
      "verification": [
        "clean-environment installation",
        "CPU smoke test",
        "GPU logits match within documented tolerance",
        "cache and multimodal regression fixtures"
      ]
    },
    {
      "id": "fused-situ-attnres",
      "priority": 2,
      "scope": "kernels",
      "change": "Fuse SiTU-GLU and implement online-softmax Attention Residual kernels.",
      "expectedImpact": "high",
      "risk": "medium",
      "verification": [
        "BF16 and FP8 parity",
        "temporary-allocation count",
        "kernel latency",
        "training-gradient parity on a tiny model"
      ]
    },
    {
      "id": "hierarchical-context",
      "priority": 2,
      "scope": "agent",
      "change": "Use hierarchical summaries, structured state and evidence retrieval before reaching the raw context limit.",
      "expectedImpact": "high",
      "risk": "medium",
      "verification": [
        "long-horizon task success",
        "evidence recall",
        "context tokens per completed task",
        "regression against full-context baseline"
      ]
    },
    {
      "id": "agent-stopping-verification",
      "priority": 2,
      "scope": "agent",
      "change": "Train and evaluate explicit stopping, evidence checking and recovery policies for long autonomous tasks.",
      "expectedImpact": "high",
      "risk": "medium",
      "verification": [
        "false-completion rate",
        "successful recovery after tool failure",
        "evidence-backed answer rate",
        "tokens and tool calls per completed task"
      ]
    },
    {
      "id": "adaptive-compute",
      "priority": 3,
      "scope": "architecture",
      "change": "Evaluate fixed compute profiles with different expert, depth and reasoning budgets.",
      "expectedImpact": "high",
      "risk": "high",
      "verification": [
        "quality-cost Pareto curve",
        "static-shape compatibility",
        "expert load balance",
        "latency distribution"
      ]
    },
    {
      "id": "distilled-variant",
      "priority": 3,
      "scope": "model",
      "change": "Distill agentic behavior into a smaller model with external memory.",
      "expectedImpact": "very-high",
      "risk": "high",
      "verification": [
        "coding and agentic benchmarks",
        "memory footprint",
        "single-node feasibility",
        "long-context workflow retention"
      ]
    }
  ]
}