tinymind-native-8b-remote-handoff/bundle/model/runtime.py

"""Runtime mode selection for TinyMind Omega++."""

from __future__ import annotations

from enum import Enum
import json
from pathlib import Path

from .config import OmegaConfig


class RuntimeMode(str, Enum):
    BF16_QUALITY = "bf16_quality"
    INT4_SPARSE_FAST = "int4_sparse_fast"
    INT6_BRIDGE_IMMA_FAST = "int6_bridge_imma_fast"
    AUTO = "auto"


def int6_bridge_available(
    report_path: str | Path = "reports/int6_bridge_imma_eval/int6_bridge_imma_eval_report.json",
    min_hardware_imma_tops: float = 100.0,
) -> bool:
    path = Path(report_path)
    if not path.exists():
        return False
    try:
        report = json.loads(path.read_text(encoding="utf-8-sig"))
    except (OSError, json.JSONDecodeError):
        return False
    gate = report.get("claim_gate", {})
    metrics = report.get("metrics", {})
    return (
        gate.get("fused_two_pass_kernel_measured") is True
        and gate.get("imma_sp_sass_observed") is True
        and float(metrics.get("avg_hardware_imma_tops", 0.0)) >= min_hardware_imma_tops
    )


def resolve_runtime_mode(
    cfg: OmegaConfig,
    cuda_available: bool,
    sparse_artifact_available: bool,
    int6_bridge_artifact_available: bool = False,
    requested: str | None = None,
) -> RuntimeMode:
    mode = requested or cfg.precision_mode
    if mode == RuntimeMode.INT6_BRIDGE_IMMA_FAST.value:
        if cuda_available and int6_bridge_artifact_available:
            return RuntimeMode.INT6_BRIDGE_IMMA_FAST
        if cuda_available and sparse_artifact_available:
            return RuntimeMode.INT4_SPARSE_FAST
        return RuntimeMode.BF16_QUALITY
    if mode == RuntimeMode.INT4_SPARSE_FAST.value:
        if cuda_available and sparse_artifact_available:
            return RuntimeMode.INT4_SPARSE_FAST
        return RuntimeMode.BF16_QUALITY
    if mode == RuntimeMode.AUTO.value:
        if cuda_available and int6_bridge_artifact_available:
            return RuntimeMode.INT6_BRIDGE_IMMA_FAST
        if cuda_available and sparse_artifact_available:
            return RuntimeMode.INT4_SPARSE_FAST
        return RuntimeMode.BF16_QUALITY
    return RuntimeMode.BF16_QUALITY