tools/mega_factory/spec.py

"""MegaSpec — everything that varies between megakernel tasks.

The megakernel family cannot use `_factory/spec.py`: those tasks grade ONE pure function over shape
tuples, whereas these grade a STATEFUL, multi-step workload under measured fusion gates with a
throughput reward.

Two sub-families share this spec and this builder:

* `family="e1"` — whole-model decode megakernels. Reward is tokens/s; the gates are correctness,
  kernels-per-step and dominant-kernel share, and the kernel-count gate is what makes it a megakernel
  task rather than a speed task.
* `family="e2"` — enabling primitives (a device-wide barrier, an instruction interpreter, a fused
  layer, a warp-specialised GEMV...). These are NOT whole models, so the kernels-per-*step* gate is
  not automatically the right thing to grade; each e2 spec sets its own limits and must justify them
  in `gates_md`.

Numbers below that look arbitrary are measured — see CALIBRATION.md.
"""
from dataclasses import dataclass, field


@dataclass
class MegaSpec:
    # ---- identity ---------------------------------------------------------------------------------
    name: str
    title: str
    blurb: str
    keywords: list = field(default_factory=list)
    family: str = "e1"              # "e1" whole-model decode | "e2" enabling primitive

    # ---- architecture -----------------------------------------------------------------------------
    # wdtype: "bf16" | "fp8" | "int4" | "nvfp4".  Every quantised form ships weights ALREADY quantised
    # with their scales and the reference dequantises those exact bytes (see CALIBRATION.md §6).
    cfg: dict = field(default_factory=dict)

    # Reference/fixture source embedded verbatim into both environment/reference.py and the private
    # grader. Empty -> the shared Llama-shaped decoder in model.py.
    model_src: str = ""

    # ---- decode regime ----------------------------------------------------------------------------
    batch: int = 1                  # small by design: this family lives in the latency regime
    prefill_len: int = 2048         # KV already holds this many tokens when the timed loop starts
    max_seq: int = 4096
    decode_steps: int = 32          # timed steps per rep
    correct_steps: int = 8          # steps compared against the reference
    prof_steps: int = 4             # steps profiled for the kernel-count gate

    # ---- gates ------------------------------------------------------------------------------------
    tol: float = 3e-2               # full-output RELATIVE error. NOT top-k agreement: with random
                                    # weights logits are near-uniform, top-1 ties flip on noise and a
                                    # correct kernel fails (measured top1 0.79-0.92). relerr is stable.
    max_kernels_per_step: float = 8.0
    min_dominant_share: float = 0.90

    # ---- reward -----------------------------------------------------------------------------------
    reward_metric: str = "tokens/s"
    reward_work: float = 0.0        # numerator per step; 0 -> `batch` (one token per sequence per step)

    # ---- entry points -----------------------------------------------------------------------------
    entry_build: str = "build_model"
    entry_step: str = "decode_step"
    step_sig: str = "handle, token_ids, pos"
    step_doc: str = ("One decode step for the whole batch; append this position's K/V into the cache."
                     "\n\n    token_ids : (B,) int64      pos : int, absolute position being written"
                     "\n    returns   : (B, vocab) logits\n    ")
    step_ret: str = "logits"
    unfused_kernels: int = 0        # measured kernel launches/call for the eager reference, if known
    arg_doc: str = ("weights   : dict from the reference's make_weights (see /app/reference.py)"
                    "\n    kv_cache  : list of (k, v) per layer, each (B, n_kv, max_seq_len, hd) bf16,"
                    " prefilled")

    # ---- roofline ---------------------------------------------------------------------------------
    bytes_per_step: float = 0.0     # override; 0 -> weight_bytes() + kv_bytes() (Llama-shaped only)
    floor_us_override: float = 0.0  # override the roofline entirely (compute-bound tasks)

    # ---- prose ------------------------------------------------------------------------------------
    intro_md: str = ""
    spec_md: str = ""
    contract_md: str = ""
    regime_md: str = ""
    perf_md: str = ""
    precision_md: str = ""
    correctness_md: str = ""
    gates_md: str = ""              # REQUIRED for family e2: why these gate values are the right ones
    faithfulness_md: str = ""

    # ---- environment ------------------------------------------------------------------------------
    base_image: str = "pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel"
    pip_extra: str = "einops nvidia-cutlass-dsl"
    module: str = "megakernel.py"
    gpus: int = 1
    agent_timeout_sec: float = 14400.0
    verifier_timeout_sec: float = 2700.0
    memory_mb: int = 65536

    ELT = {"bf16": 2, "fp8": 1, "int4": 0.5, "nvfp4": 0.5}

    def weight_bytes(self):
        """Bytes of weight read per decode step -- the roofline. lm_head is read in full at bs=1.

        Only meaningful for the plain Llama-shaped configs; anything else (MoE, hybrid, primitives)
        sets `bytes_per_step` explicitly."""
        c = self.cfg
        d, ffn, n_q, n_kv, hd = c["d"], c["ffn"], c["n_q"], c["n_kv"], c["hd"]
        per_layer = n_q * hd * d + 2 * n_kv * hd * d + d * n_q * hd + 3 * ffn * d
        elt = self.ELT[c["wdtype"]]
        return (c["layers"] * per_layer + c["vocab"] * d) * elt

    def kv_bytes(self):
        """Bytes of KV read per decode step at the deepest position."""
        c = self.cfg
        return (2 * self.batch * c["layers"] * c["n_kv"]
                * (self.prefill_len + self.decode_steps) * c["hd"] * 2)

    def total_bytes(self):
        if self.bytes_per_step:
            return float(self.bytes_per_step)
        return self.weight_bytes() + self.kv_bytes()

    def floor_us(self, hbm_bw=4.8e12):
        if self.floor_us_override:
            return float(self.floor_us_override)
        return self.total_bytes() / hbm_bw * 1e6

    def work_per_step(self):
        return self.reward_work or float(self.batch)

    def validate(self):
        assert self.name and self.cfg, "name/cfg required"
        assert self.family in ("e1", "e2"), self.family
        assert "wdtype" in self.cfg and self.cfg["wdtype"] in self.ELT, self.cfg.get("wdtype")
        if self.family == "e1":
            for k in ("layers", "d", "n_q", "n_kv", "hd", "vocab", "eps", "theta"):
                assert k in self.cfg, f"cfg missing {k}"
            assert self.cfg["n_q"] % self.cfg["n_kv"] == 0, "n_q must be a multiple of n_kv"
            assert self.batch <= 8, "whole-model megakernels are a SMALL-batch family by construction"
            assert self.prefill_len + self.decode_steps * max(1, int(self.cfg.get("tokens_per_step", 1))) \
                <= self.max_seq, "decode would overrun max_seq"
        else:
            assert self.gates_md.strip(), (
                "family e2 must justify its gates in gates_md -- the kernels-per-step gate is not "
                "automatically the right contract for a primitive")
        # the task is only meaningful if there is headroom to fuse into
        assert self.floor_us() > 50, f"roofline {self.floor_us():.0f}us too small to be measurable"
        return self