k2_pack/k2/parameter_flops.py

"""Emit the frozen CPU parameter/symbolic/profiler FLOP receipt."""

from __future__ import annotations

import argparse
import os
import platform
from dataclasses import asdict

import numpy as np
import torch

from .config import load_full_config
from .constants import A6_PROTOCOL_SHA256, ATTENTION_MODES
from .io import file_record, write_json_exclusive
from .model import K2Student, StudentSpec, allocated_parameter_count, state_dict_tensor_bytes


def symbolic_forward_flops(spec: StudentSpec, batch: int, length: int) -> dict[str, int]:
    b, l, d, z = batch, length, spec.d_model, spec.latent_dim
    per_block = {
        "qkv_and_output_projections": 8 * b * l * d * d,
        "attention_qk_and_av_matmuls": 4 * b * l * l * d,
        "mlp_linears": 4 * b * l * d * spec.mlp_hidden,
    }
    components = {
        "input_projection": 2 * b * l * z * d,
        "time_mlp": 2 * b * (128 * d + d * d),
        "transformer_blocks": spec.blocks * sum(per_block.values()),
        "head_a_h": 2 * b * l * d * spec.within_token_hidden,
        "head_a_z": 2 * b * l * z * spec.within_token_hidden,
        "head_a_t_once_per_sequence": 2 * b * d * spec.within_token_hidden,
        "head_b": 2 * b * l * spec.within_token_hidden * z,
        "head_c_h": 2 * b * l * d * z,
        "head_c_z": 2 * b * l * z * z,
    }
    return {
        "total": int(sum(components.values())),
        "components": {key: int(value) for key, value in components.items()},
        "per_transformer_block_components": {
            key: int(value) for key, value in per_block.items()
        },
    }


def profiler_forward_flops(model: K2Student, mode: str) -> int:
    z = torch.zeros((1, 64, 16), dtype=torch.float32)
    t = torch.full((1, 1, 1), 0.5, dtype=torch.float32)
    with torch.no_grad(), torch.profiler.profile(
        activities=[torch.profiler.ProfilerActivity.CPU], with_flops=True,
        record_shapes=True,
    ) as profile:
        model(z, t, mode)
    return int(sum(int(event.flops or 0) for event in profile.key_averages()))


def graph_active_parameter_count(model: K2Student, mode: str) -> int:
    model.zero_grad(set_to_none=True)
    z = torch.zeros((1, 64, 16), dtype=torch.float32)
    t = torch.full((1, 1, 1), 0.5, dtype=torch.float32)
    model(z, t, mode).sum().backward()
    active = sum(
        parameter.numel() for parameter in model.parameters()
        if parameter.grad is not None
    )
    model.zero_grad(set_to_none=True)
    return int(active)


def generate_receipt(config_path: str, output: str) -> dict:
    config = load_full_config(config_path)
    if str(torch.__version__) != "2.7.0+cu126" or np.__version__ != "1.26.4":
        raise RuntimeError(
            f"receipt requires torch 2.7.0+cu126/numpy 1.26.4; "
            f"got {torch.__version__}/{np.__version__}"
        )
    if torch.cuda.is_available():
        raise RuntimeError("public CPU receipt requires CUDA to be unavailable")
    torch.manual_seed(0)
    torch.use_deterministic_algorithms(True)
    model = K2Student(StudentSpec.from_protocol(config.protocol)).cpu().eval()
    parameter_count = allocated_parameter_count(model)
    profiler = {mode: profiler_forward_flops(model, mode) for mode in ATTENTION_MODES}
    active = {mode: graph_active_parameter_count(model, mode) for mode in ATTENTION_MODES}
    symbolic_b1 = symbolic_forward_flops(model.spec, 1, 64)
    symbolic_b256 = symbolic_forward_flops(model.spec, 256, 64)
    config_record = file_record(config.path)
    config_record.pop("path")
    config_record["relative_path"] = "config/k2_full_config.json"
    receipt = {
        "schema_version": 1,
        "receipt": "TCFM-A6-K2-parameter-and-flop-v1",
        "status": "PASS" if (
            profiler["full"] == profiler["prefix"]
            and active == {"full": parameter_count, "prefix": parameter_count}
        ) else "FAIL",
        "scope": "CPU architecture receipt; no training/evaluation metric",
        "hardware": "CPU",
        "h200_claim": False,
        "environment": {
            "python": platform.python_version(),
            "torch": str(torch.__version__),
            "numpy": np.__version__,
            "cuda_used": False,
            "cuda_available": torch.cuda.is_available(),
            "cuda_visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES"),
            "deterministic_algorithms": torch.are_deterministic_algorithms_enabled(),
        },
        "a6_protocol_sha256": A6_PROTOCOL_SHA256,
        "full_config": config_record,
        "architecture_closure": config.raw["architecture_closure"],
        "student_spec": asdict(model.spec),
        "allocated_parameter_count": parameter_count,
        "active_parameter_count_by_mode": active,
        "active_parameter_method": "parameter has non-None gradient after a complete B=1,L=64 forward/backward",
        "state_dict_tensor_bytes": state_dict_tensor_bytes(model),
        "only_mode_difference": "equal-shaped boolean attention mask entries",
        "symbolic_flop_convention": (
            "one multiply plus one add equals two FLOPs; counts dense Linear and "
            "attention matmuls only; bias, normalization, elementwise, softmax, "
            "trigonometric and loss operations excluded"
        ),
        "symbolic_forward_flops_batch1_length64": symbolic_b1,
        "symbolic_forward_flops_batch256_length64": symbolic_b256,
        "cpu_torch_profiler_forward_flops_batch1_length64": profiler,
        "profiler_parity_required": True,
    }
    if receipt["status"] != "PASS":
        raise AssertionError(f"full/prefix CPU profiler FLOPs differ: {profiler}")
    write_json_exclusive(output, receipt)
    return receipt


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", required=True)
    parser.add_argument("--output", required=True)
    args = parser.parse_args()
    receipt = generate_receipt(args.config, args.output)
    print(receipt["status"])


if __name__ == "__main__":
    main()