optimize_decoder_fp16_matmul.py

from __future__ import annotations

import argparse
import hashlib
import json
import re
from collections import defaultdict
from pathlib import Path

import numpy as np
import onnx
import onnxruntime as ort
from onnx import TensorProto, helper, numpy_helper


ROOT = Path(__file__).resolve().parent
BASELINE = ROOT / ".work/models/shared/decoder_unified_gather_qdq_int8.onnx"
DEFAULT_SOURCE = ROOT / ".work/models/model-opt/decoder_gather_before_dq_int8.onnx"
DEFAULT_DESTINATION = ROOT / ".work/models/model-opt/decoder_static_fp16_matmul.onnx"
DEFAULT_REPORT = ROOT / ".work/reports/model-fp16-matmul-optimization.json"
VISION_TOKENS = 256
KEEP_QDQ_MATMUL_NAMES = {"/lm_head/MatMul"}


def sha256(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as source:
        for chunk in iter(lambda: source.read(1024 * 1024), b""):
            digest.update(chunk)
    return digest.hexdigest()


def dequantize_to_fp16(
    quantized: onnx.TensorProto,
    scale: onnx.TensorProto,
    zero_point: onnx.TensorProto,
    axis: int,
) -> np.ndarray:
    values = numpy_helper.to_array(quantized).astype(np.float32)
    scales = numpy_helper.to_array(scale).astype(np.float32)
    zeros = numpy_helper.to_array(zero_point).astype(np.float32)
    broadcast_shape = [1] * values.ndim
    broadcast_shape[axis] = scales.size
    return ((values - zeros.reshape(broadcast_shape)) * scales.reshape(broadcast_shape)).astype(np.float16)


def rewrite_matmul_weights(model: onnx.ModelProto, selected_layers: set[int]) -> dict:
    graph = model.graph
    initializers = {value.name: value for value in graph.initializer}
    consumers: dict[str, list[onnx.NodeProto]] = defaultdict(list)
    for node in graph.node:
        for name in node.input:
            consumers[name].append(node)

    dq_by_matmul_name: dict[str, onnx.NodeProto] = {}
    fp16_initializers: list[onnx.TensorProto] = []
    removable_initializers: set[str] = set()
    quantized_elements = 0
    for node in graph.node:
        if node.op_type != "DequantizeLinear" or len(node.input) < 3:
            continue
        quantized = initializers.get(node.input[0])
        scale = initializers.get(node.input[1])
        zero_point = initializers.get(node.input[2])
        if quantized is None or scale is None or zero_point is None:
            continue
        matches = [consumer for consumer in consumers[node.output[0]] if consumer.op_type == "MatMul"]
        if not matches:
            continue
        if len(matches) != 1:
            raise RuntimeError(f"Expected one MatMul consumer for {node.name}, found {len(matches)}")
        matmul = matches[0]
        if matmul.name in KEEP_QDQ_MATMUL_NAMES:
            continue
        layer_match = re.search(r"/decoder/layers\.(\d+)/", matmul.name)
        if layer_match is None:
            raise RuntimeError(f"Unexpected non-layer MatMul {matmul.name}")
        if int(layer_match.group(1)) not in selected_layers:
            continue
        if matmul.input[1] != node.output[0]:
            raise RuntimeError(f"Quantized weight is not RHS input for {matmul.name}")
        axis = next((attribute.i for attribute in node.attribute if attribute.name == "axis"), 1)
        fp16_name = f"{quantized.name}_static_fp16"
        fp16 = dequantize_to_fp16(quantized, scale, zero_point, axis)
        fp16_initializers.append(numpy_helper.from_array(fp16, fp16_name))
        matmul.input[1] = fp16_name
        dq_by_matmul_name[matmul.name] = node
        removable_initializers.update(node.input[:3])
        quantized_elements += fp16.size

    expected_matmuls = 7 * len(selected_layers)
    if len(dq_by_matmul_name) != expected_matmuls:
        raise RuntimeError(f"Expected {expected_matmuls} selected MatMuls, found {len(dq_by_matmul_name)}")

    rewritten: list[onnx.NodeProto] = []
    removed_dq_names = {node.name for node in dq_by_matmul_name.values()}
    for node in graph.node:
        if node.name in removed_dq_names:
            continue
        if node.name not in dq_by_matmul_name:
            rewritten.append(node)
            continue
        input_fp16 = f"{node.input[0]}__for_{node.name.replace('/', '_')}_fp16"
        output_float = node.output[0]
        output_fp16 = f"{output_float}__fp16"
        rewritten.append(
            helper.make_node(
                "Cast",
                [node.input[0]],
                [input_fp16],
                name=f"{node.name}/CastInputToFp16",
                to=TensorProto.FLOAT16,
            )
        )
        node.input[0] = input_fp16
        node.output[0] = output_fp16
        rewritten.append(node)
        rewritten.append(
            helper.make_node(
                "Cast",
                [output_fp16],
                [output_float],
                name=f"{node.name}/CastOutputToFp32",
                to=TensorProto.FLOAT,
            )
        )

    retained = [value for value in graph.initializer if value.name not in removable_initializers]
    del graph.initializer[:]
    graph.initializer.extend(retained)
    graph.initializer.extend(fp16_initializers)
    del graph.node[:]
    graph.node.extend(rewritten)
    return {
        "matmuls_rewritten": len(dq_by_matmul_name),
        "layers_rewritten": sorted(selected_layers),
        "matmuls_kept_qdq": sorted(KEEP_QDQ_MATMUL_NAMES),
        "weight_elements": quantized_elements,
        "runtime_fp32_weight_bytes_avoided": quantized_elements * 4,
        "static_fp16_weight_bytes": quantized_elements * 2,
        "source_int8_weight_bytes_removed": quantized_elements,
        "estimated_live_weight_bytes_avoided": quantized_elements * 3,
    }


def feeds(session: ort.InferenceSession, past_length: int, *, prefill: bool) -> dict[str, np.ndarray]:
    rng = np.random.default_rng(20260717 + past_length)
    result: dict[str, np.ndarray] = {}
    for value in session.get_inputs():
        if value.name == "vision_embeds":
            length = VISION_TOKENS if prefill else 0
            result[value.name] = rng.normal(0, 0.2, [1, length, 512]).astype(np.float32)
        elif value.name == "token_ids":
            result[value.name] = np.array([[1 if prefill else 4]], dtype=np.int32)
        elif value.name == "position_ids":
            result[value.name] = (
                np.arange(VISION_TOKENS + 1, dtype=np.int32)[None, :]
                if prefill
                else np.array([[past_length]], dtype=np.int32)
            )
        else:
            result[value.name] = rng.normal(0, 0.02, [1, 2, past_length, 64]).astype(np.float32)
    return result


def validate_cpu(baseline: Path, candidate: Path) -> dict:
    baseline_session = ort.InferenceSession(str(baseline), providers=["CPUExecutionProvider"])
    candidate_session = ort.InferenceSession(str(candidate), providers=["CPUExecutionProvider"])
    checks = {}
    for label, past_length, prefill in (
        ("prefill", 0, True),
        ("step_257", 257, False),
        ("step_383", 383, False),
    ):
        model_feeds = feeds(baseline_session, past_length, prefill=prefill)
        expected = baseline_session.run(None, model_feeds)
        actual = candidate_session.run(None, model_feeds)
        max_abs = [float(np.max(np.abs(left - right))) for left, right in zip(expected, actual)]
        checks[label] = {
            "logits_max_abs": max_abs[0],
            "all_outputs_max_abs": max(max_abs),
            "top_token_baseline": int(expected[0][0, -1].argmax()),
            "top_token_candidate": int(actual[0][0, -1].argmax()),
        }
        if checks[label]["top_token_baseline"] != checks[label]["top_token_candidate"]:
            raise RuntimeError(f"CPU top-token parity failed for {label}: {checks[label]}")
    return checks


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--baseline", type=Path, default=BASELINE)
    parser.add_argument("--source", type=Path, default=DEFAULT_SOURCE)
    parser.add_argument("--destination", type=Path, default=DEFAULT_DESTINATION)
    parser.add_argument("--report", type=Path, default=DEFAULT_REPORT)
    parser.add_argument(
        "--layers",
        default="1,2,3",
        help="Comma-separated decoder layers whose seven MatMuls execute in FP16",
    )
    arguments = parser.parse_args()
    arguments.destination.parent.mkdir(parents=True, exist_ok=True)
    arguments.report.parent.mkdir(parents=True, exist_ok=True)

    selected_layers = {int(value) for value in arguments.layers.split(",") if value != ""}
    if not selected_layers.issubset(set(range(6))):
        raise ValueError(f"Invalid decoder layers {sorted(selected_layers)}")
    model = onnx.load(arguments.source)
    optimization = rewrite_matmul_weights(model, selected_layers)
    model.producer_name = "vibe-manga-baberu-webgpu-static-fp16-matmul"
    model.producer_version = "1"
    onnx.checker.check_model(model)
    onnx.save(model, arguments.destination)
    parity = validate_cpu(arguments.baseline, arguments.destination)
    report = {
        "source": {
            "path": str(arguments.source.relative_to(ROOT)),
            "bytes": arguments.source.stat().st_size,
            "sha256": sha256(arguments.source),
        },
        "optimized": {
            "path": str(arguments.destination.relative_to(ROOT)),
            "bytes": arguments.destination.stat().st_size,
            "sha256": sha256(arguments.destination),
        },
        "capability": {
            "layers": 6,
            "hidden_size": 512,
            "kv_heads": 2,
            "vocabulary": 14630,
            "max_new_tokens": 128,
            "architecture_changed": False,
        },
        "static_fp16_matmul": optimization,
        "cpu_validation": parity,
    }
    arguments.report.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
    print(json.dumps(report, indent=2))


if __name__ == "__main__":
    main()