p2o 0.1.1

A PaddlePaddle New IR (PIR) to ONNX model converter.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159

#!/usr/bin/env python3
"""Probe internal Paddle tensors by temporarily retargeting fetch ops.

This is a maintained diagnostic helper for regression triage, not a one-off
spike. It reuses regression_suite input generation/run helpers so CPU/GPU or
Paddle/ONNX discrepancies can be localized to selected tensor ids.
"""

from __future__ import annotations

import argparse
import json
import tempfile
from dataclasses import replace
from pathlib import Path
from typing import Any

from regression import infer_input_info, load_manifest, make_input_tensors, run_paddle


REPO_ROOT = Path(__file__).resolve().parents[1]


def tensor_meta_by_id(ops: list[dict[str, Any]]) -> dict[int, dict[str, Any]]:
    meta: dict[int, dict[str, Any]] = {}
    for op in ops:
        outputs = op.get("O", [])
        if isinstance(outputs, dict):
            outputs = [outputs]
        for output in outputs:
            if isinstance(output, dict) and "%" in output:
                meta[int(output["%"])] = output
    return meta


def rewrite_fetch_model(model_dir: Path, tensor_ids: list[int], out_dir: Path) -> None:
    model_json = model_dir / "inference.json"
    model_params = model_dir / "inference.pdiparams"
    data = json.loads(model_json.read_text())
    ops = data["program"]["regions"][0]["blocks"][0]["ops"]
    fetch_indices = [idx for idx, op in enumerate(ops) if op.get("#") == "1.fetch"]
    if len(tensor_ids) > len(fetch_indices):
        raise ValueError(f"Requested {len(tensor_ids)} tensors, but model only has {len(fetch_indices)} fetch ops")

    tensor_meta = tensor_meta_by_id(ops)
    for fetch_idx, tensor_id in zip(fetch_indices, tensor_ids):
        if tensor_id not in tensor_meta:
            raise ValueError(f"Tensor %{tensor_id} not found in inference.json")
        fetch_op = ops[fetch_idx]
        fetch_op["I"] = [{"%": tensor_id}]
        fetch_output = fetch_op["O"]
        if isinstance(fetch_output, list) and fetch_output and isinstance(fetch_output[0], dict):
            fetch_output[0]["TT"] = tensor_meta[tensor_id]["TT"]

    out_dir.mkdir(parents=True, exist_ok=True)
    (out_dir / "inference.json").write_text(json.dumps(data, separators=(",", ":")))
    target = out_dir / "inference.pdiparams"
    if target.exists() or target.is_symlink():
        target.unlink()
    target.symlink_to(model_params)


def summarize_diff(reference, candidate) -> dict[str, Any]:
    import numpy as np

    ref = np.asarray(reference)
    cand = np.asarray(candidate)
    if ref.shape != cand.shape:
        return {
            "shape_match": False,
            "reference_shape": list(ref.shape),
            "candidate_shape": list(cand.shape),
            "reference_dtype": str(ref.dtype),
            "candidate_dtype": str(cand.dtype),
        }

    if np.issubdtype(ref.dtype, np.integer) or np.issubdtype(ref.dtype, np.bool_):
        diff = ref.astype(np.int64, copy=False) - cand.astype(np.int64, copy=False)
    else:
        diff = ref.astype(np.float64, copy=False) - cand.astype(np.float64, copy=False)
    abs_diff = np.abs(diff)
    ref_abs_mean = float(np.mean(np.abs(ref.astype(np.float64, copy=False)))) if ref.size else 0.0
    return {
        "shape_match": True,
        "reference_shape": list(ref.shape),
        "candidate_shape": list(cand.shape),
        "reference_dtype": str(ref.dtype),
        "candidate_dtype": str(cand.dtype),
        "max_abs_diff": float(np.max(abs_diff)) if abs_diff.size else 0.0,
        "mean_abs_diff": float(np.mean(abs_diff)) if abs_diff.size else 0.0,
        "reference_mean_abs": ref_abs_mean,
        "relative_mean_abs_diff": float(np.mean(abs_diff) / (ref_abs_mean + 1e-12)) if abs_diff.size else 0.0,
        "reference_nan_count": int(np.isnan(ref).sum()) if np.issubdtype(ref.dtype, np.floating) else 0,
        "candidate_nan_count": int(np.isnan(cand).sum()) if np.issubdtype(cand.dtype, np.floating) else 0,
    }


def main() -> int:
    parser = argparse.ArgumentParser(description="Probe internal Paddle tensors by rewriting fetch ops.")
    parser.add_argument("--manifest", type=Path, default=Path("scripts/regression_models.json"))
    parser.add_argument("--model", default="ppdoclayoutv3")
    parser.add_argument("--seed", type=int, default=20260412)
    parser.add_argument("--devices", nargs=2, default=["cpu", "gpu"])
    parser.add_argument("--tensors", nargs="+", type=int, required=True)
    parser.add_argument("--json-out", type=Path)
    args = parser.parse_args()

    specs = load_manifest(args.manifest)
    if args.model not in specs:
        raise ValueError(f"Unknown model '{args.model}'")
    model_spec = specs[args.model]

    input_info = infer_input_info(model_spec)
    inputs = make_input_tensors(model_spec, input_info, args.seed)

    with tempfile.TemporaryDirectory(prefix=f"{args.model}_probe_", dir=REPO_ROOT / ".tmp_regression") as tmp_raw:
        tmp_dir = Path(tmp_raw)
        rewrite_fetch_model(model_spec.model_dir, args.tensors, tmp_dir)
        probe_spec = replace(model_spec, model_dir=tmp_dir)

        outputs_by_device: dict[str, list[Any]] = {}
        for device in args.devices:
            outputs_by_device[device] = run_paddle(probe_spec, inputs, device)

    reference_name, candidate_name = args.devices
    results = []
    for index, tensor_id in enumerate(args.tensors):
        results.append(
            {
                "tensor_id": tensor_id,
                reference_name: {
                    "shape": list(outputs_by_device[reference_name][index].shape),
                    "dtype": str(outputs_by_device[reference_name][index].dtype),
                },
                candidate_name: {
                    "shape": list(outputs_by_device[candidate_name][index].shape),
                    "dtype": str(outputs_by_device[candidate_name][index].dtype),
                },
                "diff": summarize_diff(
                    outputs_by_device[reference_name][index],
                    outputs_by_device[candidate_name][index],
                ),
            }
        )

    report = {
        "model": args.model,
        "seed": args.seed,
        "devices": args.devices,
        "results": results,
    }
    if args.json_out is not None:
        args.json_out.write_text(json.dumps(report, indent=2))
    print(json.dumps(report, indent=2))
    return 0


if __name__ == "__main__":
    raise SystemExit(main())