ONNX
onnxruntime
onnx-mlir
quantization
fp32
File size: 16,352 Bytes
ed3aeeb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
#!/usr/bin/env python3
"""Independently validate all non-AD01 T60 batch records and artifacts."""

from __future__ import annotations

import argparse
import json
import shlex
import sys
import re
from collections import Counter
from pathlib import Path
from typing import Any

REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
    sys.path.insert(0, str(REPO_ROOT))

from scripts import run_mlir_batch as batch


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--report",
        type=Path,
        default=batch.BATCH_LOG_DIR / "validation_report.json",
    )
    parser.add_argument(
        "--artifact-manifest",
        type=Path,
        default=batch.BATCH_LOG_DIR / "artifact_manifest.json",
    )
    return parser.parse_args()


def add_error(errors: list[str], message: str) -> None:
    errors.append(message)


def main() -> int:
    args = parse_args()
    rows = batch.eligible_rows()
    expected_ids = {row["model_id"] for row in rows}
    errors: list[str] = []
    config_count = 0
    result_count = 0
    variant_count = 0
    stage_count = 0
    command_count = 0
    log_count = 0
    artifact_records: dict[str, dict[str, Any]] = {}
    required = Counter()
    variants = Counter()
    failures = Counter()
    optional_tosa = Counter()
    optional_stablehlo = Counter()
    import_compiler_attempts = 0
    prerequisite_attempts = 0
    forbidden_operations = 0
    command_policy_passes = 0
    quant_source_representation = Counter()
    quant_onnx_preservation = Counter()
    quant_low_level = Counter()
    required_failure_signatures = Counter()
    blockers: list[dict[str, Any]] = []

    for model_id in sorted(expected_ids):
        config_path = batch.CONFIG_DIR / f"{model_id}_mlir.json"
        if not config_path.is_file():
            add_error(errors, f"{model_id}: config missing: {config_path}")
            continue
        config = json.loads(config_path.read_text())
        try:
            batch.validate(batch.CONFIG_SCHEMA, config)
        except Exception as error:
            add_error(errors, f"{model_id}: config schema: {error}")
            continue
        config_count += 1
        result_path = batch.resolve_config_path(config["model_dir"]) / "mlir" / "mlir_batch_run_result.json"
        if not result_path.is_file():
            add_error(errors, f"{model_id}: result missing: {result_path}")
            continue
        result = json.loads(result_path.read_text())
        try:
            batch.validate(batch.RESULT_SCHEMA, result)
        except Exception as error:
            add_error(errors, f"{model_id}: result schema: {error}")
            continue
        result_count += 1
        if result.get("config_sha256") != batch.sha256(config_path):
            add_error(errors, f"{model_id}: config checksum mismatch")
        if result.get("model_id") != model_id:
            add_error(errors, f"{model_id}: result model_id mismatch")
        forbidden_operations += len(result.get("forbidden_operations_performed", []))
        flattened_expected = sum(len(item.get("stages", [])) for item in result["variants"].values())
        if len(result.get("stages", [])) != flattened_expected:
            add_error(errors, f"{model_id}: flattened stage record count mismatch")

        for variant_name, variant in result["variants"].items():
            variant_count += 1
            variants[variant["status"]] += 1
            required[variant["required_path_status"]] += 1
            if variant.get("failure_code"):
                failures[variant["failure_code"]] += 1
            if variant_name == "public_quantized":
                quant = variant.get("quantization", {})
                quant_source_representation[quant.get("source", {}).get("representation", "MISSING")] += 1
                quant_onnx_preservation[quant.get("onnx_dialect_preservation", "BLOCKED")] += 1
                quant_low_level[quant.get("low_level_status", "BLOCKED")] += 1
            input_record = variant["input"]
            input_path = Path(input_record["path"])
            if input_record["exists"]:
                if not input_path.is_file() or batch.sha256(input_path) != input_record["sha256"]:
                    add_error(errors, f"{model_id}/{variant_name}: input checksum mismatch")
            if variant.get("input_integrity_unchanged") is False:
                add_error(errors, f"{model_id}/{variant_name}: input mutation detected")

            stages = variant["stages"]
            import_stage = next((item for item in stages if item["stage"] == "onnx_to_onnx_dialect"), None)
            if import_stage is None:
                add_error(errors, f"{model_id}/{variant_name}: no T60 prerequisite/import attempt")
            else:
                prerequisite_attempts += 1
                if import_stage.get("command_argv", [None])[0] == str(batch.ONNX_MLIR):
                    import_compiler_attempts += 1

            required_nonpass = next(
                (
                    item for item in stages
                    if not item.get("optional") and item.get("status") != "PASS"
                ),
                None,
            )
            if required_nonpass:
                stderr_path = Path(required_nonpass.get("stderr_log", ""))
                stderr_text = stderr_path.read_text(errors="replace") if stderr_path.is_file() else ""
                if required_nonpass.get("exit_code") == 134 and "ONNXDequantizeLinear" in stderr_text:
                    signature = "ONNXDequantizeLinear per-axis operand assertion (signal 6)"
                elif required_nonpass.get("exit_code") == 134 and "expected only ranked shapes" in stderr_text:
                    signature = "KrnlTypeConverter unranked tensor assertion (signal 6)"
                elif required_nonpass.get("exit_code") == 134 and "cast<Ty>() argument of incompatible type" in stderr_text:
                    signature = "Krnl lowering RankedTensorType cast assertion (signal 6)"
                else:
                    match = re.search(r"failed to legalize operation '([^']+)'", stderr_text)
                    if match:
                        signature = f"unlegalized {match.group(1)}"
                    elif required_nonpass.get("secondary_failure_code") == "MISSING_ONNX_PREREQUISITE":
                        signature = "missing T40 public-quantized ONNX prerequisite"
                    else:
                        signature = str(required_nonpass.get("failure_code") or "UNKNOWN")
                required_failure_signatures[signature] += 1
                if required_nonpass.get("status") == "BLOCKED":
                    blockers.append({
                        "model_id": model_id,
                        "variant": variant_name,
                        "stage": required_nonpass["stage"],
                        "failure_code": required_nonpass.get("failure_code"),
                        "secondary_failure_code": required_nonpass.get("secondary_failure_code"),
                        "stderr_log": required_nonpass.get("stderr_log"),
                        "resolution_condition": (
                            "A T40 exporter must produce the public quantized ONNX from the "
                            "unchanged public checkpoint, then checker/runtime/quantization-preservation "
                            "validation and checksum recording must pass. No self-quantization is allowed."
                        ),
                    })

            for stage in stages:
                stage_count += 1
                command_count += bool(stage.get("command"))
                command_argv = stage.get("command_argv", [])
                executable = Path(command_argv[0]).name if command_argv else ""
                if executable not in {"onnx-mlir", "onnx-mlir-opt", "mlir-opt", "test"}:
                    add_error(errors, f"{model_id}/{variant_name}/{stage['stage']}: unexpected executable {executable}")
                else:
                    command_policy_passes += 1
                command_text = " ".join(command_argv).lower()
                forbidden_tokens = (
                    "representative_dataset", "calibration", "quantize_static",
                    "quantize_dynamic", "fine_tune", "finetune", "optimizer.step",
                    "backward()",
                )
                if any(token in command_text for token in forbidden_tokens):
                    add_error(errors, f"{model_id}/{variant_name}/{stage['stage']}: forbidden command token")
                if stage.get("failure_code"):
                    failures[stage["failure_code"]] += 1
                for key in ("stdout_log", "stderr_log", "resource_log", "command_log"):
                    path_value = stage.get(key)
                    if not path_value or not Path(path_value).is_file():
                        add_error(errors, f"{model_id}/{variant_name}/{stage['stage']}: missing {key}")
                    else:
                        log_count += 1
                for record in stage.get("inputs", []) + stage.get("outputs", []):
                    if not record.get("exists"):
                        continue
                    path = Path(record["path"])
                    if not path.is_file():
                        add_error(errors, f"{model_id}/{variant_name}/{stage['stage']}: artifact missing {path}")
                    elif batch.sha256(path) != record.get("sha256") or path.stat().st_size != record.get("bytes"):
                        add_error(errors, f"{model_id}/{variant_name}/{stage['stage']}: artifact checksum/size mismatch {path}")

            for name, record in variant.get("artifacts", {}).items():
                if not record.get("exists"):
                    continue
                path = Path(record["path"])
                actual = batch.file_record(path)
                if actual != record:
                    add_error(errors, f"{model_id}/{variant_name}/{name}: canonical artifact record mismatch")
                    continue
                prior = artifact_records.get(str(path))
                if prior and prior != record:
                    add_error(errors, f"{model_id}/{variant_name}/{name}: conflicting artifact record")
                artifact_records[str(path)] = record

            routes = variant.get("optional_routes", {})
            if "tosa" in routes:
                optional_tosa[routes["tosa"]["status"]] += 1
            if "stablehlo" in routes:
                optional_stablehlo[routes["stablehlo"]["status"]] += 1

            if variant["required_path_status"] == "PASS":
                suffix = "fp32" if variant_name == "fp32" else "quantized"
                ir_root = batch.resolve_config_path(config["model_dir"]) / "mlir" / suffix
                for filename in ("onnx.mlir", "krnl.mlir", "affine_scf_memref.mlir", "llvm.mlir"):
                    if not (ir_root / filename).is_file() or (ir_root / filename).stat().st_size == 0:
                        add_error(errors, f"{model_id}/{variant_name}: PASS required route lacks {filename}")

        mlir_root = batch.resolve_config_path(config["model_dir"]) / "mlir"
        supporting = {
            "central_config": config_path,
            "model_config": (
                batch.resolve_config_path(config["model_dir"])
                / "config"
                / "mlir_batch_config.json"
            ),
            "result": result_path,
            "stage_matrix": mlir_root / "mlir_stage_matrix.json",
            "quantization_preservation": mlir_root / "mlir_quantization_preservation.json",
            "artifact_manifest": mlir_root / "artifact_manifest.json",
            "conversion_log": mlir_root / "mlir_conversion.log",
        }
        for name, path in supporting.items():
            if not path.is_file() or path.stat().st_size == 0:
                add_error(errors, f"{model_id}: supporting artifact missing/empty {name}: {path}")
                continue
            record = batch.file_record(path)
            artifact_records[str(path)] = record
            if path.suffix == ".json":
                try:
                    json.loads(path.read_text())
                except (OSError, json.JSONDecodeError) as error:
                    add_error(errors, f"{model_id}: invalid JSON {name}: {error}")
        local_config = supporting["model_config"]
        if local_config.is_file() and json.loads(local_config.read_text()) != config:
            add_error(errors, f"{model_id}: model-local config differs from central config")
        model_manifest = supporting["artifact_manifest"]
        if model_manifest.is_file():
            document = json.loads(model_manifest.read_text())
            if len(document.get("artifacts", [])) != 8:
                add_error(errors, f"{model_id}: required artifact manifest must contain 8 records")
            for record in document.get("artifacts", []):
                if not record.get("exists"):
                    continue
                actual = batch.file_record(Path(record["path"]))
                expected = {key: record.get(key) for key in ("path", "exists", "sha256", "bytes")}
                if actual != expected:
                    add_error(errors, f"{model_id}: per-model manifest mismatch {record['path']}")

    if expected_ids != {path.stem.split("_", 1)[0] for path in batch.CONFIG_DIR.glob("*_mlir.json")}:
        add_error(errors, "config model ID set does not exactly match 21 eligible models excluding AD01")
    if "AD01" in expected_ids:
        add_error(errors, "AD01 was not excluded")
    if forbidden_operations:
        add_error(errors, f"forbidden operation records present: {forbidden_operations}")

    manifest = {
        "schema_version": "1.0",
        "generated_at": batch.utc_now(),
        "artifact_count": len(artifact_records),
        "artifacts": [artifact_records[path] for path in sorted(artifact_records)],
    }
    batch.write_json(args.artifact_manifest, manifest)
    report = {
        "schema_version": "1.0",
        "generated_at": batch.utc_now(),
        "status": "PASS" if not errors else "FAIL",
        "errors": errors,
        "expected_model_count": 21,
        "config_schema_pass_count": config_count,
        "result_schema_pass_count": result_count,
        "variant_count": variant_count,
        "variant_status_counts": dict(sorted(variants.items())),
        "required_route_status_counts": dict(sorted(required.items())),
        "failure_code_counts": dict(sorted(failures.items())),
        "optional_tosa_route_status_counts": dict(sorted(optional_tosa.items())),
        "optional_stablehlo_route_status_counts": dict(sorted(optional_stablehlo.items())),
        "quant_source_representation_counts": dict(sorted(quant_source_representation.items())),
        "quant_onnx_dialect_preservation_counts": dict(sorted(quant_onnx_preservation.items())),
        "quant_low_level_status_counts": dict(sorted(quant_low_level.items())),
        "required_failure_signature_counts": dict(sorted(required_failure_signatures.items())),
        "blockers": blockers,
        "t60_prerequisite_attempt_count": prerequisite_attempts,
        "actual_onnx_mlir_import_command_count": import_compiler_attempts,
        "stage_record_count": stage_count,
        "exact_command_record_count": command_count,
        "command_policy_pass_count": command_policy_passes,
        "log_file_validation_count": log_count,
        "canonical_artifact_checksum_pass_count": len(artifact_records),
        "artifact_manifest": str(args.artifact_manifest),
        "artifact_manifest_sha256": batch.sha256(args.artifact_manifest),
        "forbidden_operations_performed": forbidden_operations,
        "ad01_excluded": "AD01" not in expected_ids,
        "registry_sha256": batch.sha256(batch.REGISTRY),
        "toolchain_lock_sha256": batch.sha256(batch.TOOLCHAIN_LOCK),
        "command_argv": [sys.executable, *sys.argv],
        "command": shlex.join([sys.executable, *sys.argv]),
        "working_directory": str(batch.REPO_ROOT),
    }
    batch.write_json(args.report, report)
    print(json.dumps(report, indent=2, sort_keys=True))
    return 0 if not errors else 1


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