File size: 17,177 Bytes
6fad3f9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 | """Reviewed Kimi K3 native self-benchmark import contract.
DCP overlaps the TP group and is recorded explicitly in the v6 identity.
The reviewed original and sanitized source hashes bind every imported point.
"""
from __future__ import annotations
import hashlib
import json
import math
from pathlib import Path
import pyarrow as pa
import pyarrow.parquet as pq
POLICY = "dynamo-selfbench-kimi-k3-20260916-v1"
CAMPAIGN = "kimi-k3-selfbench-20260916"
SCHEMA_NAME = "aic_fpm_forward_perf"
SCHEMA_VERSION = 6
METHOD = "dynamo_native_self_benchmark"
LEAF = "data/moonshotai--Kimi-K3/gb300/vllm/0.29.0/tp8-dcp8"
# Original SHA-256, Dynamo revision, row count, reviewed sanitized SHA-256.
SOURCES = {
"2201004": (
"9e902768fc79f03aaa78554baf05a9ef3a7a47eae08ca2460a0109dec81bb010",
"54174204af65eec5661fb1ba451c19012dac9306",
451,
"b1d5ba27b6b6a643cdfe1f8c052d249707722077b9d75d1a0767141da19b90a3",
),
"2209472": (
"da7a7611d9ee166c1f16dde5610e42cfa2f94f4284078ac28d0418899223c2e9",
"e8382179d91636626e7844b2da9f99066af2f639",
288,
"38e875570738119c8f6b13d5ed76c2b960d85341e7e36d6d04e7afbdc0dbe1ae",
),
}
IDENTITY = {
"model_id": "moonshotai/Kimi-K3",
"model_revision": None,
"system": "gb300",
"gpu_family": "GB300",
"framework": "vllm",
"framework_version": "0.29.0",
"parallelism": "tp8-dcp8",
"parallel_strategy": "pure_tp",
"tp": 8,
"pp": 1,
"dp": 1,
"moe_tp": 8,
"moe_ep": 1,
"cp": 1,
"dcp": 8,
"physical_gpu_count": 8,
"weight_quantization": "mxfp4",
"kv_cache_dtype": "fp8",
"collection_method": METHOD,
}
# The sidecar uses canonical FPM fields, not the richer manifest identity.
FPM_SELECTOR = {
"model_path": IDENTITY["model_id"],
"system": IDENTITY["system"],
"backend": IDENTITY["framework"],
"backend_version": IDENTITY["framework_version"],
**{
key: IDENTITY[key]
for key in (
"weight_quantization",
"kv_cache_dtype",
"parallel_strategy",
"tp",
"pp",
"dp",
"moe_tp",
"moe_ep",
"cp",
"dcp",
)
},
}
def read(path: Path):
return json.loads(path.read_text())
def sha(path: Path):
return hashlib.sha256(path.read_bytes()).hexdigest()
def require(condition, message):
if not condition:
raise ValueError(message)
def checked_file(root: Path, record):
path = root / record["path"]
require(
path.resolve().is_relative_to(root.resolve()), "provenance path escapes dataset"
)
require(sha(path) == record["sha256"], f"source hash mismatch: {path}")
return path
def arrow_schema(root):
spec = read(root / "schemas/aisim-fpm-forward-v6.schema.json")
types = {
"string": pa.string(),
"int64": pa.int64(),
"float64": pa.float64(),
"bool": pa.bool_(),
}
return pa.schema(
[
pa.field(f["name"], types[f["arrow_type"]], nullable=f["nullable"])
for f in spec["fields"]
]
)
def classify(point, regime):
rows = point.get("rows")
balanced = not point.get("partition") and (
not rows or len({tuple(r) for r in rows}) == 1
)
if regime == "fake_attention_kv_random_kda":
return "synthetic_attention_boundary"
return "primary" if balanced else "nonuniform_prefill"
def canonical_row(source, entry):
point = entry["point"]
fpm = entry["fpms"][0]
repeats = fpm.get("kvwarm_giant_median_of", 1)
return {
"cell_id": "kimi-k3-gb300-vllm-0.29.0-tp8-dcp8",
"model_path": IDENTITY["model_id"],
"system": "gb300",
"backend": "vllm",
"backend_version": "0.29.0",
"weight_quantization": "mxfp4",
"gemm_quant_mode": "bfloat16",
"moe_quant_mode": "w4a16_mxfp4",
"fmha_quant_mode": None,
"fmha_resolution": "runtime_not_resolved",
"comm_quant_mode": None,
"kv_cache_dtype": "fp8",
"parallel_strategy": "pure_tp",
"tp": 8,
"pp": 1,
"dp": 1,
"moe_tp": 8,
"moe_ep": 1,
"cp": 1,
"dcp": 8,
"moe_backend": "auto",
"attention_backend": "FLASHINFER_MLA",
"enable_wideep": False,
"enable_eplb": False,
"workload_kind": point["point_type"],
"batch_size": point["batch_size"],
"total_prefill_tokens": point["total_prefill_tokens"],
"total_kv_read_tokens": point["total_kv_read_tokens"],
"partition_policy": "balanced_v1",
"kv_seed_regime": entry["kv_seed_regime"],
"latency_ms": fpm["wall_time"] * 1000,
"global_warmup_iterations": source["config"]["warmup_iterations"],
"warmup_repeats": None,
"measurement_repeats": repeats,
"measurement_policy": "kvwarm_median_of_3"
if repeats == 3
else "dynamo_native_single_sample_v1",
"model_support_level": None,
"model_template_id": None,
"model_template_version": None,
"aic_database_version": None,
"source_plan_sha256": None,
"collector_attempt_id": None,
"runtime_run_id": source["run_id"],
"runtime_grid_digest": source["grid_digest"],
}
def validate_sources(root: Path, provenance):
require(
provenance["policy_id"] == POLICY, "unreviewed self-benchmark import policy"
)
require(
provenance["source_kind"] == "self_benchmark", "wrong collection source kind"
)
require(provenance["collection_method"] == METHOD, "wrong collection method")
require(provenance["timing_scope"] == "scheduler_wall_time", "wrong latency scope")
require(
provenance["independent_evaluation_truth"] is False,
"calibration is not independent truth",
)
require(
{s["job_id"] for s in provenance["sources"]} == set(SOURCES),
"unexpected source set",
)
outputs = {
"primary": [],
"synthetic_attention_boundary": [],
"nonuniform_prefill": [],
}
seen = set()
for record in provenance["sources"]:
original_hash, commit, count, sanitized_hash = SOURCES[record["job_id"]]
require(
record["original_sha256"] == original_hash
and record["dynamo_commit"] == commit,
"unreviewed source revision",
)
source_path = checked_file(root, record)
require(
sha(source_path) == sanitized_hash,
f"source differs from reviewed sanitized bytes: {record['job_id']}",
)
data = read(source_path)
require(
data["status"] == "complete"
and data["valid"]
and data["usable"]
and data["timing_valid"],
"incomplete source",
)
require(
not data["error"] and not data["skipped_points"],
"source has skipped/failed points",
)
require(
len(data["results"]) == len(data["iteration_groups"]) == count,
"source count mismatch",
)
require(
data["coverage"]["completed_points"] == count, "source coverage mismatch"
)
require(
data["limits"]["max_num_running_reqs"] == 32
and data["limits"]["max_model_len"] == 1048576,
"engine limit mismatch",
)
require(
data["recurrent_state"]
== {
"initialization": "random",
"policy": "uniform-request-layer-rank-v1",
"uniform_bound": 0.01,
},
"unknown KDA state policy",
)
for entry, group in zip(data["results"], data["iteration_groups"]):
point = entry["point"]
require(
group["benchmark_id"] == point["benchmark_id"], "benchmark ID mismatch"
)
require(len(entry["fpms"]) == 1, "expected one attention-DP observation")
fpm = entry["fpms"][0]
phase, batch = point["point_type"], point["batch_size"]
require(fpm["dp_rank"] == 0, "unexpected attention-DP rank")
require(
phase in ("prefill", "decode") and 0 < batch <= 32,
"invalid phase or batch",
)
require(
math.isfinite(fpm["wall_time"]) and fpm["wall_time"] > 0,
"invalid latency",
)
require(
fpm.get("kvwarm_giant_median_of", 1) in (1, 3),
"unexpected repeat policy",
)
metrics = fpm["scheduled_requests"]
require(
metrics[f"num_{phase}_requests"] == batch, "scheduled batch mismatch"
)
require(
metrics["sum_prefill_tokens"] == point["total_prefill_tokens"],
"scheduled prefill mismatch",
)
kv_field = (
"sum_decode_kv_tokens" if phase == "decode" else "sum_prefill_kv_tokens"
)
require(
metrics[kv_field] == point["total_kv_read_tokens"],
"scheduled KV mismatch",
)
require(
group["point"] == point
and group["complete"]
and group["expected_dp_ranks"] == [0],
"invalid synchronization",
)
require(
group["rank_results"] == [{"dp_rank": 0, "fpms": entry["fpms"]}]
and group["wall_time"] == fpm["wall_time"],
"rank timing mismatch",
)
if phase == "decode":
require(
entry["kv_seed_regime"]
in ("real_attention_kv_random_kda", "fake_attention_kv_random_kda"),
"unknown decode KV regime",
)
else:
require(
entry["kv_seed_regime"] in ("not_applicable", "real_prefix"),
"unknown prefill KV regime",
)
key = (data["run_id"], point["benchmark_id"])
require(key not in seen, "duplicate source point")
seen.add(key)
classification = classify(point, entry["kv_seed_regime"])
if classification == "nonuniform_prefill":
require(
phase == "prefill" and point["rows"],
"nonuniform point needs explicit prefill rows",
)
require(
len(point["rows"]) == batch
and sum(r[0] for r in point["rows"])
== point["total_prefill_tokens"]
and sum(r[1] for r in point["rows"])
== point["total_kv_read_tokens"],
"nonuniform rows disagree with totals",
)
outputs[classification].append(
{
"runtime_run_id": data["run_id"],
"source_job_id": record["job_id"],
"result": entry,
}
)
else:
outputs[classification].append(canonical_row(data, entry))
require(
{k: len(v) for k, v in outputs.items()}
== {
"primary": 669,
"synthetic_attention_boundary": 14,
"nonuniform_prefill": 56,
},
"classification count mismatch",
)
for key in ("primary", "synthetic_attention_boundary"):
coords = [
tuple(
r[k]
for k in (
"workload_kind",
"batch_size",
"total_prefill_tokens",
"total_kv_read_tokens",
)
)
for r in outputs[key]
]
require(len(set(coords)) == len(coords), "duplicate canonical coordinates")
return outputs
def validate_snapshot(root: Path, manifest):
require(manifest["configuration_path"] == LEAF, "unreviewed self-benchmark leaf")
require(
all(manifest.get(k) == v for k, v in IDENTITY.items()),
"DCP configuration identity mismatch",
)
require(
manifest["tp"] * manifest["pp"] * manifest["dp"] * manifest["cp"]
== manifest["physical_gpu_count"],
"DCP must not multiply GPU count",
)
require(manifest["tp"] % manifest["dcp"] == 0, "DCP must divide TP")
require(
manifest["measurements"]["file_count"] == 0
and manifest["measurements"]["protocol_id"] is None,
"self-calibration cannot be independent measurement truth",
)
provenance_path = root / LEAF / "fpm/provenance/import.json"
provenance = read(provenance_path)
expected = validate_sources(root, provenance)
engine = read(root / LEAF / "fpm/provenance/engine-config.json")
require(
engine["physical_gpu_count"] == 8 and engine["model_runner"] == "V1",
"engine topology/runner mismatch",
)
require(
engine["enable_expert_parallel"] is False
and engine["speculative_decoding"] is False,
"unexpected EP or speculation",
)
for key, value in {
"tensor-parallel-size": "8",
"pipeline-parallel-size": "1",
"decode-context-parallel-size": "8",
"max-num-seqs": "32",
"max-model-len": "1048576",
"kv-cache-dtype": "fp8",
}.items():
require(engine["flags"][key] == value, f"engine flag mismatch: {key}")
require(
len(manifest["fpm"]) == 2, "expected primary and synthetic-boundary libraries"
)
claimed = {provenance_path.relative_to(root).as_posix()}
claimed.update(s["path"] for s in provenance["sources"])
for source in provenance["supporting_files"]:
checked_file(root, source)
claimed.add(source["path"])
found_roles = set()
for record in manifest["fpm"]:
data_path = checked_file(root, record)
metadata = read(root / record["metadata_path"])
category = metadata["sample_classification"]
require(
category in ("primary", "synthetic_attention_boundary"),
"invalid library class",
)
require(category not in found_roles, "duplicate library class")
found_roles.add(category)
require(
metadata["schema_name"] == SCHEMA_NAME
and metadata["schema_version"] == SCHEMA_VERSION,
"wrong DCP schema",
)
require(
metadata["import_policy"] == POLICY
and metadata["collection_method"] == METHOD
and metadata["source_kind"] == "self_benchmark",
"missing collection policy",
)
require(
metadata["provenance_sha256"] == sha(provenance_path),
"import provenance hash mismatch",
)
require(
record["role"] == ("primary" if category == "primary" else "comparator"),
"synthetic data promoted to primary",
)
table = pq.read_table(data_path)
require(
table.schema.remove_metadata().equals(arrow_schema(root)),
"physical DCP schema mismatch",
)
require(
table.to_pylist() == expected[category], "published rows differ from source"
)
require(
metadata["configuration_selector"] == FPM_SELECTOR,
"metadata identity mismatch",
)
nonuniform = read(root / LEAF / "fpm/provenance/nonuniform-prefill.json")
require(
nonuniform == expected["nonuniform_prefill"],
"nonuniform request layouts were changed",
)
return claimed
def validate_catalog_record(root: Path, record):
metadata = read(root / record["metadata_path"])
require(
record["schema_name"] == SCHEMA_NAME
and record["schema_version"] == SCHEMA_VERSION,
"unreviewed external FPM schema",
)
require(
record["configuration_path"] == LEAF
and all(record.get(k) == v for k, v in IDENTITY.items()),
"catalog DCP identity mismatch",
)
require(metadata["import_policy"] == POLICY, "catalog import policy mismatch")
require(
record.get("source_kind") == "self_benchmark"
and record.get("timing_scope") == "scheduler_wall_time",
"catalog method/timing mismatch",
)
manifest = read(root / LEAF / "manifest.json")
entry = next(item for item in manifest["fpm"] if item["path"] == record["path"])
for field in ("path", "metadata_path", "sha256", "row_count", "role", "phases"):
require(record[field] == entry[field], f"catalog/manifest mismatch: {field}")
catalog_row = next(
row
for row in pq.read_table(root / "catalog/fpm.parquet").to_pylist()
if row["path"] == record["path"]
)
require(
{k: v for k, v in catalog_row.items() if v is not None}
== {k: v for k, v in record.items() if v is not None},
"DCP catalog Parquet/JSON mismatch",
)
return SCHEMA_VERSION
|