TiGa-RCE's picture
Finalize reproducibility bundle and locked revisions
f3a14ee verified
Raw
History Blame Contribute Delete
12.2 kB
#!/usr/bin/env python3
"""Portable entry point for the published embedding quantization evidence."""
from __future__ import annotations
import argparse
import hashlib
import json
import subprocess
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parent
LOCK_PATH = ROOT / "models.lock.json"
DATASET_PATH = ROOT / "inputs/retrieval_pairs.json"
EVAL_SCRIPT = ROOT / "scripts/frozen-original/qwen3_embedding_retrieval_smoke.py"
MIXED_SCRIPT = ROOT / "scripts/frozen-original/mixed_index_compatibility.py"
FAMILIES = (
"qwen3-embedding-0.6b",
"gte-qwen2-1.5b",
"qwen3-embedding-8b",
)
VARIANTS = ("bf16", "q4", "oq4", "oq4e", "q6", "oq6", "oq6e", "q8", "oq8", "oq8e")
QUICK_VARIANTS = ("bf16", "q4", "oq4e", "q6", "oq6e", "q8", "oq8e")
CUDA_VARIANTS = ("bf16", "bnb-int8", "bnb-nf4")
DEFAULT_SPACE = "TiGa-RCE/embedding-quantization-cuda-control"
def run(command: list[str]) -> None:
print("+", " ".join(command), flush=True)
subprocess.run(command, check=True)
def load_lock(require_complete: bool = True) -> dict:
data = json.loads(LOCK_PATH.read_text())
records = data.get("models", [])
if len(records) != 30:
raise RuntimeError(f"models.lock.json contains {len(records)} records, expected 30")
if require_complete:
incomplete = [item["repo_id"] for item in records if not item.get("revision")]
if incomplete:
raise RuntimeError(
"model publication is incomplete; missing locked Hub revisions: "
+ ", ".join(incomplete)
)
return data
def model_records(family: str, profile: str) -> list[dict]:
variants = QUICK_VARIANTS if profile == "quick" else VARIANTS
records = {
(item["family"], item["variant"]): item
for item in load_lock()["models"]
}
return [records[(family, variant)] for variant in variants]
def mlx_command(args: argparse.Namespace) -> None:
try:
from huggingface_hub import snapshot_download
except ImportError as error:
raise RuntimeError("install the MLX environment with: uv sync --extra mlx") from error
if sys.platform != "darwin":
raise RuntimeError("the MLX reproduction lane requires macOS on Apple Silicon")
output_root = Path(args.output).resolve()
model_root = Path(args.models).resolve()
families = FAMILIES if args.family == "all" else (args.family,)
for family in families:
records = model_records(family, args.profile)
quality_dir = output_root / family / "quality"
quality_dir.mkdir(parents=True, exist_ok=True)
for record in records:
variant = record["variant"]
model_dir = model_root / family / variant
snapshot_download(
repo_id=record["repo_id"],
revision=record["revision"],
local_dir=model_dir,
)
vectors = quality_dir / f"{variant}.npz"
if args.force or not vectors.exists():
run([
sys.executable,
str(EVAL_SCRIPT),
"encode",
"--model",
str(model_dir),
"--dataset",
str(DATASET_PATH),
"--output",
str(vectors),
])
reference = quality_dir / "bf16.npz"
for record in records:
variant = record["variant"]
if variant == "bf16":
continue
run([
sys.executable,
str(EVAL_SCRIPT),
"compare",
"--reference",
str(reference),
"--candidate",
str(quality_dir / f"{variant}.npz"),
"--output",
str(quality_dir / f"compare-{variant}.json"),
])
def mixed_command(args: argparse.Namespace) -> None:
results_root = Path(args.results_root).resolve()
output = Path(args.output).resolve()
run([
sys.executable,
str(MIXED_SCRIPT),
"--results-root",
str(results_root),
"--output",
str(output),
])
def cuda_command(args: argparse.Namespace) -> None:
try:
from gradio_client import Client
from huggingface_hub import get_token
except ImportError as error:
raise RuntimeError(
"install the ZeroGPU client environment with: uv sync --extra zerogpu"
) from error
client = Client(args.space, hf_token=get_token())
families = FAMILIES if args.family == "all" else (args.family,)
variants = CUDA_VARIANTS if args.variant == "all" else (args.variant,)
output_root = Path(args.output).resolve()
output_root.mkdir(parents=True, exist_ok=True)
for family in families:
for variant in variants:
if family == "qwen3-embedding-0.6b":
if variant == "bf16":
result = client.predict(api_name="/run_bf16_control")
else:
result = client.predict(variant, api_name="/run_quantized_control")
elif variant == "bf16":
result = client.predict(family, api_name="/run_family_bf16_control")
else:
result = client.predict(
family,
variant,
api_name="/run_family_quantized_control",
)
if not isinstance(result, dict):
raise RuntimeError(
f"unexpected Space response for {family}/{variant}: "
f"{type(result).__name__}"
)
if "diagnostic_error" in result:
raise RuntimeError(
f"Space control failed for {family}/{variant}: "
f"{result['diagnostic_error']}: {result.get('diagnostic_message', '')}"
)
destination = output_root / family / f"cuda-{variant}.json"
destination.parent.mkdir(parents=True, exist_ok=True)
destination.write_text(json.dumps(result, indent=2) + "\n")
print(destination)
def sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for block in iter(lambda: handle.read(1024 * 1024), b""):
digest.update(block)
return digest.hexdigest()
def verify_command(_: argparse.Namespace) -> None:
load_lock()
checksum_path = ROOT / "checksums.sha256"
if not checksum_path.exists():
raise RuntimeError("checksums.sha256 is missing")
failures = []
for line in checksum_path.read_text().splitlines():
if not line.strip():
continue
expected, relative = line.split(" ", 1)
path = ROOT / relative
if not path.is_file():
failures.append(f"missing: {relative}")
elif sha256(path) != expected:
failures.append(f"checksum: {relative}")
quality_root = ROOT / "local-results/full-q4-q6-q8"
expected_pass = {
"qwen3-embedding-0.6b": {"q8", "oq8", "oq8e"},
"gte-qwen2-1.5b": {"q6", "oq6", "oq6e", "q8", "oq8", "oq8e"},
"qwen3-embedding-8b": {"q6", "oq6", "oq6e", "q8", "oq8", "oq8e"},
}
for family in FAMILIES:
for variant in VARIANTS:
path = quality_root / family / "quality" / f"{variant}.npz"
if not path.is_file():
failures.append(f"missing result: {path.relative_to(ROOT)}")
if variant != "bf16":
comparison_path = quality_root / family / "quality" / f"compare-{variant}.json"
if not comparison_path.is_file():
failures.append(f"missing comparison: {comparison_path.relative_to(ROOT)}")
continue
passed = bool(json.loads(comparison_path.read_text())["smoke_gate"]["passed"])
if passed != (variant in expected_pass[family]):
failures.append(f"unexpected MLX gate: {family}/{variant}={passed}")
mixed = json.loads((ROOT / "local-results/mixed-index-compatibility.json").read_text())
if set(mixed.get("families", {})) != set(FAMILIES):
failures.append("mixed-index result does not contain all three families")
direction_count = 0
maximum_false_negatives = 0
for family in mixed.get("families", {}).values():
for candidate in family["candidates"].values():
for direction in candidate["directions"].values():
direction_count += 1
if direction["retrieval"]["top1"] != 1.0:
failures.append("a mixed-index direction did not retain Top-1")
maximum_false_negatives = max(
maximum_false_negatives,
int(direction["fixed_threshold"]["false_negative"]),
)
if direction_count != 54:
failures.append(f"mixed-index direction count is {direction_count}, expected 54")
if maximum_false_negatives != 5:
failures.append(
f"maximum mixed-index false negatives is {maximum_false_negatives}, expected 5"
)
cuda_expectations = {
"qwen3-embedding-0.6b": {"bnb-int8": True, "bnb-nf4": False},
"gte-qwen2-1.5b": {"bnb-int8": False, "bnb-nf4": False},
"qwen3-embedding-8b": {"bnb-int8": True, "bnb-nf4": False},
}
for family, variants in cuda_expectations.items():
for variant, expected in variants.items():
result = json.loads(
(ROOT / "cloud-results" / family / f"cuda-{variant}.json").read_text()
)
if result["metrics"]["top1"] != 1.0:
failures.append(f"CUDA Top-1 changed: {family}/{variant}")
if family == "gte-qwen2-1.5b" and variant == "bnb-int8":
corrected = json.loads(
(ROOT / "cloud-results" / family / "cuda-bnb-int8-corrected-comparison.json").read_text()
)
cosine = corrected["minimum_aligned_cosine_int8_vs_direct_bf16"]
else:
cosine = result["cuda_bf16_comparison"]["minimum_aligned_cosine_vs_cuda_bf16"]
if (cosine >= 0.99) != expected:
failures.append(f"unexpected CUDA gate: {family}/{variant} cosine={cosine}")
if failures:
raise RuntimeError("bundle verification failed:\n- " + "\n- ".join(failures))
print(
"PASS: locked revisions, checksums, 30 MLX vectors and gates, "
"54 mixed-index directions, and six CUDA controls verified"
)
def parser() -> argparse.ArgumentParser:
root = argparse.ArgumentParser(description=__doc__)
commands = root.add_subparsers(dest="command", required=True)
mlx = commands.add_parser("mlx", help="download locked MLX models and rerun the gate")
mlx.add_argument("--profile", choices=("quick", "full"), default="quick")
mlx.add_argument("--family", choices=("all",) + FAMILIES, default="qwen3-embedding-0.6b")
mlx.add_argument("--models", default="./work/models")
mlx.add_argument("--output", default="./work/mlx")
mlx.add_argument("--force", action="store_true")
mlx.set_defaults(func=mlx_command)
mixed = commands.add_parser("mixed-index", help="rerun the frozen migration test")
mixed.add_argument(
"--results-root",
default=str(ROOT / "local-results/full-q4-q6-q8"),
)
mixed.add_argument("--output", default="./work/mixed-index-compatibility.json")
mixed.set_defaults(func=mixed_command)
cuda = commands.add_parser("cuda", help="invoke the CUDA controls on ZeroGPU")
cuda.add_argument("--family", choices=("all",) + FAMILIES, default="qwen3-embedding-0.6b")
cuda.add_argument("--variant", choices=("all",) + CUDA_VARIANTS, default="all")
cuda.add_argument("--space", default=DEFAULT_SPACE)
cuda.add_argument("--output", default="./work/cuda")
cuda.set_defaults(func=cuda_command)
verify = commands.add_parser("verify", help="verify the published evidence bundle")
verify.set_defaults(func=verify_command)
return root
if __name__ == "__main__":
arguments = parser().parse_args()
arguments.func(arguments)