File size: 12,856 Bytes
c83c617 | 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 | """Convert, gate, compare and package the independent V2 KV-cache experiment."""
import argparse
import json
import shutil
import sys
import time
from pathlib import Path
ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(ROOT / "src"))
import numpy as np
import torch
from vimeml.deployment import kv_cache as kv, v2
from vimeml.deployment.bundle import fresh_directory, read_json, write_json
from vimeml.deployment.validation import numeric_error
from vimeml.training.data import file_sha
from vimeml.training.evaluate_ime import common_prefix_length
def align(bundle, directory, reference, output, baseline=None):
runtime = kv.Runtime(bundle, directory)
old = v2.CoreMLLM(bundle, baseline) if baseline else None
errors, invariants, saved = [], [], {}
info = read_json(reference / "reference.json")
if info.get("format") != "vimeml_v2_mac_preparation_reference_v1" or info["checkpoint_step"] != runtime.metadata["checkpoint_step"]:
raise ValueError("Invalid frozen reference identity.")
if old and any(old.metadata[key] != runtime.metadata[key] for key in ("checkpoint_sha256", "tokenizer_sha256", "model")):
raise ValueError("Stateless baseline model/tokenizer identity differs.")
with np.load(reference / "logits.npz", allow_pickle=False) as arrays:
for name, case in info["cases"].items():
ids = arrays[name + "_input_ids"][0]
expected = old.model(torch.tensor([ids.tolist()])).numpy() if old else arrays[name + "_logits"]
patterns = [[len(ids)], [1] * len(ids)]
if len(ids) > 3: patterns.append([3, len(ids)-3])
for chunks in patterns:
parts, past, state = [], 0, None
for size in chunks:
logits, state = runtime.step(ids[past:past+size], state, past)
parts.append(logits); past += size
actual = np.concatenate(parts, axis=1)
valid = case["valid_prefix"]
error = numeric_error(expected[:, :valid], actual[:, :valid], 3e-4, 3e-4)
errors.append({"case": name, "chunks": chunks, **error})
if chunks == [len(ids)]: saved[name] = actual
last, _ = runtime.step(ids, last_only=True)
errors.append({"case": name, "last_only": True, **numeric_error(expected[:, -1:], last, 3e-4, 3e-4)})
for name in ("right_pad", "changed_future"):
invariants.append({"case": name, **numeric_error(saved["short"], saved[name][:, :16], 3e-4, 3e-4)})
# Immutable shared snapshots, stale suffix overwrite, and A/B/A branch
# order catch errors that fresh-prefill comparisons cannot detect.
prefix = arrays["short_input_ids"][0][:7]
_, state = runtime.step(prefix, last_only=True)
original = tuple(a.copy() for a in state)
for suffix in ([9], [47, 81, 102], [9]):
actual, _ = runtime.step(suffix, state, len(prefix))
expected = (old.model(torch.tensor([[*prefix.tolist(), *suffix]])).numpy() if old else
v2.BundleLM(bundle).model(torch.tensor([[*prefix.tolist(), *suffix]])).detach().numpy())[:, len(prefix):]
errors.append({"branch": suffix, **numeric_error(expected, actual, 3e-4, 3e-4)})
immutable = all(np.array_equal(a, b) for a, b in zip(state, original))
dirty = tuple(a.copy() for a in state)
for a in dirty: a[:, :, :, len(prefix):, :] = 999
a, _ = runtime.step([9], state, len(prefix))
b, _ = runtime.step([9], dirty, len(prefix))
invariants.append({"case": "unused_tail", **numeric_error(a, b, 0, 0)})
report = {"format": "vimeml_kv_alignment_v1", "passed": immutable and all(e["outside_tolerance"] == 0 for e in errors+invariants),
"coreml_manifest_sha256": runtime.metadata["coreml_manifest_sha256"], "model": runtime.metadata,
"comparison": "same-weight stateless Core ML" if old else "frozen Windows FP32",
"baseline_manifest_sha256": old.metadata["coreml_manifest_sha256"] if old else None,
"tolerances": {"atol": 3e-4, "rtol": 3e-4}, "immutable_branch_inputs": immutable,
"errors": errors, "invariants": invariants}
fresh_directory(output); write_json(output / "alignment.json", report)
return report
class CachedForward:
"""Evaluation adapter: share exact batch token prefixes within one call."""
def __init__(self, runtime):
self.runtime, self.config = runtime, runtime.config
def __call__(self, inputs):
ids = inputs.numpy().astype(np.int32)
common = common_prefix_length(ids.tolist())
if common:
prefix, state = self.runtime.step(ids[0, :common])
result = []
for row in ids:
if common == len(row): logits = prefix
elif common:
suffix, _ = self.runtime.step(row[common:], state, common)
logits = np.concatenate((prefix, suffix), axis=1)
else: logits, _ = self.runtime.step(row)
result.append(logits)
return torch.from_numpy(np.concatenate(result, axis=0))
class CachedLM(v2.BundleLM):
def __init__(self, bundle, directory):
super().__init__(bundle)
self.runtime = kv.Runtime(bundle, directory)
self.model, self.metadata = CachedForward(self.runtime), self.runtime.metadata
self._generating = False
self._ids = self._state = None
@torch.inference_mode()
def next_logits(self, ids):
if self._generating and self._ids and ids[:len(self._ids)] == self._ids and len(ids) > len(self._ids):
logits, state = self.runtime.step(ids[len(self._ids):], self._state, len(self._ids), last_only=True)
else:
logits, state = self.runtime.step(ids, last_only=True)
if self._generating: self._ids, self._state = list(ids), state
return torch.from_numpy(logits[0,-1])
def generate(self, *args, **kwargs):
self._generating, self._ids, self._state = True, None, None
try: return super().generate(*args, **kwargs)
finally: self._generating, self._ids, self._state = False, None, None
def timing(work, repeats=20):
for _ in range(3): work()
times = []
for _ in range(repeats):
start = time.perf_counter(); work(); times.append((time.perf_counter()-start)*1000)
return {"n": len(times), "p50_ms": float(np.median(times)), "p95_ms": float(np.percentile(times, 95)), "max_ms": max(times)}
def benchmark(bundle, directory, baseline, output):
runtime, old = kv.Runtime(bundle, directory), v2.CoreMLLM(bundle, baseline)
rng = np.random.default_rng(716)
rows = []
for length in (8, 32, 64, 120):
ids = rng.integers(4, 16384, length).tolist(); ids[0] = 2
_, state = runtime.step(ids[:-1], last_only=True)
rows.append({"prefix_tokens": length-1,
"stateless": timing(lambda: old.model(torch.tensor([ids]))),
"cached_decode": timing(lambda: runtime.step(ids[-1:], state, length-1, last_only=True)),
"prefill_last_only": timing(lambda: runtime.step(ids, last_only=True))})
def generated(cached, ids, tokens=8):
sequence = list(ids)
if cached: logits, state = runtime.step(sequence, last_only=True)
else: logits = old.model(torch.tensor([sequence])).numpy()[:, -1:]
for i in range(tokens):
token = int(logits[0,-1].argmax()); sequence.append(token)
if i == tokens-1: break
if cached: logits, state = runtime.step([token], state, len(sequence)-1, last_only=True)
else: logits = old.model(torch.tensor([sequence])).numpy()[:, -1:]
return sequence
generation = []
for length in (8, 32, 64):
ids = [2, *rng.integers(4, 16384, length-1).tolist()]
generation.append({"prefix_tokens": length, "new_tokens": 8,
"tokens_equal": generated(True,ids) == generated(False,ids),
"cached_total": timing(lambda: generated(True,ids), 10),
"stateless_total": timing(lambda: generated(False,ids), 10)})
fresh_directory(output)
write_json(output / "benchmark.json", {"scope": "Mac CPU_ONLY Python/Core ML; includes Python and array-validation overhead, not iPhone or keyboard extension",
"model": runtime.metadata, "decode": rows, "generation": generation,
"cache_bytes_per_snapshot": runtime.manifest["cache_bytes_per_snapshot"]})
def package(bundle, source, compiled, comparison, output):
m = kv.verify_package(source)
gate = read_json(comparison)
if m["kind"] != "linear8_fp32_compute" or not (compiled / "coremldata.bin").is_file():
raise ValueError("Expected compiled INT8 KV resources.")
if (gate.get("format") != "vimeml_kv_alignment_v1" or not gate.get("passed")
or gate.get("comparison") != "same-weight stateless Core ML"
or gate["coreml_manifest_sha256"] != file_sha(source / "manifest.json")):
raise ValueError("Passed same-weight stateless comparison required before packaging.")
fresh_directory(output)
resources = output / "Resources"; resources.mkdir()
shutil.copytree(compiled, resources / "TinyJapaneseV21KVINT8.mlmodelc")
# Reuses the exact V2 tokenizer already shipped with the baseline, not a
# second tokenizer copy. Installer checks this existing identity.
write_json(resources / "VimeLMManifestV21KV.json", {"format": "vime_ios_lm_v2_kv_v1",
"architecture": v2.ARCHITECTURE, "model_version": "2.1-extend5-step40000-int8-b32-kv-v1",
"minimum_ios": 18, "vocab_size": 16384, "context_length": 128, "compute_units": "CPU_ONLY",
"special_ids": {"pad": 0, "unk": 1, "bos": 2, "eos": 3},
"tokenizer_sha256": m["bundle"]["tokenizer_sha256"], "checkpoint_sha256": m["bundle"]["checkpoint_sha256"],
"cache_shape": m["cache_shape"], "cache_dtype": "float32",
"source_manifest_sha256": file_sha(source / "manifest.json"),
"stateless_alignment_sha256": file_sha(comparison),
"compiled_files_sha256": {name: item["sha256"] for name,item in v2.inventory(compiled,hashes=True).items()},
"policy": "Experimental opt-in; explicit FP32 request-local immutable caches; baseline remains default."})
write_json(output / "manifest.json", {"format": "vimeml_kv_client_resources_v1", "model": m,
"files": v2.inventory(output)})
def main():
p = argparse.ArgumentParser(description=__doc__)
p.add_argument("--threads", type=int, default=4)
stages = p.add_subparsers(dest="stage", required=True)
for name in ("convert", "align", "compress", "evaluate", "samples", "benchmark", "package"):
s = stages.add_parser(name); s.add_argument("--output", type=Path, required=True)
if name != "compress": s.add_argument("--bundle", type=Path, required=True)
if name in ("align", "evaluate", "samples", "benchmark"): s.add_argument("--model", type=Path, required=True)
if name in ("compress", "package"): s.add_argument("--source", type=Path, required=True)
if name == "compress": s.add_argument("--alignment", type=Path, required=True)
if name == "align":
s.add_argument("--reference", type=Path, required=True); s.add_argument("--baseline", type=Path)
if name in ("evaluate", "samples", "benchmark"): s.add_argument("--baseline", type=Path, required=True)
if name == "evaluate": s.add_argument("--benchmark", type=Path, required=True)
if name == "package":
s.add_argument("--compiled", type=Path, required=True)
s.add_argument("--comparison", type=Path, required=True)
args = p.parse_args()
if args.output.exists() or args.threads < 1: p.error("Fresh output and positive thread count required.")
torch.set_num_threads(args.threads)
if args.stage == "convert": kv.convert(args.bundle,args.output)
elif args.stage == "compress": kv.compress(args.source,args.output,args.alignment)
elif args.stage == "align":
report = align(args.bundle,args.model,args.reference,args.output,args.baseline)
print(json.dumps({"passed": report["passed"], "max_abs": max(e["max_abs"] for e in report["errors"])}, indent=2))
if not report["passed"]: raise SystemExit("Alignment failed; evidence preserved.")
elif args.stage == "benchmark": benchmark(args.bundle,args.model,args.baseline,args.output)
elif args.stage == "package": package(args.bundle,args.source,args.compiled,args.comparison,args.output)
else:
lm = CachedLM(args.bundle,args.model)
if args.stage == "evaluate": v2.evaluate(lm,args.benchmark,args.baseline,args.output)
else: v2.samples(lm,args.baseline,args.output)
print(f"Output: {args.output.resolve()}")
if __name__ == "__main__": main()
|