Automatic Speech Recognition
MLX
English
apple-silicon
speech-to-text
asr
stt
low-bit
ternary
quantization-aware-training
on-device
streaming
Instructions to use FermionResearch/Phonon-1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use FermionResearch/Phonon-1 with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir Phonon-1 FermionResearch/Phonon-1
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
File size: 10,503 Bytes
bfafebe | 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 | #!/usr/bin/env python3
"""Byte-plane-split release packer/unpacker.
A BF16 array interleaves a highly predictable exponent byte with a near-random
mantissa byte. Compressed interleaved, zstd models neither population well.
Split into per-byte-position planes, the exponent plane compresses hard. U8
code planes are left untouched.
This is a *transport* transform only:
* the transform is applied per tensor, using the dtype in the safetensors
header, so it never has to guess;
* bytes not covered by any tensor (the header, alignment padding) are copied
verbatim;
* ``unpack`` reconstructs the original file and the manifest carries the
original SHA-256 and length of every member, so an install proves
byte-identity before anything is used.
Runtime tensors are therefore provably unchanged and accuracy cannot move.
python package_release_bps.py pack parity --level 19
python package_release_bps.py unpack local_stt/releases_bps/phonon-parity.bps.tar.zst DEST
python package_release_bps.py verify parity # full roundtrip proof
"""
from __future__ import annotations
import argparse
import hashlib
import io
import json
import subprocess
import tarfile
import tempfile
import time
from pathlib import Path
import numpy as np
ROOT = Path(__file__).resolve().parent
PROFILES = {
"parity": ROOT / "model_v18_mlx_quint5",
"micro": ROOT / "model_v18_mlx_hybrid4_quint5",
"audio6": ROOT / "model_v18_mlx_head8audio6_quint5",
}
ITEMSIZE = {"BOOL": 1, "U8": 1, "I8": 1, "U16": 2, "I16": 2, "F16": 2, "BF16": 2,
"U32": 4, "I32": 4, "F32": 4, "U64": 8, "I64": 8, "F64": 8, "F8_E4M3": 1,
"F8_E5M2": 1}
BLOCK = 16 << 20
FORMAT = "phonon-byteplane-tar-zstd-v1"
def sha256_bytes(data: bytes) -> str:
return hashlib.sha256(data).hexdigest()
def sha256_file(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(BLOCK), b""):
digest.update(chunk)
return digest.hexdigest()
def tensor_spans(raw: np.ndarray):
"""Return (base, [(start, end, itemsize), ...]) sorted, non-overlapping."""
header_len = int.from_bytes(raw[:8].tobytes(), "little")
header = json.loads(raw[8:8 + header_len].tobytes())
base = 8 + header_len
spans = []
for name, meta in header.items():
if name == "__metadata__":
continue
start, end = meta["data_offsets"]
itemsize = ITEMSIZE[meta["dtype"]]
if itemsize > 1 and (end - start) % itemsize == 0:
spans.append((int(start), int(end), itemsize))
spans.sort()
merged = []
last_end = 0
for start, end, itemsize in spans:
if start < last_end: # overlapping/aliased tensors
continue
merged.append((start, end, itemsize))
last_end = end
return base, merged
def split_file(path: Path) -> tuple[bytes, dict]:
raw = np.fromfile(path, dtype=np.uint8)
base, spans = tensor_spans(raw)
out = io.BytesIO()
out.write(raw[:base].tobytes()) # header verbatim
cursor = 0
plan = []
for start, end, itemsize in spans:
if start > cursor: # padding / uncovered bytes
out.write(raw[base + cursor: base + start].tobytes())
chunk = raw[base + start: base + end]
for i in range(itemsize):
out.write(chunk[i::itemsize].tobytes())
plan.append([start, end, itemsize])
cursor = end
tail = raw[base + cursor:]
if tail.size:
out.write(tail.tobytes())
meta = {
"base": base,
"plan": plan,
"payload_bytes": int(raw.size - base),
"original_bytes": int(raw.size),
"original_sha256": sha256_bytes(raw.tobytes()),
}
return out.getvalue(), meta
def join_file(data: bytes, meta: dict) -> bytes:
raw = np.frombuffer(data, dtype=np.uint8)
base = meta["base"]
out = np.empty(meta["original_bytes"], dtype=np.uint8)
out[:base] = raw[:base]
src = base
cursor = 0
for start, end, itemsize in meta["plan"]:
if start > cursor:
width = start - cursor
out[base + cursor: base + start] = raw[src: src + width]
src += width
n = end - start
per = n // itemsize
block = raw[src: src + n].reshape(itemsize, per)
out[base + start: base + end] = block.T.reshape(-1)
src += n
cursor = end
remaining = meta["payload_bytes"] - cursor
if remaining:
out[base + cursor:] = raw[src: src + remaining]
return out.tobytes()
def tar_info(name: str, size: int) -> tarfile.TarInfo:
info = tarfile.TarInfo(name)
info.size = size
info.mtime = 0
info.mode = 0o644
info.uid = info.gid = 0
info.uname = info.gname = ""
return info
def pack(profile: str, level: int, out_dir: Path) -> dict:
source = PROFILES[profile]
out_dir.mkdir(parents=True, exist_ok=True)
archive = out_dir / f"phonon-{profile}.bps.tar.zst"
members = sorted(p for p in source.rglob("*") if p.is_file())
manifest = {
"release_format": FORMAT,
"profile": profile,
"compression": {"codec": "zstd", "level": level},
"transform": "byte-plane-split-per-tensor-v1",
"files": [],
}
payloads: list[tuple[str, bytes]] = []
for path in members:
rel = str(path.relative_to(source))
blob = path.read_bytes()
entry = {"path": rel, "original_bytes": len(blob),
"original_sha256": sha256_bytes(blob)}
if path.name.startswith("model-") and path.suffix == ".safetensors":
transformed, meta = split_file(path)
entry["transform"] = meta
entry["stored_bytes"] = len(transformed)
payloads.append((rel + ".bps", transformed))
else:
entry["stored_bytes"] = len(blob)
payloads.append((rel, blob))
manifest["files"].append(entry)
manifest_bytes = (json.dumps(manifest, indent=2, sort_keys=True) + "\n").encode()
started = time.perf_counter()
args = ["zstd", "-q", f"-{level}", "-T0", "-f", "-o", str(archive), "-"]
if level >= 20:
args.insert(1, "--ultra")
process = subprocess.Popen(args, stdin=subprocess.PIPE)
assert process.stdin is not None
with tarfile.open(fileobj=process.stdin, mode="w|") as tar:
tar.addfile(tar_info("bps_manifest.json", len(manifest_bytes)),
io.BytesIO(manifest_bytes))
for name, blob in payloads:
tar.addfile(tar_info(name, len(blob)), io.BytesIO(blob))
process.stdin.close()
if process.wait() != 0:
raise RuntimeError("zstd failed")
pack_s = time.perf_counter() - started
return {"profile": profile, "archive": str(archive),
"archive_bytes": archive.stat().st_size,
"source_bytes": sum(p.stat().st_size for p in members),
"level": level, "pack_seconds": pack_s,
"archive_sha256": sha256_file(archive)}
def unpack(archive: Path, dest: Path) -> dict:
dest.mkdir(parents=True, exist_ok=True)
started = time.perf_counter()
process = subprocess.Popen(["zstd", "-q", "-d", "-c", str(archive)],
stdout=subprocess.PIPE)
assert process.stdout is not None
manifest = None
written = []
with tarfile.open(fileobj=process.stdout, mode="r|") as tar:
for member in tar:
handle = tar.extractfile(member)
if handle is None:
continue
blob = handle.read()
if member.name == "bps_manifest.json":
manifest = json.loads(blob)
index = {row["path"]: row for row in manifest["files"]}
continue
if manifest is None:
raise RuntimeError("bps_manifest.json must be the first member")
rel = member.name[:-4] if member.name.endswith(".bps") else member.name
row = index[rel]
data = join_file(blob, row["transform"]) if "transform" in row else blob
got = sha256_bytes(data)
if got != row["original_sha256"] or len(data) != row["original_bytes"]:
raise RuntimeError(f"checksum mismatch on {rel}")
target = dest / rel
target.parent.mkdir(parents=True, exist_ok=True)
target.write_bytes(data)
written.append(rel)
if process.wait() != 0:
raise RuntimeError("zstd decompression failed")
missing = {row["path"] for row in manifest["files"]} - set(written)
if missing:
raise RuntimeError(f"archive is missing members: {sorted(missing)}")
return {"files": len(written), "unpack_seconds": time.perf_counter() - started}
def verify(profile: str, level: int, out_dir: Path) -> dict:
"""Pack, unpack to a temporary directory, and prove byte-identity."""
packed = pack(profile, level, out_dir)
source = PROFILES[profile]
with tempfile.TemporaryDirectory() as tmp:
stats = unpack(Path(packed["archive"]), Path(tmp))
mismatched = []
for path in sorted(p for p in source.rglob("*") if p.is_file()):
rel = path.relative_to(source)
other = Path(tmp) / rel
if not other.exists() or sha256_file(other) != sha256_file(path):
mismatched.append(str(rel))
packed.update(stats)
packed["roundtrip_byte_identical"] = not mismatched
packed["mismatched"] = mismatched
return packed
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("command", choices=("pack", "unpack", "verify"))
ap.add_argument("target")
ap.add_argument("dest", nargs="?")
ap.add_argument("--level", type=int, default=19)
ap.add_argument("--out-dir", type=Path, default=ROOT / "releases_bps")
args = ap.parse_args()
if args.command == "unpack":
print(json.dumps(unpack(Path(args.target), Path(args.dest)), indent=2))
return
fn = pack if args.command == "pack" else verify
result = fn(args.target, args.level, args.out_dir)
ratio = 100 * result["archive_bytes"] / result["source_bytes"]
result["percent_of_source"] = ratio
print(json.dumps(result, indent=2, sort_keys=True))
print(f"{args.target}: {result['source_bytes']/1e6:.1f} MB -> "
f"{result['archive_bytes']/1e6:.1f} MB ({ratio:.2f}%)")
if __name__ == "__main__":
main()
|