File size: 10,825 Bytes
7845694 | 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 | """Apple Silicon option readout using MLX-LM's native Qwen3.5 model and caches.
Scores remain conditional on the declared options, not calibrated confidence.
Each suffix owns an independent copy of both attention and recurrent state.
"""
from __future__ import annotations
import copy
import hashlib
from importlib.metadata import distribution, version
import json
from pathlib import Path
import platform
import re
import time
from .core import softmax
from .direct import PROMPT_VERSION, encode_prompt
from .shared import _state_prefix
DEFAULT_CACHE_LIMIT_MIB = 256
def load_model(source: str, revision: str, bits: int | None = None, *,
cache_limit_mib: int = DEFAULT_CACHE_LIMIT_MIB):
"""Load a pinned checkpoint strictly; optional affine quantization is in memory.
Use the upstream sanitizer for text weights. Custom model code is prohibited,
just as it is for the Torch loader. Hash the actual source artifacts, including
local checkpoints, so a user-supplied local revision is not the only provenance.
"""
if platform.system() != "Darwin" or platform.machine() != "arm64":
raise RuntimeError("MLX backend requires macOS on Apple Silicon")
local = Path(source).is_dir()
if not revision or (not local and not re.fullmatch(r"[0-9a-f]{40}", revision)):
raise ValueError("Remote models require a pinned 40-character revision; local models require a revision label")
if bits not in (None, 4, 8):
raise ValueError("MLX quantization must be 4 or 8 bits")
if type(cache_limit_mib) is not int or cache_limit_mib < 0:
raise ValueError("MLX cache limit must be a nonnegative integer in MiB")
try:
import mlx.core as mx
import mlx.nn as nn
from mlx.utils import tree_flatten
from mlx_lm import load
except ImportError as error:
raise RuntimeError("Install the MLX extra: pip install -e '.[test,mlx]'") from error
from huggingface_hub import snapshot_download
if not mx.metal.is_available():
raise RuntimeError("MLX Metal GPU is unavailable")
mx.set_default_device(mx.gpu)
# MLX defaults to caching almost all system RAM. Bound inactive allocations
# so repeated variable-length prompts can coexist with other local models.
cache_limit = cache_limit_mib * 1024 * 1024
mx.set_cache_limit(cache_limit)
path = Path(source) if local else Path(snapshot_download(
source, revision=revision,
allow_patterns=["*.json", "model*.safetensors", "*.jinja", "*.txt", "*.model"],
))
config = json.loads((path / "config.json").read_text())
if config.get("model_file") or config.get("model_type") not in {"qwen3_5"}:
raise ValueError("MLX backend supports native Qwen3.5 text scoring only; custom model code is not allowed")
if bits and (config.get("quantization") or config.get("quantization_config")):
raise ValueError("In-memory quantization requires an unquantized source checkpoint")
artifacts = {}
for file in sorted(path.iterdir()):
if file.is_file() and file.suffix in {".json", ".safetensors", ".jinja", ".txt", ".model"}:
with file.open("rb") as stream:
checksum = hashlib.sha256()
for chunk in iter(lambda: stream.read(8 * 1024 * 1024), b""):
checksum.update(chunk)
artifacts[file.name] = checksum.hexdigest()
model, tokenizer = load(path, tokenizer_config={"trust_remote_code": False})
if bits:
nn.quantize(model, group_size=64, bits=bits, mode="affine")
model.eval()
mx.eval(model.parameters())
mx.synchronize()
metadata = {
"source": source, "revision": revision, "backend": "mlx",
"mlx_version": version("mlx"), "mlx_lm_version": version("mlx-lm"),
"transformers_version": version("transformers"),
"mlx_lm_source": json.loads(distribution("mlx-lm").read_text("direct_url.json") or "null"),
"allocator_cache_limit_bytes": cache_limit,
"dtype": sorted({str(value.dtype) for _, value in tree_flatten(model.parameters())}),
"quantization": {"bits": bits, "group_size": 64, "mode": "affine"} if bits else config.get("quantization"),
"source_artifact_sha256": artifacts,
}
return model, tokenizer, metadata
def _result(row, encoded, selected, metadata, mode):
ids, slots, prompt_hash = encoded
return {
"id": row["id"], "option_ids": [option["id"] for option in row["options"]],
"probabilities": softmax(selected), "option_logits": selected,
"answer_token_ids": slots, "input_tokens": len(ids),
"input_ids_sha256": hashlib.sha256(json.dumps(ids).encode()).hexdigest(),
"prompt_sha256": prompt_hash, "prompt_version": PROMPT_VERSION,
"model": {**metadata, "serving_config": f"mlx-{mode}-v1"},
"readout": "native last-position logits restricted to declared answer slots; no generated tokens",
"probability_status": "conditional option score; uncalibrated as decision confidence",
}
def score(model, tokenizer, row, metadata, max_tokens=4096):
import mlx.core as mx
mx.synchronize()
started = time.perf_counter()
encoded = encode_prompt(tokenizer, row, max_tokens)
ids, slots, _ = encoded
mark = time.perf_counter()
logits = model(mx.array([ids]))[0, -1].astype(mx.float32)
selected = logits[mx.array(slots)].tolist()
mx.synchronize()
result = _result(row, encoded, selected, metadata, "direct")
result.update(forward_seconds=time.perf_counter() - mark, total_seconds=time.perf_counter() - started)
return result
def _prefill(model, prefix):
import mlx.core as mx
from mlx_lm.models.cache import make_prompt_cache
cache = make_prompt_cache(model)
# Evaluate the cache, not the unused vocabulary projection for the prefix.
model(mx.array([prefix]), cache=cache)
mx.eval([entry.state for entry in cache])
mx.synchronize()
return cache
def _check_prefix(prefix, encoded):
if not prefix or any(ids[:len(prefix)] != prefix or len(ids) <= len(prefix) for ids, _, _ in encoded):
raise ValueError("The fixed state prefix does not match every full prompt")
class SerialPrefixScorer:
"""Reuse only the current exact state; never mutate the retained prefix."""
def __init__(self, model, tokenizer, metadata, max_tokens=4096):
self.model, self.tokenizer, self.metadata = model, tokenizer, metadata
self.max_tokens = max_tokens
self.prefix = self.cache = None
def score(self, row):
import mlx.core as mx
mx.synchronize()
started = time.perf_counter()
encoded = encode_prompt(self.tokenizer, row, self.max_tokens)
prefix = _state_prefix(self.tokenizer, row["state"])
hit = self.cache is not None and prefix == self.prefix
_check_prefix(prefix, [encoded])
prefill_seconds = 0.0
if not hit:
self.cache = self.prefix = None
mark = time.perf_counter()
cache = _prefill(self.model, prefix)
prefill_seconds = time.perf_counter() - mark
self.cache, self.prefix = cache, prefix
mark = time.perf_counter()
branch = copy.deepcopy(self.cache)
mx.eval([entry.state for entry in branch])
mx.synchronize()
copy_seconds = time.perf_counter() - mark
ids, slots, _ = encoded
mark = time.perf_counter()
logits = self.model(mx.array([ids[len(prefix):]]), cache=branch)[0, -1].astype(mx.float32)
selected = logits[mx.array(slots)].tolist()
mx.synchronize()
suffix_seconds = time.perf_counter() - mark
result = _result(row, encoded, selected, self.metadata, "serial")
result.update(cache_hit=hit, prefix_tokens=len(prefix), prefill_seconds=prefill_seconds,
copy_seconds=copy_seconds, suffix_forward_seconds=suffix_seconds,
forward_seconds=prefill_seconds + suffix_seconds,
total_seconds=time.perf_counter() - started)
return result
def score_shared(model, tokenizer, rows, metadata, max_tokens=4096):
"""Prefill once, merge independent cache branches, then score padded suffixes.
Right padding is masked by native recurrent caches. Causal attention prevents
later padding from affecting real positions. Read each row's last *real* token;
the disposable branch caches do not need to be finalized for further decoding.
"""
import mlx.core as mx
if not rows or any(row["state"] != rows[0]["state"] for row in rows):
raise ValueError("Shared scoring requires one nonempty exact state")
if len({row["id"] for row in rows}) != len(rows):
raise ValueError("Decision IDs must be unique")
mx.synchronize()
started = time.perf_counter()
encoded = [encode_prompt(tokenizer, row, max_tokens) for row in rows]
prefix = _state_prefix(tokenizer, rows[0]["state"])
_check_prefix(prefix, encoded)
suffixes = [ids[len(prefix):] for ids, _, _ in encoded]
lengths = [len(ids) for ids in suffixes]
width = max(lengths)
pad = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
if pad is None:
raise ValueError("Tokenizer requires a padding or EOS token")
tokens = mx.array([ids + [pad] * (width - len(ids)) for ids in suffixes])
encode_seconds = time.perf_counter() - started
mark = time.perf_counter()
cache = _prefill(model, prefix)
prefill_seconds = time.perf_counter() - mark
mark = time.perf_counter()
branches = [entry.merge([entry] * len(rows)) for entry in cache]
for entry in branches:
entry.prepare(lengths=lengths, right_padding=[width - size for size in lengths])
mx.eval([entry.state for entry in branches])
mx.synchronize()
replicate_seconds = time.perf_counter() - mark
mark = time.perf_counter()
logits = model(tokens, cache=branches)
selected = [logits[i, lengths[i] - 1, mx.array(slots)].astype(mx.float32)
for i, (_, slots, _) in enumerate(encoded)]
mx.eval(selected)
mx.synchronize()
suffix_seconds = time.perf_counter() - mark
results = [_result(row, enc, values.tolist(), metadata, "shared")
for row, enc, values in zip(rows, encoded, selected)]
timing = {
"total_seconds": time.perf_counter() - started, "encode_seconds": encode_seconds,
"prefix_tokens": len(prefix), "prefill_seconds": prefill_seconds,
"replicate_seconds": replicate_seconds, "suffix_forward_seconds": suffix_seconds,
"batch_size": len(rows), "true_suffix_tokens": sum(lengths),
"padded_suffix_tokens": width * len(rows),
}
return results, timing
|