Datasets:
File size: 5,464 Bytes
2288fa5 | 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 | #!/usr/bin/env python3
"""Quantize a converted DFlash speculator to int4 like the official Mirai draft.
`lalamo speculator convert` only writes float dtypes, so a converted draft stays
bf16 (about 2.2GB for the 9B draft) while the official one is 755MB. This
script reloads the draft as a lalamo module, compresses every Linear with a
4-bit group-32 hybrid spec (the same scheme as the target model) and writes the
result back in the uzu-loadable layout.
The speculator directory also needs `shapes.json`, which the converter does not
write; pass --shapes to copy one (or create an Argmax chain by default).
Usage:
python scripts/quantize_speculator.py \
--source /tmp/spec_guglxni --output /tmp/spec_guglxni_int4 \
--bits 4 --group-size 32
"""
from __future__ import annotations
import argparse
import json
import shutil
import time
from pathlib import Path
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--source", type=Path, required=True, help="converted speculator directory")
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--bits", type=int, default=4, choices=(4, 8))
parser.add_argument("--group-size", type=int, default=32)
parser.add_argument("--block-size", type=int, default=32)
parser.add_argument(
"--symmetric",
action="store_true",
help="omit zero points (required for int8; uzu never compiles asymmetric int8)",
)
parser.add_argument(
"--no-rewrite",
action="store_true",
help="skip the uzu metadata rewrite (the result will not load in uzu)",
)
parser.add_argument(
"--shapes",
type=Path,
default=None,
help="shapes.json to copy; default writes an Argmax chain with budget 16",
)
args = parser.parse_args()
import equinox as eqx
import jax
import jax.numpy as jnp
from lalamo.compressed.hybrid import HybridSpec
from lalamo.compressed.incoherence import HadamardSpec, IncoherenceProcessingMode
from lalamo.compressed.int import IntSpec
from lalamo.models.speculator import SpeculatorModel
from lalamo.modules.linear import Linear
from lalamo.utils.sharding import ShardingConfig
from lalamo.utils.surgery import map_nodes_of_type_with_path
from lalamo.weight_matrix import Layout
if args.bits == 8 and not args.symmetric:
raise SystemExit("int8 must be symmetric: uzu only compiles zero-point dequant for 4 bits")
sharding = ShardingConfig.replicated()
started = time.time()
model = SpeculatorModel.load(args.source, sharding, dtype=jnp.bfloat16)
print(f"loaded {args.source} in {time.time() - started:.1f}s", flush=True)
spec = HybridSpec(
quantization_spec=IntSpec(
bits=args.bits,
group_size=args.group_size,
is_symmetric=args.symmetric,
layout=Layout.OUTPUT_INPUT,
),
adapter_spec=None,
incoherence_spec=None
if args.block_size <= 0
else HadamardSpec(
block_size=args.block_size,
processing_mode=IncoherenceProcessingMode.INPUT_OUTPUT,
),
)
quantized = 0
skipped: list[tuple[str, tuple[int, ...]]] = []
key = jax.random.key(0)
def quantize_linear(path: tuple[str, ...], linear: Linear) -> Linear:
nonlocal quantized
shape = tuple(int(x) for x in linear.weights.shape)
if len(shape) != 2 or any(dimension % args.group_size for dimension in shape):
skipped.append(("/".join(path), shape))
return linear
started_at = time.time()
new_weights = spec.compress(
linear.weights.decompress(),
key=key,
sharding_config=linear.weights.sharding_config,
is_sharded=linear.weights.is_sharded,
)
quantized += 1
print(f" {'.'.join(path)} {shape} {time.time() - started_at:.1f}s", flush=True)
return eqx.tree_at(lambda x: x.weights, linear, new_weights)
model = map_nodes_of_type_with_path(Linear, quantize_linear, model)
print(f"quantized {quantized} Linear layers, skipped {len(skipped)}", flush=True)
for path, shape in skipped:
print(f" skipped {path} {shape}")
args.output.mkdir(parents=True, exist_ok=True)
model.save(args.output)
if args.shapes is not None:
shutil.copy2(args.shapes, args.output / "shapes.json")
else:
(args.output / "shapes.json").write_text(
json.dumps(
{
"default": {
"tree_budget": 16,
"max_tree_depth": 16,
"dflash_depth_override": None,
"construction_method": {"type": "Argmax"},
}
},
indent=2,
)
+ "\n",
encoding="utf-8",
)
if not args.no_rewrite:
import sys
sys.path.insert(0, str(Path(__file__).resolve().parent))
from rewrite_lalamo_hybrid_spec import rewrite
# uzu only understands the flat HybridSpec form; lalamo writes the nested one.
rewrite(args.output / "model.safetensors", backup=False)
total = sum(path.stat().st_size for path in args.output.iterdir())
print(f"wrote {args.output} ({total / 1e6:.0f}MB) in {time.time() - started:.0f}s")
if __name__ == "__main__":
main()
|