File size: 6,303 Bytes
6b56858 | 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 | #!/usr/bin/env python3
"""Sample the pinned SDTT KLD round-7 student from local server assets only."""
from __future__ import annotations
import argparse
from pathlib import Path
from _distilled_runtime import (
benchmark_model,
configure_for_sampling,
install_upstream,
load_config,
load_tokenizer,
load_tokenizer_binding,
materialize_model,
offline_huggingface,
parse_bool,
require_directory,
require_file,
seed_everything,
checkpoint_state,
validate_embedded_config,
validate_sampling_config,
write_capture_atomic,
)
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--upstream-root", type=Path, required=True)
parser.add_argument("--checkpoint", type=Path, required=True)
parser.add_argument("--config", type=Path, required=True)
parser.add_argument("--checkpoint-sha256", required=True)
parser.add_argument("--config-sha256", required=True)
parser.add_argument("--data-config", type=Path, required=True)
parser.add_argument("--downloads-manifest", type=Path, required=True)
parser.add_argument("--dataset", choices=("lm1b", "owt"), required=True)
parser.add_argument("--tokenizer-snapshot", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--sample-count", type=int, required=True)
parser.add_argument("--batch-size", type=int, required=True)
parser.add_argument("--num-steps", type=int, required=True)
parser.add_argument("--seq-len", type=int, required=True)
parser.add_argument("--seed", type=int, required=True)
parser.add_argument("--sampler", choices=("ancestral",), required=True)
parser.add_argument("--loss", choices=("kld",), required=True)
parser.add_argument("--round", type=int, choices=(7,), required=True)
parser.add_argument("--teacher-family", choices=("masked_mdlm",), required=True)
parser.add_argument("--generation-mode", choices=("unconditional", "conditional_prefix"), default="unconditional")
parser.add_argument("--conditioning-manifest", type=Path)
parser.add_argument("--conditioning-manifest-sha256")
parser.add_argument("--conditioning-config-sha256")
parser.add_argument("--prefix-length", type=int)
parser.add_argument("--evaluation-continuation-length", type=int)
parser.add_argument("--prompt-count", type=int)
parser.add_argument("--diversity-prompt-count", type=int)
parser.add_argument("--completions-per-diversity-prompt", type=int)
parser.add_argument("--completion-schedule")
parser.add_argument("--offline", type=parse_bool, required=True)
parser.add_argument("--benchmark-output", type=Path)
parser.add_argument("--benchmark-metadata", type=Path)
parser.add_argument("--benchmark-precision", choices=("author",))
args = parser.parse_args(argv)
benchmark_values = (args.benchmark_output, args.benchmark_metadata, args.benchmark_precision)
if any(value is not None for value in benchmark_values) and not all(
value is not None for value in benchmark_values
):
parser.error("benchmark arguments must be provided together")
upstream = require_directory(args.upstream_root, "pinned SDTT source")
require_file(upstream / "src/sdtt/main.py", "pinned SDTT entrypoint")
install_upstream(upstream)
checkpoint = require_file(args.checkpoint, "SDTT student checkpoint")
config_path = require_file(args.config, "SDTT student config")
binding = load_tokenizer_binding(
args.data_config,
args.downloads_manifest,
args.dataset,
args.tokenizer_snapshot,
)
with offline_huggingface(args.offline):
from sdtt.core.distill.multi_round_sdtt import MultiRoundSDTT
config = load_config(config_path, args.config_sha256)
state, embedded_config = checkpoint_state(
checkpoint, args.checkpoint_sha256
)
validate_sampling_config(
config,
binding=binding,
sequence_length=args.seq_len,
require_di4c=False,
)
if embedded_config is not None:
validate_embedded_config(config, embedded_config)
configure_for_sampling(
config, tokenizer_snapshot=binding.snapshot, seq_len=args.seq_len
)
tokenizer = load_tokenizer(binding.snapshot)
seed_everything(args.seed)
model = materialize_model(
model_type=MultiRoundSDTT,
config=config,
tokenizer=tokenizer,
state=state,
strict=True,
)
if args.benchmark_output is not None:
benchmark_model(
model=model,
output=args.benchmark_output,
metadata_path=args.benchmark_metadata,
precision=args.benchmark_precision,
num_steps=args.num_steps,
seq_len=args.seq_len,
sampler=args.sampler,
generation_mode=args.generation_mode,
conditioning_manifest=args.conditioning_manifest,
conditioning_manifest_sha256=args.conditioning_manifest_sha256,
prompt_count=args.prompt_count,
diversity_prompt_count=args.diversity_prompt_count,
completions_per_diversity_prompt=args.completions_per_diversity_prompt,
)
else:
write_capture_atomic(
args.output,
model=model,
tokenizer=tokenizer,
sample_count=args.sample_count,
batch_size=args.batch_size,
num_steps=args.num_steps,
seq_len=args.seq_len,
sampler=args.sampler,
generation_mode=args.generation_mode,
conditioning_manifest=args.conditioning_manifest,
conditioning_manifest_sha256=args.conditioning_manifest_sha256,
prompt_count=args.prompt_count,
diversity_prompt_count=args.diversity_prompt_count,
completions_per_diversity_prompt=args.completions_per_diversity_prompt,
)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|