File size: 13,177 Bytes
784f572 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 | """Runtime helpers for zero-shot hard-prefix conditional generation."""
from __future__ import annotations
from contextlib import contextmanager
from dataclasses import dataclass
import json
from pathlib import Path
from typing import Iterator, Sequence
from dlb.conditional_prompts import PromptManifest, PromptRecord
from dlb.io import SampleValidationError, sha256_file
@dataclass(frozen=True)
class ConditioningBatch:
"""A device-materialized batch of immutable prompt conditioning tensors."""
prompt_ids: tuple[int, ...]
source_indices: tuple[int, ...]
prefix_token_ids: object
reference_token_ids: object
completion_id: int
@property
def prefix(self) -> object:
return self.prefix_token_ids
@property
def reference(self) -> object:
return self.reference_token_ids
def _require_torch():
try:
import torch
except ImportError as error: # pragma: no cover - server environments install torch.
raise RuntimeError("conditional runtime helpers require torch") from error
return torch
def _shape(value: object) -> tuple[int, ...]:
return tuple(int(part) for part in value.shape) # type: ignore[attr-defined]
def clamp_token_prefix(state, prefix_ids):
"""Return a copy of a rank-2 token state with the prompt prefix restored."""
if getattr(state, "ndim", None) != 2 or getattr(prefix_ids, "ndim", None) != 2:
raise ValueError("token state and prefix must be rank two")
if state.shape[0] != prefix_ids.shape[0] or state.shape[1] < prefix_ids.shape[1]:
raise ValueError("token state is incompatible with prefix batch")
result = state.clone()
result[:, : prefix_ids.shape[1]] = prefix_ids.to(device=state.device, dtype=state.dtype)
return result
def clamp_vocab_prefix(state, prefix_ids):
"""Return a copy of a rank-3 vocabulary state with clean one-hot prefix rows."""
torch = _require_torch()
if getattr(state, "ndim", None) != 3:
raise ValueError("vocabulary state must be rank three")
if getattr(prefix_ids, "ndim", None) != 2:
raise ValueError("vocabulary prefix must be rank two")
if state.shape[0] != prefix_ids.shape[0] or state.shape[1] < prefix_ids.shape[1]:
raise ValueError("vocabulary state is incompatible with prefix batch")
if int(prefix_ids.max().item()) >= state.shape[-1] or int(prefix_ids.min().item()) < 0:
raise ValueError("prefix token is outside vocabulary dimension")
clean = torch.nn.functional.one_hot(prefix_ids, num_classes=state.shape[-1]).to(state)
result = state.clone()
result[:, : prefix_ids.shape[1], :] = clean
return result
def base_n_digits(token_ids, *, base: int, digits_per_token: int):
"""Encode rank-2 token IDs into RDLM's flattened base-n digit rows."""
torch = _require_torch()
if getattr(token_ids, "ndim", None) != 2:
raise ValueError("token IDs must be rank two")
if type(base) is not int or base <= 1:
raise ValueError("base must be greater than one")
if type(digits_per_token) is not int or digits_per_token <= 0:
raise ValueError("digits_per_token must be positive")
powers = base ** torch.arange(digits_per_token, device=token_ids.device)
digits = (token_ids.unsqueeze(-1) // powers) % base
return digits.reshape(token_ids.shape[0], -1)
def clamp_rdlm_prefix(state, prefix_ids, *, base: int, digits_per_token: int):
"""Return a copy of an RDLM manifold state with clean base-n prefix digits."""
torch = _require_torch()
if getattr(state, "ndim", None) != 3:
raise ValueError("RDLM state must be rank three")
prefix_digits = base_n_digits(
prefix_ids.to(device=state.device),
base=base,
digits_per_token=digits_per_token,
)
if state.shape[0] != prefix_digits.shape[0] or state.shape[1] < prefix_digits.shape[1]:
raise ValueError("RDLM state is incompatible with prefix batch")
if int(prefix_digits.max().item()) >= state.shape[-1] or int(prefix_digits.min().item()) < 0:
raise ValueError("RDLM prefix digit is outside state dimension")
clean = torch.nn.functional.one_hot(prefix_digits, num_classes=state.shape[-1]).to(state)
result = state.clone()
result[:, : prefix_digits.shape[1], :] = clean
return result
def clamp_embedding_prefix(state, clean_embeddings):
"""Return a copy of an embedding state with clean prefix embeddings restored."""
if getattr(state, "ndim", None) != 3 or getattr(clean_embeddings, "ndim", None) != 3:
raise ValueError("embedding state and clean prefix must be rank three")
if (
state.shape[0] != clean_embeddings.shape[0]
or state.shape[1] < clean_embeddings.shape[1]
or state.shape[2] != clean_embeddings.shape[2]
):
raise ValueError("embedding state is incompatible with clean prefix embeddings")
result = state.clone()
result[:, : clean_embeddings.shape[1], :] = clean_embeddings.to(state)
return result
def token_project_fn(prefix_ids):
"""Build a repeatable token-prefix projector closure."""
if getattr(prefix_ids, "ndim", None) != 2:
raise ValueError("token prefix must be rank two")
prefix = prefix_ids.clone()
def project(state):
return clamp_token_prefix(state, prefix)
project.conditioning_implementation = "zero_shot_runtime_projection" # type: ignore[attr-defined]
return project
def vocab_project_fn(prefix_ids):
"""Build a repeatable one-hot-prefix projector closure."""
if getattr(prefix_ids, "ndim", None) != 2:
raise ValueError("vocabulary prefix must be rank two")
prefix = prefix_ids.clone()
def project(state):
return clamp_vocab_prefix(state, prefix.to(device=state.device))
project.conditioning_implementation = "zero_shot_runtime_projection" # type: ignore[attr-defined]
return project
def rdlm_project_fn(prefix_ids, *, base: int, digits_per_token: int):
"""Build a repeatable RDLM base-n digit prefix projector closure."""
if getattr(prefix_ids, "ndim", None) != 2:
raise ValueError("RDLM prefix must be rank two")
prefix = prefix_ids.clone()
def project(state):
return clamp_rdlm_prefix(
state,
prefix.to(device=state.device),
base=base,
digits_per_token=digits_per_token,
)
project.conditioning_implementation = "zero_shot_runtime_projection" # type: ignore[attr-defined]
return project
def embedding_project_fn(clean_embeddings):
"""Build a repeatable embedding-prefix projector closure."""
if getattr(clean_embeddings, "ndim", None) != 3:
raise ValueError("clean embedding prefix must be rank three")
clean = clean_embeddings.clone()
def project(state):
return clamp_embedding_prefix(state, clean.to(device=state.device, dtype=state.dtype))
project.conditioning_implementation = "zero_shot_runtime_projection" # type: ignore[attr-defined]
return project
def candi_prompt_conditioning(prefix_ids, *, sequence_length: int):
"""Return full-canvas prompt tokens and numeric mask for CANDI's sampler."""
torch = _require_torch()
if getattr(prefix_ids, "ndim", None) != 2:
raise ValueError("CANDI prompt prefix must be rank two")
if type(sequence_length) is not int or sequence_length < int(prefix_ids.shape[1]):
raise ValueError("CANDI prompt sequence_length cannot be shorter than prefix")
prompt_tokens = torch.zeros(
(prefix_ids.shape[0], sequence_length),
dtype=prefix_ids.dtype,
device=prefix_ids.device,
)
prompt_tokens[:, : prefix_ids.shape[1]] = prefix_ids
prompt_mask = torch.zeros(
(prefix_ids.shape[0], sequence_length),
dtype=torch.float32,
device=prefix_ids.device,
)
prompt_mask[:, : prefix_ids.shape[1]] = 1.0
return prompt_tokens, prompt_mask
def adapt_candi_generate_sample_prompt(original):
"""Adapt CANDI's prompt hook across its mixed numeric/bool mask uses."""
torch = _require_torch()
def generate_sample_prompt(prompt_tokens, x, prompt_mask):
numeric_mask = prompt_mask.to(device=x.device, dtype=x.dtype)
prompt_tokens = prompt_tokens.to(device=x.device)
prompted, clean_mask = original(prompt_tokens, x, numeric_mask)
if clean_mask.dtype != torch.bool:
clean_mask = clean_mask.to(dtype=torch.bool)
return prompted, clean_mask
return generate_sample_prompt
@contextmanager
def patched_attribute(owner: object, name: str, replacement: object) -> Iterator[None]:
"""Temporarily replace an attribute and always restore the original value."""
original = getattr(owner, name)
setattr(owner, name, replacement)
try:
yield
finally:
setattr(owner, name, original)
def _unique_object(pairs: list[tuple[str, object]]) -> dict[str, object]:
result: dict[str, object] = {}
for key, value in pairs:
if key in result:
raise SampleValidationError(f"duplicate JSON key {key!r}")
result[key] = value
return result
def _read_prompt_records(manifest: PromptManifest, root: Path) -> list[PromptRecord]:
prompt_path = root / manifest.prompt_file
if prompt_path.is_symlink() or not prompt_path.is_file():
raise ValueError(f"conditional prompt file is missing or unsafe: {prompt_path}")
if sha256_file(prompt_path) != manifest.prompt_file_sha256:
raise ValueError("conditional prompt file SHA-256 differs from manifest")
records: list[PromptRecord] = []
with prompt_path.open("r", encoding="utf-8") as handle:
for index, raw_line in enumerate(handle):
if not raw_line.strip():
raise ValueError(f"conditional prompt record {index} is blank")
value = json.loads(raw_line, object_pairs_hook=_unique_object)
record = PromptRecord.model_validate(value)
if record.prompt_id != index:
raise ValueError(f"expected prompt_id {index}, found {record.prompt_id}")
records.append(record)
if len(records) != manifest.prompt_count:
raise ValueError(
f"expected {manifest.prompt_count} prompt records, found {len(records)}"
)
return records
def _schedule_limit(manifest: PromptManifest, completion_id: int) -> int:
if type(completion_id) is not int or completion_id < 0 or completion_id > 4:
raise ValueError("completion_id must be an integer in [0, 4]")
if completion_id == 0:
return manifest.prompt_count
return min(256, manifest.prompt_count)
def load_conditioning_batch(
manifest_path: Path,
expected_manifest_sha256: str,
*,
completion_id: int,
prompt_start: int,
batch_size: int,
device: str,
vocab_size: int,
) -> ConditioningBatch:
"""Load one exact prompt/completion batch from a SHA-bound prompt manifest."""
torch = _require_torch()
manifest_path = manifest_path.resolve()
if manifest_path.is_symlink() or not manifest_path.is_file():
raise ValueError(f"conditional prompt manifest is missing or unsafe: {manifest_path}")
if sha256_file(manifest_path) != expected_manifest_sha256:
raise ValueError("conditional prompt manifest SHA-256 differs from request")
if type(prompt_start) is not int or prompt_start < 0:
raise ValueError("prompt_start must be a non-negative integer")
if type(batch_size) is not int or batch_size <= 0:
raise ValueError("batch_size must be a positive integer")
if type(vocab_size) is not int or vocab_size <= 0:
raise ValueError("vocab_size must be a positive integer")
manifest = PromptManifest.model_validate(
json.loads(manifest_path.read_text(encoding="utf-8"), object_pairs_hook=_unique_object)
)
manifest_vocab_size = int(manifest.vocabulary_size)
if vocab_size < manifest_vocab_size:
raise ValueError("conditioning runtime vocabulary size is smaller than prompt manifest")
limit = _schedule_limit(manifest, completion_id)
if prompt_start + batch_size > limit:
raise ValueError("conditioning batch crosses its completion schedule boundary")
records = _read_prompt_records(manifest, manifest_path.parents[2])
selected = records[prompt_start : prompt_start + batch_size]
prefix_rows = [record.prefix_token_ids for record in selected]
reference_rows = [record.reference_token_ids for record in selected]
for row in (*prefix_rows, *reference_rows):
if len(row) != 64 or any(token_id >= manifest_vocab_size for token_id in row):
raise ValueError("conditioning prompt tokens violate length or vocabulary bounds")
return ConditioningBatch(
prompt_ids=tuple(record.prompt_id for record in selected),
source_indices=tuple(record.source_index for record in selected),
prefix_token_ids=torch.as_tensor(prefix_rows, dtype=torch.long, device=device),
reference_token_ids=torch.as_tensor(reference_rows, dtype=torch.long, device=device),
completion_id=completion_id,
)
|