areneau's picture
Update Clairic Transformers runtime
24191a8 verified
Raw History Blame Contribute Delete
21.8 kB
"""Constrained extraction and typed readouts using native Transformers models."""
from __future__ import annotations
import asyncio
import concurrent.futures
import copy
import json
import threading
from collections.abc import Callable, Iterator
from contextlib import contextmanager, nullcontext
from collections import defaultdict
from pathlib import Path
from typing import Any
import torch
from . import model as cm
from .readout import load_head, read_record
from .math import ReadoutMath, install as install_math
from .vision_cache import VisionBatchCache
from .decoder_graph import prepare as prepare_decoder_graphs
_FLASH_ATTENTION = 'kernels-community/flash-attn3@e29f138fc363b396e5d2706c8a5f6fa7d36f41e0'
def _batch_defaults(backbone: Any) -> tuple[int, int]:
config = getattr(backbone.config, "text_config", None)
if (backbone.device.type == "cuda" and backbone.dtype == torch.bfloat16
and config is not None and config._attn_implementation == _FLASH_ATTENTION):
if config.hidden_size == 4096:
return 64, 65536
if config.hidden_size == 5120:
return 32, 32768
return 8, 8192
_ATTENTION_KERNEL_LOCK = threading.RLock()
@contextmanager
def _default_attention_kernel() -> Iterator[None]:
"""Select the pinned public attention build while initializing its loader."""
import os
from huggingface_hub import snapshot_download
from transformers.modeling_flash_attention_utils import lazy_import_flash_attention
repository, revision = _FLASH_ATTENTION.rsplit("@", 1)
build_name = "torch-stable-abi210-cu130-x86_64-linux"
snapshot = snapshot_download(repository, revision=revision,
allow_patterns=[f"build/{build_name}/**"])
build = Path(snapshot) / "build" / build_name
with _ATTENTION_KERNEL_LOCK:
previous = os.environ.get("LOCAL_KERNELS")
entries = previous.split(":") if previous else []
entries = [entry for entry in entries if entry.partition("=")[0] != repository]
os.environ["LOCAL_KERNELS"] = ":".join([*entries, f"{repository}={build}"])
try:
yield
lazy_import_flash_attention(implementation=_FLASH_ATTENTION)
finally:
if previous is None:
os.environ.pop("LOCAL_KERNELS", None)
else:
os.environ["LOCAL_KERNELS"] = previous
class _ConstrainedScores:
"""Capture probabilities after applying the schema's grammar constraints."""
def __init__(self, processor: Any, score_ids: list[int], size: int) -> None:
self.processor = processor
self.score_ids = score_ids
self.chosen: list[list[tuple[float, int, None]]] = [[] for _ in range(size)]
self.first: list[list[tuple[float, int, None]]] = [[] for _ in range(size)]
self._tokens: list[torch.Tensor] = []
self._logprobs: list[torch.Tensor] = []
self._first_values: torch.Tensor | None = None
def __call__(self, input_ids: torch.Tensor, scores: torch.Tensor) -> torch.Tensor:
if self.processor.prefilled and input_ids.is_cuda:
input_ids = input_ids[:, -1:].cpu()
masked = self.processor(input_ids, scores.clone())
logprobs = torch.log_softmax(masked.float(), dim=-1)
if not self._tokens:
self._first_values = logprobs[:, self.score_ids]
choices = masked.argmax(dim=-1)
self._tokens.append(choices)
self._logprobs.append(logprobs.gather(1, choices[:, None])[:, 0])
return masked
def finalize(self) -> None:
if not self._tokens:
return
tokens = torch.stack(self._tokens, dim=1).cpu().tolist()
values = torch.stack(self._logprobs, dim=1).cpu().tolist()
self.chosen = [
[(float(value), int(token), None)
for token, value in zip(row_tokens, row_values, strict=True)]
for row_tokens, row_values in zip(tokens, values, strict=True)
]
if self._first_values is not None:
selected = self._first_values.cpu().tolist()
self.first = [
[
(float(value), token, None)
for token, value in zip(self.score_ids, row, strict=True)
]
for row in selected
]
self._tokens.clear()
self._logprobs.clear()
self._first_values = None
class _CancellationCriteria:
def __init__(self, cancelled: threading.Event) -> None:
self.cancelled = cancelled
def __call__(
self, input_ids: torch.Tensor, scores: torch.Tensor, **kwargs: Any
) -> torch.Tensor:
return torch.full(
(input_ids.shape[0],),
self.cancelled.is_set(),
dtype=torch.bool,
device=input_ids.device,
)
class TransformersEngine:
"""Share one Transformers backbone between extraction and typed readouts."""
def __init__(
self,
backbone: Any,
head: cm.ClairicHead,
tokenizer: Any,
*,
batch_size: int | None = None,
text_batch_size: int | None = None,
batch_token_budget: int | None = None,
) -> None:
import xgrammar as xgr
self.backbone = backbone.eval()
self.head = head.to(backbone.device).float().eval()
self.tokenizer = tokenizer
default_size, default_budget = _batch_defaults(backbone)
self.batch_size = max(1, int(default_size if batch_size is None else batch_size))
self.text_batch_size = (
self.batch_size if text_batch_size is None else max(1, int(text_batch_size))
)
self.batch_token_budget = max(1, int(
default_budget if batch_token_budget is None else batch_token_budget
))
self.device, self.dtype = backbone.device, backbone.dtype
self.eos = tokenizer.convert_tokens_to_ids("<|im_end|>")
self.image_token = int(backbone.config.image_token_id)
info = xgr.TokenizerInfo.from_huggingface(
tokenizer,
vocab_size=backbone.get_output_embeddings().weight.shape[0],
stop_token_ids=[self.eos],
)
self.compiler = xgr.GrammarCompiler(info, max_threads=2)
self._worker = concurrent.futures.ThreadPoolExecutor(
max_workers=1, thread_name_prefix="clairic-model"
)
self._vision_cache = VisionBatchCache()
self._uncached_image_features = self.backbone.model.get_image_features
self.backbone.model.get_image_features = self._image_features
self._restore_math: Callable[[], None] | None = None
self._decoder_graphs: Any = None
self._decoder_cleanup: Any = None
def begin_request(self) -> None:
self._vision_cache.begin()
async def end_request(self) -> None:
await asyncio.get_running_loop().run_in_executor(
self._worker, self._vision_cache.clear
)
def _image_features(
self,
pixel_values: torch.Tensor,
image_grid_thw: torch.Tensor | None = None,
**kwargs: Any,
) -> Any:
return self._vision_cache.features(
self._uncached_image_features,
self.backbone.model.visual,
pixel_values,
image_grid_thw,
**kwargs,
)
@classmethod
def load(
cls,
path: Path,
tokenizer: Any,
*,
device: str = "cuda",
dtype: Any = None,
batch_size: int | None = None,
text_batch_size: int | None = None,
batch_token_budget: int | None = None,
**kwargs: Any,
) -> TransformersEngine:
from transformers import Qwen3_5ForConditionalGeneration
class _NativeQwen3_5ForConditionalGeneration(Qwen3_5ForConditionalGeneration):
_keep_in_fp32_modules_strict: list[str] = ["A_log"]
if dtype is None:
dtype = torch.bfloat16 if str(device).startswith("cuda") else torch.float32
if isinstance(dtype, str):
dtype = getattr(torch, dtype)
kwargs.setdefault("device_map", {"": device})
default_flash = False
if "attn_implementation" not in kwargs:
attention = "sdpa"
if str(device).startswith("cuda") and dtype == torch.bfloat16:
if torch.cuda.get_device_capability(torch.device(device))[0] == 9:
attention = _FLASH_ATTENTION
default_flash = True
kwargs["attn_implementation"] = {
"text_config": attention, "vision_config": attention,
}
with _default_attention_kernel() if default_flash else nullcontext():
backbone = _NativeQwen3_5ForConditionalGeneration.from_pretrained(
path,
dtype=dtype,
local_files_only=True,
**kwargs,
)
engine = cls(
backbone,
load_head(
path,
match_bfloat16=(
backbone.device.type == "cuda" and backbone.dtype == torch.bfloat16
),
),
tokenizer,
batch_size=batch_size,
text_batch_size=text_batch_size,
batch_token_budget=batch_token_budget,
)
try:
engine._restore_math = install_math(backbone, engine.head)
engine._decoder_graphs = engine._worker.submit(
prepare_decoder_graphs, backbone, max(engine.batch_size, engine.text_batch_size)
).result()
except BaseException:
engine.shutdown()
raise
return engine
def _inputs(
self, sequences: list[list[int]], media: list[Any]
) -> dict[str, torch.Tensor]:
width = max(map(len, sequences))
pad = self.tokenizer.pad_token_id
if pad is None:
pad = self.eos
inputs = {
"input_ids": torch.tensor(
[[pad] * (width - len(ids)) + ids for ids in sequences],
dtype=torch.long,
device=self.device,
),
"attention_mask": torch.tensor(
[[0] * (width - len(ids)) + [1] * len(ids) for ids in sequences],
dtype=torch.long,
device=self.device,
),
}
mappings = [mapping for payload in media if payload for mapping in payload]
if mappings:
for name in cm.MEDIA_BATCH_KEYS:
inputs[name] = torch.cat(
[mapping[name] for mapping in mappings], dim=0
).to(
device=self.device,
dtype=self.dtype if name == "pixel_values" else torch.long,
)
inputs["mm_token_type_ids"] = (
inputs["input_ids"] == self.image_token
).long()
self._vision_cache.prepare(mappings, inputs, self.backbone.model.visual.dtype)
return inputs
def _grammar(self, parameters: dict[str, Any]) -> Any:
if "json_schema" in parameters:
value = parameters["json_schema"]
return self.compiler.compile_json_schema(
json.loads(value) if isinstance(value, str) else value,
any_whitespace=True,
)
if "regex" in parameters:
return self.compiler.compile_regex(parameters["regex"])
if "ebnf" in parameters:
return self.compiler.compile_grammar(parameters["ebnf"])
raise ValueError(
"Extraction continuations require a JSON schema, regex, or grammar"
)
@torch.inference_mode()
def _generate(
self,
*,
input_ids: list[int] | list[list[int]],
image_data: Any = None,
sampling_params: dict[str, Any] | list[dict[str, Any]],
rid: Any = None,
token_ids_logprob: list[int] | None = None,
_cancelled: threading.Event | None = None,
**kwargs: Any,
) -> Any:
from xgrammar.contrib.hf import LogitsProcessor
if _cancelled is not None and _cancelled.is_set():
raise RuntimeError("Inference cancelled")
if not input_ids:
return []
single = isinstance(input_ids[0], int)
sequences = [input_ids] if single else input_ids
size = len(sequences)
if not size:
return []
parameters = (
[sampling_params] * size
if isinstance(sampling_params, dict)
else sampling_params
)
media = [image_data] if single else (image_data or [None] * size)
ids = [rid] if single else (rid or [None] * size)
if not (len(parameters) == len(media) == len(ids) == size):
raise ValueError("Extraction batch arguments have inconsistent lengths")
batch_size = self.batch_size if any(media) else self.text_batch_size
results: list[Any] = [None] * size
groups: dict[tuple[int, tuple[int, ...]], list[int]] = defaultdict(list)
for index, params in enumerate(parameters):
if params.get("temperature", 0) != 0:
raise ValueError("Clairic extraction uses deterministic decoding")
groups[
(
int(params.get("max_new_tokens", 256)),
tuple(params.get("stop_token_ids", [self.eos])),
)
].append(index)
for (maximum, stop_ids), indices in groups.items():
pending: list[int] = []
width = 0
chunks: list[list[int]] = []
for index in indices:
length = len(sequences[index])
if pending and (
len(pending) >= batch_size
or max(width, length) * (len(pending) + 1) > self.batch_token_budget
):
chunks.append(pending)
pending, width = [], 0
pending.append(index)
width = max(width, length)
if pending:
chunks.append(pending)
for chunk in chunks:
if _cancelled is not None and _cancelled.is_set():
raise RuntimeError("Inference cancelled")
inputs = self._inputs(
[sequences[i] for i in chunk], [media[i] for i in chunk]
)
scores = _ConstrainedScores(
LogitsProcessor([self._grammar(parameters[i]) for i in chunk]),
token_ids_logprob or [],
len(chunk),
)
config = copy.deepcopy(self.backbone.generation_config)
config.do_sample, config.num_beams = False, 1
config.temperature, config.top_p, config.top_k = None, None, None
config.max_new_tokens, config.use_cache = maximum, True
config.eos_token_id = list(stop_ids)
config.pad_token_id = self.tokenizer.pad_token_id or self.eos
scope = (self._decoder_graphs.request(len(chunk))
if self._decoder_graphs is not None else nullcontext(None))
with scope as native_cache:
generated = self.backbone.generate(
**inputs,
**({"past_key_values": native_cache} if native_cache is not None else {}),
generation_config=config,
logits_processor=[scores],
stopping_criteria=[_CancellationCriteria(_cancelled)]
if _cancelled is not None
else [],
)
if _cancelled is not None and _cancelled.is_set():
raise RuntimeError("Inference cancelled")
scores.finalize()
continuations = (
generated[:, inputs["input_ids"].shape[1] :].cpu().tolist()
)
for offset, (index, continuation) in enumerate(
zip(chunk, continuations, strict=True)
):
stop = next(
(
i
for i, token in enumerate(continuation)
if token in stop_ids
),
None,
)
output = (
continuation[: stop + 1] if stop is not None else continuation
)
logprobs = scores.chosen[offset][: len(output)]
if [item[1] for item in logprobs] != output:
raise RuntimeError(
"Generated tokens differ from constrained confidence scores"
)
text_ids = output[:-1] if stop is not None else output
results[index] = {
"output_ids": output,
"text": self.tokenizer.decode(
text_ids,
skip_special_tokens=False,
clean_up_tokenization_spaces=False,
),
"meta_info": {
"id": ids[index],
"completion_tokens": len(output),
"finish_reason": {"type": "stop", "matched": output[-1]}
if stop is not None
else {"type": "length"},
"output_token_logprobs": logprobs,
"output_token_ids_logprobs": [scores.first[offset]],
},
}
return results[0] if single else results
async def async_generate(self, **kwargs: Any) -> Any:
return await self._submit(
lambda cancelled: self._generate(_cancelled=cancelled, **kwargs)
)
async def _submit(self, operation: Callable[[threading.Event], Any]) -> Any:
cancelled = threading.Event()
future = asyncio.get_running_loop().run_in_executor(
self._worker, lambda: operation(cancelled)
)
try:
return await asyncio.shield(future)
except asyncio.CancelledError:
cancelled.set()
# Admission stays occupied until the worker has actually stopped.
# A prefill cannot be interrupted; generation checks between tokens.
while not future.done():
try:
await asyncio.shield(future)
except asyncio.CancelledError:
continue
except Exception:
break
if not future.cancelled():
future.exception()
raise
@torch.inference_mode()
def _read_typed(
self,
item: cm.EncodedRecord,
media: Any,
_cancelled: threading.Event | None = None,
) -> dict[str, dict[str, Any]]:
if _cancelled is not None and _cancelled.is_set():
raise RuntimeError("Inference cancelled")
inputs = self._inputs([list(item.input_ids)], [media])
output = self.backbone.model(**inputs, use_cache=False, return_dict=True)
if _cancelled is not None and _cancelled.is_set():
raise RuntimeError("Inference cancelled")
scope = (
ReadoutMath()
if str(self.device).startswith("cuda") and self.dtype == torch.bfloat16
else nullcontext()
)
with scope:
values = read_record(
self.head,
output.last_hidden_state[0],
item,
self.backbone.get_output_embeddings().weight,
)
values = cm.expand_output(
cm.compact_output(values).cpu(),
item,
bins=self.head.bins.out_features,
estimate_lexical=self.head.estimate_lexical,
span_trim=self.head.span_trim,
number_groups=self.head.number_groups,
number_pointer=self.head.number_pointer,
)
return cm.decode_record(values, item, self.tokenizer)
async def read_typed(
self, item: cm.EncodedRecord, media: Any = None
) -> dict[str, dict[str, Any]]:
return await self._submit(
lambda cancelled: self._read_typed(item, media, _cancelled=cancelled)
)
def _decoder_diagnostics(self) -> Any:
if self._decoder_graphs is not None:
return self._decoder_graphs.diagnostics()
return self._decoder_cleanup
def shutdown(self) -> None:
try:
manager, self._decoder_graphs = self._decoder_graphs, None
def close_graphs() -> None:
try:
manager.close()
finally:
self._decoder_cleanup = manager.diagnostics()
try:
if manager is not None:
if threading.get_ident() == manager.thread:
close_graphs()
else:
self._worker.submit(close_graphs).result()
finally:
self._worker.shutdown(wait=True, cancel_futures=True)
self._vision_cache.clear()
if self.backbone.model.get_image_features == self._image_features:
self.backbone.model.get_image_features = self._uncached_image_features
finally:
restore, self._restore_math = self._restore_math, None
if restore is not None:
restore()