Image-Text-to-Text
Transformers
Safetensors
qwen3_5
clairic
cloudflare
systemone
multimodal
structured-output
information-extraction
custom-code
conversational
Instructions to use Cloudflare/clairic-flash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Cloudflare/clairic-flash with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="Cloudflare/clairic-flash") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("Cloudflare/clairic-flash") model = AutoModelForMultimodalLM.from_pretrained("Cloudflare/clairic-flash", device_map="auto") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Cloudflare/clairic-flash with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Cloudflare/clairic-flash" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Cloudflare/clairic-flash", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/Cloudflare/clairic-flash
- SGLang
How to use Cloudflare/clairic-flash with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Cloudflare/clairic-flash" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Cloudflare/clairic-flash", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Cloudflare/clairic-flash" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Cloudflare/clairic-flash", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use Cloudflare/clairic-flash with Docker Model Runner:
docker model run hf.co/Cloudflare/clairic-flash
File size: 15,808 Bytes
dc7d8aa | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 | from __future__ import annotations
from collections.abc import Callable
from typing import TYPE_CHECKING, Any, cast
if TYPE_CHECKING:
from types import ModuleType
from . import ar, image_rows, schema
def unused_prediction(result: dict[str, Any]) -> dict[str, Any]:
return {
"value": None,
"valid": False,
"complete": False,
"tokens": int(result.get("meta_info", {}).get("completion_tokens", 0)),
"_unused_row_proposal": True,
}
def _empty_inventory(result: dict[str, Any], eos: int, tokenizer: Any) -> bool:
finish = result.get("meta_info", {}).get("finish_reason", {}) or {}
if finish.get("type") != "stop" or finish.get("matched") != eos:
return False
ids = result.get("output_ids")
if ids is not None:
ids = ids[:-1] if ids and ids[-1] == eos else ids
text = tokenizer.decode(
ids, skip_special_tokens=False, clean_up_tokenization_spaces=False
)
else:
text = result.get("text", "").removesuffix(ar.EOS)
return text.strip() == "null"
async def _generate(
engine: Any,
record: dict[str, Any],
tasks: list[tuple[str, list[int], Any]],
eos: int,
tokenizer: Any,
**kwargs: Any,
) -> list[dict[str, Any]]:
return await engine.async_generate(**kwargs)
_BATCHED_KEYS = {"input_ids", "image_data", "rid", "cache_salt", "sampling_params"}
_SHARED_KEYS = {"return_logprob", "logprob_start_len", "token_ids_logprob"}
def _select_kwargs(
kwargs: dict[str, Any], indices: list[int], size: int
) -> dict[str, Any]:
if set(kwargs) - _BATCHED_KEYS - _SHARED_KEYS:
raise ValueError("Unknown native argument for first-item scheduling")
result = {}
for key, value in kwargs.items():
if key in _BATCHED_KEYS and isinstance(value, list):
if len(value) != size:
raise ValueError("Native batch argument cardinality differs: " + key)
result[key] = [value[index] for index in indices]
else:
result[key] = value
return result
def _ordinary_lists(
record: dict[str, Any],
tasks: list[tuple[str, list[int], Any]],
item_name: Callable[[str, int], str],
) -> dict[str, tuple[int, ...]]:
if record.get("_null_recovery"):
return {}
fields = record["fields"]
annotations = (
record.get("_meta", {}).get("extraction_contract", {}).get("fields", {})
)
names = [name for (name, _, _) in tasks]
if len(set(names)) != len(names):
raise ValueError("AR task names must be distinct")
indices = {name: index for (index, name) in enumerate(names)}
(claimed, groups) = (set(), {})
forbidden = {"_physical_row_scope", "_image_row_inventory", "_image_row_proposal"}
plans = (
record.get("_financial_line_plans", {}),
record.get("_list_source_bounds", {}),
)
for parent, annotation in annotations.items():
spec = fields.get(parent, {})
queries = annotation.get("query_fields")
item = spec.get("items")
kind = item.get("type") if isinstance(item, dict) else item
if (
spec.get("type") != "list"
or kind not in ("text", "number", "date")
or spec.get("knowledge")
or (set(annotation) != {"query_fields"})
or (not isinstance(queries, list))
or (len(queries) < 2)
or any((spec.get(key) for key in forbidden))
or (parent in indices)
or any((parent in plan for plan in plans))
):
continue
valid = True
for ordinal, name in enumerate(queries, 1):
child = fields.get(name, {})
ownership = annotations.get(name, {})
if (
name != item_name(parent, ordinal)
or name not in indices
or name in claimed
or (ownership != {"parent": parent, "index": ordinal})
or (child.get("type") not in ("text", "number", "date"))
or (child.get("nullable", True) is not True)
or any((child.get(key) for key in forbidden))
or (name in record.get("_source_row_gates", ()))
):
valid = False
break
if "enum" in child and None not in child["enum"]:
valid = False
break
if valid:
groups[parent] = tuple((indices[name] for name in queries))
claimed.update(queries)
return groups
def _completed_null(result: dict[str, Any], eos: int, tokenizer: Any) -> bool:
meta = result.get("meta_info", {})
finish = meta.get("finish_reason", {}) or {}
ids = result.get("output_ids")
if (
finish.get("type") != "stop"
or finish.get("matched") != eos
or (not isinstance(ids, list))
or (not ids)
or (ids[-1] != eos)
or (meta.get("completion_tokens") != len(ids))
):
return False
return (
tokenizer.decode(
ids[:-1], skip_special_tokens=False, clean_up_tokenization_spaces=False
).strip()
== "null"
)
def _row_cohorts(
record: dict[str, Any],
tasks: list[tuple[str, list[int], Any]],
item_name: Callable[[str, int], str],
) -> dict[str, dict[str, Any]]:
fields = record["fields"]
annotations = (
record.get("_meta", {}).get("extraction_contract", {}).get("fields", {})
)
indices = {name: index for (index, (name, _, _)) in enumerate(tasks)}
result = {}
for parent, annotation in annotations.items():
(queries, spec) = (annotation.get("image_row_fields"), fields.get(parent, {}))
if (
set(annotation) != {"image_row_fields"}
or not spec.get("_image_row_inventory")
or parent not in indices
or (not isinstance(queries, list))
or (len(queries) <= 8)
or (parent in record.get("_list_source_bounds", {}))
or (parent in record.get("_financial_line_plans", {}))
or (not isinstance(spec.get("items"), dict))
or (spec["items"].get("type") != "object")
):
continue
text_keys = [
key
for (key, child) in spec["items"].get("properties", {}).items()
if child.get("type") == "text"
]
if not text_keys:
continue
if any(
(
name != item_name(parent, ordinal)
or name not in indices
or annotations.get(name) != {"parent": parent, "index": ordinal}
or (fields.get(name, {}).get("type") != "object")
or (not fields[name].get("_image_row_proposal"))
for (ordinal, name) in enumerate(queries, 1)
)
):
continue
result[parent] = {
"inventory": indices[parent],
"proposals": tuple((indices[name] for name in queries)),
"text_keys": text_keys,
}
return result
def _cohort_proof(
record: dict[str, Any],
tasks: list[tuple[str, list[int], Any]],
results: list[dict[str, Any]],
parent: str,
cohort: dict[str, Any],
tokenizer: Any,
scalar_result: Callable[..., dict[str, Any]],
image_rows: ModuleType,
normalize: Callable[
[dict[str, Any], dict[str, dict[str, Any]]], dict[str, dict[str, Any]]
],
) -> dict[str, Any] | None:
spec = record["fields"][parent]
native = results[cohort["inventory"]]
inventory = scalar_result(native, spec, tokenizer, record.get("input", ""), None)
if not inventory.get("valid") or not inventory.get("complete"):
return None
value = inventory.get("value")
if value is not None and (
not isinstance(value, list)
or any(
(
not isinstance(row, dict)
or set(row) != {"label"}
or (not isinstance(row["label"], str))
for row in value
)
)
):
return None
labels = [row["label"] for row in value or []]
if len(labels) > 8 or len(labels) > spec.get("max_items", 32):
return None
retained = {parent: inventory}
for index in cohort["proposals"][: len(labels)]:
name = tasks[index][0]
result = results[index]
if result.get("_clair_unused_row_proposal"):
return None
parsed = scalar_result(
result, record["fields"][name], tokenizer, record.get("input", ""), None
)
retained[name] = parsed
normalized = normalize(record, retained)
normalized_inventory = normalized[parent]
if (
not normalized_inventory.get("valid")
or not normalized_inventory.get("complete")
or normalized_inventory.get("value") != value
):
return None
proposals = []
for index in cohort["proposals"][: len(labels)]:
parsed = normalized[tasks[index][0]]
if (
not parsed.get("valid")
or not parsed.get("complete")
or (not isinstance(parsed.get("value"), dict))
or (not any((value is not None for value in parsed["value"].values())))
):
return None
proposals.append(parsed)
(selected, conflicts) = image_rows._select(labels, proposals, cohort["text_keys"])
if conflicts or any(
(row is not proposals[index] for (index, row) in enumerate(selected))
):
return None
return {
"inventory_labels": labels,
"selected_ordinals": list(range(1, len(labels) + 1)),
"assignment_cost": 0,
"cohort": 8,
"original_proposal_capacity": len(cohort["proposals"]),
"inventory_valid_complete": True,
"same_ordinal_matches": True,
}
def _collector_record(
record: dict[str, Any], tasks: list[tuple[str, list[int], Any]]
) -> dict[str, Any]:
annotations = record["_meta"]["extraction_contract"]["fields"]
names = {name for (name, _, _) in tasks}
updated = dict(annotations)
for parent, annotation in annotations.items():
if parent in names and annotation.get("image_row_fields"):
updated[parent] = {
**annotation,
"image_row_fields": [
name for name in annotation["image_row_fields"] if name in names
],
}
return {
**record,
"_meta": {
**record["_meta"],
"extraction_contract": {
**record["_meta"]["extraction_contract"],
"fields": updated,
},
},
}
def _normalize_retained(
record: dict[str, Any],
predictions: dict[str, dict[str, Any]],
normalize: Callable[
[dict[str, Any], dict[str, dict[str, Any]]], dict[str, dict[str, Any]]
],
) -> dict[str, dict[str, Any]]:
meta = record.get("_meta", {})
view = {
**record,
"_meta": {
**meta,
**{
kind: {
name: spec
for (name, spec) in meta.get(kind, {}).items()
if name in predictions
}
for kind in ("printed_temporal", "printed_numbers")
},
},
}
return normalize(view, predictions)
async def generate(
engine: Any,
record: dict[str, Any],
tasks: list[tuple[str, list[int], Any]],
eos: int,
tokenizer: Any,
*,
parse_result: Callable[..., dict[str, Any]] | None = None,
normalize: Callable[
[dict[str, Any], dict[str, dict[str, Any]]], dict[str, dict[str, Any]]
]
| None = None,
**kwargs: Any,
) -> list[dict[str, Any]]:
if (
not record.get("_first_item_list_gating")
or record.get("_null_recovery")
or set(kwargs) - _BATCHED_KEYS - _SHARED_KEYS
):
return await _generate(engine, record, tasks, eos, tokenizer, **kwargs)
groups = _ordinary_lists(record, tasks, schema.list_item_name)
cohorts = (
_row_cohorts(record, tasks, schema.list_item_name)
if parse_result is not None and normalize is not None
else {}
)
if not groups and (not cohorts):
return await _generate(engine, record, tasks, eos, tokenizer, **kwargs)
size = len(tasks)
rids = kwargs.get("rid")
if (
not isinstance(rids, list)
or len(rids) != size
or len(set(rids)) != size
or (kwargs.get("input_ids") != [ids for (_, ids, _) in tasks])
):
raise ValueError("Cohort scheduling requires the original distinct AR batch")
_select_kwargs(kwargs, list(range(size)), size)
tails = {index for group in groups.values() for index in group[1:]}
tails.update(
(index for cohort in cohorts.values() for index in cohort["proposals"][8:])
)
results: list[Any] = [None] * size
async def submit(indices: list[int]) -> None:
if not indices:
return
selected = _select_kwargs(kwargs, indices, size)
subset = [tasks[index] for index in indices]
view = _collector_record(record, subset) if cohorts else record
received = await _generate(engine, view, subset, eos, tokenizer, **selected)
if "_completed_inventory_pruning" in view:
record["_completed_inventory_pruning"] = view[
"_completed_inventory_pruning"
]
if len(received) != len(indices) or any(
(
result.get("meta_info", {}).get("id") != rid
for (result, rid) in zip(received, selected["rid"], strict=True)
)
):
raise RuntimeError("Native completion differs from its cohort submission")
for index, result in zip(indices, received, strict=True):
results[index] = result
def omit(index: int, marker: str) -> None:
(name, ids, _) = tasks[index]
unexecuted[name] = len(ids)
results[index] = {
"_clair_unused_row_proposal": True,
marker: True,
"meta_info": {"completion_tokens": 0},
}
await submit([index for index in range(size) if index not in tails])
unexecuted: dict[str, int]
continued: list[int]
(unexecuted, continued) = ({}, [])
for group in groups.values():
if _completed_null(results[group[0]], eos, tokenizer):
for index in group[1:]:
omit(index, "_first_item_unexecuted")
else:
continued.extend(group[1:])
proofs = []
for parent, cohort in cohorts.items():
proof = _cohort_proof(
record,
tasks,
results,
parent,
cohort,
tokenizer,
cast(Callable[..., dict[str, Any]], parse_result),
image_rows,
lambda record, predictions: _normalize_retained(
record,
predictions,
cast(
Callable[
[dict[str, Any], dict[str, dict[str, Any]]],
dict[str, dict[str, Any]],
],
normalize,
),
),
)
proofs.append({"parent": parent, "proof": proof, "fallback": proof is None})
if proof is None:
continued.extend(cohort["proposals"][8:])
else:
for index in cohort["proposals"][8:]:
omit(index, "_row_cohort_unexecuted")
await submit(sorted(continued))
if any((result is None for result in results)):
raise RuntimeError("Cohort scheduling left an unaccounted result slot")
record["_unexecuted_list_items"] = unexecuted
record["_row_cohort_gating"] = proofs
return results
|