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: 8,203 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 | from __future__ import annotations
import base64
import io
import json
from typing import TYPE_CHECKING
from .api import ClairImage
from . import ar, grounding, identifier_spelling, image_focus
from PIL import Image
if TYPE_CHECKING:
from collections.abc import Callable, Sequence
from typing import Any
from .api import PredictRequest
from .backend import ARTask, EncodedRequest
from transformers import PreTrainedTokenizerBase
VERSION = "27b-po-all-content-horizontal2-supplement-v2"
MAX_PIXELS = 1600000
def supplement(images: Sequence[Image.Image]) -> ClairImage:
views: list[Image.Image] = []
for image in images:
focused = image_focus.focus_document(image)
views.append(
focused.resize(
(round(focused.width * 2.0), focused.height), Image.Resampling.BICUBIC
)
)
board = Image.new(
"RGB",
(
max((image.width for image in views)),
sum((image.height for image in views)) + 32 * (len(views) - 1),
),
"white",
)
y = 0
for image in views:
board.paste(image, (0, y))
y += image.height + 32
scale = min(1.0, (MAX_PIXELS / (board.width * board.height)) ** 0.5)
if scale < 1:
board = board.resize(
(max(1, int(board.width * scale)), max(1, int(board.height * scale))),
Image.Resampling.BICUBIC,
)
buffer = io.BytesIO()
board.save(buffer, format="PNG")
return ClairImage.model_validate(
{
"content_type": "image/png",
"base64": base64.b64encode(buffer.getvalue()).decode(),
}
)
def plain_task(
task: ARTask,
wire_spec: dict[str, Any],
public_spec: dict[str, Any],
tokenizer: PreTrainedTokenizerBase,
max_length: int,
) -> tuple[ARTask, dict[str, Any], dict[str, str]]:
(name, ids, media) = task
source_ids = list(ids)
text = tokenizer.decode(
ids, skip_special_tokens=False, clean_up_tokenization_spaces=False
)
if tokenizer.encode(text, add_special_tokens=False) != source_ids:
raise ValueError(
"PO scalar source prompt cannot make a lossless tokenizer roundtrip"
)
before = identifier_spelling.INSTRUCTION + "\n"
if text.count(before) != 1 or wire_spec.get("instructions", "").count(before) != 1:
raise ValueError(
"PO scalar source must contain exactly one spelling instruction"
)
scalar = {
**wire_spec,
"instructions": wire_spec["instructions"].replace(before, ""),
}
for key in ("minLength", "maxLength", "pattern"):
if key in public_spec:
scalar[key] = public_spec[key]
(old_schema, schema) = (ar.schema(wire_spec), ar.schema(scalar))
text = text.replace(before, "")
if schema != old_schema:
old = "\nSchema: " + json.dumps(
old_schema, ensure_ascii=False, separators=(",", ":")
)
new = "\nSchema: " + json.dumps(
schema, ensure_ascii=False, separators=(",", ":")
)
if text.count(old) != 1:
raise ValueError("PO scalar source must contain exactly one native schema")
text = text.replace(old, new)
plain_ids = tokenizer.encode(text, add_special_tokens=False)
if len(plain_ids) > max_length:
raise ValueError("PO scalar detail exceeds the complete model context")
return ((name, plain_ids, media), scalar, {"json_schema": json.dumps(schema)})
def selected_fields(
request: PredictRequest, encoded: EncodedRequest
) -> dict[str, dict[str, Any]]:
image_count = len(request.images)
if image_count not in (1, 2, 3) or request.documents:
return {}
(record, _, tasks, _, _) = encoded
contracts = record.get("_meta", {}).get("identifier_spelling", {})
active = {name for (name, _, media) in tasks if media is not None}
return {
name: request.fields[name]
for (name, contract) in contracts.items()
if name in active
and contract.get("wire") == identifier_spelling.VERSION
and (not request.fields[name].get("knowledge"))
and (not grounding.state_only(request.fields[name]))
}
def prepare(
request: PredictRequest,
encoded: EncodedRequest,
*,
encode_child: Callable[[PredictRequest], EncodedRequest],
tokenizer: PreTrainedTokenizerBase,
max_length: int,
media_groups: Callable[..., list[dict[str, Any]]],
) -> EncodedRequest:
selected = selected_fields(request, encoded)
if not selected:
return encoded
image_count = len(request.images)
(record, forecasts, tasks, typed, eos) = encoded
contracts = record["_meta"]["identifier_spelling"]
originals = record.get("_original_images", ())
if len(originals) != image_count:
raise ValueError("PO supplement must retain every original image in order")
fields = image_focus.preserve_scalar_fields(selected)
supplements = (
tuple((supplement([image]) for image in originals))
if image_count == 2
else (supplement(originals),)
)
child_request = request.model_copy(
update={"fields": fields, "images": (*request.images, *supplements)}
)
(child, _, replacements, child_typed, child_eos) = encode_child(child_request)
by_name = {task[0]: task for task in replacements}
if (
child_typed is not None
or child_eos != eos
or set(by_name) != set(selected)
or (len(replacements) != len(selected))
):
raise ValueError(
"PO supplement must retain the complete selected scalar task contract"
)
plain = image_count == 1
constraints: dict[str, dict[str, Any]]
remaining: dict[str, Any] | list[str]
(fields, constraints) = ({}, {})
for name, public_spec in selected.items():
if plain:
(by_name[name], fields[name], constraints[name]) = plain_task(
by_name[name], child["fields"][name], public_spec, tokenizer, max_length
)
else:
fields[name] = child["fields"][name]
constraints[name] = child["_ar_constraints"][name]
updated = [by_name.get(task[0], task) for task in tasks]
salts = dict(record.get("_ar_task_salts", {}))
for name in selected:
record["fields"][name] = fields[name]
record["_ar_constraints"][name] = constraints[name]
salts[name] = child.get("_ar_task_salts", {}).get(name, child["_cache_salt"])
if plain:
remaining = {
name: contract
for (name, contract) in contracts.items()
if name not in selected
}
if remaining:
record["_meta"]["identifier_spelling"] = remaining
else:
record["_meta"].pop("identifier_spelling", None)
for key in ("model_document_scalar_view", "identifier_spelling_image_view"):
previous = record["_meta"].get(key)
if previous is None:
continue
remaining = [name for name in previous["fields"] if name not in selected]
if not remaining:
record["_meta"].pop(key)
else:
retained = {**previous, "fields": remaining}
if "po_fields" in retained:
retained["po_fields"] = [
name for name in retained["po_fields"] if name not in selected
]
record["_meta"][key] = retained
record["_meta"]["model_po_supplement"] = {
"version": VERSION,
"fields": list(selected),
"original_image_count": image_count,
"native_image_count": len(child_request.images),
"horizontal_scale": 2.0,
"supplement_max_pixels": MAX_PIXELS,
"wire": "native_scalar" if plain else identifier_spelling.VERSION,
}
record.update(
_ar_tasks=updated, _ar_task_salts=salts, _ar_prefix=[], _ar_group_prefixes=[]
)
record["_ar_media_groups"] = media_groups(
updated,
default_salt=record["_cache_salt"],
task_salts=salts,
annotations=record["_meta"]["extraction_contract"]["fields"],
)
return (record, forecasts, updated, typed, eos)
|