Image-Text-to-Text
Transformers
Safetensors
qwen3_5
clairic
cloudflare
systemone
multimodal
structured-output
information-extraction
custom-code
conversational
Instructions to use Cloudflare/clairic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Cloudflare/clairic with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="Cloudflare/clairic") 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") model = AutoModelForMultimodalLM.from_pretrained("Cloudflare/clairic", 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 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Cloudflare/clairic" # 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", "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
- SGLang
How to use Cloudflare/clairic 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" \ --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", "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" \ --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", "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 with Docker Model Runner:
docker model run hf.co/Cloudflare/clairic
File size: 7,605 Bytes
524fd9d | 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 | """Reuse pinned FA3 unpadding metadata within one owned decode forward."""
from __future__ import annotations
from collections.abc import Callable, Iterator
from contextlib import contextmanager
from contextvars import ContextVar
import threading
from typing import Any
import torch
from transformers.integrations.flash_attention import flash_attention_forward
from transformers.modeling_flash_attention_utils import _get_unpad_data, _index_first_axis
_PINNED_FA3 = "kernels-community/flash-attn3@e29f138fc363b396e5d2706c8a5f6fa7d36f41e0"
_VARLEN_KEYS = frozenset({"cu_seq_lens_q", "cu_seq_lens_k", "max_length_q", "max_length_k"})
class _Entry:
def __init__(self) -> None:
self.key: tuple[Any, ...] | None = None
self.mask: torch.Tensor | None = None
self.indices: torch.Tensor | None = None
self.cu_k: torch.Tensor | None = None
self.cu_q: torch.Tensor | None = None
self.maximum = 0
self.bytes = 0
def clear(self) -> None:
self.key, self.mask, self.indices, self.cu_k, self.cu_q = None, None, None, None, None
self.maximum, self.bytes = 0, 0
class DecodeMetadata:
def __init__(self, max_bytes: int = 8 * 1024**2) -> None:
if max_bytes <= 0:
raise ValueError("Expected a positive metadata bound")
self.max_bytes = max_bytes
self._lock = threading.Lock()
self._current: ContextVar[_Entry | None] = ContextVar("clairic_decode_metadata", default=None)
self._active: _Entry | None = None
self._closed = False
self.scopes = 0
self.materializations = 0
self.hits = 0
self.peak_bytes = 0
@contextmanager
def scope(self, args: tuple[Any, ...], kwargs: dict[str, Any]) -> Iterator[None]:
token = self._current.set(None)
acquired = False
entry = None
try:
inputs = kwargs.get("input_ids", args[0] if args else None)
if inputs is None:
inputs = kwargs.get("inputs_embeds")
eligible = (not self._closed and not torch.is_grad_enabled()
and torch.is_inference_mode_enabled() and not torch.compiler.is_compiling()
and isinstance(inputs, torch.Tensor) and inputs.is_cuda
and inputs.ndim in (2, 3) and inputs.shape[1] == 1)
if eligible and self._lock.acquire(blocking=False):
acquired = True
if not self._closed:
entry = _Entry()
self._active = entry
self._current.set(entry)
self.scopes += 1
yield
finally:
if entry is not None:
entry.clear()
self._active = None
self._current.reset(token)
if acquired:
self._lock.release()
def _metadata(self, mask: torch.Tensor) -> _Entry | None:
entry = self._current.get()
if self._closed or entry is None:
return None
batch = mask.shape[0]
if mask.numel() * 8 + 2 * (batch + 1) * 4 > self.max_bytes:
entry.clear()
return None
try:
version = mask._version
except RuntimeError:
version = None
key = (id(mask), version, mask.data_ptr(), mask.storage_offset(), mask.device,
mask.dtype, tuple(mask.shape), tuple(mask.stride()),
torch.cuda.current_stream(mask.device).cuda_stream)
if entry.mask is mask and entry.key == key:
self.hits += 1
return entry
entry.clear()
indices, cu_k, maximum = _get_unpad_data(mask)
cu_q = torch.arange(batch + 1, dtype=torch.int32, device=mask.device)
retained = sum(value.numel() * value.element_size() for value in (indices, cu_k, cu_q))
if retained > self.max_bytes:
return None
entry.key, entry.mask = key, mask
entry.indices, entry.cu_k, entry.cu_q = indices, cu_k, cu_q
entry.maximum, entry.bytes = int(maximum.item()), retained
self.materializations += 1
self.peak_bytes = max(self.peak_bytes, retained)
return entry
def attend(self, interface: Callable[..., tuple[torch.Tensor, torch.Tensor | None]],
module: Any, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor,
mask: torch.Tensor | None, *, dropout: float, scaling: float,
**kwargs: Any) -> tuple[torch.Tensor, torch.Tensor | None]:
eligible = (not self._closed and self._current.get() is not None
and not torch.is_grad_enabled() and torch.is_inference_mode_enabled()
and not torch.compiler.is_compiling() and interface is flash_attention_forward
and module.config._attn_implementation == _PINNED_FA3
and dropout == 0.0
and getattr(module.config, "sliding_window", None) is None
and kwargs.get("sliding_window") is None and not (_VARLEN_KEYS & kwargs.keys())
and mask is not None and mask.is_cuda and mask.ndim == 2
and mask.dtype == torch.bool and mask.numel() > 0
and all(item.is_cuda and item.dtype == torch.bfloat16 and item.ndim == 4
and item.device == mask.device and item.stride(-1) == 1
for item in (query, key, value)))
if eligible:
assert mask is not None
positions = kwargs.get("position_ids")
eligible = (query.shape[0] == key.shape[0] == value.shape[0] == mask.shape[0]
and query.shape[2] == 1 and key.shape[2] == value.shape[2] == mask.shape[1]
and mask.shape[1] > 1
and query.shape[1] in (16, 24) and key.shape[1] == value.shape[1] == 4
and query.shape[3] == key.shape[3] == value.shape[3] == 256
and (positions is None or (isinstance(positions, torch.Tensor)
and positions.ndim == 2 and tuple(positions.shape) == (query.shape[0], 1))))
entry = self._metadata(mask) if eligible and mask is not None else None
if entry is None:
active = self._current.get()
if active is not None:
active.clear()
return interface(module, query, key, value, mask, dropout=dropout, scaling=scaling, **kwargs)
assert entry.indices is not None and entry.cu_q is not None and entry.cu_k is not None
packed_key = _index_first_axis(key.transpose(1, 2), entry.indices).unsqueeze(0).transpose(1, 2)
packed_value = _index_first_axis(value.transpose(1, 2), entry.indices).unsqueeze(0).transpose(1, 2)
return interface(module, query, packed_key, packed_value, None, dropout=dropout, scaling=scaling,
cu_seq_lens_q=entry.cu_q, cu_seq_lens_k=entry.cu_k,
max_length_q=1, max_length_k=entry.maximum, **kwargs)
def snapshot(self) -> dict[str, int | bool]:
return {"scopes": self.scopes, "materializations": self.materializations, "hits": self.hits,
"peak_bytes": self.peak_bytes, "retained_bytes": self._active.bytes if self._active else 0,
"closed": self._closed}
def close(self) -> None:
self._closed = True
with self._lock:
if self._active is not None:
self._active.clear()
self._active = None
|