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: 11,134 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 | """Owned single-layer graph capture and strict resource lifecycle."""
from __future__ import annotations
import threading
from typing import Any
import weakref
def tensor_key(value: Any, *, version: bool) -> tuple[Any, ...]:
try:
counter = value._version if version else None
except RuntimeError:
counter = None
storage = value.untyped_storage()
return (id(value), counter, value.data_ptr(), storage.data_ptr(), storage.nbytes(),
value.storage_offset(), str(value.device), str(value.dtype),
tuple(value.shape), tuple(value.stride()))
def callable_key(value: Any) -> tuple[int, int | None]:
return (id(getattr(value, "__func__", value)),
id(value.__self__) if hasattr(value, "__self__") else None)
def input_key(value: Any) -> tuple[Any, ...]:
return str(value.device), str(value.dtype), tuple(value.shape), tuple(value.stride())
def owner_key(torch: Any, owner: Any, cache: Any, operation: Any) -> tuple[Any, ...]:
modules = []
for name, module in owner.named_modules():
if module._forward_hooks or module._forward_pre_hooks:
raise RuntimeError("Forward hooks invalidate decoder capture")
scalars = tuple(sorted((key, value) for key, value in vars(module).items()
if not key.startswith("_") and isinstance(value, (str, bool, int, float))))
callables = tuple(sorted((key, callable_key(value)) for key, value in vars(module).items()
if not key.startswith("_") and callable(value)))
modules.append((name, id(module), callable_key(module.forward), scalars, callables))
registry = torch.nn.modules.module
if registry._global_forward_hooks or registry._global_forward_pre_hooks:
raise RuntimeError("Global forward hooks invalidate decoder capture")
parameters = tuple((name, tensor_key(value, version=True))
for name, value in owner.named_parameters())
if any(item[1][1] is None for item in parameters):
raise RuntimeError("Versioned parameters required")
buffers = tuple((name, tensor_key(value, version=True))
for name, value in owner.named_buffers())
layer = cache.layers[0]
if not (cache.has_previous_state(0) and layer.is_conv_states_initialized
and layer.is_recurrent_states_initialized):
raise RuntimeError("Already initialized live decode cache required")
states = tuple(tensor_key(value, version=False)
for value in (layer.conv_states, layer.recurrent_states))
return (id(owner), id(cache), id(layer), callable_key(operation), tuple(modules),
parameters, buffers, states, callable_key(cache.update_recurrent_state),
callable_key(cache.has_previous_state), callable_key(layer.update_recurrent_state))
def describe_pool(segments: list[dict[str, Any]], pool: tuple[int, int]) -> dict[str, Any]:
blocks = [block for segment in segments for block in segment["blocks"]]
pending = {"active_awaiting_free", "active_pending_free"}
return {"pool": pool, "segments": segments,
"foreign_pool": any(tuple(segment["segment_pool_id"]) != pool for segment in segments),
"reserved_bytes": sum(segment["total_size"] for segment in segments),
"allocated_bytes": sum(segment["allocated_size"] for segment in segments),
"active_bytes": sum(segment["active_size"] for segment in segments),
"active_allocated_blocks": sum(block["state"] == "active_allocated" for block in blocks),
"pending_blocks": sum(block["state"] in pending for block in blocks),
"unknown_block_states": sorted({block["state"] for block in blocks}
- pending - {"active_allocated", "inactive"})}
class DecoderGraph:
"""One exact bucket/dtype graph with owned input and cloned FP32 outputs.
Capture/replay must run on the engine's exclusive CUDA worker. Mutation is
intentional only for the two captured cache tensors. Any owner, parameter,
callable, stream, runtime, cache storage or hook change rejects the session.
Capture/replay failures propagate; callers must close the engine rather than
retrying eager work on a possibly invalid CUDA context.
"""
def __init__(self, backend: Any, owner: Any, cache: Any, value: Any,
operation: Any, *, exclusive_process: bool) -> None:
if not exclusive_process:
raise ValueError("Exclusive process required")
self.backend, self.owner, self.cache, self.operation = backend, owner, cache, operation
self.thread, self.closed = threading.get_ident(), False
self.context = backend.context(value)
self.input = input_key(value)
self.key = owner_key(backend.torch, owner, cache, operation)
self.entry = backend.capture(value, operation, cache)
self.check(value)
def check(self, value: Any) -> None:
if self.closed or self.thread != threading.get_ident():
raise RuntimeError("Closed graph or foreign thread")
if self.backend.context(value) != self.context or input_key(value) != self.input:
raise RuntimeError("Input layout, dtype, stream or runtime changed")
if owner_key(self.backend.torch, self.owner, self.cache, self.operation) != self.key:
raise RuntimeError("Decoder owner, parameter, callable or cache changed")
def run(self, value: Any) -> Any:
self.check(value)
return self.direct(value)
def direct(self, value: Any) -> Any:
# The padded manager executes complete guards before direct replay.
return self.backend.replay_clone(self.entry, value)
def close(self) -> dict[str, Any]:
if self.closed or self.thread != threading.get_ident():
raise RuntimeError("Closed graph or foreign thread")
details = self.backend.release(self.entry)
self.owner = self.cache = self.operation = self.key = None
self.closed = True
return details
class TorchBackend:
def __init__(self, torch: Any, runtime_key: Any) -> None:
self.torch, self.runtime_key = torch, runtime_key
def context(self, value: Any) -> tuple[Any, ...]:
torch = self.torch
if (torch.is_grad_enabled() or torch.compiler.is_compiling() or not value.is_cuda
or torch.cuda.is_current_stream_capturing() or not value.is_contiguous()
or value.dtype not in (torch.bfloat16, torch.float32)
or value.ndim != 3 or value.shape[1] != 1 or min(value.shape) < 1):
raise RuntimeError("Unsupported decoder graph context/input")
return (torch.cuda.current_stream(value.device).cuda_stream, self.runtime_key())
def capture(self, value: Any, operation: Any, cache: Any) -> dict[str, Any]:
torch = self.torch
current = torch.cuda.current_stream(value.device)
side = torch.cuda.Stream(device=value.device)
side.wait_stream(current)
layer = cache.layers[0]
states = (layer.conv_states, layer.recurrent_states)
saved = tuple(state.clone(memory_format=torch.preserve_format) for state in states)
side.wait_stream(current)
def restore(current_states: Any, before_states: Any) -> None:
for state, before in zip(current_states, before_states, strict=True):
state.copy_(before)
graph = torch.cuda.CUDAGraph()
with torch.cuda.stream(side):
static_input = torch.empty_like(value)
static_input.copy_(value)
for _ in range(2):
warm_output = operation(static_input)
side.synchronize()
del warm_output
restore(states, saved)
side.synchronize()
graph.capture_begin(capture_error_mode="global")
try:
static_output = operation(static_input)
finally:
graph.capture_end()
side.synchronize()
restore(states, saved)
current.wait_stream(side)
current.synchronize()
assert static_output.dtype == torch.float32
static_input.record_stream(current)
static_output.record_stream(current)
for state, before in zip(states, saved, strict=True):
assert torch.equal(state.contiguous().view(torch.int16), before.contiguous().view(torch.int16))
del saved, states
pool = tuple(graph.pool())
details = describe_pool(torch.cuda.memory_snapshot(mempool_id=pool, include_traces=False), pool)
assert not details["foreign_pool"] and details["reserved_bytes"] > 0
return {"graph": graph, "static_input": static_input, "static_output": static_output,
"stream": current, "side": side, "pool": pool,
"pool_at_capture": details,
"static_input_storage_bytes": static_input.untyped_storage().nbytes(),
"capture_mutation_restored": True}
def replay_clone(self, entry: dict[str, Any], value: Any) -> Any:
if self.torch.cuda.current_stream(value.device).cuda_stream != entry["stream"].cuda_stream:
raise RuntimeError("Replay stream changed")
entry["static_input"].copy_(value)
entry["graph"].replay()
return entry["static_output"].clone()
def release(self, entry: dict[str, Any]) -> dict[str, Any]:
torch = self.torch
refs = {key: weakref.ref(entry[key]) for key in ("static_input", "static_output", "graph")}
events = []
def snapshot(stage: str) -> None:
item = describe_pool(torch.cuda.memory_snapshot(
mempool_id=entry["pool"], include_traces=False), entry["pool"])
item.update(stage=stage, weak_alive={key: ref() is not None for key, ref in refs.items()})
events.append(item)
entry["stream"].synchronize()
entry["side"].synchronize()
snapshot("before_drop")
entry["graph"].reset()
entry["static_input"] = entry["static_output"] = entry["graph"] = None
snapshot("after_drop")
entry["stream"].synchronize()
entry["side"].synchronize()
with torch.cuda.stream(entry["stream"]):
probe = torch.empty((1,), device=entry["stream"].device, dtype=torch.uint8)
del probe
entry["stream"].synchronize()
snapshot("after_fence_allocator_poll")
final = events[-1]
checks = {"no_graph_refs": not any(final["weak_alive"].values()),
"correct_pool": not final["foreign_pool"],
"pool_active_zero": final["active_bytes"] == 0,
"pool_allocated_zero": final["allocated_bytes"] == 0,
"no_active_blocks": final["active_allocated_blocks"] == 0,
"no_pending_blocks": final["pending_blocks"] == 0,
"known_block_states": not final["unknown_block_states"]}
return {"passed": all(checks.values()), "checks": checks, "events": events}
|