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
Download clairic_runtime/gdn/adapter.py from Cloudflare/clairic-flash: direct link, hf CLI and curl.
- Browser
- Download file 5.78 kB
-
https://huggingface.co/Cloudflare/clairic-flash/resolve/main/clairic_runtime/gdn/adapter.py
- Command line
-
hf download hf://Cloudflare/clairic-flash/clairic_runtime/gdn/adapter.py
-
curl -L -o adapter.py https://huggingface.co/Cloudflare/clairic-flash/resolve/main/clairic_runtime/gdn/adapter.py
5.78 kB
| """Pack native inputs and preserve native cache axes around selected kernels.""" | |
| from __future__ import annotations | |
| from typing import Any | |
| import torch | |
| from .. import recurrent_metadata | |
| def _packed_indices(mixed: torch.Tensor, mask: torch.Tensor | None) -> tuple[torch.Tensor, torch.Tensor]: | |
| """Return valid-token indices and one boundary per native row.""" | |
| batch, length = mixed.shape[:2] | |
| if mask is None: | |
| indices = torch.arange(batch * length, device=mixed.device) | |
| boundaries = torch.arange(batch + 1, device=mixed.device, dtype=torch.int32) * length | |
| return indices, boundaries | |
| if mask.ndim != 2 or mask.shape[0] != batch or mask.shape[1] < length: | |
| raise ValueError("Expected a two-dimensional mask covering current tokens") | |
| valid = mask[:, -length:].to(device=mixed.device, dtype=torch.bool) | |
| lengths = valid.sum(dim=1, dtype=torch.int32) | |
| if bool((lengths == 0).any()): | |
| raise ValueError("Each recurrent sequence must contain a current token") | |
| boundaries = torch.cat((lengths.new_zeros(1), lengths.cumsum(0, dtype=torch.int32))) | |
| return valid.reshape(-1).nonzero().flatten(), boundaries | |
| def packed_indices(mixed: torch.Tensor, mask: torch.Tensor | None) -> tuple[torch.Tensor, torch.Tensor]: | |
| """Reuse exact packing metadata within an owned inference forward.""" | |
| return recurrent_metadata.packed_indices(mixed, mask, _packed_indices) | |
| def source_state(native: torch.Tensor | None, batch: int, heads: int, key_dim: int, | |
| value_dim: int, reference: torch.Tensor) -> torch.Tensor: | |
| """Map native K,V cache dimensions to the selected kernel's V,K layout.""" | |
| if native is None: | |
| return torch.zeros(batch, heads, value_dim, key_dim, | |
| device=reference.device, dtype=torch.float32) | |
| if tuple(native.shape) != (batch, heads, key_dim, value_dim): | |
| raise ValueError("Unexpected native recurrent cache shape") | |
| if native.dtype != torch.float32: | |
| raise ValueError("Recurrent states must retain float32 precision") | |
| return native.transpose(-1, -2).contiguous() | |
| def native_state(source: torch.Tensor) -> torch.Tensor: | |
| """Expose K,V cache axes without narrowing or copying source state.""" | |
| return source.transpose(-1, -2) | |
| def store_state(cache: Any, layer_idx: int, state: torch.Tensor) -> torch.Tensor: | |
| """Initialize FP32 recurrent storage while retaining native cache lifecycle.""" | |
| if state.dtype != torch.float32: | |
| raise ValueError("Expected float32 recurrent state") | |
| layer = cache.layers[layer_idx] | |
| if not layer.is_recurrent_states_initialized: | |
| layer.recurrent_states = torch.empty_like(state) | |
| if not torch.compiler.is_compiling(): | |
| torch._dynamo.mark_static_address(layer.recurrent_states) | |
| layer.is_recurrent_states_initialized = True | |
| elif layer.recurrent_states.dtype != torch.float32: | |
| raise ValueError("Cannot resume a narrowed recurrent cache") | |
| return cache.update_recurrent_state(state, layer_idx) | |
| def prefill(mixed: torch.Tensor, a: torch.Tensor, b: torch.Tensor, A_log: torch.Tensor, | |
| dt_bias: torch.Tensor, key_heads: int, value_heads: int, key_dim: int, | |
| value_dim: int, initial_state: torch.Tensor | None, | |
| attention_mask: torch.Tensor | None) -> tuple[torch.Tensor, torch.Tensor]: | |
| """Run packed prefill and scatter valid outputs to native padded shape.""" | |
| from .gating import fused_gdn_gating | |
| from .index import prepare_chunk_indices | |
| from .l2norm import l2norm_fwd | |
| from .prefill import chunk_gated_delta_rule_fwd | |
| batch, length, _ = mixed.shape | |
| indices, boundaries = packed_indices(mixed, attention_mask) | |
| packed = mixed.flatten(0, 1).index_select(0, indices) | |
| pa = a.flatten(0, 1).index_select(0, indices) | |
| pb = b.flatten(0, 1).index_select(0, indices) | |
| query, key, value = torch.split(packed, | |
| (key_heads * key_dim, key_heads * key_dim, value_heads * value_dim), dim=-1) | |
| query = l2norm_fwd(query.reshape(1, -1, key_heads, key_dim).contiguous()) | |
| key = l2norm_fwd(key.reshape(1, -1, key_heads, key_dim).contiguous()) | |
| value = value.reshape(1, -1, value_heads, value_dim).contiguous() | |
| gate, beta = fused_gdn_gating(A_log, pa, pb, dt_bias) | |
| state = source_state(initial_state, batch, value_heads, key_dim, value_dim, mixed) | |
| slots = recurrent_metadata.sequence_slots(mixed) | |
| _, output, *_ = chunk_gated_delta_rule_fwd(query, key, value, gate, beta, | |
| key_dim ** -0.5, state, slots, boundaries, prepare_chunk_indices(boundaries, 64)) | |
| padded = mixed.new_zeros(batch * length, value_heads, value_dim) | |
| padded.index_copy_(0, indices, output.squeeze(0)) | |
| return padded.reshape(batch, length, value_heads, value_dim), native_state(state) | |
| def decode(mixed: torch.Tensor, a: torch.Tensor, b: torch.Tensor, A_log: torch.Tensor, | |
| dt_bias: torch.Tensor, initial_state: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: | |
| """Use the selected packed single-token recurrent operation.""" | |
| from .recurrent import fused_recurrent_gated_delta_rule_packed_decode | |
| batch, length, _ = mixed.shape | |
| if length != 1: | |
| raise ValueError("Packed recurrent decode requires one token per row") | |
| _, heads, key_dim, value_dim = initial_state.shape | |
| state = source_state(initial_state, batch, heads, key_dim, value_dim, mixed) | |
| output = mixed.new_empty(batch, 1, heads, value_dim) | |
| slots = recurrent_metadata.sequence_slots(mixed) | |
| fused_recurrent_gated_delta_rule_packed_decode(mixed[:, 0].contiguous(), | |
| a[:, 0].contiguous(), b[:, 0].contiguous(), A_log, dt_bias, key_dim ** -0.5, | |
| state, output, slots, use_qk_l2norm_in_kernel=True) | |
| return output, native_state(state) | |