Instructions to use tiny-random/sensenova-loop with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tiny-random/sensenova-loop with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="tiny-random/sensenova-loop", trust_remote_code=True)
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 AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("tiny-random/sensenova-loop", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use tiny-random/sensenova-loop with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tiny-random/sensenova-loop"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tiny-random/sensenova-loop",
		"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/tiny-random/sensenova-loop

SGLang

How to use tiny-random/sensenova-loop 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 "tiny-random/sensenova-loop" \
    --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": "tiny-random/sensenova-loop",
		"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 "tiny-random/sensenova-loop" \
        --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": "tiny-random/sensenova-loop",
		"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 tiny-random/sensenova-loop with Docker Model Runner:
```
docker model run hf.co/tiny-random/sensenova-loop
```

sensenova-loop / .meta /create_model.py

yujiepan

Upload folder using huggingface_hub

2c1a8e3 verified 2 days ago

Raw History Blame Contribute Delete

6 kB

def create_model(source_model_id: str, source_revision: str, save_folder: str) -> str:
    import copy
    import json
    from pathlib import Path

    import torch
    from huggingface_hub import hf_hub_download
    from safetensors.torch import save_file
    from transformers import AutoConfig, AutoModelForImageTextToText, set_seed

    folder = Path(save_folder)
    folder.mkdir(parents=True, exist_ok=True)
    # Transformers 5.18 loads this custom Loop/mHC wrapper around the native Qwen3.5 backbone; ship it pinned.
    for name in (
        "modeling_sensenova_loop.py",
        "chat_template.jinja",
        "processor_config.json",
        "preprocessor_config.json",
        "video_preprocessor_config.json",
        "tokenizer.json",
        "tokenizer_config.json",
    ):
        (folder / name).write_bytes(Path(hf_hub_download(source_model_id, name, revision=source_revision)).read_bytes())

    config = json.loads(Path(hf_hub_download(source_model_id, "config.json", revision=source_revision)).read_text())
    # Pre-loop 0-3, looped block 4-7 executed twice, post-loop 8-11: twelve physical layers / sixteen executions,
    # four complete linear/linear/linear/full cycles. Residual width 8, head dim 32 with eight Q / four KV heads.
    text_config = config["text_config"]
    text_config.update(
        head_dim=32,
        hidden_size=8,
        intermediate_size=32,
        num_hidden_layers=12,
        num_attention_heads=8,
        num_key_value_heads=4,
        linear_key_head_dim=32,
        linear_num_key_heads=4,
        linear_num_value_heads=8,
        linear_value_head_dim=32,
    )
    text_config["layer_types"] = ["linear_attention", "linear_attention", "linear_attention", "full_attention"] * 3
    text_config["rope_parameters"]["mrope_section"] = [1, 1, 2]
    # Two-stream mHC spans exactly the looped block [4, 8), matching the source [4, 28) cadence.
    config["loop_settings"] = {"counts": [2], "end": 8, "eval_count": 2, "start": 4}
    config["mhc"] = {
        "end": 8,
        "fixed_identity": False,
        "iterations": 20,
        "pre_scale": 1.0,
        "start": 4,
        "streams": 2,
    }
    config["hidden_size"] = 8
    # Four vision blocks keep the native Qwen3.5 tower and the fixed patch/temporal/merge and position contracts.
    config["vision_config"].update(
        hidden_size=64,
        intermediate_size=128,
        num_heads=2,
        out_hidden_size=8,
        depth=4,
    )
    (folder / "config.json").write_text(json.dumps(config, indent=2) + "\n")

    set_seed(42)
    model = AutoModelForImageTextToText.from_config(
        AutoConfig.from_pretrained(folder, trust_remote_code=True),
        trust_remote_code=True,
        dtype=torch.bfloat16,
    )
    # The source keeps the hybrid linear-attention SSM parameters FP32 even under a BF16 load.
    from transformers.models.qwen3_5.modeling_qwen3_5 import Qwen3_5GatedDeltaNet

    for layer in model.model.language_model.layers:
        linear = getattr(layer, "linear_attn", None)
        if isinstance(linear, Qwen3_5GatedDeltaNet):
            with torch.no_grad():
                linear.A_log = torch.nn.Parameter(linear.A_log.float())
                linear.dt_bias = torch.nn.Parameter(linear.dt_bias.float())
                linear.norm.weight = torch.nn.Parameter(linear.norm.weight.float())
                linear.conv1d.weight = torch.nn.Parameter(linear.conv1d.weight.float())
    # One BF16 MTP draft layer reusing a full-attention backbone layer outside the mHC range.
    full_layer = model.model.language_model.layers[11]
    model.mtp = torch.nn.ModuleDict(
        {
            "pre_fc_norm_embedding": torch.nn.RMSNorm(8),
            "fc": torch.nn.Linear(16, 8, bias=False),
            "layers": torch.nn.ModuleList([copy.deepcopy(full_layer)]),
            "norm": torch.nn.RMSNorm(8),
            "pre_fc_norm_hidden": torch.nn.RMSNorm(8),
        }
    )
    model = model.cpu()
    representation = str(model)

    state = model.state_dict()
    backbone, draft = {}, {}
    for name, tensor in state.items():
        if name.startswith("mtp."):
            draft[name] = tensor.to(torch.bfloat16).contiguous()
        else:
            # The shipped checkpoint stores the complete backbone FP32, mHC routers included.
            backbone[name] = tensor.to(torch.float32).contiguous()
    save_file(backbone, str(folder / "model.safetensors"), metadata={"format": "pt"})
    save_file(draft, str(folder / "mtp.safetensors"), metadata={"format": "pt"})
    # Rewrite the edited config (save was manual) and pin the source generation config bytes.
    (folder / "config.json").write_text(json.dumps(config, indent=2) + "\n")
    (folder / "generation_config.json").write_bytes(
        Path(hf_hub_download(source_model_id, "generation_config.json", revision=source_revision)).read_bytes()
    )

    expected_draft = {
        "mtp.fc.weight",
        "mtp.pre_fc_norm_embedding.weight",
        "mtp.pre_fc_norm_hidden.weight",
        "mtp.norm.weight",
        "mtp.layers.0.input_layernorm.weight",
        "mtp.layers.0.post_attention_layernorm.weight",
        "mtp.layers.0.self_attn.q_proj.weight",
        "mtp.layers.0.self_attn.k_proj.weight",
        "mtp.layers.0.self_attn.v_proj.weight",
        "mtp.layers.0.self_attn.o_proj.weight",
        "mtp.layers.0.self_attn.q_norm.weight",
        "mtp.layers.0.self_attn.k_norm.weight",
        "mtp.layers.0.mlp.gate_proj.weight",
        "mtp.layers.0.mlp.up_proj.weight",
        "mtp.layers.0.mlp.down_proj.weight",
    }
    assert set(draft) == expected_draft, sorted(set(draft) ^ expected_draft)
    assert all(
        name.count("mhc_attn") + name.count("mhc_ffn") == 0 or backbone[name].dtype == torch.float32
        for name in backbone
    )
    return representation


create_model(
    source_model_id="sensenova/SenseNova-Loop-9B-Preview",
    source_revision="9937e91c7301b7dcbb5bf17e3e4e95c499a11e99",
    save_folder="/tmp/sensenova-loop",
)

	def create_model(source_model_id: str, source_revision: str, save_folder: str) -> str:
	import copy
	import json
	from pathlib import Path

	import torch
	from huggingface_hub import hf_hub_download
	from safetensors.torch import save_file
	from transformers import AutoConfig, AutoModelForImageTextToText, set_seed

	folder = Path(save_folder)
	folder.mkdir(parents=True, exist_ok=True)
	# Transformers 5.18 loads this custom Loop/mHC wrapper around the native Qwen3.5 backbone; ship it pinned.
	for name in (
	"modeling_sensenova_loop.py",
	"chat_template.jinja",
	"processor_config.json",
	"preprocessor_config.json",
	"video_preprocessor_config.json",
	"tokenizer.json",
	"tokenizer_config.json",
	):
	(folder / name).write_bytes(Path(hf_hub_download(source_model_id, name, revision=source_revision)).read_bytes())

	config = json.loads(Path(hf_hub_download(source_model_id, "config.json", revision=source_revision)).read_text())
	# Pre-loop 0-3, looped block 4-7 executed twice, post-loop 8-11: twelve physical layers / sixteen executions,
	# four complete linear/linear/linear/full cycles. Residual width 8, head dim 32 with eight Q / four KV heads.
	text_config = config["text_config"]
	text_config.update(
	head_dim=32,
	hidden_size=8,
	intermediate_size=32,
	num_hidden_layers=12,
	num_attention_heads=8,
	num_key_value_heads=4,
	linear_key_head_dim=32,
	linear_num_key_heads=4,
	linear_num_value_heads=8,
	linear_value_head_dim=32,
	)
	text_config["layer_types"] = ["linear_attention", "linear_attention", "linear_attention", "full_attention"] * 3
	text_config["rope_parameters"]["mrope_section"] = [1, 1, 2]
	# Two-stream mHC spans exactly the looped block [4, 8), matching the source [4, 28) cadence.
	config["loop_settings"] = {"counts": [2], "end": 8, "eval_count": 2, "start": 4}
	config["mhc"] = {
	"end": 8,
	"fixed_identity": False,
	"iterations": 20,
	"pre_scale": 1.0,
	"start": 4,
	"streams": 2,
	}
	config["hidden_size"] = 8
	# Four vision blocks keep the native Qwen3.5 tower and the fixed patch/temporal/merge and position contracts.
	config["vision_config"].update(
	hidden_size=64,
	intermediate_size=128,
	num_heads=2,
	out_hidden_size=8,
	depth=4,
	)
	(folder / "config.json").write_text(json.dumps(config, indent=2) + "\n")

	set_seed(42)
	model = AutoModelForImageTextToText.from_config(
	AutoConfig.from_pretrained(folder, trust_remote_code=True),
	trust_remote_code=True,
	dtype=torch.bfloat16,
	)
	# The source keeps the hybrid linear-attention SSM parameters FP32 even under a BF16 load.
	from transformers.models.qwen3_5.modeling_qwen3_5 import Qwen3_5GatedDeltaNet

	for layer in model.model.language_model.layers:
	linear = getattr(layer, "linear_attn", None)
	if isinstance(linear, Qwen3_5GatedDeltaNet):
	with torch.no_grad():
	linear.A_log = torch.nn.Parameter(linear.A_log.float())
	linear.dt_bias = torch.nn.Parameter(linear.dt_bias.float())
	linear.norm.weight = torch.nn.Parameter(linear.norm.weight.float())
	linear.conv1d.weight = torch.nn.Parameter(linear.conv1d.weight.float())
	# One BF16 MTP draft layer reusing a full-attention backbone layer outside the mHC range.
	full_layer = model.model.language_model.layers[11]
	model.mtp = torch.nn.ModuleDict(
	{
	"pre_fc_norm_embedding": torch.nn.RMSNorm(8),
	"fc": torch.nn.Linear(16, 8, bias=False),
	"layers": torch.nn.ModuleList([copy.deepcopy(full_layer)]),
	"norm": torch.nn.RMSNorm(8),
	"pre_fc_norm_hidden": torch.nn.RMSNorm(8),
	}
	)
	model = model.cpu()
	representation = str(model)

	state = model.state_dict()
	backbone, draft = {}, {}
	for name, tensor in state.items():
	if name.startswith("mtp."):
	draft[name] = tensor.to(torch.bfloat16).contiguous()
	else:
	# The shipped checkpoint stores the complete backbone FP32, mHC routers included.
	backbone[name] = tensor.to(torch.float32).contiguous()
	save_file(backbone, str(folder / "model.safetensors"), metadata={"format": "pt"})
	save_file(draft, str(folder / "mtp.safetensors"), metadata={"format": "pt"})
	# Rewrite the edited config (save was manual) and pin the source generation config bytes.
	(folder / "config.json").write_text(json.dumps(config, indent=2) + "\n")
	(folder / "generation_config.json").write_bytes(
	Path(hf_hub_download(source_model_id, "generation_config.json", revision=source_revision)).read_bytes()
	)

	expected_draft = {
	"mtp.fc.weight",
	"mtp.pre_fc_norm_embedding.weight",
	"mtp.pre_fc_norm_hidden.weight",
	"mtp.norm.weight",
	"mtp.layers.0.input_layernorm.weight",
	"mtp.layers.0.post_attention_layernorm.weight",
	"mtp.layers.0.self_attn.q_proj.weight",
	"mtp.layers.0.self_attn.k_proj.weight",
	"mtp.layers.0.self_attn.v_proj.weight",
	"mtp.layers.0.self_attn.o_proj.weight",
	"mtp.layers.0.self_attn.q_norm.weight",
	"mtp.layers.0.self_attn.k_norm.weight",
	"mtp.layers.0.mlp.gate_proj.weight",
	"mtp.layers.0.mlp.up_proj.weight",
	"mtp.layers.0.mlp.down_proj.weight",
	}
	assert set(draft) == expected_draft, sorted(set(draft) ^ expected_draft)
	assert all(
	name.count("mhc_attn") + name.count("mhc_ffn") == 0 or backbone[name].dtype == torch.float32
	for name in backbone
	)
	return representation


	create_model(
	source_model_id="sensenova/SenseNova-Loop-9B-Preview",
	source_revision="9937e91c7301b7dcbb5bf17e3e4e95c499a11e99",
	save_folder="/tmp/sensenova-loop",
	)