LFM2.5-VL-1.6B-ONNX

<div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/2b08LKpev0DNEk6DlnWkY.png" alt="Liquid AI" style="width: 100%; max-width: 100%; height: auto; display: inline-block; margin-bottom: 0.5em; margin-top: 0.5em;" /> <div style="display: flex; justify-content: center; gap: 0.5em; margin-bottom: 1em;"> <a href="https://playground.liquid.ai/"><strong>Try LFM</strong></a> • <a href="https://docs.liquid.ai/lfm"><strong>Documentation</strong></a> • <a href="https://leap.liquid.ai/"><strong>LEAP</strong></a> </div> </div>

LFM2.5-VL-1.6B-ONNX

ONNX export of LFM2.5-VL-1.6B for cross-platform inference.

Recommended Variants

• WebGPU: Use FP16 encoder + Q4 decoder (Q8 not supported)
• Server: FP16+Q4 for efficiency, FP16+FP16 for quality

Model Files

onnx/
├── embed_tokens.onnx           # Token embeddings (FP32)
├── embed_tokens.onnx_data
├── embed_tokens_fp16.onnx      # Token embeddings (FP16)
├── embed_tokens_fp16.onnx_data
├── embed_images.onnx           # Vision encoder (FP32)
├── embed_images.onnx_data*
├── embed_images_fp16.onnx      # Vision encoder (FP16)
├── embed_images_fp16.onnx_data*
├── embed_images_q4.onnx        # Vision encoder (Q4)
├── embed_images_q4.onnx_data
├── embed_images_q8.onnx        # Vision encoder (Q8)
├── embed_images_q8.onnx_data
├── decoder.onnx                # Language decoder (FP32)
├── decoder.onnx_data*
├── decoder_fp16.onnx           # Language decoder (FP16)
├── decoder_fp16.onnx_data*
├── decoder_q4.onnx             # Language decoder (Q4)
├── decoder_q4.onnx_data
├── decoder_q8.onnx             # Language decoder (Q8)
└── decoder_q8.onnx_data

* Large models (>2GB) split weights across multiple files:
  decoder.onnx_data, decoder.onnx_data_1, decoder.onnx_data_2, etc.
  All data files must be in the same directory as the .onnx file.

Python

Installation

pip install onnxruntime transformers pillow torch huggingface_hub
# or with GPU support:
pip install onnxruntime-gpu transformers pillow torch huggingface_hub

Inference

import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download
from transformers import AutoProcessor
from PIL import Image

# Download model files (fp16 encoder + q4 decoder recommended)
model_id = "LiquidAI/LFM2.5-VL-1.6B-ONNX"
embed_tokens_path = hf_hub_download(model_id, "onnx/embed_tokens_fp16.onnx")
embed_images_path = hf_hub_download(model_id, "onnx/embed_images_fp16.onnx")
decoder_path = hf_hub_download(model_id, "onnx/decoder_q4.onnx")

# Download all data files (handles multiple splits for large models)
from huggingface_hub import list_repo_files
for f in list_repo_files(model_id):
    if any(f.startswith(f"onnx/{name}") for name in [
        "embed_tokens_fp16.onnx_data",
        "embed_images_fp16.onnx_data",
        "decoder_q4.onnx_data"
    ]):
        hf_hub_download(model_id, f)

# Load ONNX sessions
embed_tokens = ort.InferenceSession(embed_tokens_path)
embed_images = ort.InferenceSession(embed_images_path)
decoder = ort.InferenceSession(decoder_path)

# Load processor
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Prepare input
image = Image.open("photo.jpg")
messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "What is in this image?"}
]}]

# Process inputs
prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(images=[image], text=prompt, return_tensors="pt")

# Convert to numpy with correct dtypes
pixel_values = inputs["pixel_values"].numpy().astype(np.float32)
pixel_attention_mask = inputs["pixel_attention_mask"].numpy().astype(np.int64)
spatial_shapes = inputs["spatial_shapes"].numpy().astype(np.int64)
input_ids = inputs["input_ids"].numpy().astype(np.int64)

# Get image embeddings
image_outputs = embed_images.run(None, {
    "pixel_values": pixel_values,
    "pixel_attention_mask": pixel_attention_mask,
    "spatial_shapes": spatial_shapes,
})
image_embeds = image_outputs[0]

# Get token embeddings
token_outputs = embed_tokens.run(None, {"input_ids": input_ids})
token_embeds = token_outputs[0]

# Replace <image> tokens with image embeddings
image_token_id = processor.tokenizer.convert_tokens_to_ids("<image>")
image_positions = np.where(input_ids[0] == image_token_id)[0]
for i, pos in enumerate(image_positions):
    if i < len(image_embeds):
        token_embeds[0, pos] = image_embeds[i]

# Initialize KV cache for stateful decoding
ONNX_DTYPE = {"tensor(float)": np.float32, "tensor(float16)": np.float16, "tensor(int64)": np.int64}
cache = {}
for inp in decoder.get_inputs():
    if inp.name in {"inputs_embeds", "attention_mask", "position_ids"}:
        continue
    shape = [d if isinstance(d, int) else 1 for d in inp.shape]
    for i, d in enumerate(inp.shape):
        if isinstance(d, str) and "sequence" in d.lower():
            shape[i] = 0
    cache[inp.name] = np.zeros(shape, dtype=ONNX_DTYPE.get(inp.type, np.float32))

# Generate tokens
seq_len = token_embeds.shape[1]
generated_tokens = []

for step in range(100):  # max tokens
    if step == 0:
        embeds = token_embeds.astype(np.float32)
    else:
        last_token = np.array([[generated_tokens[-1]]], dtype=np.int64)
        embeds = embed_tokens.run(None, {"input_ids": last_token})[0].astype(np.float32)

    attn_mask = np.ones((1, seq_len + len(generated_tokens)), dtype=np.int64)
    feed = {"inputs_embeds": embeds, "attention_mask": attn_mask, **cache}

    outputs = decoder.run(None, feed)
    next_token = int(np.argmax(outputs[0][0, -1]))
    generated_tokens.append(next_token)

    # Update cache
    for i, out in enumerate(decoder.get_outputs()[1:], 1):
        name = out.name.replace("present_conv", "past_conv").replace("present.", "past_key_values.")
        if name in cache:
            cache[name] = outputs[i]

    if next_token == processor.tokenizer.eos_token_id:
        break

print(processor.tokenizer.decode(generated_tokens, skip_special_tokens=True))

WebGPU (Browser)

Installation

npm install onnxruntime-web @huggingface/transformers

Enable WebGPU

WebGPU is required for browser inference. To enable:

1. Chrome/Edge: Navigate to chrome://flags/#enable-unsafe-webgpu, enable, and restart
2. Verify: Check chrome://gpu for "WebGPU" status
3. Test: Run navigator.gpu.requestAdapter() in DevTools console

Inference

import * as ort from "onnxruntime-web/webgpu";
import { AutoTokenizer } from "@huggingface/transformers";

// Check WebGPU availability
if (!navigator.gpu) {
  throw new Error("WebGPU not available. Enable at chrome://flags/#enable-unsafe-webgpu");
}
const adapter = await navigator.gpu.requestAdapter();
if (!adapter) {
  throw new Error("WebGPU adapter not found. Check chrome://gpu for status.");
}

ort.env.wasm.numThreads = 1;

const modelId = "LiquidAI/LFM2.5-VL-1.6B-ONNX";
const modelBase = `https://huggingface.co/${modelId}/resolve/main`;

// Load tokenizer
const tokenizer = await AutoTokenizer.from_pretrained(modelId);

// Load ONNX sessions with external data
// For models with multiple data files (>2GB), add additional entries to externalData array
async function loadSession(name, dataFiles = 1) {
  const onnxPath = `${modelBase}/onnx/${name}.onnx`;
  const externalData = [];
  for (let i = 0; i < dataFiles; i++) {
    const suffix = i === 0 ? "" : `_${i}`;
    const fileName = `${name}.onnx_data${suffix}`;
    externalData.push({ path: fileName, data: `${modelBase}/onnx/${fileName}` });
  }
  return ort.InferenceSession.create(onnxPath, {
    executionProviders: ["webgpu"],
    externalData,
  });
}

const embedTokens = await loadSession("embed_tokens_fp16");
const embedImages = await loadSession("embed_images_fp16");
const decoder = await loadSession("decoder_q4");

// Model config (from config.json)
const hiddenSize = 1536;
const numKVHeads = 12;
const headDim = 128;

// Get text embeddings helper
async function getTextEmbeddings(ids) {
  const tensor = new ort.Tensor("int64", new BigInt64Array(ids.map(BigInt)), [1, ids.length]);
  const out = await embedTokens.run({ input_ids: tensor });
  return out.inputs_embeds;
}

// Initialize KV cache
function initCache() {
  const cache = {};
  for (const name of decoder.inputNames) {
    if (name.startsWith("past_conv")) {
      cache[name] = new ort.Tensor("float32", new Float32Array(hiddenSize * 3), [1, hiddenSize, 3]);
    } else if (name.startsWith("past_key_values")) {
      cache[name] = new ort.Tensor("float32", new Float32Array(0), [1, numKVHeads, 0, headDim]);
    }
  }
  return cache;
}

// Update cache from outputs
function updateCache(cache, outputs) {
  for (const [name, tensor] of Object.entries(outputs)) {
    if (name.startsWith("present_conv")) {
      cache[name.replace("present_conv", "past_conv")] = tensor;
    } else if (name.startsWith("present.")) {
      cache[name.replace("present.", "past_key_values.")] = tensor;
    }
  }
}

// Build prompt and tokenize
const prompt = tokenizer.apply_chat_template(messages, { add_generation_prompt: true, tokenize: false });
const inputIds = tokenizer.encode(prompt);

// Get embeddings (for VL: merge image embeddings at <image> token positions)
let inputsEmbeds = await getTextEmbeddings(inputIds);

// Generation loop
const cache = initCache();
const eosTokenId = tokenizer.eos_token_id;
const generatedTokens = [];
let curLen = inputsEmbeds.dims[1];
let embeds = inputsEmbeds;

for (let step = 0; step < 256; step++) {
  const attentionMask = new ort.Tensor("int64", new BigInt64Array(curLen).fill(1n), [1, curLen]);

  const outputs = await decoder.run({ inputs_embeds: embeds, attention_mask: attentionMask, ...cache });

  // Greedy decode: argmax of last token logits
  const logits = outputs.logits;
  const vocabSize = logits.dims[2];
  const lastLogits = logits.data.slice((logits.dims[1] - 1) * vocabSize);
  const nextToken = lastLogits.indexOf(Math.max(...lastLogits));

  generatedTokens.push(nextToken);
  if (nextToken === eosTokenId) break;

  updateCache(cache, outputs);
  embeds = await getTextEmbeddings([nextToken]);
  curLen++;
}

console.log(tokenizer.decode(generatedTokens, { skip_special_tokens: true }));

WebGPU Notes

• Recommended: embed_images_fp16.onnx + decoder_q4.onnx
• For higher quality: embed_images_fp16.onnx + decoder_fp16.onnx
• Image preprocessing requires tiling (512×512), patch extraction (16×16), and normalization
• Large models (>2GB) split weights across multiple files (e.g., .onnx_data, .onnx_data_1). Use loadSession(name, dataFiles) with the number of data files
• int64 tensors require BigInt64Array

License

This model is released under the LFM 1.0 License.