--- library_name: onnxruntime license: other license_name: lfm1.0 license_link: LICENSE pipeline_tag: any-to-any tags: - liquid - lfm3 - onnx language: - en - ar - zh - fr - de - it - ja - ko - pt - es - vi - th - id - hi - ru - pl - da - el - fi - he - nl - sv - tr base_model: LiquidAI/LFM3-Omni-2B base_model_relation: quantized ---
Liquid AI
API Try Docs Blog
# LFM3-Omni-2B-ONNX ONNX version of [LFM3-Omni-2B](https://huggingface.co/LiquidAI/LFM3-Omni-2B) for cross-platform inference with ONNX Runtime. Find architecture and usage details in the [original model card](https://huggingface.co/LiquidAI/LFM3-Omni-2B). ## Files
Each precision pairs a decoder with a vision encoder, an audio encoder and a codec decoder. Every precision also runs the float32 [`depth_decoder`](https://huggingface.co/LiquidAI/LFM3-Omni-2B-ONNX/blob/main/onnx/depth_decoder.onnx), which the sizes include.
Precision Decoder Vision encoder Audio encoder Codec decoder Size Device
q4 decoder_q4 vision_encoder_q8 audio_encoder codec_decoder 2.91 GB CPU
q8 decoder_q8 vision_encoder_fp16 audio_encoder codec_decoder 3.73 GB CPU
q4f16 decoder_q4f16 vision_encoder_fp16 audio_encoder_fp16 codec_decoder_fp16 2.47 GB GPU
q8f16 decoder_q8f16 vision_encoder_fp16 audio_encoder_fp16 codec_decoder_fp16 3.12 GB GPU
fp16 decoder_fp16 vision_encoder_fp16 audio_encoder_fp16 codec_decoder_fp16 4.68 GB GPU
fp32 decoder vision_encoder audio_encoder codec_decoder 9.03 GB Reference
The fp16, q4f16 and q8f16 decoders run with fp16 activations, the others with fp32. q4 and q4f16 keep the embedding, the LM head and 27 of the 153 layer projections in int8 and the rest in int4. Every decoder keeps the speech-frame codebook table in float. ## Use with ONNX Runtime
The vision encoder maps each image, or each temporal patch of a video, to projected features, and the audio encoder maps the log-mel features of 16 kHz audio to audio features. The decoder takes them at the `<|image|>`, `<|video|>` and `<|audio|>` placeholders of `input_ids` (`image_features`, `video_features`, `audio_features`), and the codes of generated speech frames at the `<|audio_out|>` placeholders (`audio_out_codes`). Each step also takes `attention_mask`, `position_ids`, `logits_to_keep` and the convolution and attention caches (`past_conv.N`, `past_key_values.N.key` and `.value`). It returns `logits`, the final-norm `hidden_states` and the updated caches (`present_conv.N`, `present.N.key` and `.value`). The fp16, q4f16 and q8f16 decoders keep their caches in fp16. For speech, the depth decoder predicts the 12 codes of each speech frame from the decoder's hidden state in 3 MaskGIT passes, and the codec decoder maps the codes to STFT features that an inverse STFT turns into 24 kHz audio. For preprocessing and the inverse STFT, use the values in `processor_config.json` and `audio_tokenizer/config.json`. On GPU, use fp16, q4f16 or q8f16. The CUDA execution provider has no fp32 GroupQueryAttention kernel, so the other graphs run attention on the CPU. In the browser, use `import * as ort from 'onnxruntime-web/webgpu'`: the default `onnxruntime-web` import cannot run these graphs. On WebGPU, pair `decoder_q4f16` with the float32 side graphs of the q4 bundle (`vision_encoder_q8`, `audio_encoder`, `codec_decoder`), and prefill in chunks of at most 1,024 tokens. Recommended sampling: greedy decoding for text, and `audio_temperature=0.4, audio_repetition_penalty=1.2, audio_repetition_window=64` for speech or interleaved output. More options in the [ONNX guide](https://docs.liquid.ai/deployment/on-device/onnx).