Skip to content

[Feature Request] Support for INT4 (q4) Quantization in Rust/Candle Backend #400

Description

@BenjaminMarsili

Due diligence

  • I have done my due diligence in trying to find the answer myself.

Topic

The Rust implementation

Question

Could the team prioritize adding INT4 (Q4_0/Q4_K) quantization support specifically for the moshi-backend Rust implementation, alongside optimized ARM NEON/SIMD kernels for the Mimi codec and Temporal Transformer layers?

I am developing an offline, full-duplex assistant running on a Raspberry Pi 5 paired with the Hailo-10H NPU (40 TOPS).

To achieve a great user experience, the system must maintain Moshi's native 12.5Hz (80ms per frame) cadence without thermal throttling or audio stutter. While the Hailo-10H features 8GB of dedicated memory, the 34 GB/s LPDDR4 bandwidth creates a hard physical bottleneck for a 7.6B model in INT8 (8GB * 12.5Hz$ far exceeds the bus capacity).

To maintain real-time streaming, we must reduce the weight-fetch payload to a 4-bit profile.

Excited to hear from you.

Metadata

Metadata

Assignees

No one assigned

    Labels

    questionFurther information is requested

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions