phi-3.5-moe-cpu-offload-gguf

Phi-3.5-MoE Instruct - F16 GGUF with MoE CPU Offloading Support

F16 GGUF conversion of microsoft/Phi-3.5-MoE-instruct with Rust bindings for llama.cpp's MoE CPU offloading functionality.

Model Details

• Base Model: microsoft/Phi-3.5-MoE-instruct
• Format: GGUF F16 precision
• File Size: 79GB
• Parameters: 41.9B total (6.6B active per token)
• Architecture: 32 layers, 16 experts per layer, 2 active experts per token
• Context Length: 131K tokens
• Converted by: MikeKuykendall

MoE CPU Offloading

This model supports MoE CPU offloading via llama.cpp (implemented in PR #15077). Shimmy provides Rust bindings for this functionality, enabling:

• VRAM Reduction: 96.5% (77.7GB → 2.8GB measured on GH200)
• Performance Trade-off: 3.1x slower generation (13.8 → 4.5 TPS)
• Use Case: Running 42B parameter MoE on consumer GPUs (<10GB VRAM)

Controlled Baseline (NVIDIA GH200, N=3)

Trade-off: Memory for speed. Best for VRAM-constrained scenarios where generation speed is less critical than model size.

Download

huggingface-cli download MikeKuykendall/phi-3.5-moe-cpu-offload-gguf \
  --include "phi-3.5-moe-f16.gguf" \
  --local-dir ./models

Usage

llama.cpp (CPU Offloading)

# Standard loading (requires ~80GB VRAM)
./llama-server -m phi-3.5-moe-f16.gguf -c 4096

# With MoE CPU offloading (requires ~3GB VRAM + 80GB RAM)
./llama-server -m phi-3.5-moe-f16.gguf -c 4096 --cpu-moe

Shimmy (Rust Bindings)

# Install Shimmy
cargo install --git https://github.com/Michael-A-Kuykendall/shimmy --features llama-cuda

# Standard loading
shimmy serve --model phi-3.5-moe-f16.gguf

# With MoE CPU offloading
shimmy serve --model phi-3.5-moe-f16.gguf --cpu-moe

# Query the API
curl http://localhost:11435/api/generate \
  -d '{
    "model": "phi-3.5-moe",
    "prompt": "Explain mixture of experts in simple terms",
    "max_tokens": 256,
    "stream": false
  }'

Prompt Format

<|system|>
You are a helpful assistant.<|end|>
<|user|>
Your question here<|end|>
<|assistant|>

Performance Notes

Standard GPU Loading:

• VRAM: 77.7GB
• Speed: 13.8 TPS
• Latency: 730ms TTFT
• Use when: VRAM is plentiful, speed is critical

CPU Offloading:

• VRAM: 2.8GB (96.5% reduction)
• Speed: 4.5 TPS (3.1x slower)
• Latency: 2,251ms TTFT
• Use when: Limited VRAM, speed less critical

Original Model

• Developers: Microsoft
• License: MIT
• Paper: Phi-3 Technical Report
• Blog: Phi-3.5-MoE Announcement

Technical Validation

Full validation report with controlled baselines: Shimmy MoE CPU Offloading Technical Report

Citation

@techreport{abdin2024phi,
  title={Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone},
  author={Abdin, Marah and others},
  year={2024},
  institution={Microsoft}
}

---

GGUF conversion and MoE offloading validation by MikeKuykendall