Skip to content

Latest commit

 

History

History
181 lines (140 loc) · 5.27 KB

GLM-5-Tutorial.md

File metadata and controls

181 lines (140 loc) · 5.27 KB

Running GLM-5 with SGLang and KT-Kernel

This tutorial demonstrates how to run GLM-5 model inference using SGLang integrated with KT-Kernel for CPU-GPU heterogeneous inference. This setup enables efficient deployment of large MoE models by offloading experts to CPU. KT-Kernel supports both BF16 and FP8 precision backends, allowing you to choose between maximum quality and reduced memory footprint.

Table of Contents

Prerequisites

Before starting, ensure you have:

  1. SGLang installed

    Install the kvcache-ai fork of SGLang (one of):

    # Option A: One-click install (from ktransformers root)
./install.sh

# Option B: pip install
pip install sglang-kt

  2. KT-Kernel installed

    git clone https://github.com/kvcache-ai/ktransformers.git
git submodule update --init --recursive
cd kt-kernel && ./install.sh

  3. transformers reinstalled

    pip install git+https://github.com/huggingface/transformers.git

  4. CUDA toolkit - CUDA 12.0+ recommended (12.8+ for best FP8 support)

  5. Hugging Face CLI - For downloading models:

    pip install -U huggingface-hub

Step 1: Download Model Weights

Download the GLM-5 weights from Hugging Face.

# FP8
hf download zai-org/GLM-5-FP8 \
  --local-dir /path/to/GLM-5-FP8

# BF16
hf download zai-org/GLM-5 \
  --local-dir /path/to/GLM-5

Note: Replace /path/to/ with your actual storage path throughout this tutorial.

Step 2: Launch SGLang Server

Start the SGLang server with KT-Kernel integration for CPU-GPU heterogeneous inference.

# FP8 Precision
export PYTORCH_ALLOC_CONF=expandable_segments:True
export SGLANG_ENABLE_JIT_DEEPGEMM=0

python -m sglang.launch_server \
  --host 0.0.0.0 \
  --port 30000 \
  --model /path/to/GLM-5-FP8 \
  --kt-weight-path /path/to/GLM-5-FP8 \
  --kt-cpuinfer 96 \
  --kt-threadpool-count 2 \
  --kt-num-gpu-experts 30 \
  --kt-method FP8 \
  --kt-gpu-prefill-token-threshold 1024 \
  --kt-enable-dynamic-expert-update \
  --kt-expert-placement-strategy uniform \
  --trust-remote-code \
  --mem-fraction-static 0.75 \
  --served-model-name GLM5 \
  --enable-mixed-chunk \
  --tensor-parallel-size 8 \
  --enable-p2p-check \
  --disable-shared-experts-fusion \
  --chunked-prefill-size 16384 \
  --max-running-requests 4 \
  --max-total-tokens 128000 \
  --attention-backend flashinfer \
  --fp8-gemm-backend cutlass \
  --kv-cache-dtype bf16 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --watchdog-timeout 3000

# BF16 Precision
export PYTORCH_ALLOC_CONF=expandable_segments:True
export SGLANG_ENABLE_JIT_DEEPGEMM=0

python -m sglang.launch_server \
  --host 0.0.0.0 \
  --port 30000 \
  --model /path/to/GLM-5 \
  --kt-weight-path /path/to/GLM-5 \
  --kt-cpuinfer 96 \
  --kt-threadpool-count 2 \
  --kt-num-gpu-experts 10 \
  --kt-method BF16 \
  --kt-gpu-prefill-token-threshold 1024 \
  --kt-enable-dynamic-expert-update \
  --kt-expert-placement-strategy uniform \
  --trust-remote-code \
  --mem-fraction-static 0.75 \
  --served-model-name GLM5 \
  --enable-mixed-chunk \
  --tensor-parallel-size 8 \
  --enable-p2p-check \
  --disable-shared-experts-fusion \
  --chunked-prefill-size 16384 \
  --max-running-requests 4 \
  --max-total-tokens 128000 \
  --attention-backend flashinfer \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --watchdog-timeout 3000

Layerwise prefill requires one extra MoE layer's worth of VRAM.

If you encounter OOM, adjust --kt-num-gpu-experts, --chunked-prefill-size, --mem-fraction-static and --max-total-tokens when launching the server.

If you encounter other issues, try kt doctor to diagnose your setup.

See KT-Kernel Parameters for detailed parameter tuning guidelines.

Step 3: Send Inference Requests

Once the server is running (default: http://localhost:30000), you can interact with the model in several ways:

Option A: Interactive Chat with KT CLI

The easiest way to chat with the model:

kt chat

This opens an interactive terminal chat session. Type your messages and press Enter to send. Use Ctrl+C to exit.

Option B: OpenAI-Compatible API

The server exposes an OpenAI-compatible API at http://localhost:30000/v1.

curl example (streaming):

curl http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "GLM5",
    "messages": [{"role": "user", "content": "hi, who are you?"}],
    "stream": true
  }'

Additional Resources