0xSero / GLM-4.7-185B-W4A16

Support this work → · X · GitHub · REAP paper · Cerebras REAP

GLM-4.7-185B-W4A16

W4A16 quantization of 0xSero/GLM-4.7-185B .

At a glance

Which variant should I pick?

𓌳 REAP 𓌳 the Experts: Why Pruning Prevails for One-Shot MoE Compression
📄 Paper • 💻 Code • 📝 Blog

GLM-4.7-REAP-50-W4A16

✨ Highlights

50% Expert-Pruned + INT4 Quantized — Double compression for efficient deployment.

• ~6.5x Total Compression : 700GB → ~92GB
• REAP + AutoRound : Expert pruning + weight quantization
• Optimized for Code & Tools : Calibrated on code generation and function calling
• Lower VRAM : Fits on 2-4x fewer GPUs than BF16

📋 Model Specifications

Compression Pipeline

GLM-4.7 (358B, 700GB)
        │
        ▼  REAP 50% expert pruning
        │
GLM-4.7-REAP-50 (179B)
        │
        ▼  AutoRound W4A16 quantization
        │
GLM-4.7-REAP-50-W4A16 (~92GB)  ◀── This model

Total: ~6.5x compression

🔬 Calibration Dataset: Deep Dive

REAP's effectiveness depends critically on calibration data that represents the target use case . We specifically optimized for code generation , function/tool calling , and agentic workflows .

Why These 3 Datasets?

The Science Behind Dataset Selection

REAP Algorithm:
1. Forward pass calibration samples through model
2. Record which experts activate and their magnitudes
3. Compute saliency = router_weight × activation_norm
4. Prune lowest-saliency experts

Key Insight: Experts are TASK-SPECIFIC
├── Some experts specialize in natural language
├── Some experts specialize in code syntax
├── Some experts specialize in JSON/structured output
└── Some experts specialize in multi-turn context

If calibration lacks code → code-specialized experts appear "unused" → get pruned → model loses coding ability

Cerebras' Original Mix (from paper)

Cerebras used the same 3 datasets in their GLM-4.6 REAP experiments:

• evol-codealpaca-v1 for code generation
• xlam-function-calling-60k for tool calling
• SWE-smith-trajectories for agentic tasks

We followed this exact recipe for reproducibility.

Combined Dataset

Our calibration mix: 0xSero/glm47-reap-calibration-v2

🚀 Deployment

vLLM (Recommended)

vllm serve 0xSero/GLM-4.7-185B-W4A16 \
    --tensor-parallel-size 4 \
    --trust-remote-code \
    --quantization gptq

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "0xSero/GLM-4.7-185B-W4A16",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("0xSero/GLM-4.7-185B-W4A16", trust_remote_code=True)

🧩 Reproduction

Step 1: REAP Pruning

#!/usr/bin/env python3
"""
REAP Pruning Script for MoE Models
Adapted from: https://github.com/CerebrasResearch/reap
"""

import subprocess
import sys

def run_reap(
    model_path: str,
    compression_ratio: float,
    dataset: str = "0xSero/glm47-reap-calibration-v2",
    samples: int = 1360,
    seed: int = 42,
    distance: str = "angular",
    reuse_observations: str = None,
):
    """
    Run REAP expert pruning.

    Args:
        model_path: Path to base model
        compression_ratio: 0.30 = prune 30%, keep 70%
        dataset: Calibration dataset (code + tools + agentic)
        samples: Number of calibration samples
        seed: Random seed for reproducibility
        distance: Distance metric for expert clustering
        reuse_observations: Path to pre-computed observations for instant pruning
    """
    cmd = [
        sys.executable, "src/reap/prune.py",
        "--model-name", model_path,
        "--dataset-name", dataset,
        "--compression-ratio", str(compression_ratio),
        "--prune-method", "reap",
        "--seed", str(seed),
        "--samples_per_category", str(samples),
        "--model_max_length", "2048",
        "--distance_measure", distance,
        "--record_pruning_metrics_only", "true",
    ]

    if reuse_observations:
        # Instant pruning: skip calibration, reuse precomputed expert scores
        cmd.extend(["--load_observations", reuse_observations])

    subprocess.run(cmd, check=True)

# Example: Create 40% pruned model
run_reap(
    model_path="/path/to/GLM-4.7",
    compression_ratio=0.40,  # Prune 40% of experts
)

Step 2: AutoRound Quantization

#!/usr/bin/env python3
"""
AutoRound W4A16 Quantization
Intel's state-of-the-art weight quantization using signed gradient descent.
"""

from auto_round import AutoRound

def quantize_w4a16(
    model_path: str,
    output_dir: str,
    bits: int = 4,
    group_size: int = 128,
    format: str = "auto_gptq",
):
    """
    Quantize model to INT4 weights with FP16 activations.

    Args:
        model_path: Path to REAP-pruned model
        output_dir: Output directory
        bits: Weight bit width (4 for W4A16)
        group_size: Quantization group size (128 is optimal)
        format: Output format (auto_gptq for vLLM compatibility)
    """
    ar = AutoRound(
        model_path,
        scheme="W4A16",
        device="cuda",
        device_map="auto",
        trust_remote_code=True,
        batch_size=1,
        seqlen=512,
        nsamples=64,
    )
    ar.quantize_and_save(output_dir, format=format)

# Example: Quantize REAP-40 to W4A16
quantize_w4a16(
    model_path="./GLM-4.7-REAP-40",
    output_dir="./GLM-4.7-REAP-40-W4A16",
)

⚖️ License

Apache 2.0

License & citation

License inherited from the base model.

@misc{lasby2025reap,
  title  = {REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
  author = {Mike Lasby and Ivan Lazarevich and Nish Sinnadurai and Sean Lie and Yani Ioannou and Vithursan Thangarasa},
  year   = {2025}, eprint = {2510.13999}, archivePrefix = {arXiv}
}

Sponsors

Made possible by NVIDIA · TNG Technology · Lambda · Prime Intellect · Hot Aisle .