---
license: other
license_name: glm-5.3
license_link: LICENSE
language:
  - en
  - zh
library_name: transformers
tags:
  - jang
  - fp8
  - vllm
  - glm
  - glm-moe-dsa
  - moe
  - hopper
  - h200
  - mtp
pipeline_tag: text-generation
base_model: zai-org/GLM-5.3
base_model_relation: quantized
---

<p align="center"><img src="jangq-logo.png" width="160"/></p>

# GLM-5.3-FP8

**GLM-5.3 — ~756 GB on disk** (down from the ~1.5 TB BF16 source) — **FP8 E4M3
block-128** quantization, weight-only with dynamic activations. Runs **natively
on Hopper (H100 / H200)** in vLLM via the DeepSeek-V3-style block-FP8 kernels —
no Blackwell required.

- **Source:** [zai-org/GLM-5.3](https://huggingface.co/zai-org/GLM-5.3-BF16)
  (GLM-MoE-DSA architecture, 78 layers, 256 routed experts top-8, 1 shared
  expert, MLA + DeepSeek Sparse Attention, MTP head, 1M context)
- **Quantization:** FP8 E4M3, weight-only **static** with per-**128×128-block**
  scales (`weight_scale_inv`, fp32); **dynamic** activation quantization at
  runtime — bit-for-bit the same scheme as the official `zai-org/GLM-5.3` FP8
  release and the DeepSeek-V3 block-FP8 format
- **Coverage:** the exact quantized-tensor set of the official FP8 release
  (**59,044** tensors). Verified 1:1 tensor-name match against both the BF16 and
  official-FP8 indexes
- **MTP included:** the multi-token-prediction layer (`model.layers.78.*`,
  incl. `eh_proj` / `enorm` / `hnorm` / `shared_head`) is present and quantized
  identically to the official release
- **Bundle size:** **~756 GB** across 282 shards
- **Runs on:** 8× H200 (or H100) with vLLM tensor-parallel

## What's in the bundle

| Module | Source dtype | Bundle dtype |
|---|---|---|
| Routed experts (256 × 3 mats × sparse layers) | BF16 | **FP8 E4M3** + F32 `weight_scale_inv` (block 128×128) |
| Attention (MLA q/kv proj, o_proj) | BF16 | **FP8 E4M3** + F32 block scales |
| Dense MLP (first 3 layers) | BF16 | **FP8 E4M3** + F32 block scales |
| MTP layer 78 (linear weights) | BF16 | **FP8 E4M3** + F32 block scales |
| RMSNorms, router gates, `e_score_correction_bias` | BF16 / F32 | **passthrough** (BF16 / F32) |
| DSA indexer norms / proj, shared-expert gate | BF16 | **passthrough** |
| `embed_tokens`, `lm_head` | BF16 | **passthrough** |

Dequant per block: `w = weight_fp8 * weight_scale_inv`, `scale = amax / 448.0`.

## Usage (vLLM)

```bash
vllm serve JANGQ-AI/GLM-5.3-FP8 \
  --tensor-parallel-size 8 \
  --trust-remote-code
```

Enable MTP speculative decoding per vLLM's GLM / DeepSeek MTP docs for a decode
speedup (the MTP weights are shipped here).

## Method

Converted from the BF16 release with a streaming shard-by-shard quantizer (one
tensor in memory at a time), so no GPU and no large host RAM are needed. The set
of tensors to quantize is taken directly from the official FP8 index rather than
re-derived heuristically, and the output `config.json` (`quantization_config`
with `modules_to_not_convert`) is a verbatim copy of the official FP8 config — so
vLLM treats this identically to the upstream release. The scheme is data-free
(dynamic activation scales), so no calibration set is involved.

---
*Quantized by [JANGQ-AI](https://huggingface.co/JANGQ-AI).*