About this checkpoint
This is a PTPC (Per-Token activation + Per-Channel weight) FP8 variant of GLM-5.2,
requantized from the block-scaled baseline zai-org/GLM-5.2-FP8
for AMD Instinct MI300X / MI308X / MI325X (gfx942) inference with SGLang + aiter rowwise FP8 GEMM.
-
Quantization: w8a8_fp8 — per-channel static weight scales (FP8 E4M3) + dynamic per-token activation scales
-
Derived from: block-scaled zai-org/GLM-5.2-FP8 (weight_block_size=[128,128]) via offline
per-channel requantization (weight_block_size removed). Every block-quantized Linear was
dequantized to float and re-quantized with one FP8 scale per output channel; the activation
path is quantized dynamically per token at runtime by aiter.
-
Quantized modules: all MoE routed/shared experts, dense MLP, attention projections
(q_a_proj, q_b_proj, kv_a_proj_with_mqa, kv_b_proj, o_proj) and DSA indexer
projections (indexer.wk, indexer.wq_b). Kept in BF16: layernorms, router gate (+bias),
indexer.weights_proj, indexer.k_norm, kv_a_layernorm, q_a_layernorm, the MTP/next-token
layer norms (eh_proj, enorm, hnorm, shared_head.norm), lm_head, model.embed_tokens.
-
Validated (GSM8K, 5-shot, lm-eval-harness, exact_match):
Table with columns: Checkpoint, flexible-extract, strict-match| Checkpoint | flexible-extract | strict-match |
|---|
zai-org/GLM-5.2-FP8 (block-scale source) | 0.9507 ± 0.006 | 0.9507 ± 0.006 |
| GLM-5.2-FP8-PTPC (this model) | 0.9462 ± 0.0062 | 0.9462 ± 0.0062 |
The −0.45 pp delta is within the confidence interval → no meaningful regression.
Measured on 8× AMD Instinct MI308X (gfx942) with the SGLang PTPC stack.
Serving with SGLang
PTPC routes the MoE/dense FP8 GEMMs through aiter's tuned rowwise (per-token × per-channel)
FP8 kernels. Use SGLang with --quantization w8a8_fp8 and SGLANG_USE_AITER_FP8_PER_TOKEN=1:
export SGLANG_USE_AITER=1
export SGLANG_USE_AITER_FP8_PER_TOKEN=1
export SGLANG_ADAPTIVE_FP8_DISPATCH=1
python3 -m sglang.launch_server \
--model-path ginsongsong/GLM-5.2-FP8-PTPC --tp-size 8 \
--quantization w8a8_fp8 --attention-backend aiter \
--nsa-prefill-backend tilelang --nsa-decode-backend tilelang \
--mem-fraction-static 0.85 --served-model-name glm-5.2-fp8 \
--trust-remote-code --disable-custom-all-reduce --disable-radix-cache \
--port 30000
On gfx942 the 8 GPUs must be in SPX compute-partition mode for TP=8
(rocm-smi --showcomputepartition). CPX mode wedges RCCL init at TP=8.
GLM-5.2
👋 Join the GLM Discord community.
📖 See the GLM-5.2 blog and GLM-5 Technical report.
📍 Use GLM-5.2 API services on the Z.ai API Platform.
[Paper] [GitHub]
Introduction
GLM-5.2 is Z.ai's latest flagship model for long-horizon tasks, delivering a substantial leap in
long-horizon capability over GLM-5.1 on a solid 1M-token context. Highlights:
- Solid 1M context that stably sustains long-horizon work.
- Advanced coding with flexible effort — multiple thinking-effort levels to balance
performance and latency.
- Improved architecture — IndexShare reuses the same indexer across every four sparse
attention layers (−2.9× per-token FLOPs at 1M context), and an improved MTP layer raises
speculative-decoding acceptance length by up to 20%.
- Pure open — MIT license.
GlmMoeDsaForCausalLM: 78 layers, 256 routed experts (+1 shared), MLA + DSA sparse attention
with IndexShare, 1 MTP (next-token-prediction) layer.
Benchmark (upstream BF16, for reference)
Table with columns: Benchmark, GLM-5.2, GLM-5.1| Benchmark | GLM-5.2 | GLM-5.1 |
|---|
| HLE | 40.5 | 31.0 |
| AIME 2026 | 99.2 | 95.3 |
| GPQA-Diamond | 91.2 | 86.2 |
| SWE-bench Pro | 62.1 | 58.4 |
| Terminal-Bench 2.1 (Terminus-2) | 81.0 | 63.5 |
| MCP-Atlas (Public) | 76.8 |
See the original model card for the full benchmark suite
and evaluation settings.
Citation
@misc{glm5team2026glm5vibecodingagentic,
title={GLM-5: from Vibe Coding to Agentic Engineering},
author={GLM-5-Team},
year={2026},
eprint={2602.15763},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2602.15763},
}
Acknowledgement
PTPC requantization and AMD ROCm (SGLang + aiter) validation methodology follow the AMD ROCm blog
"Further Accelerating Kimi-K2.5 on AMD Instinct MI325X: W4A8 & W8A8 Quantization with AMD Quark"
and the GLM-5.1-FP8-PTPC recipe. Original model © Z.ai, released under the MIT license.