zai-org

/

GLM-4.7-Flash

like 1.44k

Follow

Z.ai 8.63k

Description

For example, let's take a conversation:

> User: Hello! Write me simple shell script to get SHA256 checksum of files in folder.
> Model: <think>step_1; step_2; step_3;</think> response

It works great and has no issues. But then:

> User: Translate it to Russian
> Model: <think> step_4;step_5;step_6;</think> PERFECT response IN Russian without errors

Still works great. Try other option (request in Greek to translate to Russian):

> User: Μεταφράστε το στα Ρωσική
> Model: <think> step_4; step_4; step_5; step_4; step_6;</think> Response in step_6; Russ##&*an with step_7;step_4; errors

Now it broken completely. Try other option:

> User: Переведи на русский
> Model: <think> step_4; step_4; step_5; step_4; step_6;</think> Response in step_6; Russ##&*an with step_7;step_4; errors

Observations

• There are no such issues with GGUF model using llama.cpp or ik_llama.cpp, no matter which quant to use. Only VLLM and SGLang are affected;
• Same issue happening with Glm 4.5 Air AWQ, but I though it was because of lobotomized AWQ;
• Qwen models (both instruct and thinking, 30B and 80B, dense and MoE, BF16, FP8 and AWQ) are not affected at all;
• Noticed [gMASK] in jinja, tried to ask model itself "What is [gMASK] %template content%" and got broken output immediately;
• Changing sampling parameters (or enabling speculative decoding, which disables min_p) may help. For example, it gives adequate responses (SOME TIMES) with top_p: 0.9;
• Inference from Zai.org in model card works flawless, no issues.

May it be caused by specific UTF-8 issues with jinja? Or my configuration with 4 GPUs where 3 3090 and 1 4090? Or VLLM has issues with default sampling params?