how to quant fp16 to fp8?

[dongyibo]

I use llm-compressors but found quantization_config in config.json generated just like this:

"quantization_config": {
...
"format": "float-quantized",
"global_compression_ratio": null,
"ignore": [
"lm_head"
],
"quant_method": "compressed-tensors",
"quantization_status": "compressed",
},

I notice that quant_method is compressed-tensors,it's performance is poor than quant_method of fp8 by many tests in sglang.

at the sametime, I found quant_method in deepseekv3's config.json is fp8. So, how should we quantify it to obtain a quantized model with quant_method as fp8?