Modelpublic

ukkathva/TelecomGPT-R1-27B-FP8-Dynamic

sourceHugging Faceapache-2.0updated 11d agoView on Hugging Face

0likes50downloads

Model Card

TelecomGPT-R1-27B-FP8-Dynamic

FP8 Dynamic quantization of KU-DFI/TelecomGPT-R1.

Quantization

—Scheme: FP8_DYNAMIC
—Serialization: compressed-tensors
—Target modules: Linear
—lm_head: unquantized
—Calibration dataset: none
—Source precision: BF16

Intended use

Telecom reasoning, alarm analysis, root-cause analysis, protocol reasoning, and evaluation against operator-specific incident datasets.

A100 note

NVIDIA A100 is an Ampere GPU and does not provide native Hopper-style FP8 Tensor Core execution. The FP8 checkpoint still reduces model-weight memory, and vLLM can use its supported Ampere execution path when loading the model.