jamesdborin/glm-5.2-redhat-dspark-sglang

Red Hat GLM-5.2 DSpark for SGLang

SGLang-compatible configuration of the Red Hat GLM-5.2 DSpark reference checkpoint.

This repository is an inference-only SGLang packaging of RedHatAI/GLM-5.2-speculator.dspark for nvidia/GLM-5.2-NVFP4. The weights are not converted. The Speculators configuration was flattened and annotated for SGLang's Qwen3DSparkModel loader; the original is retained as speculators_config.json.

Weights are unchanged from the Apache-2.0 licensed Red Hat source model; only the loader configuration is adapted.

SGLang usage

python -m sglang.launch_server \
  --model-path nvidia/GLM-5.2-NVFP4 \
  --tp 4 \
  --quantization modelopt_fp4 \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path jamesdborin/glm-5.2-redhat-dspark-sglang \
  --speculative-draft-model-quantization unquant \
  --speculative-dspark-block-size 8 \
  --speculative-num-draft-tokens 9 \
  --trust-remote-code

Validated with SGLang v0.5.17 on NVIDIA B200 GPUs.

License and attribution

Released under Apache-2.0. The reference speculator originates from RedHatAI/GLM-5.2-speculator.dspark, which is Apache-2.0 licensed. The target model is MIT licensed.