Qwopus3.8-27B-Flash NVFP4 + MTP

Text-only NVIDIA NVFP4 quantization of Jackrong/Qwopus3.8-27B-Flash, packaged for single-GPU SGLang inference on a 32 GB RTX 5090. The model's native MTP layer is preserved for EAGLE/NEXTN speculative decoding.

Important: text only

This checkpoint does not support image or video input. The source model is multimodal, but this conversion intentionally removes its 333 vision tensors to reserve VRAM for the 27B language model, native MTP draft, and KV cache. SGLang supports Qwen vision models; this particular artifact does not.

Conversion

• Source revision: 44d24e8cb20ceb3cdf4fe200b5a0afd970ee748a
• Quantizer: NVIDIA ModelOpt 0.46
• Quantization: NVFP4 weights and activations, group size 16
• Calibration: CNN/DailyMail, 512 samples, sequence length 512, batch size 1
• KV metadata: FP8 E4M3
• Quantized modules: 400
• Native MTP tensors: 15, retained bit-for-bit from the source checkpoint
• Vision tensors removed: 333
• Safetensors: 2 shards, 19,671,389,312 tensor bytes

The export passed safetensors index/header validation, complete NVFP4 scale triplet checks, MTP equality checks, SGLang load validation, API generation, and long-context request tests.

Qualified RTX 5090 profile

The tested single-stream profile uses:

python3 -m sglang.launch_server \
  --model-path /model \
  --served-model-name qwopus3.8-27b-nvfp4-mtp \
  --host 0.0.0.0 --port 1234 \
  --kv-cache-dtype fp8_e4m3 \
  --attention-backend flashinfer \
  --context-length 131072 \
  --max-total-tokens 129241 \
  --chunked-prefill-size 1024 \
  --mamba-radix-cache-strategy extra_buffer \
  --mamba-ssm-dtype bfloat16 \
  --max-mamba-cache-size 1 \
  --mem-fraction-static 0.96 \
  --max-running-requests 1 \
  --speculative-algorithm EAGLE \
  --speculative-draft-model-path /model \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  --disable-radix-cache \
  --disable-prefill-cuda-graph \
  --weight-loader-drop-cache-after-load \
  --random-seed 42 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

Do not add --language-only: the checkpoint already declares language_model_only=true, and the extra runtime flag is incompatible with this wrapper configuration in the qualified SGLang build.

Reproducible setup scripts and the pinned SGLang build are in shiftedx/Qwen3.8-27B-RTX5090-SGLang-DFlash.

Measured results

Measured on one RTX 5090 with one active request. Results depend on prompt and MTP acceptance and are not universal performance guarantees.

At the qualified 131K profile, GPU residency was approximately 27.6 GB and the SGLang container used approximately 4.9 GiB of host RAM. No CPU weight offload was enabled.

Limitations

• Text only; no vision encoder is present.
• Experimental ModelOpt NVFP4 format; use the pinned SGLang profile above.
• No broad post-quantization quality evaluation has been completed yet.
• SGLang reports that no explicit FP8 KV scaling factors are present and falls back to 1.0; evaluate quality for your workload before production use.
• The physical token pool is 129,241 tokens, so prompt plus generated output must fit within that total even though the logical context is 131,072.

License and attribution

Apache-2.0, following the source checkpoint. Credit belongs to the Qwopus authors and the Qwen team; this repository only provides the quantized, text-only packaging and measured RTX 5090 serving profile.