Qwen3.8-27B · FP8
This page aggregates 1 real-world runs of Qwen3.8-27B (FP8) on NVIDIA RTX PRO 5000 Blackwell with SGLang, contributed by 1 independent source platforms; metrics are averages of published measurements.
31.5 tok/s
Decode
Decode speed
5,089 tok/s
Prefill
Prefill speed
— s
TTFT
Time to first token
43.2 GB
VRAM
VRAM usage
1
Measured runs
1
Independent sources
Other
Source platforms
5 days ago
Last verified
Performance
- SGLang · FP8 (current)Decode 31.5 · Prefill 5089 ·
Core figures
- Decode (avg)
- 31.5 tok/s
- Prefill (avg)
- 5,089 tok/s
- TTFT (avg)
- — s
- VRAM (avg)
- 43.2 GB
- MTP acceptance rate
- —
- TTFB
- — GB
- Power draw
- — W
Configuration
Member-level fields are taken from the most recent run
- Model
- Qwen3.8-27B
- Quantization
- FP8
- Framework
- SGLang
- Version
- 未知
- Context length
- 226000 tokens
- Batch size
- 1
- Flash Attention
- On
Hardware
Nominal and measured figures are shown side by side; whether it runs is the reader's call
- GPU
- NVIDIA RTX PRO 5000 Blackwell
- Nominal VRAM
- 48 GB
- Measured VRAM (avg)
- 43.2 GB
- OS
- 未知
Get started
Original model resources
Base model resources may not match this quantization. Use the verified deployment weights above when available.
Startup command
Command from the latest benchmark runpython -m sglang.launch_server --model-path Qwen3.8-27B-FP8 --attention-backend flashinfer --kv-cache-dtype fp8_e4m3 --mamba-full-memory-ratio 0.2 --chunked-prefill-size 2048 --context-length 262144 --mem-fraction-static 0.90Deployment guide
https://lcz.me/topic/1193Sources & evidence
1 measured records in total, each traceable to its original source
未知 · 未知 · 226000 ctx
31.5 tok/s
Decode
5,089 tok/s
Prefill
— s
TTFT
43.2 GB
VRAM
—
MTP
— W
Power draw
bench_one_batch：batch=1、input=4096，prefill 5,006→5,089 t/s，decode 31.3→31.5 t/s。