Skip to main content

Deployment

For all methods and hardware platforms, see the official SGLang installation guide. The two paths below match the Python / Docker toggle in the command panel.
Command
Then run the Python output of the command panel below in that environment.
Pick your card + checkpoint precision to generate the launch command. The model runs single-GPU on every supported card. The panel opens on the configuration we serve and evaluate DaisyRose-27B with: DSpark speculative decoding and the High-Throughput strategy.
--mamba-full-memory-ratio is the one sizing flag that matters for throughput on hybrid GDN models: the default (0.9) over-provisions the KV pool and silently clamps concurrency. Set your average request length in the Mamba ratio calculator below; the computed value is pinned into the command.
DaisyRose-27B has the same architecture and tensor shapes as Qwen3.8-27B, so every cell is the Qwen3.8-27B recipe for the same card and precision. The GB300 cells with DSPARK are measured with DaisyRose-27B weights on v0.5.23: all three checkpoints serve text, tool-calling and image requests, and DSpark’s draft is accepted for 3.4-3.5 tokens per step in long-context thinking workloads. The other cells have not been re-measured with DaisyRose-27B weights and carry the unverified badge.

Mamba ratio calculator

Hybrid GDN models split post-weight memory into a worst-case-reserved GDN state pool (sets the concurrency ceiling) and a paged attention KV pool, divided by --mamba-full-memory-ratio. Every parameter below except L and the target concurrency is read live from the Deploy panel and Playground selection; the balanced value is the per-request cost ratio:
Formula
  • S — state slots per running request: extra_buffer=5 (default), extra_buffer_lazy=4, no_buffer=3, disabled radix cache =1. For the two extra_buffer strategies, SGLANG_OPT_MAMBA_SKIP_DECODE_LOCK=1 frees one slot, and extra_buffer frees one more with the overlap scheduler off; the calculator reads both knobs.
  • D — verify intermediate states under speculative decoding: --speculative-num-draft-tokens for EAGLE/MTP (4 at the recommended 3/1/4) and for DFLASH (8, DFlash2’s block size); --speculative-dspark-block-size + 1 for DSPARK, where the block size falls back to the draft checkpoint’s block_size when the flag is omitted (7 for RadixArk/Qwen3.8-27B-DSpark, so D = 8); 0 with speculation off or with --enable-linear-replayssm-spec, which keeps the verify intermediates on a fixed ring instead of per-request slots.
  • state_bytes — one state slot, from the fixed geometry (48 GDN layers x 48 heads x 128 x 128 at --mamba-ssm-dtype, plus bf16 conv state): 153.9 MB at fp32, 78.4 MB at bf16.
  • kv_bytes_per_token — 16 attention layers x GQA 4 x 256 x K+V: 32.8 KB at fp8, 65.5 KB at bf16.
  • L — average total request length in tokens: input + output.
--max-mamba-cache-size = target_concurrency x S is the equivalent explicit pin and overrides the ratio; the calculator emits it alongside. D is not a term here: the engine divides the state pool by S alone and sizes the speculative verify buffer separately, so folding D into the pin would over-provision the pool. After boot, verify with the max_running_requests line in the server log — it should not be capped below your target concurrency.

Playground

The Playground is where you experiment with SGLang features beyond the recipes above. The Deploy panel emits this model’s documented launch recipes; the Playground lets you turn on additional knobs on top of whichever cell the Deploy panel is currently showing.

1. Model Introduction

DaisyRose-27B is RadixArk’s first model: Qwen3.8-27B post-trained with reinforcement learning only, without supervised fine-tuning or distillation, on about 1,200 long-horizon terminal tasks. It thinks before it answers and is aimed at agentic coding, terminal and software-engineering work. Evaluation results are on the model card. It keeps the base model’s architecture, tokenizer and chat template: a dense hybrid Gated Delta Networks (GDN) vision-language model with 48 linear-attention and 16 full-attention layers, an in-checkpoint MTP head, native image input and a 262,144-token context. The vision encoder and the MTP head are the base model’s, unchanged. It is a drop-in replacement for Qwen3.8-27B, and the Qwen3.8-27B page covers the architecture and platform tuning in depth. License: Apache 2.0.
ModelQuantizationWeights
DaisyRose-27BBF16RadixArk/DaisyRose-27B
DaisyRose-27B-FP8FP8 (blockwise)RadixArk/DaisyRose-27B-FP8
DaisyRose-27B-NVFP4NVFP4 W4A4 + FP8 projections, lm_head packed to FP4RadixArk/DaisyRose-27B-NVFP4
The NVFP4 checkpoint is a NVIDIA Model Optimizer export with the same quantized-layer map as nvidia/Qwen3.8-27B-NVFP4: FP8 attention and GDN projections, NVFP4 MLPs, and an FP4-packed lm_head. It needs FP4 tensor cores, so it has no H200 cell.

2. Configuration Tips

  • DSpark: the draft model is a separate checkpoint trained on Qwen3.8-27B, and it carries over to DaisyRose-27B — add --speculative-algorithm DSPARK --speculative-draft-model-path RadixArk/Qwen3.8-27B-DSpark. DSpark does not take --speculative-num-draft-tokens; its verify window comes from the draft checkpoint (D = 8), which the calculator folds into --mamba-full-memory-ratio.
  • MTP: --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 uses the in-checkpoint MTP head and needs no extra download.
  • Hardware fit: BF16 weights ~52GB, FP8 ~29GB, NVFP4 ~23GB. On a 32GB RTX 5090 only NVFP4 is serviceable, and the recipe pins a single stream.
  • SM120/SM121 (RTX PRO 6000, RTX 5090, DGX Spark): use --attention-backend flashinfer. On DGX Spark keep --mem-fraction-static 0.80; the 128GB is shared with the host.
  • Parsers: every recipe carries --reasoning-parser qwen3 --tool-call-parser qwen3_coder. Without them a client receives thinking and tool calls as raw text.
  • Platform notes, GDN state precision and prefill-chunk guidance are the same as for the base model: see Qwen3.8-27B → Configuration Tips.

3. Advanced Usage

The examples use the openai Python client against the server started above, with --served-model-name DaisyRose-27B added to the launch command.

3.1 Reasoning

Thinking is on by default. The qwen3 reasoning parser returns it in reasoning_content and the answer in content.
Example
Output Example:
Output
Turn thinking off per request with enable_thinking:
Example
Output Example:
Output

3.2 Tool Calling

Example
Output Example:
Output