license: apache-2.0
base_model: RadixArk/Qwen3.8-27B-DSpark
base_model_relation: finetune
library_name: transformers
language:
- en
- zh
tags: - qwen3.8
- dspark
- speculative-decoding
- rtx-5090
- sglang
- bfloat16
Qwen3.8-27B Uncensored DSpark RTX 5090
Target-matched BF16 DSpark speculative drafter for fleetml/Qwen3.8-27B-Uncensored-NVFP4-RTX5090.
The drafter was initialized from RadixArk/Qwen3.8-27B-DSpark and trained on hidden states captured from the exact OrcaRouter-derived NVFP4 target.
Combined profile
| Field | Value |
|---|---|
| Combined disk size | 24.31 GB |
| Combined GPU weight allocation | 22.44 GB |
| Complete optimized GPU use | Approximately 29.39 GB |
| GPU memory remaining | 3.22 GB |
| Architecture context window | 262,144 tokens |
| RTX 5090 launch context | 65,536 tokens |
| Optimized active token pool | 16,384 tokens |
Measured result
The matched pair reached 139.32 tokens per second median decode on one RTX 5090 with a 16,384-token pool. The paired result measured 1.18 times the public warm start, 1.41 times native MTP depth 5, and 2.79 times standard decoding in the same deterministic prompt suite.
Median accepted length was 4.70 tokens with a median acceptance rate of 0.53.
Drafter details
| Field | Value |
|---|---|
| Architecture | DSparkDraftModel |
| Precision | BF16 |
| Parameters | 1,857,358,337 |
| Layers | 5 |
| Block size | 7 |
| Target layers | 5, 19, 33, 47, 61 |
| Export size | 3,714,723,322 bytes |
| Safetensors SHA-256 | cb2591889afbbf51aae4545ae8f1569fda17a21448a44e2750fa8182fab53fd5 |
Training
- Exact-target teacher features: 126 records
- Maximum training sequence length: 1,024 tokens
- Training length: 64 optimizer steps
- Training precision: BF16
- Dataset source: HuggingFaceH4/ultrachat_200k
- Training framework: SpecForge at commit
2fc993077c3c53df14ebdd5d414c854864476af8
Serving
Use this repository as --speculative-draft-model-path with the target repository. The complete RTX 5090 launch command is provided on the target model page.
Use
This drafter is intended for controlled local research, evaluation, and agent development. Use it responsibly and comply with the Apache 2.0 license and applicable law.