Downloads · 30 days
43
17% of all-time downloads
Laguna-S-2.1-MXFP4-dspark
olka-fi/Laguna-S-2.1-MXFP4-dspark
Laguna-S-2.1-MXFP4-dspark is a text generation model from olka-fi. Use it when you need the model to write or continue text. It is set up for speculators. The card lists the license as openmdw-1.1.
A DSpark speculative-decoding draft head for olka-fi/Laguna-S-2.1-MXFP4 (118B-A8B MoE verifier, MXFP4). Trained with vllm-project/speculators.
All-time downloads
258
Public
Parameters
941M
1.9 GB on disk
Likes
2
Public
Repo makeup
Click a slice to open those files.
.safetensors1.9 GB · 100%
Parameter types
How the weights are stored.
BF16941M · 100%
Try a prompt
Base models
Model card
From the Hugging Face model README
Laguna-S-2.1-MXFP4 — DSpark speculator head
A DSpark speculative-decoding draft head for
olka-fi/Laguna-S-2.1-MXFP4
(118B-A8B MoE verifier, MXFP4). Trained with
vllm-project/speculators.
Usage (vLLM)
Exact command used to serve this head and produce the numbers above (B300, single GPU):
vllm serve olka-fi/Laguna-S-2.1-MXFP4 \
--speculative-config '{"model":"olka-fi/Laguna-S-2.1-MXFP4-dspark","method":"dspark","num_speculative_tokens":4}' \
--port 8000 --gpu-memory-utilization 0.9
DSpark extends DFlash's anchored-block drafting with a Markov head (conditions each draft position on the previously drafted token within the block) and a confidence head (predicts per-position acceptance probability).
Evaluation
| Metric | Result | Notes |
|---|---|---|
| GSM8K (full test, 1319, temp 0) | 95.5% | chat-template CoT harness, num_speculative_tokens=4; matches base verifier (~94.5%) — spec decode is lossless |
| Decode speedup | ~2× | vs. no-draft baseline (cudagraphs, B300) |
Head architecture
| Algorithm | dspark (anchored block, Markov + confidence heads) |
| Draft decoder | 3 × Qwen3-style layers, hidden 3072, sliding-window 512 on all draft layers |
| Block size | 8 |
| Aux hidden layers (from verifier) | [2, 24, 45] + final |
| Draft vocab | 32000 (EAGLE-3 vocab compression) |
| Markov rank | 256 (vanilla) |
| Params | ~941M (bf16), ~1.88 GB |
Training
- Data:
mlabonne/open-perfectblend, 22k samples, seq-len 16000. - Regime: offline hidden-state distillation (teacher-forced / off-policy), 10 epochs.
- Loss:
{"ce": 0.1, "tv": 0.9}; Muon + AdamW.
Requirements / caveats
- Needs a vLLM build with
dsparksupport - On NVIDIA Blackwell B300 (sm100)
VLLM_FORCE_MXFP4_MARLIN=1is required — the CUTLASS MXFP4 MoE kernel is fine for plain generation but faults under DSpark's draft-block shapes; the Marlin MXFP4 backend is stable.
- Task
- Text Generation
- Library
- speculators
- License
- openmdw-1.1
- Created
- Jul 25, 2026
- Updated
- Jul 25, 2026