A DSpark speculator for the Inkling-Small NVFP4 target, enabling faster inference through speculative decoding. DSpark extends the DFlash parallel-draft backbone with a Markov logit-bias head and a per-position confidence head. This checkpoint was trained with SpecForge using hidden states from a live SGLang target engine.

Model Specifications

Base model: thinkingmachines/Inkling-Small-NVFP4; the exact matching target checkpoint and tokenizer are required.
Format: Safetensors (single-file BF16, 1.39B trainable parameters; draft weights only).
Draft: 6 layers (Qwen3-style GQA), hidden 4096, 32 heads / 8 KV heads, head_dim 128, FFN 12288, rope_theta 8000000, block_size=7.
RoPE: YaRN, factor 128, original_max_position_embeddings=8192, max_position_embeddings=1048576 (matches the target addressable range); config carries both rope_scaling and rope_parameters schemas.
Vocabulary: 200,058 tokenizer entries and 201,024 padded weight rows; mask_token_id=200064.
DSpark heads: Markov rank 256 (vanilla) and confidence head (with-Markov).
Aux hidden-state layers: [1, 6, 12, 17, 23, 28, 34, 39].
Trained context: 8,192 native; long-context adaptation with sequences up to 65,536.
Target weights: target embedding and unembedding weights are not included in this checkpoint.

Evaluation Results

Acceptance length over the full DeepSpec workload datasets. DSPARK block size 7, thinking effort 0.99, temperature 0 and temperature 1.0 / top-p 0.95:

Dataset	n	acc_len @ T=0	acc_len @ T=1
GSM8K	1,319	5.4254	5.2700
MATH500	500	4.8009	4.6787
MBPP	257	4.1515	4.0105
HumanEval	164	3.9911	3.9084
MT-Bench	80	3.7309	3.6076
LBPP	162	3.6420	3.5093
AIME24	30	3.5456	3.3864
AIME25	30	3.4172	3.3192
LiveCodeBench	1,055	3.4154	3.3139
Alpaca	52,002	3.3504	3.2087
Arena-Hard-v2	750	3.2091	3.0265
SWE-Bench	300	3.1914	2.9975
Mean (12)		3.8226	3.6864

RULER V2 at the 1M input configuration (50 prompts per partition, actual prompts 1,022,335–1,045,000 tokens):

Partition	acc_len @ T=0	acc_len @ T=1
MK (multi-key)	4.2665	3.9417
MV (multi-value)	4.2324	3.9666
QA	3.4938	3.2632

Long-context acceptance is flat on held-out 16K–64K agentic trajectories (67 rows, temperature 0): 3.592 (8–16K), 3.560 (16–32K), 3.534 (32K+).

Serving with SGLang

Requires a SGLang build with DSpark support

bash

SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 \
sglang serve --trust-remote-code \
  --model-path thinkingmachines/Inkling-Small-NVFP4 --tp 8 \
  --quantization modelopt_fp4 --attention-backend fa4 --page-size 128 \
  --fp4-gemm-backend flashinfer_trtllm --moe-runner-backend flashinfer_trtllm_routed \
  --enable-torch-symm-mem --mamba-radix-cache-strategy extra_buffer \
  --mem-fraction-static 0.60 --swa-full-tokens-ratio 0.1 --mamba-full-memory-ratio 0.1 \
  --max-running-requests 68 --reasoning-parser inkling --tool-call-parser inkling \
  --skip-server-warmup --speculative-algorithm DSPARK \
  --speculative-draft-model-path RadixArk/Inkling-Small-DSpark \
  --speculative-draft-model-quantization unquant \
  --speculative-dspark-block-size 7 \
  --chunked-prefill-size 8192 --cuda-graph-max-bs-prefill 8192 \
  --disable-flashinfer-autotune --host 0.0.0.0 --port 30000

Training Details

Framework: SpecForge online distillation with hidden states captured from a frozen Inkling-Small target served by a colocated per-node TP4 SGLang engine; KV injection of fused target features into every draft layer with block-local bidirectional attention.
Loss: 0.1 CE + 0.9 L1 distillation + 1.0 confidence BCE, 512 sampled anchors per sequence, block_size=7, within-block decay gamma 4.