Inkling DSpark speculator

Overview

A DSpark speculator for the Inkling 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.

Preview: This is a preview checkpoint. Official training and evaluation are in progress.

Model Specifications

• Base model: Inkling NVFP4; the exact matching target checkpoint and tokenizer are required.
• Format: Safetensors (single-file BF16, 1.93B trainable parameters + embedding + lm_head).
• Draft: 5 layers (Qwen3-style GQA), hidden 6144, 64 heads / 16 KV heads, head_dim 64, FFN 12288, rope_theta 8000000, block_size=7.
• 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: [5, 17, 35, 47, 59].
• Trained context: sequence length 4096.
• Target weights: target embedding and unembedding weights are not included in this checkpoint.

Evaluation Results

Acceptance length (acc_len) at temperature 0:

Serving with SGLang

Requires a SGLang build (docker pull lmsysorg/sglang:dev-cu13-inkling-dspark) with DSpark support:

SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 \
python -m sglang.launch_server \
  --trust-remote-code \
  --model-path thinkingmachines/Inkling-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 0.68 \
  --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-DSpark-Preview \
  --speculative-draft-model-quantization unquant \
  --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 the frozen Inkling target.
• Data: 400K Inkling regenerations from open-perfectblend.
• Schedule: up to 10 epochs, AdamW, peak learning rate 6e-4, cosine decay, 4% warmup, and gradient clipping at 1.0.
• Loss: 0.1 CE + 0.9 L1 distillation + 1.0 confidence BCE, with 512 sampled anchors per sequence and block_size=7.