README.md
| 1 | --- |
| 2 | library_name: transformers |
| 3 | pipeline_tag: text-generation |
| 4 | tags: |
| 5 | - speculative-decoding |
| 6 | - dspark |
| 7 | - dflash |
| 8 | - specforge |
| 9 | - sglang |
| 10 | - long-context |
| 11 | inference: false |
| 12 | --- |
| 13 | |
| 14 | # Kimi K3 DSpark speculator |
| 15 | |
| 16 | ## Overview |
| 17 | |
| 18 | A long-context DSpark speculator for |
| 19 | [Kimi K3](https://huggingface.co/moonshotai/Kimi-K3). It supports context lengths of up to 1 million tokens. |
| 20 | |
| 21 | A DSpark speculator for the Kimi K3 target, enabling faster inference through speculative decoding. |
| 22 | 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](https://github.com/sgl-project/SpecForge/) using hidden states from a live [SGLang](https://github.com/sgl-project/sglang) target engine. |
| 23 | |
| 24 | |
| 25 | ## Model Specifications |
| 26 | |
| 27 | - **Base model:** `moonshotai/Kimi-K3` |
| 28 | - **Format:** Safetensors (single-file BF16, 2,249,289,601 parameters) |
| 29 | - **Draft:** 5 full-attention Qwen3-style GQA layers, hidden size 7168, |
| 30 | 64 query heads / 16 KV heads, and `block_size=7` |
| 31 | - **Verification width:** 1 current token + 7 draft tokens |
| 32 | - **Auxiliary target layers:** `[7, 23, 51, 67, 83]` |
| 33 | - **Trained context:** 65,536 tokens |
| 34 | - **Target weights:** embedding and unembedding weights are not included |
| 35 | |
| 36 | ## Evaluation Results |
| 37 | |
| 38 | `acc_len` is SGLang's histogram-native request acceptance length, averaged |
| 39 | within each question and then equally across questions. |
| 40 | |
| 41 | | Dataset | Questions | acc_len | |
| 42 | |---|---:|---:| |
| 43 | | SWE-Rebench | 50 | **4.6594** | |
| 44 | | GSM8K | 1,319 | **5.4176** | |
| 45 | | MATH500 | 500 | **4.1329** | |
| 46 | | HumanEval | 164 | **5.5121** | |
| 47 | | MBPP | 257 | **5.1980** | |
| 48 | | MT-Bench | 80 | **3.9342** | |
| 49 | | AIME26 | 30 | **2.9893** | |
| 50 | | RULER V2 1M (MK/MV/QA) | 150 (50 per partition) | **4.2553** | |
| 51 | |
| 52 | RULER V2 uses the 1M input configuration. Actual prompts span |
| 53 | 1,000,432–1,047,925 tokens; partition acc_len is 4.4658 for MK, 4.3081 for |
| 54 | MV, and 3.9919 for QA. |
| 55 | |
| 56 | ### AIME26 acc_len by output length |
| 57 | |
| 58 | | Output-token bucket | Questions | Actual output range | acc_len | |
| 59 | |---|---:|---:|---:| |
| 60 | | 0–1K | 13 | 192–885 | **3.1310** | |
| 61 | | 1–2K | 5 | 1,359–1,828 | **2.5773** | |
| 62 | | 2–4K | 6 | 2,210–3,732 | **2.5632** | |
| 63 | | 4–8K | 4 | 5,187–7,750 | **2.7174** | |
| 64 | | 8–16K | 0 | — | — | |
| 65 | | 16–32K | 0 | — | — | |
| 66 | | 32K+ | 2 | 54,545–224,703 | **4.9194** | |
| 67 | |
| 68 | |
| 69 | ## Serving with SGLang |
| 70 | |
| 71 | [SGLang Cookbook](https://lmsysorg.mintlify.app/cookbook/autoregressive/Moonshotai/Kimi-K3#hw=b300&pdMode=unified&strategy=low-latency&spec=dspark&hicache=off) provides Kimi K3 deployment recipes. |
| 72 | ```bash |
| 73 | sglang serve \ |
| 74 | --trust-remote-code \ |
| 75 | --model-path moonshotai/Kimi-K3 \ |
| 76 | --tp-size 8 \ |
| 77 | --dcp-size 8 \ |
| 78 | --mem-fraction-static 0.85 \ |
| 79 | --max-mamba-cache-size 160 \ |
| 80 | --max-running-requests 32 \ |
| 81 | --cuda-graph-max-bs-decode 32 \ |
| 82 | --reasoning-parser kimi_k3 \ |
| 83 | --tool-call-parser kimi_k3 \ |
| 84 | --host 0.0.0.0 \ |
| 85 | --port 30000 \ |
| 86 | --speculative-algorithm DSPARK \ |
| 87 | --speculative-draft-model-path RadixArk/Kimi-K3-DSpark \ |
| 88 | --speculative-dspark-block-size 7 \ |
| 89 | --speculative-draft-attention-backend trtllm_mha \ |
| 90 | --enable-linear-replayssm-spec \ |
| 91 | --context-length 1048576 \ |
| 92 | --chunked-prefill-size 16384 |
| 93 | ``` |
| 94 | |
| 95 | YaRN-16 is enabled in the published draft config by default with |
| 96 | `original_max_position_embeddings=65536` and |
| 97 | `max_position_embeddings=1048576`; no separate draft config override is |
| 98 | required. |
| 99 | |
| 100 | ## Training Details |
| 101 | |
| 102 | - **Framework:** SpecForge online distillation, with hidden states captured from a frozen Kimi K3 target served by a live SGLang engine. Draft trained from random initialization. |
| 103 | - **Loss:** `0.1 CE + 0.9 L1 distillation + 1.0 confidence BCE`, decay gamma 4.0, with 512 sampled anchors per sequence and `block_size=7`. |
| 104 | - **Topology:** 4 nodes × 4 GB300 (16 ranks) — 2 × TP8 target replicas, DP2 sampler, FSDP16 `SHARD_GRAD_OP` on the draft, TP-batch scatter. Batch 8 per replica × 32 accumulation steps × 2 replicas = global batch 512. |
| 105 | |
| 106 | |
| 107 | |