Update README.md

README.md

@@ -3,13 +3,15 @@ library_name: sglang
 pipeline_tag: text-generation
 license: mit
 base_model:
-  - inclusionAI/Ling-3.0-Flash
+  - inclusionAI/Ling-3.0-flash
 tags:
   - speculative-decoding
   - tree-speculative-decoding
   - ubtree
+  - dflash
   - sglang
   - ling3
+  - arxiv:2609.39972
 language:
   - zh
   - en
@@ -17,99 +19,100 @@ language:
 
 # Ling3-Flash-UBTree-B16
 
-Ling3-Flash-UBTree-B16 is a tree speculative-decoding draft model for
-`inclusionAI/Ling-3.0-Flash`. It combines a five-layer DFlash draft backbone
-with a learned UBTree selector. The draft predicts multiple future positions
-in parallel, the selector builds a prefix-closed candidate tree, and the
-Ling3-Flash target model verifies the tree before any token is committed.
+[Paper](https://arxiv.org/abs/2609.39972) · [Target model](https://huggingface.co/inclusionAI/Ling-3.0-flash)
 
-This repository contains a speculative-decoding draft checkpoint, not a
-standalone language model. It must be used together with the compatible
-Ling3-Flash target model and a UBTree-enabled SGLang runtime.
+**Ling3-Flash-UBTree-B16** is a UBTree speculative-decoding draft model for
+[inclusionAI/Ling-3.0-flash](https://huggingface.co/inclusionAI/Ling-3.0-flash). It combines a
+five-layer DFlash backbone with a learned bigram selector to propose multiple
+candidate continuations for target-model verification.
 
-## Model structure
+> **SGLang integration — stay tuned!** Our UBTree adaptation for Ling3 has not
+> yet been released to the open-source SGLang community. Public installation
+> instructions and serving recipes will be added here when it becomes available.
 
-| Item | Value |
-|---|---|
-| Target model | `inclusionAI/Ling-3.0-Flash` |
-| Architecture | `Ling3UBTreeModel` |
-| Draft backbone | 5-layer DFlash |
-| Draft block size | 16 tokens |
-| Maximum verification tree | 65 positions per verification step |
-| Benchmark verification tree | 32 positions per verification step |
-| Training | Joint draft-backbone and tree-selector training |
-| Precision | BF16 |
-
-### Configuration compatibility
-
-`config.json` identifies this checkpoint as `Ling3UBTreeModel`. The
-`model_type: qwen3` field is retained for compatibility with the shared dense
-draft configuration used by the runtime; it does not mean that this checkpoint
-targets a Qwen3 model.
+This checkpoint is used alongside the matching target model; it cannot serve
+as a standalone text-generation model.
 
-## Intended use
+## How UBTree works
 
-The model is intended to accelerate Ling3-Flash inference in text generation,
-mathematical reasoning, code generation, and long-form generation workloads.
-It can also support research on tree-based speculative decoding.
+![UBTree architecture and tree-native training pipeline](assets/ubtree_pipeline.png)
 
-It is not intended to replace the target model or to generate outputs without
-target verification.
+*UBTree pipeline*
 
-## Runtime requirements
+UBTree pairs a **unigram proposer** with a **bigram selector**:
 
-- `inclusionAI/Ling-3.0-Flash` as the target model.
-- A UBTree-enabled SGLang runtime that registers `Ling3UBTreeModel`.
-- This repository as the speculative draft-model path.
+1. A DFlash backbone proposes candidate tokens for future positions in parallel.
+2. The selector scores transitions between adjacent candidates in parallel.
+3. Tree construction retains promising branches within a verification budget.
+4. The target verifies the tree and commits the accepted continuation.
 
-The checkpoint is not directly loadable as a standalone
-`transformers.AutoModelForCausalLM` model.
+The paper also introduces tree-native training with renormalized KL supervision
+on high-temperature target traces, encouraging plausible alternative branches.
+See the [paper](https://arxiv.org/abs/2609.39972) for the method and experiments.
 
-## Performance reference
+## Model specifications
 
-The following results are from an optimized SGLang evaluation performed on
-**2026-09-23** using H200 GPUs. The setup used Ling3-Flash with tensor
-parallelism 4, greedy decoding, a fixed 32-position verification tree, overlap
-scheduling, and tuned Triton MoE kernels. The concurrency benchmark additionally
-used Full Decode CUDA Graph with batch sizes 1 through 32 captured. Throughput
-is measured as client-visible output tokens divided by end-to-end wall time.
-
-Each task used the same ordered inputs, one warm-up request, temperature 0,
-thinking disabled, EOS enabled, and `max_new_tokens=2048`.
-
-| Dataset | Samples | Output tokens/s |
+| Item | Value |
+|---|---|
+| Target model | [inclusionAI/Ling-3.0-flash](https://huggingface.co/inclusionAI/Ling-3.0-flash) |
+| Draft architecture | `Ling3UBTreeModel` |
+| Draft backbone | 5 full-attention Transformer layers |
+| Parameters stored in this checkpoint | 481,983,774 (482.0M) |
+| Weight dtype | BF16 |
+| Hidden size | 2,560 |
+| Attention heads / KV heads | 32 / 32 |
+| Target auxiliary feature layers (zero-based) | 1, 11, 23, 29, 35 |
+| Draft block size | 16 (anchor-inclusive runtime convention) |
+| Candidate top-k | 64 per draft position |
+| Default verification tree budget | 64 non-root nodes + 1 anchor |
+| Training | DFlash warm start, then joint backbone + UBTree selector training |
+
+The parameter count covers tensors distributed in this repository. The draft
+shares token embeddings and the language-model head with the target; these
+shared weights are not included here. B16 denotes the draft block size, not
+the verification tree's node budget.
+
+## Serving with SGLang
+
+UBTree support for Ling3 in SGLang is coming soon. We will share installation
+instructions and serving examples once the integration is publicly available.
+**Stay tuned!**
+
+## Benchmark results
+
+Results for this checkpoint will be added here. **TBD** means not yet reported.
+
+| Dataset | Acceptance length | Speedup |
 |---|---:|---:|
-| GSM8K | 128 | 908.50 |
-| MATH-500 | 128 | 1136.46 |
-| AIME25 | 30 | 1068.17 |
-| HumanEval | 128 | 1012.01 |
-| MBPP | 128 | 966.96 |
-| LiveCodeBench | 128 | 957.39 |
-
-The fixed-width Full Decode CUDA Graph concurrency sweep on 128 GSM8K prompts
-reported the following median throughput across three rounds:
-
-| Concurrent requests | Output tokens/s |
-|---:|---:|
-| 1 | 907.05 |
-| 4 | 1509.43 |
-| 16 | 2174.15 |
-| 32 | 2178.32 |
-
-These figures are reference measurements from the stated environment, not
-speedups over a target-only server. The original benchmark summary does not
-include an exact public software revision, so results may vary across runtime
-versions. These performance tests do not establish strict output equivalence
-or task quality.
-
-## Limitations
-
-- Requires the matching Ling3-Flash target and UBTree runtime plugin.
-- Verified performance does not automatically transfer to other tensor
-  parallel sizes, runtime versions, hardware, or concurrency levels.
-- Sampling, grammar-constrained decoding, and other serving features require
-  independent validation.
+| GSM8K | TBD | TBD |
+| MATH-500 | TBD | TBD |
+| AIME 2025 | TBD | TBD |
+| HumanEval | TBD | TBD |
+| MBPP | TBD | TBD |
+| LiveCodeBench | TBD | TBD |
+
+Acceptance length is the mean number of tokens committed per verification
+step, including the target token. Speedup is measured against target-only
+autoregressive decoding with the same hardware, inputs, and decoding settings.
+The runtime version, tree budget, concurrency, and timing protocol will be
+reported alongside the results.
+
+## Citation
+
+If you use UBTree in your work, please cite:
+
+```bibtex
+@article{liang2026ubtree,
+  title   = {UBTree: Parallel Tree Drafting via Unigram and Bigram Models for Speculative Decoding},
+  author  = {Liang, Chumeng and Wang, Linxuan and Peng, Xinyu and Liu, Huabin and Chen, Yuxin and Liu, Ge and Lin, Guang and Song, Qifan and Li, Jianguo},
+  journal = {arXiv preprint arXiv:2609.39972},
+  year    = {2026},
+  url     = {https://arxiv.org/abs/2609.39972}
+}
+```
 
 ## License
 
-This checkpoint is released under the MIT License.
+The model weights are released under the **MIT License**. The pipeline figure
+is from [Liang et al., Figure 2](https://arxiv.org/html/2609.39972v1#S2.F2),
+reproduced without modification under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).