--- title: Qwen3.5-0.8B-GDN2-LoopSpatial canonical_url: "https://www.modelscope.cn/models/tysgydk/Qwen3.5-0.8B-GDN2-LoopSpatial" md_url: "https://www.modelscope.cn/models/tysgydk/Qwen3.5-0.8B-GDN2-LoopSpatial.md" repository: tysgydk/Qwen3.5-0.8B-GDN2-LoopSpatial last_updated: 2026-09-22 license: apache-2.0 pipeline_tag: image-text-to-text tasks: - image-text-to-text model_type: - qwen3_5_gdn2 architectures: - Qwen35GDN2ForConditionalGeneration base_model: - Qwen/Qwen3.5-0.8B base_model_relation: finetune parameters: 965.7M tensor_type: - F32 - BF16 library_name: - pytorch - transformer - safetensors frameworks: - Pytorch downloads: 32 stars: 0 tags: - gated-delta-net - gdn2 - linear-attention - multimodal - spatial-reasoning --- # Qwen3.5-0.8B-GDN2-LoopSpatial > Qwen3.5-0.8B-GDN2-LoopSpatial - tysgydk 在 ModelScope 开源的模型。Qwen3.5-0.8B-GDN2-LoopSpatial tysgydk/Qwen3.5-0.8B-GDN2-LoopSpatial 是 ModelScope 魔搭社区上的 965.7M 参数image-text-to-text模型，采用 apache-2.0 许可，基于 Qwen/Qwen3.5-0.8B 构建。 - **Repository**: tysgydk/Qwen3.5-0.8B-GDN2-LoopSpatial - **License**: apache-2.0 - **Tasks**: image-text-to-text - **Parameters**: 965.7M - **Base model**: Qwen/Qwen3.5-0.8B - **Tags**: gated-delta-net, gdn2, linear-attention, multimodal, spatial-reasoning - **Downloads**: 32 - **Stars**: 0 - **Last updated**: 2026-09-22 Source: https://www.modelscope.cn/models/tysgydk/Qwen3.5-0.8B-GDN2-LoopSpatial --- # Qwen3.5-0.8B-GDN2-LoopSpatial **v4 = [GDN2 v2](https://modelscope.cn/models/tysgydk/Qwen3.5-0.8B-GDN2) 主干 + 循环（Ouro 式 K=2 权值共享）+ 空间模块（多点注入 + 闭环回写）。** 在 GDN2 改造之上， 本权重同时携带两个研究组件：一个循环机制——整个 24 层文本块每次前向权重共享地 应用 K 次，以零初始化的可学习 `loop_emb` 轮次标记区分精化轮次；以及 v3 的空间 状态模块（32 槽位、7 个门控注入点），并升级为**闭环**：SlotFeedback 把语言模型的 中间隐状态回写进槽位，槽位不再只读。 **v4 = [GDN2 v2](https://modelscope.cn/models/tysgydk/Qwen3.5-0.8B-GDN2) backbone + looping (Ouro-style K=2 weight sharing) + spatial module (multi-point injection + closed-loop write-back).** On top of the GDN2 conversion, this checkpoint carries two research components: a loop mechanism — the whole 24-layer text block is applied K times per forward with shared weights, a zero-init learned `loop_emb` marking each refinement round — and the v3 spatial state module (32 slots, 7 gated injection points), upgraded to **closed-loop**: SlotFeedback writes the LM's intermediate hidden states back into the slots, so the workspace is no longer read-only. 权重形态（v4.1 起）是**联合训练完全体**（plane，Plan E）：循环、空间、数学 CoT、 照片 VQA 从训练第 0 步同时进配比——实测证明"同时训练可共存、先后抛光必跷跷板" （先后抛光会在 '\n'→`<|im_end|>` 终止边界上互相污染）。v4.0 的嫁接版权重与评测 数字保留在审计仓库与本文"已知限制"末尾。 As of v4.1 the weight form is the **jointly-trained full model** (plane, Plan E): loop, spatial, math CoT, and photo VQA all enter the training mix from step 0 — experiments show that simultaneous co-training coexists while sequential polishing seesaws (polish-after-the-fact contaminates the '\n'→`<|im_end|>` termination boundary). The v4.0 grafted weights and numbers are kept in the audit repo and at the end of "Known limitations". ## 使用方法 | Usage K=1 默认用法与 v2/v3 完全相同——不 patch 循环时，本权重就是一个带空间模块的 GDN2 VLM： The default K=1 usage is identical to v2/v3 — without the loop patch, this checkpoint is simply a GDN2 VLM with the spatial module: ```python import torch from transformers import AutoModelForImageTextToText, AutoProcessor model = AutoModelForImageTextToText.from_pretrained( "tysgydk/Qwen3.5-0.8B-GDN2-LoopSpatial", trust_remote_code=True, dtype=torch.bfloat16, device_map="cuda") processor = AutoProcessor.from_pretrained( "tysgydk/Qwen3.5-0.8B-GDN2-LoopSpatial", trust_remote_code=True) ``` K=2 循环是可选附加件，需要仓库内的 `loop_patch.py`（Ouro 式定次块循环， `loop_emb` 权重存放在 `spatial.safetensors` 里）： The K=2 loop is an opt-in add-on via the repo's `loop_patch.py` (Ouro-style fixed-count block looping; the `loop_emb` weight lives in `spatial.safetensors`): ```python import sys; sys.path.insert(0, model_dir) from loop_patch import patch_loop patch_loop(model.model.language_model, 2) from safetensors.torch import load_file model.model.language_model.loop_emb.weight.data.copy_( load_file(f"{model_dir}/spatial.safetensors")["loop_emb.weight"]) ``` 空间模块用法同 v3：`spatial_runtime.py` 的 `attach_spatial` 构建槽位并挂上 7 个门控注入点；不构建槽位（或 `hook.clear()` 之后）空间通路完全惰性。 循环与空间可同时开启——闭环回写（`spatial_round_update`）发生在循环轮次之间。 The spatial module works as in v3: `attach_spatial` from `spatial_runtime.py` builds the slots and mounts the 7 gated injection points; with no slots built (or after `hook.clear()`) the spatial pathway is inert. Loop and spatial can be enabled together — the closed-loop write-back (`spatial_round_update`) fires between loop rounds. 环境要求：`transformers >= 5.17`、`flash-linear-attention >= 0.5.2`。 Requirements: `transformers >= 5.17`, `flash-linear-attention >= 0.5.2`. ## 能力一览（v4.1 完全体，K=1 模式实测）| Capabilities (v4.1 full model, measured, K=1 mode) > v4.1 起权重切换为**联合训练完全体**（plane，Plan E）：循环 + 空间 + 数学 CoT 从第 0 步 > 同时进配比联合训练，取代原先的嫁接版（嫁接版数字见"已知限制"末尾的历史记录）。 > > As of v4.1 the weights are the **jointly-trained full model** (plane, Plan E): loop + > spatial + math CoT co-trained from step 0, replacing the grafted build (graft numbers > kept at the end of "Known limitations" for the record). | 基准 Benchmark | 本模型 This model | 参照 Reference | |---|---|---| | TextVQA val 512（精确/宽松，标准协议）/ TextVQA val 512 (exact / soft, standard) | **58.4% / 73.0%** | v2 发布版 63.3% / 73.6% | | TextVQA val 512（stop-at-`\n` 协议*）/ TextVQA val 512 (stop-at-newline protocol*) | **63.7% / 73.4%** | v2 发布版 63.3% / 73.6% | | 语言评测 loss（96 条留出集）/ Language eval loss (96-sample holdout) | **1.931** | v2 发布版 1.776 | | 合成属性任务 Synthetic attribute task | **39/40** | v2 发布版 39/40 | | GSM8K（K=1 / K=2，n=200）/ GSM8K | **38.5% / 39.0%** | v2 发布版 K=1 36.5% | | BLINK val（K=2，真实照片）/ BLINK val (K=2, real photos) | **55.3%** | 基线总分 39.6；VC 子项 **56.4%**（基线 39.5） | | 合成遮挡·更近 closer（n=300）/ synthetic occlusion closer | **84%** | v3.1 86% | | 合成遮挡 occlude / synthetic occlusion occlude | **68%** | v3.1 67% | | 跨视角 xview_side / cross-view side | **73%** | v3.1 68% | | 遮挡 2 跳 occlude_2hop / occlusion 2-hop | **45%** | v3.1 52% | | 遮挡 3 跳 occlude_3hop / occlusion 3-hop | **80%** | — | \* stop-at-`\n`（遇换行即停）是 VQA 评测框架的标准做法。本模型的标准协议精确率与 宽松率之间的差几乎全部来自"答案正确但未在换行处停下"（终止边界效应，详见下文）， 换行截停后精确率回到 v2 发布版水平。 \* stop-at-newline is standard practice in VQA evaluation harnesses. The gap between our exact and soft scores under the standard protocol comes almost entirely from "correct answer, but generation did not stop at the newline" (a termination-boundary effect, see below); with newline cut-off the exact rate returns to the v2 release level. ## 循环：v4.1 判决与"按任务选 K" | The loop: the v4.1 verdict and per-task K v4.1 完全体上（n=200，贪心，同协议头对头）：K=2 循环 **39.0%** vs K=1 **38.5%**—— 循环在数学上**不再倒挂**（此前所有非联合训练的形态 K=2 都低于 K=1：嫁接版 31.0、 Plan C 21.5）。数学返利依赖与循环共同训练的注意力/MLP 权重，这是完全体与嫁接版的 本质差别。 On the v4.1 full model (n=200, greedy, head-to-head): K=2 loop **39.0%** vs K=1 **38.5%** — the loop is **no longer a liability** on math (in every non-jointly-trained form K=2 scored below K=1: graft 31.0, Plan C 21.5). The math dividend depends on attention/MLP weights co-trained with the loop — the essential difference between the full model and the graft. 历史峰值参考：训练腿 loop2p3 曾达 K=2 **43.0%**（n=400 配对 McNemar p=0.001，但其 照片域终止崩盘不可发布）。v4.1 的 K=2 增益（+0.5）尚属噪声级，循环的真正价值在 空间反馈闭环（中区槽位跨轮精化），不在数学。 Historical reference: the training leg loop2p3 once reached K=2 **43.0%** (n=400 paired, McNemar p=0.001) but was unreleasable due to photo-domain termination collapse. The v4.1 K=2 gain (+0.5) is still noise-level; the loop's real value lies in the spatial feedback loop (cross-round refinement of mid-region slots), not math. 结论仍是**按任务选 K**：K=2 在照片 OCR 类提取任务上仍收税（标准协议 45.7% vs K=1 58.4%；stop-at-`\n` 协议 58.6% vs 63.7%——overthinking 机制），K=1 是零税默认。 The conclusion remains **choose K per task**: K=2 still taxes photo-OCR-style extraction (45.7% vs 58.4% standard protocol; 58.6% vs 63.7% with stop-at-newline — an overthinking mechanism). K=1 is the zero-tax default. ## 已知限制 | Known limitations - 0.8B 研究型模型；空间模块在合成场景上训练。 - 0.8B research model; spatial module trained on synthetic scenes. - 标准协议（不截停）下 TextVQA 精确率 58.4%，低于 v2 纯血版（63.3%）；差距几乎 全部来自终止边界效应（答案正确但未在换行处停下），stop-at-`\n` 协议下为 63.7% （v2 为 63.3%）。语言评测 loss 略高（1.931 vs 1.776）。 - Under the standard protocol (no cut-off) TextVQA exact is 58.4%, below the v2 purebred release (63.3%); the gap is almost entirely a termination-boundary effect (correct answers that do not stop at the newline). Under the stop-at-newline protocol it is 63.7% (v2: 63.3%). Language eval loss is slightly higher (1.931 vs 1.776). - K=2 循环在照片 OCR 类任务上仍收税（见上节）；K=1 为零税默认。 - The K=2 loop still taxes photo-OCR-style tasks (see above); K=1 is the zero-tax default. - 合成榜 distance 子项较 v3.1 有回落（45% vs 61%）。 - The synthetic distance subtask regressed versus v3.1 (45% vs 61%). - 思考模式（`enable_thinking=True`）：0.8B 基底从未学会闭合 `` 块 （已验证：原版基底行为完全相同）。请使用默认的非思考模式。 - Thinking mode (`enable_thinking=True`): the 0.8B base never learned to close `` blocks (verified: original base behaves identically). Use the default non-thinking mode. - 槽位预算 32 个；更密集的场景未测试。K>2 未测试。 - Slot budget is 32; denser scenes were not tested. K>2 was not tested. **v4.0（嫁接版）历史记录 | v4.0 (grafted) historical record**：TextVQA 60.0/73.2 （标准协议；stop-at-`\n` 协议 63.3/73.4）、GSM8K K=1/K=2 = 36.5/31.0（循环增益 未穿过嫁接）、属性任务 40/40、遮挡 2 跳 45%。嫁接版权重仍在审计仓库留档。 TextVQA 60.0/73.2 (standard; 63.3/73.4 stop-at-newline), GSM8K K=1/K=2 = 36.5/31.0 (loop dividend did not transfer through the graft), attribute task 40/40, occlusion 2-hop 45%. The grafted weights remain archived in the audit repository. ## 相关仓库 | Related repositories - [tysgydk/world3d](https://modelscope.cn/datasets/tysgydk/world3d) — world3d 真 3D 空间推理数据集（本模型空间模块的训练集）/ the true-3D spatial reasoning dataset this model's spatial module was trained on - [tysgydk/Qwen3.5-0.8B-GDN2](https://modelscope.cn/models/tysgydk/Qwen3.5-0.8B-GDN2) — v2：GDN2 改造主干 / v2: the GDN2 conversion backbone - [tysgydk/Qwen3.5-0.8B-GDN2-Spatial](https://modelscope.cn/models/tysgydk/Qwen3.5-0.8B-GDN2-Spatial) — v3.1：只读空间模块 / v3.1: the read-only spatial module - [tysgydk/gdn2-audit](https://modelscope.cn/models/tysgydk/gdn2-audit) — 全部中间权重与评测的审计归档 / audit archive of all intermediate checkpoints and evals ## 引用 | Citation 基于千问 3.5（Apache-2.0）和 FLA 的 GatedDeltaNet2 构建，循环机制受 Ouro 启发。 使用本权重请引用千问、FLA/GDN 与 Ouro 的原始论文。 Built on Qwen3.5 (Apache-2.0) and FLA's GatedDeltaNet2, with the loop mechanism inspired by Ouro. Please cite the original Qwen, FLA/GDN, and Ouro papers if you use this checkpoint.