Add MiMo model support

[Belugaaaa]

Before submitting

• Lint code. If there are lint issues, please format the code first.

Basic checks passed locally:

python3 -m py_compile scripts/mimo/*.py paddleformers/transformers/mimo/*.py tests/transformers/mimo/*.py
bash -n scripts/mimo/*.sh tests/integration_test/mimo_sft_single_card.sh

Unit tests passed locally:

CUDA_VISIBLE_DEVICES=0 python -m unittest tests.transformers.mimo.test_modeling -v

Result: Ran 22 tests ... OK (skipped=3).

• Add test cases into tests folder. If there are codecov issues, please add tests cases first.

Added test coverage:

• tests/transformers/mimo/test_modeling.py
• tests/integration_test/mimo_sft_single_card.sh
• tests/config/ci/mimo_sft_single.yaml
• tests/config/benchmark/config/sft/MiMo-7B-Base.yaml
• tests/config/benchmark/config/sft/MiMo-7B-Base-Reduced-Depth-FullWidth.yaml

PR types

New features

PR changes

Models, Docs

Description

This PR adds PaddleFormers support for MiMo, including:

• MiMo model/config implementation.
• AutoConfig and AutoModelForCausalLM registration.
• HF-to-Paddle checkpoint conversion helpers.
• Forward/generation alignment helpers.
• Tiny and reduced-depth full-width asset generation helpers.
• GSM8K SFT configs for full and reduced-depth validation.
• Compiler on/off inference and training benchmark helpers.
• Unit tests and single-card SFT CI smoke config.
• MiMo acceptance and migration status documentation.

Local validation results:

• Full official MiMo FP32 forward/generation alignment passed after converting HF safetensors to Paddle native bf16 checkpoint.
• Full official FP32 logits: max_diff=0.003246307373046875, mean_diff=4.875975355389528e-05.
• Greedy generation first 10 tokens match Transformers.
• Reduced-depth full-width Paddle GSM8K 300-step SFT completed successfully.
• Paddle reduced-depth SFT final eval_loss=2.16945743560791, train_loss=3.152836615641912.
• Reduced-depth inference compiler benchmark passed: dynamic 10840.92 tokens/s, to_static 17253.67 tokens/s, speedup 59.15%.
• LoRA training compiler fallback completed dynamic/static 30-step runs with the same final train_loss=11.528600597381592; dynamic 7906.42 tokens/s, static 8369.32 tokens/s, speedup 5.85%.

Known remaining acceptance items:

• Paddle and ms-swift GSM8K loss curves both decrease, but are not numerically aligned yet. We aligned and ruled out visible differences including system prompt, weight_decay, adam_beta2, scheduler type, fused vs non-fused AdamW, sample shuffle order, MTP vs no-MTP checkpoint structure, and sampled mapped weights.
• Current explanation for the remaining loss gap is framework-level training semantics: ms-swift/HF normalizes accumulated loss by total non-ignored label tokens across the gradient accumulation window, while PaddleFormers computes micro-batch mean loss and averages gradients over gradient_accumulation_steps. On variable-length GSM8K samples this changes per-sample weighting during optimization.
• Full-parameter static training reached the optimizer step but hit local GPU memory pressure while creating optimizer states. The LoRA run is documented as a resource-constrained fallback validation and is not treated as satisfying the formal full-training 20% speedup target.
• CE tiny checkpoint still needs to be uploaded to an approved location, then CE baseline losses/generation tokens can be filled in.

[Paddle-CI-Bot]

PaddleFormers Log Analysis

Run #26796278351 · Attempt 1

日志分析报告

流水线名称	问题标签	修复建议	日志片段
Unittest GPU CI	其他（网络/图片下载失败）	在 video_processing_utils.py::_decode_and_sample_videos 中校验 HTTP 响应状态码，或将测试远程图片 URL 替换为本地 fixture	报错代码

---

失败的测试 case：

FAILED tests/transformers/qwen2_5_vl/test_processor.py::Qwen2_5_VLProcessorTest::test_apply_chat_template_video_frame_sampling

==== 1 failed, 4222 passed, 314 skipped, 15 warnings in 1357.65s (0:22:37) =====

---

根本原因分析：

测试的最后一个分支将视频输入设置为帧图片 URL 列表（"url": ["https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_images/example1.jpg", ...]），由 video_processing_utils.py:286 的 _decode_and_sample_videos 直接透传给上游 transformers.image_processing_base.fetch_images → httpx.get(url).content → PIL.Image.open(BytesIO(content))。CI 环境中 httpx 请求未抛出异常（说明 HTTP 连接本身成功），但返回的内容不是有效图片（可能是 BCEBos 的限速/错误响应页），PIL 无法识别格式，抛出 PIL.UnidentifiedImageError。本 PR 本身（Add MiMo model support）未直接修改 video_processing_utils.py 或 qwen2_5_vl/processor.py，失败点是已有路径因网络抖动暴露出的缺少响应内容校验问题。

---

修复建议：

1. 校验 HTTP 响应内容：在 video_processing_utils.py::_decode_and_sample_videos 中，当 videos 元素为 URL list（帧模式）时，检查 httpx.get() 的 response.status_code == 200 且 len(content) > 0，否则抛出带 URL 信息的明确异常，方便排查。

2. 消除测试的网络依赖：将 test_apply_chat_template_video_frame_sampling 中的 BCEBos 远程帧图片 URL，替换为用 tempfile 写入的本地合法 JPEG/PNG 文件，彻底消除网络抖动引起的偶发失败。

3. 短期 workaround：rerun 该流水线验证是否为偶发网络问题；若稳定复现，则按第 1/2 点修复。

---

🔍 准确性记录：请点击评论底部 😊 图标，选择 👍（准确）或 👎（有误），将自动记录到 CI 监控系统

_{🔄 每次 Re-run 后自动更新}