Super-squash branch 'main' using huggingface_hub

.gitattributes

@@ -0,0 +1,37 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text
+model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text

.ipynb_checkpoints/config-checkpoint.json

@@ -0,0 +1,294 @@
+{
+  "add_full_attention_sink_bias": false,
+  "add_swa_attention_sink_bias": true,
+  "architectures": [
+    "MiMoV2FlashForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_chunk_size": 128,
+  "attention_dropout": 0.0,
+  "attention_value_scale": 0.707,
+  "auto_map": {
+    "AutoConfig": "configuration_mimo_v2_flash.MiMoV2FlashConfig",
+    "AutoModel": "modeling_mimo_v2_flash.MiMoV2Model",
+    "AutoModelForCausalLM": "modeling_mimo_v2_flash.MiMoV2FlashForCausalLM"
+  },
+  "dtype": "bfloat16",
+  "head_dim": 192,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "hybrid_block_size": null,
+  "hybrid_layer_pattern": [
+    0,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0
+  ],
+  "initializer_range": 0.02,
+  "intermediate_size": 16384,
+  "layer_types": [
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "layernorm_epsilon": 1e-05,
+  "max_position_embeddings": 262144,
+  "moe_intermediate_size": 2048,
+  "moe_layer_freq": [
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1
+  ],
+  "n_group": 1,
+  "n_routed_experts": 256,
+  "n_shared_experts": null,
+  "norm_topk_prob": true,
+  "num_attention_heads": 64,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 48,
+  "num_key_value_heads": 4,
+  "partial_rotary_factor": 0.334,
+  "quantization_config": {
+    "config_groups": {
+      "group_0": {
+        "format": "pack-quantized",
+        "input_activations": null,
+        "output_activations": null,
+        "targets": [
+          "Linear"
+        ],
+        "weights": {
+          "actorder": null,
+          "block_structure": null,
+          "dynamic": false,
+          "group_size": 32,
+          "num_bits": 4,
+          "observer": "minmax",
+          "observer_kwargs": {},
+          "strategy": "group",
+          "symmetric": true,
+          "type": "int"
+        }
+      }
+    },
+    "format": "pack-quantized",
+    "global_compression_ratio": null,
+    "ignore": [
+      "model.layers.0.self_attn.q_proj",
+      "model.layers.0.self_attn.k_proj",
+      "model.layers.0.self_attn.v_proj",
+      "model.layers.0.self_attn.o_proj",
+      "model.layers.0.self_attn.qkv_proj",
+      "model.layers.0.mlp.gate_up_proj",
+      "model.layers.0.mlp.gate_proj",
+      "model.layers.0.mlp.up_proj",
+      "model.layers.0.mlp.down_proj",
+      "model.layers.1.self_attn.o_proj",
+      "model.layers.2.self_attn.o_proj",
+      "model.layers.3.self_attn.o_proj",
+      "model.layers.4.self_attn.o_proj",
+      "model.layers.5.self_attn.o_proj",
+      "model.layers.6.self_attn.o_proj",
+      "model.layers.7.self_attn.o_proj",
+      "model.layers.8.self_attn.o_proj",
+      "model.layers.9.self_attn.o_proj",
+      "model.layers.10.self_attn.o_proj",
+      "model.layers.11.self_attn.o_proj",
+      "model.layers.12.self_attn.o_proj",
+      "model.layers.13.self_attn.o_proj",
+      "model.layers.14.self_attn.o_proj",
+      "model.layers.15.self_attn.o_proj",
+      "model.layers.16.self_attn.o_proj",
+      "model.layers.17.self_attn.o_proj",
+      "model.layers.18.self_attn.o_proj",
+      "model.layers.19.self_attn.o_proj",
+      "model.layers.20.self_attn.o_proj",
+      "model.layers.21.self_attn.o_proj",
+      "model.layers.22.self_attn.o_proj",
+      "model.layers.23.self_attn.o_proj",
+      "model.layers.24.self_attn.o_proj",
+      "model.layers.25.self_attn.o_proj",
+      "model.layers.26.self_attn.o_proj",
+      "model.layers.27.self_attn.o_proj",
+      "model.layers.28.self_attn.o_proj",
+      "model.layers.29.self_attn.o_proj",
+      "model.layers.30.self_attn.o_proj",
+      "model.layers.31.self_attn.o_proj",
+      "model.layers.32.self_attn.o_proj",
+      "model.layers.33.self_attn.o_proj",
+      "model.layers.34.self_attn.o_proj",
+      "model.layers.35.self_attn.o_proj",
+      "model.layers.36.self_attn.o_proj",
+      "model.layers.37.self_attn.o_proj",
+      "model.layers.38.self_attn.o_proj",
+      "model.layers.39.self_attn.o_proj",
+      "model.layers.40.self_attn.o_proj",
+      "model.layers.41.self_attn.o_proj",
+      "model.layers.42.self_attn.o_proj",
+      "model.layers.43.self_attn.o_proj",
+      "model.layers.44.self_attn.o_proj",
+      "model.layers.45.self_attn.o_proj",
+      "model.layers.46.self_attn.o_proj",
+      "model.layers.47.self_attn.o_proj",
+      "lm_head"
+    ],
+    "kv_cache_scheme": null,
+    "quant_method": "compressed-tensors",
+    "quantization_status": "compressed",
+    "sparsity_config": {},
+    "transform_config": {},
+    "version": "0.13.1.a20251215"
+  },
+  "rope_scaling": null,
+  "rope_theta": 10000,
+  "routed_scaling_factor": null,
+  "scoring_func": "sigmoid",
+  "sliding_window": 128,
+  "sliding_window_size": 128,
+  "swa_head_dim": 192,
+  "swa_num_attention_heads": 64,
+  "swa_num_key_value_heads": 8,
+  "swa_rope_theta": 10000,
+  "swa_v_head_dim": 128,
+  "tie_word_embeddings": false,
+  "topk_group": 1,
+  "topk_method": "noaux_tc",
+  "transformers_version": "4.57.3",
+  "use_cache": true,
+  "v_head_dim": 128,
+  "vocab_size": 152576
+}

README.md

@@ -0,0 +1,372 @@
+---
+license: mit
+library_name: transformers
+base_model: XiaomiMiMo/MiMo-V2-Flash
+---
+
+# MiMo-V2-Flash AWQ - INT4
+
+## Model Details
+
+### Quantization Details
+
+- **Quantization Method:** cyankiwi AWQ v1.0
+- **Bits:** 4
+- **Group Size:** 32
+- **Calibration Dataset:** [nvidia/Llama-Nemotron-Post-Training-Dataset](https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset)
+- **Quantization Tool:** [llm-compressor](https://github.com/vllm-project/llm-compressor)
+
+### Memory Usage
+
+| **Type** | **MiMo-V2-Flash** | **MiMo-V2-Flash-AWQ-4bit** |
+|:---------------:|:----------------:|:----------------:|
+| **Memory Size** | 291.6 GB | 166.0 GB | 
+
+## Inference
+
+### Prerequisite
+
+```bash 
+git clone https://github.com/sgl-project/sglang.git
+cd sglang
+git fetch origin pull/15207/head:pr-15207
+git checkout pr-15207
+pip install -e "python"
+```
+
+### Basic Usage
+
+```bash 
+python3 -m sglang.launch_server \
+        --model-path cyankiwi/MiMo-V2-Flash-AWQ-4bit \
+        --trust-remote-code \
+        --reasoning-parser qwen3 \
+        --tool-call-parser mimo
+```
+
+## Additional Information
+
+### Known Issues
+
+- No MTP implementation
+
+### Changelog
+
+- **v0.9.0** - Initial quantized release without MTP implementation
+
+### Authors
+
+- **Name:** Ton Cao
+- **Contacts:** ton@cyan.kiwi
+
+<br/><br/>
+
+<div align="center">
+  <picture>
+    <source srcset="https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/Xiaomi_MiMo_darkmode.png?raw=true" media="(prefers-color-scheme: dark)">
+    <img src="https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/Xiaomi_MiMo.png?raw=true" width="60%" alt="Xiaomi-MiMo" />
+  </picture>
+</div>
+
+<br/>
+
+<div align="center" style="line-height: 1;">
+  |
+  <a href="https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash" target="_blank">🤗 HuggingFace</a>
+  &nbsp;|
+  <a href="https://github.com/XiaomiMiMo/MiMo-V2-Flash/blob/main/paper.pdf" target="_blank">📔 Technical Report </a>
+  &nbsp;|
+  <a href="https://mimo.xiaomi.com/blog/mimo-v2-flash" target="_blank">📰 Blog </a>
+  &nbsp;|
+  <br/><br/>
+  <strong>Play around!</strong> &nbsp;
+  <a href="https://aistudio.xiaomimimo.com" target="_blank">🗨️ Xiaomi MiMo Studio </a>
+  &nbsp;
+  <a href="https://platform.xiaomimimo.com/" target="_blank">🎨 Xiaomi MiMo API Platform </a>
+</div>
+<br/>
+
+# MiMo-V2-Flash
+
+**MiMo-V2-Flash** is a Mixture-of-Experts (MoE) language model with **309B total parameters** and **15B active parameters**. Designed for high-speed reasoning and agentic workflows, it utilizes a novel hybrid attention architecture and Multi-Token Prediction (MTP) to achieve state-of-the-art performance while significantly reducing inference costs.
+
+<p align="center">
+  <img width="80%" src="https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/MiMo-v2-flash-performance.jpg?raw=true">
+</p>
+
+-----
+
+## 1. Introduction
+
+MiMo-V2-Flash creates a new balance between long-context modeling capability and inference efficiency. Key features include:
+
+  * **Hybrid Attention Architecture**: Interleaves Sliding Window Attention (SWA) and Global Attention (GA) with a 5:1 ratio and an aggressive 128-token window. This reduces KV-cache storage by nearly 6x while maintaining long-context performance via learnable **attention sink bias**.
+  * **Multi-Token Prediction (MTP)**: Equipped with a lightweight MTP module (0.33B params/block) using dense FFNs. This triples output speed during inference and will be good to accelerates rollout in RL training.
+  * **Efficient Pre-Training**: Trained on 27T tokens using FP8 mixed precision and native 32k seq length. The context window supports up to 256k length.
+  * **Agentic Capabilities**: Post-training utilizes Multi-Teacher On-Policy Distillation (MOPD) and large-scale agentic RL, achieving superior performance on **SWE-Bench** and complex reasoning tasks.
+
+-----
+
+## 2. Model Downloads
+
+| Model                  | Total Params | Active Params | Context Length |                               Download                                |
+| :--------------------- | :----------: | :-----------: | :------------: | :-------------------------------------------------------------------: |
+| **MiMo-V2-Flash-Base** |     309B     |      15B      |      256k      | [🤗 HuggingFace](https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash-Base) |
+| **MiMo-V2-Flash**      |     309B     |      15B      |      256k      |   [🤗 HuggingFace](https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash)    |
+
+> [!IMPORTANT]
+> We also open-source the 3-layer MTP weights to foster community research.
+
+-----
+
+## 3. Evaluation Results
+
+### Base Model Evaluation
+
+MiMo-V2-Flash-Base demonstrates strong performance across standard benchmarks, surpassing models with significantly larger parameter counts.
+
+| Category         | Benchmark               | Setting/Length | MiMo-V2-Flash Base |  Kimi-K2 Base   | DeepSeek-V3.1 Base | DeepSeek-V3.2 Exp Base |
+| :--------------- | :---------------------- | :------------- | :----------------: | :-------------: | :----------------: | :--------------------: |
+| **Params**       | **#Activated / #Total** | -              |   **15B / 309B**   | **32B / 1043B** |   **37B / 671B**   |     **37B / 671B**     |
+| **General**      | BBH                     | 3-shot         |        88.5        |      88.7       |        88.2        |          88.7          |
+|                  | MMLU                    | 5-shot         |        86.7        |      87.8       |        87.4        |          87.8          |
+|                  | MMLU-Redux              | 5-shot         |        90.6        |      90.2       |        90.0        |          90.4          |
+|                  | MMLU-Pro                | 5-shot         |        73.2        |      69.2       |        58.8        |          62.1          |
+|                  | DROP                    | 3-shot         |        84.7        |      83.6       |        86.3        |          86.6          |
+|                  | ARC-Challenge           | 25-shot        |        95.9        |      96.2       |        95.6        |          95.5          |
+|                  | HellaSwag               | 10-shot        |        88.5        |      94.6       |        89.2        |          89.4          |
+|                  | WinoGrande              | 5-shot         |        83.8        |      85.3       |        85.9        |          85.6          |
+|                  | TriviaQA                | 5-shot         |        80.3        |      85.1       |        83.5        |          83.9          |
+|                  | GPQA-Diamond            | 5-shot         |        55.1        |      48.1       |        51.0        |          52.0          |
+|                  | SuperGPQA               | 5-shot         |        41.1        |      44.7       |        42.3        |          43.6          |
+|                  | SimpleQA                | 5-shot         |        20.6        |      35.3       |        26.3        |          27.0          |
+| **Math**         | GSM8K                   | 8-shot         |        92.3        |      92.1       |        91.4        |          91.1          |
+|                  | MATH                    | 4-shot         |        71.0        |      70.2       |        62.6        |          62.5          |
+|                  | AIME 24&25              | 2-shot         |        35.3        |      31.6       |        21.6        |          24.8          |
+| **Code**         | HumanEval+              | 1-shot         |        70.7        |      84.8       |        64.6        |          67.7          |
+|                  | MBPP+                   | 3-shot         |        71.4        |      73.8       |        72.2        |          69.8          |
+|                  | CRUXEval-I              | 1-shot         |        67.5        |      74.0       |        62.1        |          63.9          |
+|                  | CRUXEval-O              | 1-shot         |        79.1        |      83.5       |        76.4        |          74.9          |
+|                  | MultiPL-E HumanEval     | 0-shot         |        59.5        |      60.5       |        45.9        |          45.7          |
+|                  | MultiPL-E MBPP          | 0-shot         |        56.7        |      58.8       |        52.5        |          50.6          |
+|                  | BigCodeBench            | 0-shot         |        70.1        |      61.7       |        63.0        |          62.9          |
+|                  | LiveCodeBench v6        | 1-shot         |        30.8        |      26.3       |        24.8        |          24.9          |
+|                  | SWE-Bench (AgentLess)   | 3-shot         |        30.8        |      28.2       |        24.8        |          9.4*          |
+| **Chinese**      | C-Eval                  | 5-shot         |        87.9        |      92.5       |        90.0        |          91.0          |
+|                  | CMMLU                   | 5-shot         |        87.4        |      90.9       |        88.8        |          88.9          |
+|                  | C-SimpleQA              | 5-shot         |        61.5        |      77.6       |        70.9        |          68.0          |
+| **Multilingual** | GlobalMMLU              | 5-shot         |        76.6        |      80.7       |        81.9        |          82.0          |
+|                  | INCLUDE                 | 5-shot         |        71.4        |      75.3       |        77.2        |          77.2          |
+| **Long Context** | NIAH-Multi              | 32K            |        99.3        |      99.8       |        99.7        |         85.6*          |
+|                  |                         | 64K            |        99.9        |      100.0      |        98.6        |         85.9*          |
+|                  |                         | 128K           |        98.6        |      99.5       |        97.2        |         94.3*          |
+|                  |                         | 256K           |        96.7        |        -        |         -          |           -            |
+|                  | GSM-Infinite Hard       | 16K            |        37.7        |      34.6       |        41.5        |          50.4          |
+|                  |                         | 32K            |        33.7        |      26.1       |        38.8        |          45.2          |
+|                  |                         | 64K            |        31.5        |      16.0       |        34.7        |          32.6          |
+|                  |                         | 128K           |        29.0        |       8.8       |        28.7        |          25.7          |
+
+> \* indicates the model may fail to follow the prompt or format.
+
+### Post-training Model Evaluation
+
+Following our Post-Training Paradigm with MOPD and Agentic RL, the model achieves SOTA reasoning and agentic performance.
+
+
+
+| Benchmark                      | MiMo-V2 Flash | Kimi-K2 Thinking | DeepSeek-V3.2 Thinking | Gemini-3.0 Pro | Claude Sonnet 4.5 | GPT-5 High |
+| :----------------------------- | :-----------: | :--------------: | :--------------------: | :------------: | :---------------: | :--------: |
+| **Reasoning**                  |               |                  |                        |                |                   |            |
+| MMLU-Pro                       |     84.9      |       84.6       |          85.0          |      90.1      |       88.2        |    87.5    |
+| GPQA-Diamond                   |     83.7      |       84.5       |          82.4          |      91.9      |       83.4        |    85.7    |
+| HLE (no tools)                 |     22.1      |       23.9       |          25.1          |      37.5      |       13.7        |    26.3    |
+| AIME 2025                      |     94.1      |       94.5       |          93.1          |      95.0      |       87.0        |    94.6    |
+| HMMT Feb. 2025                 |     84.4      |       89.4       |          92.5          |      97.5      |       79.2        |    88.3    |
+| LiveCodeBench-v6               |     80.6      |       83.1       |          83.3          |      90.7      |       64.0        |    84.5    |
+| **General Writing**            |               |                  |                        |                |                   |            |
+| Arena-Hard (Hard Prompt)       |     54.1      |       71.9       |          53.4          |      72.6      |       63.3        |    71.9    |
+| Arena-Hard (Creative Writing)  |     86.2      |       80.1       |          88.8          |      93.6      |       76.7        |    92.2    |
+| **Long Context**               |               |                  |                        |                |                   |            |
+| LongBench V2                   |     60.6      |       45.1       |          58.4          |      65.6      |       61.8        |     -      |
+| MRCR                           |     45.7      |       44.2       |          55.5          |      89.7      |       55.4        |     -      |
+| **Code Agent**                 |               |                  |                        |                |                   |            |
+| SWE-Bench Verified             |     73.4      |       71.3       |          73.1          |      76.2      |       77.2        |    74.9    |
+| SWE-Bench Multilingual         |     71.7      |       61.1       |          70.2          |       -        |       68.0        |    55.3    |
+| Terminal-Bench Hard            |     30.5      |       30.6       |          35.4          |      39.0      |       33.3        |    30.5    |
+| Terminal-Bench 2.0             |     38.5      |       35.7       |          46.4          |      54.2      |       42.8        |    35.2    |
+| **General Agent**              |               |                  |                        |                |                   |            |
+| BrowseComp                     |     45.4      |        -         |          51.4          |       -        |       24.1        |    54.9    |
+| BrowseComp (w/ Context Manage) |     58.3      |       60.2       |          67.6          |      59.2      |         -         |     -      |
+| \\(\tau^2\\)-Bench                 |     80.3      |       74.3       |          80.3          |      85.4      |       84.7        |    80.2    |
+
+-----
+
+## 4. Model Architecture
+
+<p align="center">
+  <img width="80%" src="https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/MiMo-v2-flash-arch.png?raw=true">
+</p>
+
+### Hybrid Sliding Window Attention
+
+MiMo-V2-Flash addresses the quadratic complexity of long contexts by interleaving Local Sliding Window Attention (SWA) and Global Attention (GA).
+
+  * **Configuration**: Stacks of \\(M=8\\) hybrid blocks. Each block contains \\(N=5\\) SWA layers followed by 1 GA layer.
+  * **Efficiency**: SWA layers use a window size of 128 tokens, reducing KV cache significantly.
+  * **Sink Bias**: Learnable attention sink bias is applied to maintain performance despite the aggressive window size.
+
+### Lightweight Multi-Token Prediction (MTP)
+
+Unlike traditional speculative decoding, our MTP module is natively integrated for training and inference.
+
+  * **Structure**: Uses a dense FFN (instead of MoE) and SWA (instead of GA) to keep the parameter count low (0.33B per block).
+  * **Performance**: Facilitates self-speculative decoding, tripling generation speed and mitigating GPU idleness during small-batch RL training.
+
+-----
+
+## 5. Post-Training Technical Highlights
+
+MiMo-V2-Flash leverages a post-training pipeline designed to maximize reasoning and agentic capabilities through innovative distillation and reinforcement learning strategies.
+
+### 5.1 Multi-Teacher On-Policy Distillation (MOPD)
+
+We introduce **Multi-Teacher On-Policy Distillation (MOPD)**, a new paradigm that formulates knowledge distillation as a reinforcement learning process.
+* **Dense Token-Level Guidance**: Unlike methods relying on sparse sequence-level feedback, MOPD utilizes domain-specific expert models (teachers) to provide supervision at every token position.
+* **On-Policy Optimization**: The student model learns from its own generated responses rather than a fixed dataset. This eliminates exposure bias and ensures smaller, more stable gradient updates.
+* **Inherent Reward Robustness**: Rewards are derived from the distribution divergence between student and teacher, making the process naturally resistant to reward hacking.
+
+### 5.2 Scaling Agentic RL
+
+We significantly scale up the agentic training environments to improve intelligence and generalization.
+* **Massive Code Agent Environments**: We utilize real-world GitHub issues to create over 100,000 verifiable tasks. Our automated pipeline maintains a Kubernetes cluster capable of running over 10,000 concurrent pods with a 70% environment setup success rate.
+* **Multimodal Verifier for WebDev**: For web development tasks, we employ a vision-based verifier that evaluates code execution via recorded videos rather than static screenshots. This reduces visual hallucination and ensures functional correctness.
+* **Cross-Domain Generalization**: Our experiments show that large-scale RL training on code agents effectively generalizes to other domains, boosting performance in Math and General Agent tasks.
+
+### 5.3 Advanced RL Infrastructure
+
+To support high-throughput RL training for large-scale MoE models, we implemented several infrastructure optimizations on top of SGLang and Megatron-LM.
+* **Rollout Routing Replay (R3)**: Addresses numerical precision inconsistencies in MoE routing between inference and training. R3 reuses the exact routed experts from rollout during the training pass, ensuring consistency with negligible overhead.
+* **Request-Level Prefix Cache**: In multi-turn agent training, this cache stores KV states and routed experts from prior turns. It avoids re-computation and ensures sampling consistency across turns.
+* **Fine-Grained Data Scheduler**: We extend the rollout engine to schedule fine-grained sequences instead of micro-batches. Combined with partial rollout, this significantly reduces GPU idleness caused by long-tail stragglers.
+* **Toolbox & Tool Manager**: A two-layer design using Ray actor pools to handle resource contention. It eliminates cold-start delays for tool execution and isolates task logic from system policies.
+
+-----
+
+## 6. Inference & Deployment
+
+MiMo-V2-Flash supports FP8 mixed precision inference. We recommend using **SGLang** for optimal performance.
+
+### Quick Start with SGLang
+
+```bash
+pip install sglang
+
+# Launch server
+python3 -m sglang.launch_server \
+        --model-path XiaomiMiMo/MiMo-V2-Flash \
+        --served-model-name mimo-v2-flash \
+        --pp-size 1 \
+        --dp-size 2 \
+        --enable-dp-attention \
+        --tp-size 8 \
+        --moe-a2a-backend deepep \
+        --page-size 1 \
+        --host 0.0.0.0 \
+        --port 9001 \
+        --trust-remote-code \
+        --mem-fraction-static 0.75 \
+        --max-running-requests 128 \
+        --chunked-prefill-size 16384 \
+        --reasoning-parser qwen3 \
+        --tool-call-parser mimo \
+        --context-length 262144 \
+        --attention-backend fa3 \
+        --speculative-algorithm EAGLE \
+        --speculative-num-steps 3 \
+        --speculative-eagle-topk 1 \
+        --speculative-num-draft-tokens 4 \
+        --enable-mtp
+
+# Send request
+curl -i http://localhost:9001/v1/chat/completions \
+    -H 'Content-Type:application/json' \
+    -d  '{
+            "messages" : [{
+                "role": "user",
+                "content": "Nice to meet you MiMo"
+            }],
+            "model": "mimo-v2-flash",
+            "max_tokens": 4096,
+            "temperature": 0.8,
+            "top_p": 0.95,
+            "stream": true,
+            "chat_template_kwargs": {
+                "enable_thinking": true
+            }
+        }'
+```
+
+### Notifications
+
+#### 1. System prompt
+
+> [!IMPORTANT]
+> The following system prompts are **HIGHLY** recommended, please choose from English and Chinese version.
+
+English
+
+```plaintext
+You are MiMo, an AI assistant developed by Xiaomi.
+
+Today's date: {date} {week}. Your knowledge cutoff date is December 2024.
+```
+
+Chinese
+
+```plaintext
+你是MiMo（中文名称也是MiMo），是小米公司研发的AI智能助手。
+
+今天的日期：{date} {week}，你的知识截止日期是2024年12月。
+```
+
+#### 2. Sampling parameters
+
+> [!IMPORTANT]
+> Recommended sampling parameters:
+>
+> `top_p=0.95`
+> 
+> `temperature=0.8` for math, writing, web-dev
+> 
+> `temperature=0.3` for agentic taks (e.g., vibe-coding, tool-use)
+
+#### 3. Tool-use practice
+
+> [!IMPORTANT]
+> In the thinking mode with multi-turn tool calls, the model returns a `reasoning_content` field alongside `tool_calls`. To continue the conversation, the user must persist all history `reasoning_content` in the `messages` array of each subsequent request.
+
+-----
+
+## 7. Citation
+
+If you find our work helpful, please cite our technical report:
+
+```bibtex
+@misc{mimo2025flash,
+  title={MiMo-V2-Flash Technical Report},
+  author={LLM-Core Xiaomi},
+  year={2025},
+  url={https://github.com/XiaomiMiMo/MiMo-V2-Flash/paper.pdf}
+}
+```
+
+## 8. Contact
+
+Please contact us at [mimo@xiaomi.com](mailto:mimo@xiaomi.com), join our WeChat group below or open an issue if you have any questions.
+
+<p align="center">
+  <img src="https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/wechat_group/wechat1.jpg?raw=true" width="20%" />
+  <img src="https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/wechat_group/wechat2.jpg?raw=true" width="20%" />
+  <img src="https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/wechat_group/wechat3.jpg?raw=true" width="20%" />
+  <img src="https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/wechat_group/wechat4.jpg?raw=true" width="20%" />
+</p>

added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

chat_template.jinja

@@ -0,0 +1,143 @@
+{%- if not add_generation_prompt is defined -%}
+    {%- set add_generation_prompt = false -%}
+{%- endif -%}
+{%- if not enable_thinking is defined -%}
+    {%- set enable_thinking = false -%}
+{%- endif -%}
+{%- if not keep_all_reasoning is defined -%}
+    {%- set keep_all_reasoning = false -%}
+{%- endif -%}
+{%- macro render_extra_keys(json_dict, handled_keys) -%}
+    {%- if json_dict is mapping %}
+        {%- for json_key in json_dict if json_key not in handled_keys %}
+            {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
+                {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
+            {%- else %}
+                {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+{%- endmacro -%}
+{%- if messages[0]["role"] == "system" %}
+    {%- set system_message = messages[0]["content"] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{%- set ns = namespace(last_user_index=-1) %}
+{%- for m in loop_messages %}
+    {%- if m.role == 'user' %}
+        {%- set ns.last_user_index = loop.index0 -%}
+    {%- endif %}
+{%- endfor %}
+{%- if not tools is defined %}
+    {%- set tools = [] %}
+{%- endif %}
+{%- if system_message is defined %}
+    {{- "<|im_start|>system\n" + system_message }}
+{%- else %}
+    {{- "<|im_start|>system\nYou are MiMo, a helpful AI assistant engineered by Xiaomi." }}
+{%- endif %}
+{%- if tools is iterable and tools | length > 0 %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou have access to the following functions:\n\n" }}
+    {{- "<tools>" }}
+    {%- for tool in tools %}
+        {%- if tool.function is defined %}
+            {%- set tool = tool.function %}
+        {%- endif %}
+        {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
+        {%- if tool.description is defined %}
+            {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
+        {%- endif %}
+        {{- '\n<parameters>' }}
+        {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
+            {%- for param_name, param_fields in tool.parameters.properties|items %}
+                {{- '\n<parameter>' }}
+                {{- '\n<name>' ~ param_name ~ '</name>' }}
+                {%- if param_fields.type is defined %}
+                    {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
+                {%- endif %}
+                {%- if param_fields.description is defined %}
+                    {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
+                {%- endif %}
+                {%- set handled_keys = ['name', 'type', 'description'] %}
+                {{- render_extra_keys(param_fields, handled_keys) }}
+                {{- '\n</parameter>' }}
+            {%- endfor %}
+        {%- endif %}
+        {%- set handled_keys = ['type', 'properties'] %}
+        {{- render_extra_keys(tool.parameters, handled_keys) }}
+        {{- '\n</parameters>' }}
+        {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
+        {{- render_extra_keys(tool, handled_keys) }}
+        {{- '\n</function>' }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nFor each function call, output the function name and arguments in the following format:\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>value_1</parameter>\n<parameter=example_parameter_2>This is the value for the second parameter\nthat can span\nmultiple lines</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- DO NOT use function calls inside <think></think> tags.\n- The value enclosed between parameter tags is preserved exactly as-is, including newlines and spaces.\n</IMPORTANT>' }}
+{%- endif %}
+{{- '<|im_end|>' }}
+{%- for message in loop_messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if message.role == "assistant" %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- set reasoning_content = '' %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].split('<think>')[-1] %}
+                {%- set content = content.split('</think>')[-1] %}
+            {%- endif %}
+        {%- endif %}
+        {%- if (keep_all_reasoning or loop.index0 > ns.last_user_index) and reasoning_content -%}
+            {{- '<|im_start|>' + message.role + '\n<think>' + reasoning_content + '</think>' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n<think></think>' + content }}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>' }}
+    {%- elif message.role == "user" or message.role == "system"%}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>tool\n' }}
+        {%- endif %}
+        {{- '<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>\n' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>' }}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if not enable_thinking -%}
+        {{- '<think></think>' -}}
+    {%- else -%}
+        {{- '' -}}
+    {%- endif -%}
+{%- endif %}

config.json

@@ -0,0 +1,294 @@
+{
+  "add_full_attention_sink_bias": false,
+  "add_swa_attention_sink_bias": true,
+  "architectures": [
+    "MiMoV2FlashForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_chunk_size": 128,
+  "attention_dropout": 0.0,
+  "attention_value_scale": 0.707,
+  "auto_map": {
+    "AutoConfig": "configuration_mimo_v2_flash.MiMoV2FlashConfig",
+    "AutoModel": "modeling_mimo_v2_flash.MiMoV2Model",
+    "AutoModelForCausalLM": "modeling_mimo_v2_flash.MiMoV2FlashForCausalLM"
+  },
+  "dtype": "bfloat16",
+  "head_dim": 192,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "hybrid_block_size": null,
+  "hybrid_layer_pattern": [
+    0,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0
+  ],
+  "initializer_range": 0.02,
+  "intermediate_size": 16384,
+  "layer_types": [
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "layernorm_epsilon": 1e-05,
+  "max_position_embeddings": 262144,
+  "moe_intermediate_size": 2048,
+  "moe_layer_freq": [
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1
+  ],
+  "n_group": 1,
+  "n_routed_experts": 256,
+  "n_shared_experts": null,
+  "norm_topk_prob": true,
+  "num_attention_heads": 64,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 48,
+  "num_key_value_heads": 4,
+  "partial_rotary_factor": 0.334,
+  "quantization_config": {
+    "config_groups": {
+      "group_0": {
+        "format": "pack-quantized",
+        "input_activations": null,
+        "output_activations": null,
+        "targets": [
+          "Linear"
+        ],
+        "weights": {
+          "actorder": null,
+          "block_structure": null,
+          "dynamic": false,
+          "group_size": 32,
+          "num_bits": 4,
+          "observer": "minmax",
+          "observer_kwargs": {},
+          "strategy": "group",
+          "symmetric": true,
+          "type": "int"
+        }
+      }
+    },
+    "format": "pack-quantized",
+    "global_compression_ratio": null,
+    "ignore": [
+      "model.layers.0.self_attn.q_proj",
+      "model.layers.0.self_attn.k_proj",
+      "model.layers.0.self_attn.v_proj",
+      "model.layers.0.self_attn.o_proj",
+      "model.layers.0.self_attn.qkv_proj",
+      "model.layers.0.mlp.gate_up_proj",
+      "model.layers.0.mlp.gate_proj",
+      "model.layers.0.mlp.up_proj",
+      "model.layers.0.mlp.down_proj",
+      "model.layers.1.self_attn.o_proj",
+      "model.layers.2.self_attn.o_proj",
+      "model.layers.3.self_attn.o_proj",
+      "model.layers.4.self_attn.o_proj",
+      "model.layers.5.self_attn.o_proj",
+      "model.layers.6.self_attn.o_proj",
+      "model.layers.7.self_attn.o_proj",
+      "model.layers.8.self_attn.o_proj",
+      "model.layers.9.self_attn.o_proj",
+      "model.layers.10.self_attn.o_proj",
+      "model.layers.11.self_attn.o_proj",
+      "model.layers.12.self_attn.o_proj",
+      "model.layers.13.self_attn.o_proj",
+      "model.layers.14.self_attn.o_proj",
+      "model.layers.15.self_attn.o_proj",
+      "model.layers.16.self_attn.o_proj",
+      "model.layers.17.self_attn.o_proj",
+      "model.layers.18.self_attn.o_proj",
+      "model.layers.19.self_attn.o_proj",
+      "model.layers.20.self_attn.o_proj",
+      "model.layers.21.self_attn.o_proj",
+      "model.layers.22.self_attn.o_proj",
+      "model.layers.23.self_attn.o_proj",
+      "model.layers.24.self_attn.o_proj",
+      "model.layers.25.self_attn.o_proj",
+      "model.layers.26.self_attn.o_proj",
+      "model.layers.27.self_attn.o_proj",
+      "model.layers.28.self_attn.o_proj",
+      "model.layers.29.self_attn.o_proj",
+      "model.layers.30.self_attn.o_proj",
+      "model.layers.31.self_attn.o_proj",
+      "model.layers.32.self_attn.o_proj",
+      "model.layers.33.self_attn.o_proj",
+      "model.layers.34.self_attn.o_proj",
+      "model.layers.35.self_attn.o_proj",
+      "model.layers.36.self_attn.o_proj",
+      "model.layers.37.self_attn.o_proj",
+      "model.layers.38.self_attn.o_proj",
+      "model.layers.39.self_attn.o_proj",
+      "model.layers.40.self_attn.o_proj",
+      "model.layers.41.self_attn.o_proj",
+      "model.layers.42.self_attn.o_proj",
+      "model.layers.43.self_attn.o_proj",
+      "model.layers.44.self_attn.o_proj",
+      "model.layers.45.self_attn.o_proj",
+      "model.layers.46.self_attn.o_proj",
+      "model.layers.47.self_attn.o_proj",
+      "lm_head"
+    ],
+    "kv_cache_scheme": null,
+    "quant_method": "compressed-tensors",
+    "quantization_status": "compressed",
+    "sparsity_config": {},
+    "transform_config": {},
+    "version": "0.13.1.a20251215"
+  },
+  "rope_scaling": null,
+  "rope_theta": 10000,
+  "routed_scaling_factor": null,
+  "scoring_func": "sigmoid",
+  "sliding_window": 128,
+  "sliding_window_size": 128,
+  "swa_head_dim": 192,
+  "swa_num_attention_heads": 64,
+  "swa_num_key_value_heads": 8,
+  "swa_rope_theta": 10000,
+  "swa_v_head_dim": 128,
+  "tie_word_embeddings": false,
+  "topk_group": 1,
+  "topk_method": "noaux_tc",
+  "transformers_version": "4.57.3",
+  "use_cache": true,
+  "v_head_dim": 128,
+  "vocab_size": 152576
+}

configuration_mimo_v2_flash.py

@@ -0,0 +1,109 @@
+# coding=utf-8
+#
+# Copyright 2025 Xiaomi Corporation.
+# Copyright 2025 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class MiMoV2FlashConfig(PretrainedConfig):
+
+    model_type = ""
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `Hybrid`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    attribute_map = {
+        "num_local_experts": "n_routed_experts",
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        layernorm_epsilon=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_dropout=0.0,
+        hybrid_block_size=None,
+        hybrid_layer_pattern=None,
+        partial_rotary_factor=1.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.layernorm_epsilon = layernorm_epsilon
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_dropout = attention_dropout
+
+        if hybrid_block_size is not None and hybrid_layer_pattern is None:
+            hybrid_layer_pattern = [0 if ((i + 1) % hybrid_block_size == 0) else 1 for i in range(num_hidden_layers)]
+        self.hybrid_block_size = hybrid_block_size
+        self.hybrid_layer_pattern = hybrid_layer_pattern
+
+        self.partial_rotary_factor = partial_rotary_factor
+
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

generation_config.json

@@ -0,0 +1,4 @@
+{
+  "_from_model_config": true,
+  "transformers_version": "4.57.3"
+}

model-00001-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b72fe3306423aaa2547f1cce1f948e6b24e7952ac264216522d5ec0e5f70afa
+size 4996864944

model-00002-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32370938d270ca3a7b67d7b6d6291123dad139fa820d2203a3082b7b2ce6f0b7
+size 4996631272

model-00003-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c968af56b53a55fe26e58fa3634ab1d7da749efdf85a592e99b8be6777ba88d
+size 4997014192

model-00004-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8f057de876273611e661f2e55a1e07e3cafa5e130d591003e98c4503aac2bff
+size 4998784320

model-00005-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70a8d5b2f650f12b60cbbf86f107bf082779a67d1897247445d7496d2e7235d3
+size 4996631216

model-00006-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfdeb11803545c1197f0d6851843d331e62ee7a57f71d182bb3156f0755e6be0
+size 4997014248

model-00007-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7147a77aebdc931ca77137dc9088f61ff3746216fa902ebc7330e23498d4f36c
+size 4997014704

model-00008-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6113df2d0f84feb76ada86d6fcc59adf267bdff46b6c909c9f1711a5aee6fe63
+size 4998403600

model-00009-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bab63a66c83c6ee2d671b1cb17f0c540fd6477f2a23c77297792bb4541ee42b5
+size 4997017424

model-00010-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2275411b1a4b18220897a1c431cfc498416d6e51b7e5bcd293606e17b75aaaa7
+size 4997017832

model-00011-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5866ef7f23657d59c613d853d09f279a0a9bd65b26a1780937a17a8fac7b114d
+size 4996634192

model-00012-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b5b32971eb1a9ec55ea1ecb0ca168bfbc35a61cf217a0b3abaa51e0f9d794d4
+size 4997017472

model-00013-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fe306dac375b60a856bff52ebfa528c3bd58fd3ce40b8623c70e85342fef22e
+size 4998787448

model-00014-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c1e904229e6ccb1ab1621fb66f6208536dc93dee61c7ecc9d9466f11792b58a
+size 4996634136

model-00015-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7324c758422c30a65323e75981c091bd07b3775e7b6977d453b0b45c32240f42
+size 4997017528

model-00016-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09d5bbb4645ca735ac2cca7e8b3843bb08671e34e7e727e3dc5edc60ad3d475c
+size 4997017832

model-00017-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:994cfabb33ecaacbb9a5683de49ec20b355bd3025e29ba80ba374b4263b86dcd
+size 4998403696

model-00018-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdeeb3c6cb94601501f2ede6a919d66f156a583f9aaa979c3efd381d7ab7e5a1
+size 4997017584

model-00019-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1285eea168d994d02c84a325ec6dbe67b73d269de00981840d2a9370b75ddfeb
+size 4997017832

model-00020-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc5204e061786f03be2ee0f4910599f14370df8e0094943d59c40f06ef814899
+size 4996634032

model-00021-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:622c27bf3422bccd5fbb161064ea764408a149e7362679f1841b84be873f2fe1
+size 4997017632

model-00022-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3ad420476981a8a4b5a8a55d6162c234bd60a9829548aa6a1a30c978815fe06
+size 5000375328

model-00023-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6af977bd10700578417a3e53a475ef4331ae49dca7ceae10012f7c0631c40772
+size 4999765072

model-00024-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a087d3c3f6a369a66e5cf2f8164f557625c5e69da67411b5e83fbe16000a0956
+size 4997017688

model-00025-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14248f62cce116af70f7f0707e803bda83a4b8e4c41fa48ba01aed45aa6db30c
+size 4996634816

model-00026-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d61529ce02c0a5792ad53cf697a4ba3bd8cda22ca484f1817261bbacc589af1
+size 4998786552

model-00027-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb63626f046290f63c83413b5d951767ddc6a383bc98011093c99a30c764db71
+size 4997017744

model-00028-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b70cdecd1507f8fc5bb8814f862182d24d2cc30370bde495271d5cc801e202ab
+size 4996634704

model-00029-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac3b061dfa4404d4328519bb2640a23358e67cb4f723e72e80fdcec704eb7ab9
+size 4997016992

model-00030-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec6643ca0e400ddd1ca6c7b2d74ef3e98b10e2da5e778d9061bc16b54b18229f
+size 4997017792

model-00031-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45369c15e787001f14d96a3aee344e50b6d850c2a15cf7539348752f3bcfa704
+size 4998404240

model-00032-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d8f860620bca4dea514efd03c3602f9cb6ea0f15c55c3bdcf96a8cc3e43709f
+size 4997017040

model-00033-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71279c7fe6aaf06382f0b4933dc542aa04b5ec58c3eb69af416ef3f55f7a26a4
+size 4997017832

model-00034-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33cb0b845bfe841c2d9f33a73aedc1b4b4008a879289231f830f48ea1c116ccb
+size 4996634568

model-00035-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24626b14fd36cffaeb00af7dc8b13fa6a729b7d7698b9c6a26d291c1f7d75434
+size 4998786704

model-00036-of-00036.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25a1daf9240e653e5189c199511a485d613f473e076c8b225610aedaa5762a74
+size 3345154688

model.safetensors.index.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e51e5f33c35c8976f892eb95a9acc66206c7f7ae8244a345c0b950003dde3a3
+size 10570904

modeling_mimo_v2_flash.py

@@ -0,0 +1,663 @@
+# coding=utf-8
+#
+# Copyright 2025 Xiaomi Corporation.
+# Copyright 2025 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from transformers.generation import GenerationMixin
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.integrations import use_kernel_forward_from_hub
+
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+
+from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import (
+    logging,
+)
+
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple
+from .configuration_mimo_v2_flash import MiMoV2FlashConfig
+
+logger = logging.get_logger(__name__)
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    sinks: Optional[torch.Tensor] = None,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    if sinks is not None:
+        sinks = module.attention_sink_bias.reshape(1, -1, 1, 1).expand(query.shape[0], -1, query.shape[-2], -1)
+        attn_weights = torch.cat([attn_weights, sinks], dim=-1)
+
+    attn_weights = attn_weights - attn_weights.max(dim=-1, keepdim=True).values
+    probs = F.softmax(attn_weights, dim=-1, dtype=attn_weights.dtype)
+
+    if sinks is not None:
+        probs = probs[..., :-1]  # we drop the sink here
+
+    attn_weights = nn.functional.dropout(probs, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    return attn_output, attn_weights
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class MiMoV2RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+            MiMoV2RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class MiMoV2MLP(nn.Module):
+    """MiMoV2MLP matching the gate, up, and down projection layers."""
+
+    def __init__(self, config: MiMoV2FlashConfig, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
+        return down_proj
+
+
+class MiMoV2MoEGate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = (
+            config.routed_scaling_factor
+            if config.routed_scaling_factor is not None
+            else 1.0
+        )
+        self.scoring_func = config.scoring_func
+        self.topk_method = config.topk_method
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+
+        # topk selection algorithm
+        self.norm_topk_prob = config.norm_topk_prob
+        self.gating_dim = config.hidden_size
+        self.weight = nn.Parameter(
+            torch.empty((self.n_routed_experts, self.gating_dim))
+        )
+        if self.topk_method == "noaux_tc":
+            self.e_score_correction_bias = nn.Parameter(
+                torch.empty((self.n_routed_experts))
+            )
+
+    def forward(self, hidden_states):
+        bsz, seq_len, h = hidden_states.shape
+        ### compute gating score
+        hidden_states = hidden_states.view(-1, h)
+        logits = F.linear(
+            hidden_states.type(torch.float32), self.weight.type(torch.float32), None
+        )
+        if self.scoring_func == "sigmoid":
+            scores = logits.sigmoid()
+        else:
+            raise NotImplementedError(
+                f"insupportable scoring function for MoE gating: {self.scoring_func}"
+            )
+
+        ### select top-k experts
+        if self.topk_method == "noaux_tc":
+            assert not self.training
+            scores_for_choice = scores.view(bsz * seq_len, -1) + self.e_score_correction_bias.unsqueeze(0)
+            group_scores = (
+                scores_for_choice.view(bsz * seq_len, self.n_group, -1).topk(2, dim=-1)[0].sum(dim = -1)
+            )  # [n, n_group]
+            group_idx = torch.topk(
+                group_scores, k=self.topk_group, dim=-1, sorted=False
+            )[
+                1
+            ]  # [n, top_k_group]
+            group_mask = torch.zeros_like(group_scores)  # [n, n_group]
+            group_mask.scatter_(1, group_idx, 1)  # [n, n_group]
+            score_mask = (
+                group_mask.unsqueeze(-1)
+                .expand(
+                    bsz * seq_len, self.n_group, self.n_routed_experts // self.n_group
+                )
+                .reshape(bsz * seq_len, -1)
+            )  # [n, e]
+            tmp_scores = scores_for_choice.masked_fill(~score_mask.bool(), float("-inf"))  # [n, e]
+            _, topk_idx = torch.topk(
+                tmp_scores, k=self.top_k, dim=-1, sorted=False
+            )
+            topk_weight = scores.gather(1, topk_idx)
+        else:
+            raise NotImplementedError(
+                f"insupportable TopK function for MoE gating: {self.topk_method}"
+            )
+
+        ### norm gate to sum 1
+        if self.top_k > 1 and self.norm_topk_prob:
+            denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weight = topk_weight / denominator
+        topk_weight = topk_weight * self.routed_scaling_factor # must multiply the scaling factor
+
+        return topk_idx, topk_weight
+
+
+class MiMoV2MoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.experts = nn.ModuleList(
+            [
+                MiMoV2MLP(config, intermediate_size=config.moe_intermediate_size)
+                for _ in range(config.n_routed_experts)
+            ]
+        )
+        self.gate = MiMoV2MoEGate(config)
+
+    def moe(self, hidden_states: torch.Tensor, topk_indices: torch.Tensor, topk_weights: torch.Tensor):
+        r"""
+        CALL FOR CONTRIBUTION! I don't have time to optimise this right now, but expert weights need to be fused
+        to not have to do a loop here (deepseek has 256 experts soooo yeah).
+        """
+        final_hidden_states = torch.zeros_like(hidden_states, dtype=topk_weights.dtype)
+        expert_mask = torch.nn.functional.one_hot(topk_indices, num_classes=len(self.experts))
+        expert_mask = expert_mask.permute(2, 0, 1)
+
+        for expert_idx in range(len(self.experts)):
+            expert = self.experts[expert_idx]
+            mask = expert_mask[expert_idx]
+            token_indices, weight_indices = torch.where(mask)
+
+            if token_indices.numel() > 0:
+                expert_weights = topk_weights[token_indices, weight_indices]
+                expert_input = hidden_states[token_indices]
+                expert_output = expert(expert_input)
+                weighted_output = expert_output * expert_weights.unsqueeze(-1)
+                final_hidden_states.index_add_(0, token_indices, weighted_output)
+
+        # in original deepseek, the output of the experts are gathered once we leave this module
+        # thus the moe module is itelsf an IsolatedParallel module
+        # and all expert are "local" meaning we shard but we don't gather
+        return final_hidden_states.type(hidden_states.dtype)
+    
+
+    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        orig_shape = hidden_states.shape
+        topk_indices, topk_weights = self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        hidden_states = self.moe(hidden_states, topk_indices, topk_weights).view(*orig_shape)
+
+        return hidden_states
+
+
+class MiMoV2Attention(nn.Module):
+    """MiMoV2 Global Attention (pattern == 0) and Sliding Window Attention (pattern == 1)."""
+
+    def __init__(self, config: MiMoV2FlashConfig, is_swa: bool, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+
+        if is_swa:
+            self.head_dim = config.swa_head_dim
+            self.v_head_dim = config.swa_v_head_dim
+            self.num_attention_heads = config.swa_num_attention_heads
+            self.num_key_value_heads = config.swa_num_key_value_heads
+        else:
+            self.head_dim = config.head_dim
+            self.v_head_dim = config.v_head_dim
+            self.num_attention_heads = config.num_attention_heads
+            self.num_key_value_heads = config.num_key_value_heads
+
+        self.rope_dim = int(self.head_dim * config.partial_rotary_factor)
+        self.num_key_value_groups = self.num_attention_heads // self.num_key_value_heads
+        self.attention_bias = config.attention_bias
+        self.attention_dropout: float = config.attention_dropout
+        self.scaling = self.head_dim ** -0.5
+
+        # These dimensions are for the attention layers
+        q_hidden_size = self.num_attention_heads * self.head_dim
+        k_hidden_size = self.num_key_value_heads * self.head_dim
+        v_hidden_size = self.num_key_value_heads * self.v_head_dim
+        o_hidden_size = self.num_attention_heads * self.v_head_dim
+
+        self.q_proj = nn.Linear(config.hidden_size, q_hidden_size, bias=self.attention_bias)
+        self.k_proj = nn.Linear(config.hidden_size, k_hidden_size, bias=self.attention_bias)
+        self.v_proj = nn.Linear(config.hidden_size, v_hidden_size, bias=self.attention_bias)
+        self.o_proj = nn.Linear(o_hidden_size, config.hidden_size, bias=False)
+
+        self.attention_sink_bias = (
+            torch.nn.Parameter(torch.empty(config.num_attention_heads), requires_grad=False)
+            if (config.add_full_attention_sink_bias and not is_swa) or (config.add_swa_attention_sink_bias and is_swa)
+            else None
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        input_shape = hidden_states.shape[:-1]
+        qk_hidden_shape = (*input_shape, -1, self.head_dim)
+        v_hidden_shape = (*input_shape, -1, self.v_head_dim)
+
+        query_states = self.q_proj(hidden_states).view(qk_hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(qk_hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(v_hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+
+        query_rope, query_nope = query_states.split([self.rope_dim, self.head_dim - self.rope_dim], dim=-1)
+        key_rope, key_nope = key_states.split([self.rope_dim, self.head_dim - self.rope_dim], dim=-1)
+
+        query_rope, key_rope = apply_rotary_pos_emb(query_rope, key_rope, cos, sin)
+
+        query_states = torch.cat([query_rope, query_nope], dim=-1)
+        key_states = torch.cat([key_rope, key_nope], dim=-1)
+
+        if past_key_values is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sinks=self.attention_sink_bias,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class MiMoV2DecoderLayer(nn.Module):
+    """
+    MiMoV2 Decoder Layer. It dynamically chooses the correct attention
+    module based on the layer index and the `hybrid_layer_pattern`.
+    """
+
+    def __init__(self, config: MiMoV2FlashConfig, layer_idx: int):
+        super().__init__()
+
+        # This is the key logic: choose the module based on the pattern
+        is_swa_layer = config.hybrid_layer_pattern[layer_idx] == 1
+        if is_swa_layer:
+            self.attention_type = "sliding_window_attention"
+            self.self_attn = MiMoV2Attention(config, True, layer_idx)
+        else:
+            self.attention_type = "full_attention"
+            self.self_attn = MiMoV2Attention(config, False, layer_idx)
+
+        self.mlp = (
+            MiMoV2MoE(config)
+            if (
+                    getattr(config, 'n_routed_experts', None) is not None
+                    and config.moe_layer_freq[layer_idx]
+            )
+            else MiMoV2MLP(config)
+        )
+
+        self.input_layernorm = MiMoV2RMSNorm(config.hidden_size, eps=config.layernorm_epsilon)
+        self.post_attention_layernorm = MiMoV2RMSNorm(config.hidden_size, eps=config.layernorm_epsilon)
+        self.hidden_size = config.hidden_size
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # MLP or MOE
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+class MiMoV2FlashRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: MiMoV2FlashConfig, is_swa, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        if is_swa:
+            self.config.rope_theta = config.swa_rope_theta
+            self.config.head_dim = config.swa_head_dim
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+@auto_docstring
+class MiMoV2Model(PreTrainedModel):
+    """The main 'model' block, corresponding to `model.` in the weight map."""
+    config_class = MiMoV2FlashConfig
+
+    def __init__(self, config: MiMoV2FlashConfig):
+        super().__init__(config)
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList(
+            [MiMoV2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = MiMoV2RMSNorm(config.hidden_size, eps=config.layernorm_epsilon)
+        self.rotary_emb = MiMoV2FlashRotaryEmbedding(config=config, is_swa=False)
+        self.swa_rotary_emb = MiMoV2FlashRotaryEmbedding(config=config, is_swa=True)
+
+        self.has_sliding_layers = any(
+            pattern == 1 for pattern in config.hybrid_layer_pattern
+        )
+
+        # For Huggingface DynamicCache compatibility
+        self.config.layer_types = [
+            "sliding_attention" if config.hybrid_layer_pattern[i] == 1 else "full_attention"
+            for i in range(config.num_hidden_layers)
+        ]
+
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> MoeModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache(config=self.config)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        # It may already have been prepared by e.g. `generate`
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            # Prepare mask arguments
+            mask_kwargs = {
+                "config": self.config,
+                "input_embeds": inputs_embeds,
+                "attention_mask": attention_mask,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "position_ids": position_ids,
+            }
+            # Create the masks
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(**mask_kwargs),
+            }
+            # The sliding window alternating layers are not always activated depending on the config
+            if self.has_sliding_layers:
+                causal_mask_mapping["sliding_window_attention"] = create_sliding_window_causal_mask(**mask_kwargs)
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        swa_position_embeddings = self.swa_rotary_emb(hidden_states, position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask_mapping[decoder_layer.attention_type],
+                position_embeddings=(
+                    position_embeddings 
+                    if decoder_layer.attention_type == "full_attention" 
+                    else swa_position_embeddings
+                ),
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+        )
+
+
+@auto_docstring
+class MiMoV2FlashForCausalLM(PreTrainedModel,GenerationMixin):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    config_class = MiMoV2FlashConfig
+    _keys_to_ignore_on_load_unexpected = [r"model.layers\.\d+\.self_attn\.rotary_emb\.inv_freq"]
+
+    def __init__(self, config: MiMoV2FlashConfig):
+        super().__init__(config)
+        self.model = MiMoV2Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+__all__ = [
+    "MiMoV2FlashForCausalLM"
+]

recipe.yaml

@@ -0,0 +1,35 @@
+default_stage:
+  default_modifiers:
+    AWQModifier:
+      config_groups:
+        group_0:
+          targets: [Linear]
+          weights:
+            num_bits: 4
+            type: int
+            symmetric: true
+            group_size: 32
+            strategy: group
+            block_structure: null
+            dynamic: false
+            actorder: null
+            scale_dtype: null
+            zp_dtype: null
+            observer: minmax
+            observer_kwargs: {}
+          input_activations: null
+          output_activations: null
+          format: null
+      targets: [Linear]
+      ignore: [model.embed_tokens, 're:model[.]layers[.]0.*', 're:.*mlp[.]gate$', 're:.*self_attn[.]o_proj$',
+        model.norm, lm_head]
+      mappings:
+      - smooth_layer: re:.*input_layernorm$
+        balance_layers: ['re:.*q_proj$', 're:.*k_proj$', 're:.*v_proj$']
+      - smooth_layer: re:.*post_attention_layernorm$
+        balance_layers: ['re:.*gate_proj$', 're:.*up_proj$']
+      - smooth_layer: re:.*up_proj$
+        balance_layers: ['re:.*down_proj$']
+      offload_device: !!python/object/apply:torch.device [cpu]
+      duo_scaling: true
+      n_grid: 20

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
+size 11422654