feat(mimo_v25): support MiMo-V2.5-Pro

docs/fern/versions/nightly.yml

@@ -169,6 +169,8 @@ navigation:
             path: ../../model-coverage/llm/tencent/hy-mt2.mdx
           - page: "MiMo-V2-Flash"
             path: ../../model-coverage/llm/xiaomimimo/mimo-v2-flash.mdx
+          - page: "MiMo-V2.5-Pro"
+            path: ../../model-coverage/llm/xiaomimimo/mimo-v2-5-pro.mdx
           - page: "Ling 2.0"
             path: ../../model-coverage/llm/inclusionai/ling-2.mdx
       - section: "Vision Language Models"

docs/model-coverage/llm/xiaomimimo/mimo-v2-5-pro.mdx

@@ -0,0 +1,93 @@
+---
+title: "MiMo-V2.5-Pro"
+description: ""
+---
+[MiMo-V2.5-Pro](https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro) is Xiaomi's hybrid attention Mixture-of-Experts language model. It alternates full and sliding-window attention layers, uses a `sigmoid_with_bias` router with group-limited expert routing, and ships as an FP8 HF checkpoint.
+
+<Info>
+
+| | |
+|---|---|
+| **Task** | Text Generation (MoE, hybrid attention) |
+| **Architecture** | `MiMoV2ForCausalLM` |
+| **Parameters** | Approx. several hundred B total / much smaller active |
+| **HF Org** | [XiaomiMiMo](https://huggingface.co/XiaomiMiMo) |
+
+</Info>
+
+## Available Models
+
+- **MiMo-V2.5-Pro**: hybrid full/sliding-window attention with FP8 weights.
+
+## Architecture
+
+- `MiMoV2ForCausalLM`
+- Sliding-window attention using the `MiMoV2Attention(is_swa=True)` path.
+- MoE blocks use `nemo_automodel.components.moe.layers.MoE` with `score_func="sigmoid_with_bias"` and `gate_precision=fp32`.
+- FP8 round-trip in `MiMoV2StateDictAdapter` covers the bulk of attention/expert weights; layer norms, the gate, `lm_head`, and `embed_tokens` stay in bf16 per `NON_QUANTIZED_KEY_PATTERNS`.
+
+## Example HF Models
+
+| Model | HF ID |
+|---|---|
+| MiMo-V2.5-Pro | [`XiaomiMiMo/MiMo-V2.5-Pro`](https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro) |
+
+## Example Recipes
+
+| Recipe | Description |
+|---|---|
+| [mimo_v25_pro_hellaswag.yaml](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/llm_finetune/mimo_v25/mimo_v25_pro_hellaswag.yaml) | SFT: MiMo-V2.5-Pro on HellaSwag |
+
+## Try with NeMo AutoModel
+
+**1. Install** ([full instructions](/get-started/installation)):
+
+```bash
+pip install nemo-automodel
+```
+
+**2. Clone the repo** to get the example recipes:
+
+```bash
+git clone https://github.com/NVIDIA-NeMo/Automodel.git
+cd Automodel
+```
+
+**3. Run the recipe** from inside the repo:
+
+```bash
+automodel --nproc-per-node=8 examples/llm_finetune/mimo_v25/mimo_v25_pro_hellaswag.yaml
+```
+
+<Accordion title="Run with Docker">
+**1. Pull the container** and mount a checkpoint directory:
+
+```bash
+docker run --gpus all -it --rm \
+  --shm-size=8g \
+  -v $(pwd)/checkpoints:/opt/Automodel/checkpoints \
+  nvcr.io/nvidia/nemo-automodel:26.02.00
+```
+
+**2. Navigate to the AutoModel directory**:
+
+```bash
+cd /opt/Automodel
+```
+
+**3. Run the recipe**:
+
+```bash
+automodel --nproc-per-node=8 examples/llm_finetune/mimo_v25/mimo_v25_pro_hellaswag.yaml
+```
+</Accordion>
+
+See the [Installation Guide](/get-started/installation) and [LLM Fine-Tuning Guide](/recipes-e2e-examples/sft-peft).
+
+## Fine-Tuning
+
+See the [LLM Fine-Tuning Guide](/recipes-e2e-examples/sft-peft).
+
+## Hugging Face Model Cards
+
+- [XiaomiMiMo/MiMo-V2.5-Pro](https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro)

examples/llm_finetune/mimo_v25/mimo_v25_pro_hellaswag.yaml

@@ -0,0 +1,134 @@
+# Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
+#
+# 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.
+
+# 16 H100 nodes (128 GPUs):
+#   torchrun --nproc-per-node 8 -m nemo_automodel.cli.app examples/llm_finetune/mimo_v25/mimo_v25_pro_hellaswag.yaml
+
+recipe: TrainFinetuneRecipeForNextTokenPrediction
+
+seed: 1234
+
+step_scheduler:
+  global_batch_size: 256
+  local_batch_size: 8
+  ckpt_every_steps: 25
+  val_every_steps: 500
+  num_epochs: 1
+  max_steps: 100
+
+distributed:
+  strategy: fsdp2
+  tp_size: 1
+  cp_size: 1
+  pp_size: 4
+  ep_size: 32
+
+  sequence_parallel: false
+  activation_checkpointing: true
+
+  pipeline:
+    pp_schedule: interleaved1f1b
+    pp_microbatch_size: 1
+    layers_per_stage: 2
+    round_virtual_stages_to_pp_multiple: down
+    scale_grads_in_schedule: false
+    patch_inner_model: false
+    patch_causal_lm_model: false
+
+  moe:
+    reshard_after_forward: false
+    wrap_outer_model: false
+
+dist_env:
+  backend: nccl
+  timeout_minutes: 30
+
+model:
+  _target_: nemo_automodel.NeMoAutoModelForCausalLM.from_config
+  config:
+    _target_: nemo_automodel.components.models.mimo_v25.config.MiMoV2Config.from_pretrained
+    pretrained_model_name_or_path: XiaomiMiMo/MiMo-V2.5-Pro
+    name_or_path: XiaomiMiMo/MiMo-V2.5-Pro
+  trust_remote_code: false
+  load_base_model: true
+  backend:
+    _target_: nemo_automodel.components.models.common.BackendConfig
+    attn: sdpa
+    linear: torch
+    rms_norm: torch_fp32
+    rope_fusion: false
+    dispatcher: deepep
+    experts: torch_mm
+    gate_precision: float32
+    enable_hf_state_dict_adapter: true
+    enable_fsdp_optimizations: true
+
+checkpoint:
+  enabled: true
+  checkpoint_dir: checkpoints/mimo_v25_pro
+  model_save_format: safetensors
+  save_consolidated: false
+  dequantize_base_checkpoint: true
+
+loss_fn:
+  _target_: nemo_automodel.components.loss.masked_ce.MaskedCrossEntropy
+
+dataset:
+  _target_: nemo_automodel.components.datasets.llm.hellaswag.HellaSwag
+  path_or_dataset: rowan/hellaswag
+  split: train
+  tokenizer:
+    _target_: transformers.AutoTokenizer.from_pretrained
+    pretrained_model_name_or_path: XiaomiMiMo/MiMo-V2.5-Pro
+    trust_remote_code: true
+
+packed_sequence:
+  packed_sequence_size: 0
+
+dataloader:
+  _target_: torchdata.stateful_dataloader.StatefulDataLoader
+  collate_fn:
+    _target_: nemo_automodel.components.datasets.utils.default_collater
+    pad_seq_len_divisible: 64
+  shuffle: true
+
+validation_dataset:
+  _target_: nemo_automodel.components.datasets.llm.hellaswag.HellaSwag
+  path_or_dataset: rowan/hellaswag
+  split: validation
+  num_samples_limit: 64
+  tokenizer:
+    _target_: transformers.AutoTokenizer.from_pretrained
+    pretrained_model_name_or_path: XiaomiMiMo/MiMo-V2.5-Pro
+    trust_remote_code: true
+
+validation_dataloader:
+  _target_: torchdata.stateful_dataloader.StatefulDataLoader
+  collate_fn:
+    _target_: nemo_automodel.components.datasets.utils.default_collater
+    pad_seq_len_divisible: 64
+  shuffle: false
+  drop_last: true
+
+optimizer:
+  _target_: torch.optim.AdamW
+  betas: [0.9, 0.95]
+  eps: 1e-8
+  lr: 1e-5
+  weight_decay: 0.1
+
+wandb:
+  project: automodel-mimo-v25-pro
+  name: mimo_v25_pro_hellaswag_16n
+  mode: online

nemo_automodel/_transformers/registry.py

@@ -140,6 +140,10 @@
             "MiMoV2FlashForCausalLM",
             ("nemo_automodel.components.models.mimo_v2_flash.model", "MiMoV2FlashForCausalLM"),
         ),
+        (
+            "MiMoV2ForCausalLM",
+            ("nemo_automodel.components.models.mimo_v25.model", "MiMoV2ForCausalLM"),
+        ),
         (
             "Ministral3ForCausalLM",
             ("nemo_automodel.components.models.mistral3.model", "Ministral3ForCausalLM"),
@@ -282,6 +286,7 @@
     "kimi_vl": ("nemo_automodel.components.models.kimivl.model", "KimiVLConfig"),
     "llavaonevision1_5": ("nemo_automodel.components.models.llava_onevision.model", "Llavaonevision1_5Config"),
     "mimo_v2_flash": ("nemo_automodel.components.models.mimo_v2_flash.config", "MiMoV2FlashConfig"),
+    "mimo_v2": ("nemo_automodel.components.models.mimo_v25.config", "MiMoV2Config"),
     "minimax_m3_vl": ("nemo_automodel.components.models.minimax_m3_vl.config", "MiniMaxM3VLConfig"),
     "mistral4": ("nemo_automodel.components.models.mistral4.configuration", "Mistral4Config"),
     "step3p5v": ("nemo_automodel.components.models.step3p7.configuration_step3p7", "Step3p5VConfig"),

nemo_automodel/components/models/mimo_v25/__init__.py

@@ -0,0 +1,18 @@
+# Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
+#
+# 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 nemo_automodel.components.models.mimo_v25.config import MiMoV2Config
+from nemo_automodel.components.models.mimo_v25.model import MiMoV2ForCausalLM, MiMoV2Model
+
+__all__ = ["MiMoV2Config", "MiMoV2ForCausalLM", "MiMoV2Model"]