Add OpenPangu-V2 Model

docs/source/en/_toctree.yml

@@ -813,6 +813,8 @@
         title: OlmoHybrid
       - local: model_doc/openai_privacy_filter
         title: OpenAI Privacy Filter
+      - local: model_doc/openpangu_v2
+        title: OpenPanguV2
       - local: model_doc/opt
         title: OPT
       - local: model_doc/pegasus

docs/source/en/model_doc/openpangu_v2.md

@@ -0,0 +1,84 @@
+<!--Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
+Copyright 2026 The HuggingFace Inc. team. 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.
+
+⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+*This model was contributed to Hugging Face Transformers on 2026-07-06.*
+
+# OpenPanguV2
+
+OpenPanguV2 is an MoE model trained on Ascend. Its context length is 512k. During Post-training, OpenPanguV2 is 
+trained through unified SFT with slow and fast thinking capability, multiple specialist RL traning, on-policy 
+distillation combining multiple RL specialists.
+
+## Architecture
+
+OpenPanguV2 brings several major architectural improvements:
+
+* **Efficient attention**: The model retains MLA for efficient inference and combines DSA and SWA in a 1:2 layer ratio. 
+  SWA layers handle local-window modeling, while DSA layers capture sparse global context. This design lowers compute, 
+  memory footprint, and memory access costs for long-context inference while preserving accuracy.
+* **Residual topology**: TThe conventional residual path is replaced with a 4-stream mHC design, improving representation
+  diversity and generalization.
+* **Optimizer**: Training uses the Muon optimizer for faster convergence.
+
+## Usage examples
+
+The example below demonstrates how to generate text with [`Pipeline`] or the [`AutoModelForCausalLM`] class.
+
+```python
+from transformers import pipeline
+
+
+pipe = pipeline(
+    task="text-generation",
+    model="openpangu/openPangu-2.0-Flash",
+    device_map="auto"
+)
+pipe("Give me a short introduction to large language model.")
+```
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_name = "openpangu/openPangu-2.0-Flash"
+
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    device_map="auto",
+)
+input_ids = tokenizer("Give me a short introduction to large language model.", return_tensors="pt").to(model.device)
+
+output = model.generate(**input_ids, max_new_tokens=50)
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+
+## OpenPanguV2Config
+
+[[autodoc]] OpenPanguV2Config
+
+## OpenPanguV2Tokenizer
+
+[[autodoc]] OpenPanguV2Tokenizer
+
+## OpenPanguV2Model
+
+[[autodoc]] OpenPanguV2Model
+    - forward
+
+## OpenPanguV2ForCausalLM
+
+[[autodoc]] OpenPanguV2ForCausalLM
+    - forward

src/transformers/conversion_mapping.py

@@ -103,6 +103,7 @@
     "qwen3_5_moe_text": "qwen3_5_text",
     "llava_next_video": "llava_next",
     "llava_onevision": "llava_next",
+    "openpangu_v2": "qwen2_moe",
     # class-based mappings
     "PaliGemmaModel": "LlavaModel",
     "AyaVisionModel": "LlavaModel",

src/transformers/models/__init__.py

@@ -319,6 +319,7 @@
     from .oneformer import *
     from .openai import *
     from .openai_privacy_filter import *
+    from .openpangu_v2 import *
     from .opt import *
     from .ovis2 import *
     from .owlv2 import *

src/transformers/models/auto/auto_mappings.py

@@ -425,6 +425,7 @@
         ("oneformer", "OneFormerConfig"),
         ("openai-gpt", "OpenAIGPTConfig"),
         ("openai_privacy_filter", "OpenAIPrivacyFilterConfig"),
+        ("openpangu_v2", "OpenPanguV2Config"),
         ("opt", "OPTConfig"),
         ("ovis2", "Ovis2Config"),
         ("owlv2", "Owlv2Config"),

src/transformers/models/auto/modeling_auto.py

@@ -360,6 +360,7 @@ class _BaseModelWithGenerate(PreTrainedModel, GenerationMixin):
         ("oneformer", "OneFormerModel"),
         ("openai-gpt", "OpenAIGPTModel"),
         ("openai_privacy_filter", "OpenAIPrivacyFilterModel"),
+        ("openpangu_v2", "OpenPanguV2Model"),
         ("opt", "OPTModel"),
         ("ovis2", "Ovis2Model"),
         ("owlv2", "Owlv2Model"),
@@ -785,6 +786,7 @@ class _BaseModelWithGenerate(PreTrainedModel, GenerationMixin):
         ("olmo_hybrid", "OlmoHybridForCausalLM"),
         ("olmoe", "OlmoeForCausalLM"),
         ("openai-gpt", "OpenAIGPTLMHeadModel"),
+        ("openpangu_v2", "OpenPanguV2ForCausalLM"),
         ("opt", "OPTForCausalLM"),
         ("pegasus", "PegasusForCausalLM"),
         ("persimmon", "PersimmonForCausalLM"),

src/transformers/models/auto/tokenization_auto.py

@@ -252,6 +252,7 @@
         ("omdet-turbo", "CLIPTokenizer" if is_tokenizers_available() else None),
         ("oneformer", "CLIPTokenizer" if is_tokenizers_available() else None),
         ("openai-gpt", "OpenAIGPTTokenizer" if is_tokenizers_available() else None),
+        ("openpangu_v2", "OpenPanguV2Tokenizer" if is_tokenizers_available() else None),
         ("opt", "GPT2Tokenizer" if is_tokenizers_available() else None),
         ("ovis2", "Qwen2Tokenizer" if is_tokenizers_available() else None),
         ("owlv2", "CLIPTokenizer" if is_tokenizers_available() else None),

src/transformers/models/openpangu_v2/__init__.py

@@ -0,0 +1,29 @@
+# Copyright 2026 the HuggingFace Team. 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 typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_openpangu_v2 import *
+    from .modeling_openpangu_v2 import *
+    from .tokenization_openpangu_v2 import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)