1.0

.gitattributes

@@ -0,0 +1,35 @@
+*.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

LICENSE.md

@@ -0,0 +1,27 @@
+| License | Business Source License (BSL 1.1) |
+| --- | --- |
+| Licensor | M87 Labs, Inc. |
+| Licensed Work | “Moondream 3 (Preview)” including Model Weights and any Derivatives (“Derivatives” include fine-tunes, merges, quantizations, weight deltas, and other weight-level modifications or conversions.) |
+| Additional Use Grant | You may make production use of the Licensed Work, provided Your use does not include offering the Licensed Work to third parties on a hosted or embedded basis in order to compete with M87 Labs’s paid version(s) of the Licensed Work. For purposes of this license:<br><br>A “competitive offering” is a Product that is offered to third parties on a paid basis, including through paid support arrangements, that significantly overlaps with the capabilities of M87 Labs’s paid version(s) of the Licensed Work. If Your Product is not a competitive offering when You first make it generally available, it will not become a competitive offering later due to M87 Labs releasing a new version of the Licensed Work with additional capabilities. In addition, Products that are not provided on a paid basis are not competitive.<br><br>“Product” means software that is offered to end users to manage in their own environments or offered as a service on a hosted basis.<br><br>“Embedded” means including the source code or executable code from the Licensed Work in a competitive offering. “Embedded” also means packaging the competitive offering in such a way that the Licensed Work must be accessed or downloaded for the competitive offering to operate.<br><br>Hosting or using the Licensed Work(s) for internal purposes within an organization is not considered a competitive offering. M87 Labs considers your organization to include all of your affiliates under common control. |
+| Change Date | Two years after the first public release of this version of the Licensed Work |
+| Change License | Apache License, Version 2.0 |
+
+For information about alternative licensing arrangements for the Licensed Work, please contact [contact@m87.ai](mailto:contact@m87.ai).
+
+The text of the Business Source License 1.1 follows. License text copyright (c) 2020 MariaDB Corporation Ab, All Rights Reserved. “Business Source License” is a trademark of MariaDB Corporation Ab.
+
+## Terms
+
+The Licensor hereby grants you the right to copy, modify, create derivative works, redistribute, and make non-production use of the Licensed Work. The Licensor may make an Additional Use Grant, above, permitting limited production use.
+
+Effective on the Change Date, or the fourth anniversary of the first publicly available distribution of a specific version of the Licensed Work under this License, whichever comes first, the Licensor hereby grants you rights under the terms of the Change License, and the rights granted in the paragraph above terminate.
+
+If your use of the Licensed Work does not comply with the requirements currently in effect as described in this License, you must purchase a commercial license from the Licensor, its affiliated entities, or authorized resellers, or you must refrain from using the Licensed Work.
+
+All copies of the original and modified Licensed Work, and derivative works of the Licensed Work, are subject to this License. This License applies separately for each version of the Licensed Work and the Change Date may vary for each version of the Licensed Work released by Licensor.
+
+You must conspicuously display this License on each original or modified copy of the Licensed Work. If you receive the Licensed Work in original or modified form from a third party, the terms and conditions set forth in this License apply to your use of that work.
+
+Any use of the Licensed Work in violation of this License will automatically terminate your rights under this License for the current and all other versions of the Licensed Work.
+
+This License does not grant you any right in any trademark or logo of Licensor or its affiliates (provided that you may use a trademark or logo of Licensor as expressly required by this License).TO THE EXTENT PERMITTED BY APPLICABLE LAW, THE LICENSED WORK IS PROVIDED ON AN “AS IS” BASIS. LICENSOR HEREBY DISCLAIMS ALL WARRANTIES AND CONDITIONS, EXPRESS OR IMPLIED, INCLUDING (WITHOUT LIMITATION) WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, NON-INFRINGEMENT, AND TITLE.

README.md

@@ -0,0 +1,96 @@
+---
+library_name: transformers
+pipeline_tag: image-text-to-text
+license: other
+---
+
+# Moondream 3 HF
+
+**Moondream 3 HF** is a reimplementation of the [Moondream 3 (Preview)](https://huggingface.co/moondream/moondream3-preview) model using the standard Hugging Face Transformers architecture conventions.  
+
+## Overview
+
+- Multimodal vision-language model with a mixture-of-experts (MoE) text backbone  
+- Architecture and weights correspond to Moondream 3 (Preview) (approximately 9B parameters, 2B active)  
+- Implemented as standard Transformers components:
+  - `Moondream3ForConditionalGeneration`
+  - `Moondream3Model`, `Moondream3TextModel`, `Moondream3VisionModel`
+  - `Moondream3Processor`, `Moondream3ImageProcessor`, `Moondream3Config`
+
+The purpose of this repository is to make Moondream 3 interoperable with the Hugging Face ecosystem so it can be used directly with the Transformers API, including `generate()`, `Trainer`, and PEFT integrations.
+
+## Example usage
+
+Example for running multimodal inference with the `moondream3-hf` implementation:
+
+```python
+import torch
+from PIL import Image
+from transformers import AutoProcessor, AutoModelForCausalLM
+
+DEVICE="cuda:0"
+
+model = AutoModelForCausalLM.from_pretrained("NyxKrage/moondream3-hf", dtype="bfloat16", device_map=DEVICE, trust_remote_code=True)
+processor = AutoProcessor.from_pretrained("NyxKrage/moondream3-hf", use_fast=False, trust_remote_code=True)
+
+
+image1 = Image.open("image1.jpg")
+image2 = Image.open("image2.jpg")
+text = [processor.apply_chat_template("", tokenize=False)] * 2
+inputs = processor(text=text, images=[
+    image1,
+    image2,
+])
+inputs = {k: v.to(DEVICE) for k,v in inputs.items()}
+
+model.eval()
+
+with torch.no_grad():
+    outputs = model.generate(
+        **inputs,
+        use_cache=True,
+    )
+    outputs = [
+        out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs["input_ids"], outputs)
+    ]
+    for output in outputs:
+        print(processor.decode(output))
+
+```
+
+The `chat_template` uses Hugging Face’s Jinja format and accepts either a single string or a sequence of messages (`user [, assistant]`).
+
+### Training
+
+The model can be trained using `trl` and supports `peft` and `bitsandbytes` out of the box.  
+Included in the repo is also an implementation which replaces the MoE layers with a grouped_gemm implementation which has been adapted from [github:woct0rdho/transformers-qwen3-moe-fused](https://github.com/woct0rdho/transformers-qwen3-moe-fused), and can be used by importing Moondream3ForConditionalGeneration from `modeling_moondream3_fusedmoe.py` instead.
+
+
+## Prompting modes
+
+The chat template supports multiple task types via text prefixes:
+
+| Mode    | Template prefix               | Example input                             |
+| ------- | ----------------------------- | ----------------------------------------- |
+| Query   | `query:`                      | `query: What is happening in this image?` |
+| Caption | `caption:[short/normal/long]` | `caption: long`                           |
+| Detect  | `detect:`                     | `detect: dog`                             |
+| Point   | `point:`                      | `point: red car`                          |
+
+If no prefix is provided, the default mode is `query:`.
+
+This reimplementation aims to provide interoperability and ease of experimentation within the Hugging Face ecosystem. It is not an official release.
+
+## License
+
+The model weights remain under the **Business Source License 1.1 with an Additional Use Grant (No Third-Party Service)**, identical to the original [Moondream 3 Preview license](https://huggingface.co/moondream/moondream3-preview/blob/main/LICENSE.md).
+
+This allows research, personal, and most commercial use, but prohibits offering hosted or resold access that competes with M87 Labs’ paid services.
+
+All new implementation code in this repository is released under the **Apache 2.0 License**.
+
+## Credits
+
+* Original model and research: M87 Labs / Moondream AI
+* Hugging Face–compatible reimplementation: NyxKrage
+* Based on the public [Moondream 3 Preview](https://huggingface.co/moondream/moondream3-preview) release

chat_template.jinja

@@ -0,0 +1,75 @@
+{# Accepts: 
+   - messages as a single string
+   - messages as 1-2 chat dicts: user[, assistant]
+   - user.content may be a string OR a list with exactly one {type:text} and one {type:image}
+#}
+{% if messages is string -%}
+  {% set text = messages | trim -%}
+  {% set has_assistant = false -%}
+{%- else -%}
+  {% if messages | length < 1 or messages | length > 2 -%}
+    {{ raise_exception("Provide a single string or 1-2 messages (user[, assistant]).") }}
+  {%- endif -%}
+  {% if (messages[0].get('role') | default('')) != 'user' -%}
+    {{ raise_exception("First message must have role 'user'.") }}
+  {%- endif -%}
+  {% if messages | length == 2 and (messages[1].get('role') | default('')) != 'assistant' -%}
+    {{ raise_exception("Second message, if present, must have role 'assistant'.") }}
+  {%- endif -%}
+
+  {# Extract user text, supporting multimodal content #}
+  {% set ucontent = messages[0]['content'] | default('', true) -%}
+  {% if ucontent is string -%}
+    {% set text = ucontent | trim -%}
+  {%- else -%}
+    {% if ucontent | length != 2 -%}
+      {{ raise_exception("User content list must have exactly two parts: one text and one image.") }}
+    {%- endif -%}
+    {% set text_parts = ucontent | selectattr('type','equalto','text') | list -%}
+    {% set image_parts = ucontent | selectattr('type','equalto','image') | list -%}
+    {% if (text_parts | length) != 1 or (image_parts | length) != 1 -%}
+      {{ raise_exception("User content must include exactly one text and one image part.") }}
+    {%- endif -%}
+    {% set text = (text_parts[0].get('text') | default('')) | trim -%}
+  {%- endif -%}
+
+  {# Extract assistant text if present (string or list of parts) #}
+  {% set has_assistant = (messages | length == 2) -%}
+  {% if has_assistant -%}
+    {% set acontent = messages[1]['content'] | default('', true) -%}
+    {% if acontent is string -%}
+      {% set assistant_text = acontent -%}
+    {%- else -%}
+      {% set atexts = acontent | selectattr('type','equalto','text') | map(attribute='text') | list -%}
+      {% set assistant_text = (atexts | join('')) -%}
+    {%- endif -%}
+  {%- endif -%}
+{%- endif -%}
+
+{% set lower = text | lower -%}
+
+{# Routing with zero-whitespace outputs #}
+{% if text == '' -%}
+<|md_reserved_0|>describe<|md_reserved_1|>normal<|md_reserved_2|>
+{%- elif lower.startswith('caption:') -%}
+  {% set length = (text[8:] | trim | lower) -%}
+  {% if length not in ['short','normal','long'] -%}
+    {{ raise_exception("caption length must be one of: short, normal, long.") }}
+  {%- endif -%}
+<|md_reserved_0|>describe<|md_reserved_1|>{{ length }}<|md_reserved_2|>
+{%- elif lower.startswith('query:') -%}
+  {% set q = text[6:] | trim -%}
+<|md_reserved_0|>query<|md_reserved_1|>{{ q }}<|md_reserved_2|>
+{%- elif lower.startswith('detect:') -%}
+  {% set q = text[7:] | trim -%}
+<|md_reserved_0|>det<|md_reserved_1|>{{ q }}<|md_reserved_2|>
+{%- elif lower.startswith('point:') -%}
+  {% set q = text[6:] | trim -%}
+<|md_reserved_0|>point<|md_reserved_1|>{{ q }}<|md_reserved_2|>
+{%- else -%}
+  {% set q = text -%}
+<|md_reserved_0|>query<|md_reserved_1|>{{ q }}<|md_reserved_2|>
+{%- endif -%}
+{%- generation -%}
+    {%- if has_assistant -%}{{ assistant_text }}{{ eos_token }}{%- endif -%}
+{%- endgeneration -%}

config.json

@@ -0,0 +1,68 @@
+{
+  "architectures": [
+    "Moondream3ForConditonalGeneration"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_moondream3.Moondream3Config",
+    "AutoModel": "modeling_moondream3.Moondream3Model",
+    "AutoModelForCausalLM": "modeling_moondream3.Moondream3ForConditionalGeneration"
+  },
+  "model_type": "moondream3",
+  "region_config": {
+    "coord_feat_dim": 256,
+    "coord_out_dim": 1024,
+    "hidden_size": 2048,
+    "model_type": "moondream3_region",
+    "size_feat_dim": 512,
+    "size_out_dim": 2048
+  },
+  "text_config": {
+    "attention_bias": true,
+    "attention_dropout": 0.0,
+    "head_dim": 64,
+    "hidden_act": "silu",
+    "hidden_size": 2048,
+    "initializer_range": 0.02,
+    "intermediate_size": 8192,
+    "max_position_embeddings": 4096,
+    "model_type": "moondream3_text",
+    "moe_intermediate_size": 1024,
+    "moe_start_layer": 4,
+    "num_attention_heads": 32,
+    "num_experts": 64,
+    "num_experts_per_tok": 8,
+    "num_hidden_layers": 24,
+    "num_key_value_heads": 32,
+    "num_local_experts": 64,
+    "output_router_logits": false,
+    "prefix_attn": 730,
+    "bos_token_id": 0,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "rope_theta": 1500000.0,
+      "rope_type": "default"
+    },
+    "use_cache": true,
+    "vocab_size": 51200
+  },
+  "tie_word_embeddings": false,
+  "transformers_version": "5.0.0.dev0",
+  "vision_config": {
+    "attention_bias": true,
+    "attention_dropout": 0.0,
+    "crop_size": 378,
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "in_channels": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 4304,
+    "max_crops": 12,
+    "model_type": "moondream3_vision",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 27,
+    "overlap_margin": 4,
+    "patch_size": 14,
+    "proj_inner_dim": 8192,
+    "proj_out_dim": 2048
+  }
+}

configuration_moondream3.py

@@ -0,0 +1,299 @@
+# coding=utf-8
+# Copyright 2025 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.
+
+from typing import Optional, List
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+
+
+class Moondream3TextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Moondream3TextModel`]. It is used to instantiate a
+    Moondream3 model according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PreTrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 51200):
+            Vocabulary size of the Moondream3 model.
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 8192):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        num_experts (`int`, *optional*, defaults to 64):
+            Number of experts for MoE layers.
+        num_experts_per_tok (`int`, *optional*, defaults to 8):
+            Number of selected experts per token.
+        moe_intermediate_size (`int`, *optional*, defaults to 1024):
+            Intermediate size of the routed expert.
+        moe_start_layer (`int`, *optional*, defaults to 4):
+            The layer index where MoE layers start.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-6):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers.
+        head_dim (`int`, *optional*):
+            The dimension of the head. If not specified, will default to `hidden_size // num_attention_heads`.
+    """
+
+    model_type = "moondream3_text"
+    base_config_key = "text_config"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size: int = 51200,
+        hidden_size: int = 2048,
+        intermediate_size: int = 8192,
+        num_hidden_layers: int = 24,
+        num_attention_heads: int = 32,
+        num_key_value_heads: int = 32,
+        max_position_embeddings: int = 4096,
+        num_experts: int = 64,
+        num_experts_per_tok: int = 8,
+        moe_intermediate_size: int = 1024,
+        moe_start_layer: int = 4,
+        bos_id: int = 0,
+        hidden_act: str = "silu",
+        initializer_range: float = 0.02,
+        rms_norm_eps: float = 1e-5,
+        use_cache: bool = False,
+        tie_word_embeddings: bool = False,
+        attention_bias: bool = True,
+        rope_parameters: Optional[dict] = None,
+        head_dim: Optional[int] = None,
+        **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
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.attention_bias = attention_bias
+        self.head_dim = head_dim or hidden_size // num_attention_heads
+        self.bos_id = bos_id
+        
+        # MoE parameters (merged from TextMoeConfig)
+        self.num_experts = num_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_intermediate_size = moe_intermediate_size
+        self.moe_start_layer = moe_start_layer
+        
+            
+        # Try to set `rope_scaling` if available, otherwise use `rope_parameters`
+        rope_scaling = kwargs.pop("rope_scaling", None)
+        self.rope_parameters = rope_scaling or rope_parameters
+
+        # Validate the correctness of rotary position embeddings parameters
+        rope_theta = kwargs.get("rope_theta", 1500000.0)
+        rope_config_validation(self)
+        
+        # HF compatibility attributes
+        self.output_router_logits = False
+        self.output_attentions = False
+        self.output_hidden_states = False
+        self.attention_dropout = 0.0
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+
+
+class Moondream3VisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of the Moondream3 vision encoder.
+    
+    Args:
+        hidden_size (`int`, *optional*, defaults to 1152):
+            Dimension of the encoder's hidden states.
+        intermediate_size (`int`, *optional*, defaults to 4304):
+            Dimension of the encoder's MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 27):
+            Number of hidden layers in the vision encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads in the vision encoder.
+        patch_size (`int`, *optional*, defaults to 14):
+            The size of each patch in the vision encoder.
+        in_channels (`int`, *optional*, defaults to 3):
+            Number of input channels.
+        proj_out_dim (`int`, *optional*, defaults to 2048):
+            Output dimension of the projection layer.
+        crop_size (`int`, *optional*, defaults to 378):
+            Size of image crops.
+        max_crops (`int`, *optional*, defaults to 12):
+            Maximum number of crops.
+        overlap_margin (`int`, *optional*, defaults to 4):
+            Overlap margin for crops.
+        proj_inner_dim (`int`, *optional*, defaults to 8192):
+            Inner dimension of the projection MLP.
+        hidden_act (`str`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer.
+    """
+    model_type = "moondream3_vision"
+    base_config_key = "vision_config"
+
+    def __init__(
+        self,
+        hidden_size: int = 1152,
+        intermediate_size: int = 4304,
+        num_hidden_layers: int = 27,
+        num_attention_heads: int = 16,
+        patch_size: int = 14,
+        in_channels: int = 3,
+        proj_out_dim: int = 2048,
+        crop_size: int = 378,
+        max_crops: int = 12,
+        overlap_margin: int = 4,
+        proj_inner_dim: int = 8192,
+        prefix_len: int = 730,
+        hidden_act: str = "gelu_pytorch_tanh",
+        initializer_range: float = 0.02,
+        attention_bias: bool = True,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.patch_size = patch_size
+        self.in_channels = in_channels
+        self.proj_out_dim = proj_out_dim
+        self.crop_size = crop_size
+        self.max_crops = max_crops
+        self.prefix_len = prefix_len
+        self.overlap_margin = overlap_margin
+        self.proj_inner_dim = proj_inner_dim
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.attention_dropout = 0.0
+        self.attention_bias = attention_bias
+
+        super().__init__(**kwargs)
+
+
+class Moondream3RegionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of the Moondream3 region encoder for object detection and grounding.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations for region features.
+        coord_feat_dim (`int`, *optional*, defaults to 256):
+            Dimension of coordinate feature embeddings.
+        coord_out_dim (`int`, *optional*, defaults to 1024):
+            Output dimension for coordinate features.
+        size_feat_dim (`int`, *optional*, defaults to 512):
+            Dimension of size feature embeddings.
+        size_out_dim (`int`, *optional*, defaults to 2048):
+            Output dimension for size features.
+    """
+    model_type = "moondream3_region"
+    base_config_key = "region_config"
+
+    def __init__(
+        self,
+        hidden_size: int = 2048,
+        coord_feat_dim: int = 256,
+        coord_out_dim: int = 1024,
+        size_feat_dim: int = 512,
+        size_out_dim: int = 2048,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.coord_feat_dim = coord_feat_dim
+        self.coord_out_dim = coord_out_dim
+        self.size_feat_dim = size_feat_dim
+        self.size_out_dim = size_out_dim
+
+
+class Moondream3Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Moondream3Model`].
+
+    Args:
+        text_config (`Union[PreTrainedConfig, dict]`, *optional*, defaults to `Moondream3TextConfig`):
+            The config object or dictionary of the text backbone.
+        vision_config (`Union[PreTrainedConfig, dict]`, *optional*, defaults to `Moondream3VisionConfig`):
+            The config object or dictionary of the vision backbone.
+        region_config (`Union[PreTrainedConfig, dict]`, *optional*, defaults to `Moondream3RegionConfig`):
+            The config object or dictionary of the region backbone for object detection and grounding.
+        image_token_id (`int`, *optional*, defaults to 151655):
+            The image token index to encode the image prompt.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie the word embeddings.
+    """
+
+    model_type = "moondream3"
+    sub_configs = {
+        "vision_config": Moondream3VisionConfig,
+        "text_config": Moondream3TextConfig,
+        "region_config": Moondream3RegionConfig,
+    }
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        region_config=None,
+        bos_token_id=0,
+        tie_word_embeddings: bool = False,
+        **kwargs,
+    ):
+        if isinstance(vision_config, dict):
+            self.vision_config = self.sub_configs["vision_config"](**vision_config)
+        elif vision_config is None:
+            self.vision_config = self.sub_configs["vision_config"]()
+
+        if isinstance(text_config, dict):
+            self.text_config = self.sub_configs["text_config"](**text_config)
+        elif text_config is None:
+            self.text_config = self.sub_configs["text_config"]()
+
+        if isinstance(region_config, dict):
+            self.region_config = self.sub_configs["region_config"](**region_config)
+        elif region_config is None:
+            self.region_config = self.sub_configs["region_config"]()
+
+        super().__init__(**kwargs, tie_word_embeddings=tie_word_embeddings)
+
+
+__all__ = ["Moondream3Config", "Moondream3TextConfig", "Moondream3VisionConfig", "Moondream3RegionConfig"]

convert_moondream_weights_to_hf.py

@@ -0,0 +1,239 @@
+#!/usr/bin/env python3
+# coding=utf-8
+# Copyright 2025 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.
+
+import argparse
+import json
+import re
+from pathlib import Path
+from typing import Dict
+
+from safetensors.torch import load_file
+from safetensors.torch import save_file
+
+
+# Key mapping from original Moondream to HF Moondream3
+OLD_KEY_TO_NEW_KEY_MAPPING = [
+    # Text model
+    (r"model\.text\.wte", "model.text_model.embed_tokens.weight"),
+    (r"model\.text\.post_ln\.(weight|bias)", r"model.text_model.norm.\1"),
+    (r"model\.text\.lm_head\.(weight|bias)", r"lm_head.\1"),
+    (r"model\.text\.blocks\.(\d+)\.attn\.qkv\.(weight|bias)", r"model.text_model.layers.\1.self_attn.qkv.\2"),
+    (r"model\.text\.blocks\.(\d+)\.attn\.proj\.(weight|bias)", r"model.text_model.layers.\1.self_attn.o_proj.\2"),
+    (r"model\.text\.blocks\.(\d+)\.attn\.tau\.wq", r"model.text_model.layers.\1.self_attn.tau_wq.weight"),
+    (r"model\.text\.blocks\.(\d+)\.attn\.tau\.wv", r"model.text_model.layers.\1.self_attn.tau_wv.weight"),
+    (r"model\.text\.blocks\.(\d+)\.attn\.tau\.alpha", r"model.text_model.layers.\1.self_attn.tau_alpha"),
+    (r"model\.text\.blocks\.(\d+)\.ln\.(weight|bias)", r"model.text_model.layers.\1.input_layernorm.\2"),
+    (r"model\.text\.blocks\.(\d+)\.mlp\.fc1\.(weight|bias)", r"model.text_model.layers.\1.mlp.up_proj.\2"),
+    (r"model\.text\.blocks\.(\d+)\.mlp\.fc2\.(weight|bias)", r"model.text_model.layers.\1.mlp.down_proj.\2"),
+    (r"model\.text\.blocks\.(\d+)\.mlp\.router\.(weight|bias)", r"model.text_model.layers.\1.mlp.gate.\2"),
+    
+    # Vision model
+    (r"model\.vision\.patch_emb\.(weight|bias)", r"model.vision_model.embeddings.projection.\1"),
+    (r"model\.vision\.pos_emb", "model.vision_model.embeddings.position_embeddings"),
+    (r"model\.vision\.post_ln\.(weight|bias)", r"model.vision_model.post_layernorm.\1"),
+    (r"model\.vision\.blocks\.(\d+)\.attn\.qkv\.(weight|bias)", r"model.vision_model.layers.\1.self_attn.qkv.\2"),
+    (r"model\.vision\.blocks\.(\d+)\.attn\.proj\.(weight|bias)", r"model.vision_model.layers.\1.self_attn.o_proj.\2"),
+    (r"model\.vision\.blocks\.(\d+)\.ln1\.(weight|bias)", r"model.vision_model.layers.\1.input_layernorm.\2"),
+    (r"model\.vision\.blocks\.(\d+)\.ln2\.(weight|bias)", r"model.vision_model.layers.\1.post_attention_layernorm.\2"),
+    (r"model\.vision\.blocks\.(\d+)\.mlp\.fc1\.(weight|bias)", r"model.vision_model.layers.\1.mlp.up_proj.\2"),
+    (r"model\.vision\.blocks\.(\d+)\.mlp\.fc2\.(weight|bias)", r"model.vision_model.layers.\1.mlp.down_proj.\2"),
+    
+    # Vision projection
+    (r"model\.vision\.proj_mlp\.fc1\.(weight|bias)", r"model.vision_model.vision_projection.up_proj.\1"),
+    (r"model\.vision\.proj_mlp\.fc2\.(weight|bias)", r"model.vision_model.vision_projection.down_proj.\1"),
+    
+    # Region model
+    (r"model\.region\.coord_encoder\.(weight|bias)", r"model.region_encoder.coord_encoder.\1"),
+    (r"model\.region\.coord_decoder\.(weight|bias)", r"model.region_decoder.coord_decoder.\1"),
+    (r"model\.region\.size_encoder\.(weight|bias)", r"model.region_encoder.size_encoder.\1"),
+    (r"model\.region\.size_decoder\.(weight|bias)", r"model.region_decoder.size_decoder.\1"),
+    (r"model\.region\.coord_features", "model.region_encoder.coord_freq"),
+    (r"model\.region\.size_features", "model.region_encoder.size_freq"),
+]
+
+
+def rename_key(old_key: str) -> str:
+    """Convert original key name to HF key name."""
+    for pattern, new_key in OLD_KEY_TO_NEW_KEY_MAPPING:
+        if re.match(pattern, old_key):
+            return re.sub(pattern, new_key, old_key)
+    return old_key
+
+
+def convert_state_dict(original_state_dict: Dict) -> Dict:
+    """Convert original state dict to HF format."""
+    converted_state_dict = {}
+    converted_keys = []
+    for old_key, tensor in original_state_dict.items():
+        new_key = rename_key(old_key)
+        print(old_key, new_key)
+        
+        # Handle QKV weight splitting for attention
+        if "attn.qkv.weight" in old_key or "attn.qkv.bias" in old_key:
+            # Split QKV into separate Q, K, V matrices
+            layer_match = re.search(r"blocks\.(\d+)", old_key)
+            if layer_match:
+                layer_idx = int(layer_match.group(1))
+                
+                # Determine if this is text or vision model
+                if "model.text.blocks" in old_key:
+                    n_heads = 32
+                    n_kv_heads = 32
+                    head_dim = 64  # 2048 / 32
+                    base_key = f"model.text_model.layers.{layer_idx}.self_attn"
+                else:  # vision
+                    n_heads = 16
+                    n_kv_heads = 16
+                    head_dim = 72  # 1152 / 16
+                    base_key = f"model.vision_model.layers.{layer_idx}.self_attn"
+                
+                # Split tensor
+                q_dim = n_heads * head_dim
+                kv_dim = n_kv_heads * head_dim
+                
+                if "weight" in old_key:
+                    q_weight = tensor[:q_dim]
+                    k_weight = tensor[q_dim:q_dim + kv_dim]
+                    v_weight = tensor[q_dim + kv_dim:]
+                    
+                    converted_state_dict[f"{base_key}.q_proj.weight"] = q_weight
+                    converted_state_dict[f"{base_key}.k_proj.weight"] = k_weight
+                    converted_state_dict[f"{base_key}.v_proj.weight"] = v_weight
+                    converted_keys.append(old_key)
+                else:  # bias
+                    q_bias = tensor[:q_dim]
+                    k_bias = tensor[q_dim:q_dim + kv_dim]
+                    v_bias = tensor[q_dim + kv_dim:]
+                    
+                    converted_state_dict[f"{base_key}.q_proj.bias"] = q_bias
+                    converted_state_dict[f"{base_key}.k_proj.bias"] = k_bias
+                    converted_state_dict[f"{base_key}.v_proj.bias"] = v_bias
+                    converted_keys.append(old_key)
+        # Handle MoE expert weight splitting
+        elif ("mlp.fc1.weight" in old_key or "mlp.fc2.weight" in old_key) and not "proj_mlp" in old_key:
+            layer_match = re.search(r"blocks\.(\d+)", old_key)
+            if layer_match:
+                layer_idx = int(layer_match.group(1))
+                # Only process MoE layers (4+ in this model)
+                if layer_idx >= 4 and "model.text." in old_key:
+                    n_experts = 64  # From config
+                    
+                    if "fc1.weight" in old_key:
+                        # Shape: (n_experts, 2 * d_ffn, d_model) → split into individual experts
+                        for expert_idx in range(n_experts):
+                            expert_weight = tensor[expert_idx]  # Shape: (2 * d_ffn, d_model)
+                            # For GeGLU, split into gate and up projections
+                            up_weight = expert_weight[:expert_weight.shape[0]//2]  # First half
+                            gate_weight = expert_weight[expert_weight.shape[0]//2:]   # Second half
+                            
+                            converted_state_dict[f"model.text_model.layers.{layer_idx}.mlp.experts.{expert_idx}.gate_proj.weight"] = gate_weight
+                            converted_state_dict[f"model.text_model.layers.{layer_idx}.mlp.experts.{expert_idx}.up_proj.weight"] = up_weight
+                    elif "fc2.weight" in old_key:
+                        # Shape: (n_experts, d_model, d_ffn) → split into individual experts
+                        for expert_idx in range(n_experts):
+                            expert_weight = tensor[expert_idx]  # Shape: (d_model, d_ffn)
+                            converted_state_dict[f"model.text_model.layers.{layer_idx}.mlp.experts.{expert_idx}.down_proj.weight"] = expert_weight
+                else:
+                    # Dense MLP for layers < 4
+                    converted_state_dict[new_key] = tensor
+        else:
+            converted_state_dict[new_key] = tensor
+    return converted_state_dict
+
+
+def convert_moondream_weights_to_hf(
+    original_model_path: str,
+    output_file: str,
+):
+    """Convert Moondream weights to HuggingFace format."""
+    
+    # Load original state dict
+    print(f"Loading original model from {original_model_path}")
+    
+    # Find safetensors files
+    model_path = Path(original_model_path)
+    if model_path.is_file() and model_path.suffix == ".safetensors":
+        # Single file
+        original_state_dict = load_file(str(model_path))
+    elif model_path.is_dir():
+        # Directory - look for index file or single model file
+        index_path = model_path / "model.safetensors.index.json"
+        single_file_path = model_path / "model.safetensors"
+        
+        if index_path.exists():
+            with open(index_path) as f:
+                index = json.load(f)
+            
+            original_state_dict = {}
+            for filename in set(index["weight_map"].values()):
+                file_path = model_path / filename
+                if file_path.exists():
+                    state_dict = load_file(str(file_path))
+                    for k, v in state_dict.items():
+                        original_state_dict[k] = v
+                else:
+                    print(f"Warning: {file_path} not found")
+        elif single_file_path.exists():
+            original_state_dict = load_file(str(single_file_path))
+        else:
+            raise FileNotFoundError(f"Could not find model files in {original_model_path}")
+    else:
+        raise FileNotFoundError(f"Could not find model files in {original_model_path}")
+    
+    print(f"Loaded {len(original_state_dict)} tensors")
+    
+    # Convert state dict
+    print("Converting state dict...")
+    converted_state_dict = convert_state_dict(original_state_dict)
+    
+    print(f"Converted {len(converted_state_dict)} tensors")
+    
+    # Save converted weights
+    output_path = Path(output_file)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    
+    print(f"Saving converted weights to {output_path}")
+    save_file(converted_state_dict, str(output_path))
+    
+    print("Conversion complete!")
+    print(f"Converted weights saved to {output_path}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert Moondream weights to HuggingFace format")
+    parser.add_argument(
+        "--input_path",
+        type=str,
+        required=True,
+        help="Path to original Moondream model directory or safetensors file",
+    )
+    parser.add_argument(
+        "--output_file",
+        type=str,
+        required=True,
+        help="Path to save converted HuggingFace safetensors file",
+    )
+    
+    # args = parser.parse_args()
+    
+    convert_moondream_weights_to_hf(
+        "moondream/", # args.input_path,
+        "moondream3/model.safetensors", # args.output_file,
+    )
+
+
+if __name__ == "__main__":
+    main()

image_processing_moondream3.py

@@ -0,0 +1,274 @@
+# coding=utf-8
+# Copyright 2024 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.
+"""Image processor class for Moondream3."""
+
+import math
+from typing import Optional, Union
+
+import torch
+import numpy as np
+
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
+from transformers.image_utils import (
+    ImageInput,
+    make_flat_list_of_images,
+    valid_images,
+    validate_kwargs,
+)
+from transformers.processing_utils import ImagesKwargs
+from transformers.utils import TensorType, logging
+from transformers.utils.import_utils import requires_backends
+
+
+logger = logging.get_logger(__name__)
+
+
+import PIL
+
+
+class Moondream3ImageProcessorKwargs(ImagesKwargs, total=False):
+    """
+    patch_size (`Union[dict[str, int], int]` *optional*, defaults to `{"height": 16, "width": 16}`):
+        Size of the patches in the model, used to calculate the output image size. Can be overridden by `patch_size` in the `preprocess` method.
+    """
+    pass
+
+def select_tiling(
+    height: int, width: int, crop_size: int, max_crops: int
+) -> tuple[int, int]:
+    """
+    Determine the optimal number of tiles to cover an image with overlapping crops.
+    """
+    if height <= crop_size or width <= crop_size:
+        return (1, 1)
+
+    # Minimum required tiles in each dimension
+    min_h = math.ceil(height / crop_size)
+    min_w = math.ceil(width / crop_size)
+
+    # If minimum required tiles exceed max_crops, return proportional distribution
+    if min_h * min_w > max_crops:
+        ratio = math.sqrt(max_crops / (min_h * min_w))
+        return (max(1, math.floor(min_h * ratio)), max(1, math.floor(min_w * ratio)))
+
+    # Perfect aspect-ratio tiles that satisfy max_crops
+    h_tiles = math.floor(math.sqrt(max_crops * height / width))
+    w_tiles = math.floor(math.sqrt(max_crops * width / height))
+
+    # Ensure we meet minimum tile requirements
+    h_tiles = max(h_tiles, min_h)
+    w_tiles = max(w_tiles, min_w)
+
+    # If we exceeded max_crops, scale down the larger dimension
+    if h_tiles * w_tiles > max_crops:
+        if w_tiles > h_tiles:
+            w_tiles = math.floor(max_crops / h_tiles)
+        else:
+            h_tiles = math.floor(max_crops / w_tiles)
+
+    return (max(1, h_tiles), max(1, w_tiles))
+
+
+def overlap_crop_image(
+    image: np.ndarray,
+    overlap_margin: int,
+    max_crops: int,
+    base_size: tuple[int, int] = (378, 378),
+    patch_size: int = 14,
+):
+    """
+    Process an image using an overlap-and-resize cropping strategy with margin handling.
+
+    This function takes an input image and creates multiple overlapping crops with
+    consistent margins. It produces:
+    1. A single global crop resized to base_size
+    2. Multiple overlapping local crops that maintain high resolution details
+    3. A patch ordering matrix that tracks correspondence between crops
+
+    The overlap strategy ensures:
+    - Smooth transitions between adjacent crops
+    - No loss of information at crop boundaries
+    - Proper handling of features that cross crop boundaries
+    - Consistent patch indexing across the full image
+
+    Args:
+        image (np.ndarray): Input image as numpy array with shape (H,W,C)
+        base_size (tuple[int,int]): Target size for crops, default (378,378)
+        patch_size (int): Size of patches in pixels, default 14
+        overlap_margin (int): Margin size in patch units, default 4
+        max_crops (int): Maximum number of crops allowed, default 12
+
+    Returns:
+        OverlapCropOutput: Dictionary containing:
+            - crops: A numpy array containing the global crop of the full image (index 0)
+                followed by the overlapping cropped regions (indices 1+)
+            - tiling: Tuple of (height,width) tile counts
+    """
+    original_h, original_w = image.shape[:2]
+
+    # Convert margin from patch units to pixels
+    margin_pixels = patch_size * overlap_margin
+    total_margin_pixels = margin_pixels * 2  # Both sides
+
+    # Calculate crop parameters
+    crop_patches = base_size[0] // patch_size  # patches per crop dimension
+    crop_window_patches = crop_patches - (2 * overlap_margin)  # usable patches
+    crop_window_size = crop_window_patches * patch_size  # usable size in pixels
+
+    # Determine tiling
+    tiling = select_tiling(
+        original_h - total_margin_pixels,
+        original_w - total_margin_pixels,
+        crop_window_size,
+        max_crops,
+    )
+
+    # Pre-allocate crops.
+    n_crops = tiling[0] * tiling[1] + 1  # 1 = global crop
+    crops = np.zeros(
+        (n_crops, base_size[0], base_size[1], image.shape[2]), dtype=np.uint8
+    )
+
+    # Resize image to fit tiling
+    target_size = (
+        tiling[0] * crop_window_size + total_margin_pixels,
+        tiling[1] * crop_window_size + total_margin_pixels,
+    )
+
+    # if HAS_VIPS:
+    #     # Convert to vips for resizing
+    #     vips_image = pyvips.Image.new_from_array(image)
+    #     scale_x = target_size[1] / image.shape[1]
+    #     scale_y = target_size[0] / image.shape[0]
+    #     resized = vips_image.resize(scale_x, vscale=scale_y)
+    #     image = resized.numpy()
+
+    #     # Create global crop
+    #     scale_x = base_size[1] / vips_image.width
+    #     scale_y = base_size[0] / vips_image.height
+    #     global_vips = vips_image.resize(scale_x, vscale=scale_y)
+    #     crops[0] = global_vips.numpy()
+    # else:
+        # Fallback to PIL
+    pil_img = PIL.Image.fromarray(image)
+    resized = pil_img.resize(
+        (int(target_size[1]), int(target_size[0])),
+        resample=PIL.Image.Resampling.LANCZOS,
+    )
+    image = np.asarray(resized)
+
+    # Create global crop
+    global_pil = pil_img.resize(
+        (int(base_size[1]), int(base_size[0])), resample=PIL.Image.Resampling.LANCZOS
+    )
+    crops[0] = np.asarray(global_pil)
+
+    for i in range(tiling[0]):
+        for j in range(tiling[1]):
+            # Calculate crop coordinates
+            y0 = i * crop_window_size
+            x0 = j * crop_window_size
+
+            # Extract crop with padding if needed
+            y_end = min(y0 + base_size[0], image.shape[0])
+            x_end = min(x0 + base_size[1], image.shape[1])
+
+            crop_region = image[y0:y_end, x0:x_end]
+            crops[
+                1 + i * tiling[1] + j, : crop_region.shape[0], : crop_region.shape[1]
+            ] = crop_region
+
+    return {"crops": crops, "tiling": tiling}
+
+def prepare_crops(image, max_crops=12, overlap_margin=4):
+    if isinstance(image, PIL.Image.Image):
+        np_image = np.array(image.convert("RGB"))
+    elif isinstance(image, torch.Tensor):
+        np_image = image.cpu().detach().numpy()
+    else:
+        np_image = image
+    overlap_crops = overlap_crop_image(
+        np_image, max_crops=max_crops, overlap_margin=overlap_margin
+    )
+    all_crops = overlap_crops["crops"]
+    all_crops = np.transpose(all_crops, (0, 3, 1, 2))
+    all_crops = (((all_crops / 255.0) - 0.5) / 0.5)
+    return all_crops.tolist(), overlap_crops["tiling"]
+
+class Moondream3ImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a Moondream3 image processor.
+    """
+
+    model_input_names = ["pixel_values", "image_sizes"]
+    valid_kwargs = Moondream3ImageProcessorKwargs
+
+    def __init__(
+        self,
+        max_crops: int = 12,
+        overlap_margin: int = 4,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.max_crops = max_crops
+        self.overlap_margin = overlap_margin
+        self._valid_processor_keys = [
+            "max_crops",
+            "overlap_margin",
+        ]
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        max_crops: Optional[int] = None,
+        overlap_margin: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs,
+    ) -> PIL.Image.Image:
+        """
+        Preprocess an image or batch of images.
+
+        Args:
+            images (`ImageInput`):
+                Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
+                passing in images with pixel values between 0 and 1, set `do_rescale=False`.
+            max_crops (`bool`, *optional*, defaults to `self.max_crops`):
+            overlap_margin (`dict[str, int]`, *optional*, defaults to `self.overlap_margin`):
+        """
+        overlap_margin = overlap_margin if overlap_margin is not None else self.overlap_margin
+        max_crops = max_crops if max_crops is not None else self.max_crops
+
+        validate_kwargs(captured_kwargs=kwargs.keys(), valid_processor_keys=self._valid_processor_keys)
+
+        images = self.fetch_images(images)
+        images = make_flat_list_of_images(images)
+
+        if not valid_images(images[0]):
+            raise ValueError("Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, or torch.Tensor")
+
+        batch_images = []
+        batch_tiling = []
+        for image in images:
+            pixel_values, tiling = prepare_crops(image, max_crops=max_crops, overlap_margin=overlap_margin)
+            batch_images.append(pixel_values)
+            batch_tiling.append(tiling)
+
+
+        return BatchFeature(
+            data={"pixel_values": batch_images, "tiling": batch_tiling}, tensor_type=return_tensors
+        )
+
+
+__all__ = ["Moondream3ImageProcessor"]

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1b992dcf85d87a027b8af4317ceb860768cba70f921d8ab60a2de2d49d8ccf3
+size 18537874416

modeling_moondream3.py

@@ -0,0 +1,1054 @@
+# coding=utf-8
+# Copyright 2025 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.
+
+from typing import Callable, Optional, Tuple, Union
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from PIL import Image
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.processing_utils import Unpack
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.generation import GenerationMixin
+from transformers.utils import logging, TransformersKwargs
+from .configuration_moondream3 import Moondream3Config, Moondream3TextConfig, Moondream3VisionConfig, Moondream3RegionConfig
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "Moondream3Config"
+
+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.
+    """
+    rot_dim = cos.shape[-1]
+
+    def rotate_prefix(x):
+        x_rot = x[..., :rot_dim]
+        x_pass = x[..., rot_dim:]
+        x_rot = (x_rot * cos) + (rotate_half(x_rot) * sin)
+        return torch.cat([x_rot, x_pass], dim=-1)
+
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    return rotate_prefix(q), rotate_prefix(k)
+
+class Moondream3RotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: Moondream3Config, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_type = self.config.rope_parameters["rope_type"]
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = inv_freq
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: Optional[Moondream3Config] = None,
+        device: Optional["torch.device"] = None,
+        seq_len: Optional[int] = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Computes the inverse frequencies according to the original RoPE implementation
+        Args:
+            config ([`~transformers.PreTrainedConfig`]):
+                The model configuration.
+            device (`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (`int`, *optional*):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+        """
+        base = config.rope_parameters["rope_theta"]
+        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        dim = dim // 2
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        # Compute the inverse frequencies
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.float).to(device=device)[: (dim // 2)] / dim)
+        )
+        return inv_freq, attention_factor
+
+    @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)
+
+
+class Moondream3Attention(nn.Module):
+    def __init__(self, config: Moondream3TextConfig | Moondream3VisionConfig, layer_idx: Optional[int] = None, use_tau: bool = True):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = getattr(config, "head_dim", self.hidden_size // self.num_heads)
+        self.num_key_value_heads = getattr(config, "num_key_value_heads", self.num_heads)
+        attention_bias = config.attention_bias
+        self.attention_dropout = config.attention_dropout
+
+        # Initialize parameters based on config type
+        if isinstance(config, Moondream3TextConfig):
+            self.is_causal = True
+        elif isinstance(config, Moondream3VisionConfig):  # Moondream3VisionConfig
+            self.is_causal = False
+        else:
+            raise TypeError(f"Unsupported config type: {type(config)}")
+        
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.use_tau = use_tau
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=attention_bias)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=attention_bias)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=attention_bias)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=attention_bias)
+        
+        # Tau parameters for token-level attention (from original Moondream) - only for text model
+        if self.use_tau:
+            # In original, tau weights are (n_heads, qkv_dim) where qkv_dim is the combined QKV dimension
+            qkv_dim = self.num_heads * self.head_dim + 2 * self.num_key_value_heads * self.head_dim
+            self.tau_wq = nn.Linear(qkv_dim, self.num_heads, bias=False)
+            self.tau_wv = nn.Linear(qkv_dim, self.num_heads, bias=False)
+            self.tau_alpha = nn.Parameter(torch.empty(self.num_heads))
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
+        input_shape = hidden_states.shape[:-1]
+        bsz, q_len, _ = hidden_states.size()
+
+        # Get qkv combined for tau (before splitting)
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+        if self.use_tau:
+            qkv_out = torch.cat([query_states, key_states, value_states], dim=-1)
+            tok_feat = F.gelu(qkv_out)
+            tok_q = torch.tanh(self.tau_wq(tok_feat)).permute(0, 2, 1)  # (bsz, n_heads, seq_len)
+            tok_v = torch.tanh(self.tau_wv(tok_feat)).permute(0, 2, 1)  # (bsz, n_heads, seq_len)
+
+            pos = position_ids.to(tok_q.dtype) + 1
+            alpha = self.tau_alpha.to(tok_q.dtype)
+            tau_pos = 1 + (torch.sigmoid(alpha[None, :, None] * pos[:, None, :].log()) - 0.5)  # (n_heads, seq_len)
+            tau_q = (tok_q + tau_pos).unsqueeze(-1)  # (bsz, n_heads, seq_len, 1)
+            tau_v = (tok_v + tau_pos).unsqueeze(-1)  # (bsz, n_heads, seq_len, 1)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+
+        if self.use_tau:
+            query_states = query_states * tau_q
+
+            if self.num_key_value_groups > 1:
+                tau_v_repeated = tau_v.repeat(1, self.num_key_value_groups, 1, 1)[:, :self.num_key_value_heads, :, :]
+            else:
+                tau_v_repeated = tau_v
+            value_states = value_states * tau_v_repeated
+
+        cos, sin = None, None
+        if position_embeddings is not None:
+            cos, sin = position_embeddings
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_values is not None:
+
+            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)
+
+        attn_output, attn_weights = ALL_ATTENTION_FUNCTIONS["sdpa"](
+            self,
+            query_states,
+            key_states, 
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights
+
+class Moondream3MLP(nn.Module):
+    def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str = "gelu_pytorch_tanh", out_size: int | None = None, gated: bool = False, bias: bool = True):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.out_size = self.hidden_size if out_size is None else out_size
+        self.hidden_act = hidden_act
+        self.gated = gated
+        # Ungated MLP: up_proj and down_proj following HF conventions
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.out_size, bias=bias)
+        self.gate_proj = None
+        if self.gated:
+            self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
+        self.act_fn = ACT2FN[self.hidden_act]
+
+    def forward(self, x) -> torch.Tensor:
+        if self.gated:
+            h = self.act_fn(self.up_proj(x))
+            g = self.gate_proj(x)
+            x = h * (g + 1)
+        else:
+            x = self.act_fn(self.up_proj(x))
+        return self.down_proj(x)
+
+
+class Moondream3SparseMoeBlock(nn.Module):
+    def __init__(self, config: Moondream3TextConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.moe_intermediate_size = config.moe_intermediate_size
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+
+        self.gate = nn.Linear(self.hidden_size, self.num_experts, bias=True)
+        self.experts = nn.ModuleList([Moondream3MLP(hidden_size=self.hidden_size, intermediate_size=self.moe_intermediate_size, hidden_act="gelu", gated=True, bias=False) for _ in range(self.num_experts)])
+
+    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        router_logits: torch.Tensor = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float32)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            if top_x.shape[0] == 0:
+                continue
+
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits
+
+
+class Moondream3DecoderLayer(nn.Module):
+    def __init__(self, config: Moondream3TextConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.self_attn = Moondream3Attention(config, layer_idx, use_tau=True)
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.is_moe_layer = layer_idx >= config.moe_start_layer
+        if self.is_moe_layer:
+            self.mlp = Moondream3SparseMoeBlock(config)
+        else:
+            self.mlp = Moondream3MLP(self.hidden_size, self.intermediate_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,
+        output_attentions: bool = False,
+        output_router_logits: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple:
+        residual = hidden_states
+
+        # Apply layer norm like original
+        l_in = self.input_layernorm(hidden_states)
+
+        # Attention
+        hidden_states_attn, self_attn_weights = self.self_attn(
+            hidden_states=l_in,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs
+        )
+
+        # MLP
+        if self.is_moe_layer:
+            hidden_states_mlp, router_logits = self.mlp(l_in)
+        else:
+            hidden_states_mlp = self.mlp(l_in)
+            router_logits = None
+
+        # Add both attention and MLP to residual like original
+        hidden_states = residual + hidden_states_attn + hidden_states_mlp
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+class Moondream3PreTrainedModel(PreTrainedModel):
+    config_class = Moondream3Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Moondream3DecoderLayer", "Moondream3SparseMoeBlock"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    def _init_weights(self, module):
+        # Use text_config initializer_range if available, otherwise use default
+        if hasattr(self.config, 'text_config') and hasattr(self.config.text_config, 'initializer_range'):
+            std = self.config.text_config.initializer_range
+        elif hasattr(self.config, 'initializer_range'):
+            std = self.config.initializer_range
+        else:
+            std = 0.02  # Default initialization range
+            
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+class Moondream3TextModel(Moondream3PreTrainedModel):
+    config_class = Moondream3TextConfig
+
+    def __init__(self, config: Moondream3TextConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id if hasattr(config, "pad_token_id") else 0
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [Moondream3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Moondream3RotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+
+        self.post_init()
+
+    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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        hidden_states = inputs_embeds
+        batch_size = hidden_states.shape[0]
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        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)
+
+        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_logits = () if output_router_logits else None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    output_router_logits,
+                    use_cache,
+                    cache_position,
+                    position_embeddings
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_values=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            if output_router_logits and layer_outputs[-1] is not None:
+                all_router_logits += (layer_outputs[-1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = None
+        if use_cache:
+            next_cache = past_key_values
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
+                if v is not None
+            )
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class Moondream3VisionPatchEmbeddings(nn.Module):
+    def __init__(self, config: Moondream3VisionConfig):
+        super().__init__()
+        self.patch_size = config.patch_size
+        self.num_channels = config.in_channels
+        self.hidden_size = config.hidden_size
+        self.crop_size = config.crop_size
+        self.patch_size = config.patch_size
+        self.grid_size = self.crop_size // self.patch_size
+        self.num_patches = self.grid_size * self.grid_size
+
+        self.projection = nn.Linear(self.patch_size * self.patch_size * self.num_channels, self.hidden_size, bias=True)
+        self.position_embeddings = nn.Parameter(torch.zeros(1, self.num_patches, config.hidden_size))
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        B, C, H, W = pixel_values.shape
+        P1 = P2 = self.patch_size
+
+        x = pixel_values.reshape(B, C, H // P1, P1, W // P2, P2)
+
+        x = x.permute(0, 2, 4, 1, 3, 5)
+
+        x = x.reshape(B, (H // P1) * (W // P2), C * P1 * P2)
+
+        x = self.projection(x)
+        return x + self.position_embeddings
+
+class Moondream3VisionEncoderLayer(nn.Module):
+    def __init__(self, config: Moondream3VisionConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.layer_idx = layer_idx
+        
+        self.self_attn = Moondream3Attention(config, layer_idx=self.layer_idx, use_tau=False)
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=1e-5)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=1e-5)
+        self.mlp = Moondream3MLP(hidden_size=self.hidden_size, intermediate_size=self.intermediate_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states, _ = self.self_attn(hidden_states=hidden_states)
+        hidden_states = residual + hidden_states
+
+        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 Moondream3VisionModel(Moondream3PreTrainedModel):
+    config_class = Moondream3VisionConfig
+    main_input_name = "pixel_values"
+    _no_split_modules = ["Moondream3VisionEncoderLayer"]
+
+    def __init__(self, config: Moondream3VisionConfig):
+        super().__init__(config)
+        self.config = config
+        self.hidden_size = self.config.hidden_size
+        self.num_hidden_layers = self.config.num_hidden_layers
+        self.proj_inner_dim = self.config.proj_inner_dim
+        self.proj_out_dim = self.config.proj_out_dim
+
+        self.embeddings = Moondream3VisionPatchEmbeddings(config)
+        self.layers = nn.ModuleList([Moondream3VisionEncoderLayer(config,layer_idx) for layer_idx in range(self.num_hidden_layers)])
+        self.post_layernorm = nn.LayerNorm(self.hidden_size, eps=1e-5)
+        self.vision_projection = Moondream3MLP(hidden_size=self.hidden_size * 2, intermediate_size=self.proj_inner_dim, out_size=self.proj_out_dim)        
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def _reconstruct_from_crops(
+        self,
+        crops: torch.Tensor,
+        tiling: tuple[int, int],
+        overlap_margin: int = 4,
+        patch_size: int = 14,
+    ) -> torch.Tensor:
+        """
+        Reconstruct the original image from overlapping crops into a single seamless image.
+
+        Takes a list of overlapping image crops along with their positional metadata and
+        reconstructs them into a single coherent image by carefully stitching together
+        non-overlapping regions. Handles both numpy arrays and PyTorch tensors.
+
+        Args:
+            crops: List of image crops as numpy arrays or PyTorch tensors with shape
+                (H,W,C)
+            tiling: Tuple of (height,width) indicating crop grid layout
+            patch_size: Size in pixels of each patch, default 14
+            overlap_margin: Number of overlapping patches on each edge, default 4
+
+        Returns:
+            Reconstructed image as numpy array or PyTorch tensor matching input type,
+            with shape (H,W,C) where H,W are the original image dimensions
+        """
+        if isinstance(tiling, torch.Tensor):
+            tiling_h, tiling_w = tiling[0].item(), tiling[1].item()
+        else:
+            tiling_h, tiling_w = tiling
+        tiling_h, tiling_w = int(tiling_h), int(tiling_w)
+        crop_height, crop_width = crops[0].shape[:2]
+        margin_pixels = overlap_margin * patch_size
+
+        # Calculate output size (only adding margins once)
+        output_h = (crop_height - 2 * margin_pixels) * tiling_h + 2 * margin_pixels
+        output_w = (crop_width - 2 * margin_pixels) * tiling_w + 2 * margin_pixels
+        reconstructed = torch.zeros(
+            (output_h, output_w, crops[0].shape[2]),
+            device=crops[0].device,
+            dtype=crops[0].dtype,
+        )
+
+        for i, crop in enumerate(crops):
+            tile_y = i // tiling_w
+            tile_x = i % tiling_w
+
+            # For each tile, determine which part to keep
+            # Keep left margin only for first column
+            x_start = 0 if tile_x == 0 else margin_pixels
+            # Keep right margin only for last column
+            x_end = crop_width if tile_x == tiling_w - 1 else crop_width - margin_pixels
+            # Keep top margin only for first row
+            y_start = 0 if tile_y == 0 else margin_pixels
+            # Keep bottom margin only for last row
+            y_end = crop_height if tile_y == tiling_h - 1 else crop_height - margin_pixels
+
+            # Calculate where this piece belongs in the output
+            out_x = tile_x * (crop_width - 2 * margin_pixels)
+            out_y = tile_y * (crop_height - 2 * margin_pixels)
+
+            # Place the piece
+            reconstructed[
+                out_y + y_start : out_y + y_end, out_x + x_start : out_x + x_end
+            ] = crop[y_start:y_end, x_start:x_end]
+
+        return reconstructed
+
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        tiling: Tuple[int,int],
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        batch_size, num_crops = pixel_values.shape[:2]
+        # flatten batch_size and num_crops into same dim
+        pixel_values = pixel_values.view(-1, *pixel_values.shape[2:])
+        hidden_states: torch.Tensor = self.embeddings(pixel_values)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        for encoder_layer in self.layers:
+            if output_hidden_states and all_hidden_states is not None:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(encoder_layer.__call__, hidden_states)
+            else:
+                layer_outputs = encoder_layer(hidden_states)
+
+            hidden_states = layer_outputs
+
+        hidden_states = self.post_layernorm(hidden_states)
+        # B, _, _
+
+        # back out into batch_size, num_crops
+        hidden_states = hidden_states.view(batch_size, num_crops, *hidden_states.shape[1:])
+        outputs = []
+        for b in range(batch_size):
+            hs = hidden_states[b]
+            t = tiling[b]
+
+            global_features = hs[0]
+            local_features = hs[1:].view(
+                -1,
+                self.num_hidden_layers,
+                self.num_hidden_layers,
+                self.hidden_size,
+            )
+
+            reconstructed = self._reconstruct_from_crops(
+                local_features,
+                t,
+                patch_size=1,
+                overlap_margin=self.config.overlap_margin,
+            )
+
+            reconstructed = reconstructed.permute(2, 0, 1)
+            reconstructed = F.adaptive_avg_pool2d(
+                reconstructed, output_size=(self.num_hidden_layers, self.num_hidden_layers)
+            )
+            reconstructed = reconstructed.permute(1, 2, 0).view(729, self.hidden_size)
+            final_features = torch.cat([global_features, reconstructed], dim=-1)
+            outputs.append(final_features)
+        output = torch.stack(outputs, 0)
+
+        hidden_states = self.vision_projection(output)
+
+        if output_hidden_states and all_hidden_states is not None:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+        )
+
+class Moondream3RegionEncoder(nn.Module):
+    def __init__(self, config: Moondream3RegionConfig):
+        super().__init__()
+        self.coord_encoder = nn.Linear(config.coord_feat_dim, config.hidden_size)
+        self.size_encoder = nn.Linear(config.size_feat_dim, config.hidden_size)
+        
+        coord_freq = torch.randn(config.coord_feat_dim // 2, 1) * 10.0
+        size_freq = torch.randn(config.size_feat_dim // 2, 2) * 10.0
+        self.register_buffer("coord_freq", coord_freq.T)
+        self.register_buffer("size_freq", size_freq.T)
+
+    def fourier_features(self, x: torch.Tensor, w: torch.Tensor) -> torch.Tensor:
+        x_proj = torch.matmul(x, w) * 2 * torch.pi
+        return torch.cat([torch.sin(x_proj), torch.cos(x_proj)], dim=-1)
+
+    def encode_coordinate(self, coord: torch.Tensor) -> torch.Tensor:
+        fourier_features = self.fourier_features(coord, self.coord_freq)
+        return self.coord_encoder(fourier_features)
+
+    def encode_size(self, size: torch.Tensor) -> torch.Tensor:
+        fourier_features = self.fourier_features(size, self.size_freq)
+        return self.size_encoder(fourier_features)
+
+class Moondream3RegionDecoder(nn.Module):
+    def __init__(self, config: Moondream3RegionConfig):
+        super().__init__()
+        self.coord_decoder = nn.Linear(config.hidden_size, config.coord_out_dim)
+        self.size_decoder = nn.Linear(config.hidden_size, config.size_out_dim)
+
+    def decode_coordinate(self, hidden_state: torch.Tensor) -> torch.Tensor:
+        return self.coord_decoder(hidden_state)
+
+    def decode_size(self, hidden_state: torch.Tensor) -> torch.Tensor:
+        return self.size_decoder(hidden_state)
+
+class Moondream3Model(Moondream3PreTrainedModel):
+    def __init__(self, config: Moondream3Config):
+        super().__init__(config)
+        self.text_model = Moondream3TextModel(config.text_config)
+        self.vision_model = Moondream3VisionModel(config.vision_config)
+        self.vocab_size = config.text_config.vocab_size
+        
+        self.region_encoder = Moondream3RegionEncoder(config.region_config)
+        self.region_decoder = Moondream3RegionDecoder(config.region_config)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.text_model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.text_model.embed_tokens = value
+
+    def set_decoder(self, decoder):
+        self.text_model = decoder
+
+    def get_decoder(self):
+        return self.text_model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: torch.FloatTensor = None,
+        tiling: Tuple[int,int] = 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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: int = 0,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is not None) == (inputs_embeds is not None):
+            raise ValueError("Provide exactly one of input_ids or inputs_embeds.")
+
+        if not ((pixel_values is not None) ^ (tiling is None)):
+            raise ValueError("You must specify both pixel_values and tiling")
+
+        # Case A: inputs_embeds provided -> assume it already contains BOS+image+text in correct order.
+        if inputs_embeds is not None and (pixel_values is not None or tiling is not None):
+            raise ValueError(
+                "When inputs_embeds is provided, do not pass pixel_values/tiling; "
+                "inputs_embeds must already include BOS+image(+text)."
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds: torch.Tensor = self.text_model.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.Tensor = 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)
+
+        causal_mask = create_causal_mask(
+            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,
+        )
+
+        if pixel_values is not None and input_ids.shape[-1] > 1:
+            # Vision embeds
+            pixel_values = pixel_values.to(dtype=self.vision_model.embeddings.projection.weight.dtype)
+            image_embeds = self.vision_model(pixel_values, tiling=tiling)["last_hidden_state"]  # [B,P,D]
+            prefix = inputs_embeds[:, :1, :]                      # keep the first token
+            suffix = inputs_embeds[:, 1 + image_embeds.shape[1] :, :]                 # keep the rest after the image span
+            inputs_embeds = torch.cat([prefix, image_embeds, suffix], dim=1)
+
+            # N/A when doing BSZ 1 since create_causal_mask returns None in the case since theres no padding tokens
+            if causal_mask is not None:
+                img_len = image_embeds.shape[1]
+                causal_mask[:, :, :1 + img_len, :1 + img_len] = True
+                causal_mask[:, :, :1 + img_len, 1 + img_len:] = False
+
+        outputs = self.text_model(
+            input_ids=None,
+            attention_mask=causal_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+            cache_position=cache_position,
+        )
+
+        if not return_dict:
+            return tuple(v for v in [
+                outputs.last_hidden_state,
+                getattr(outputs, "past_key_values", None),
+                getattr(outputs, "hidden_states", None),
+                getattr(outputs, "attentions", None),
+            ] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=outputs.last_hidden_state,
+            past_key_values=getattr(outputs, "past_key_values", None),
+            hidden_states=getattr(outputs, "hidden_states", None),
+            attentions=getattr(outputs, "attentions", None),
+        )
+
+class Moondream3ForConditionalGeneration(Moondream3PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: Moondream3Config):
+        super().__init__(config)
+        self.model = Moondream3Model(config)
+        self.vocab_size = config.text_config.vocab_size
+        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=True)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.text_model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.text_model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model.text_model = decoder
+
+    def get_decoder(self):
+        return self.model.text_model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: torch.FloatTensor = None,
+        tiling: 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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: int = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        # Get hidden states from the base model (it already builds the multimodal prefix)
+        model_outputs = self.model(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            tiling=tiling,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=None,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+        )
+
+        hidden_states = model_outputs.last_hidden_state  # [B, T, D]
+
+        # Compute logits; only keep the tail if requested
+        if isinstance(logits_to_keep, int) and logits_to_keep > 0:
+            hs = hidden_states[:, -logits_to_keep:, :]
+        elif isinstance(logits_to_keep, slice):
+            hs = hidden_states[:, logits_to_keep, :]
+        else:
+            hs = hidden_states
+
+        logits = self.lm_head(hs)  # [B, T', V]
+
+        loss = None
+        if labels is not None:
+            # Shift if your training uses standard LM convention; here we assume labels aligned with hs
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.vocab_size)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=getattr(model_outputs, "past_key_values", None),
+            hidden_states=getattr(model_outputs, "hidden_states", None),
+            attentions=getattr(model_outputs, "attentions", None),
+        )
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+
+__all__ = [
+    "Moondream3Config",
+    "Moondream3TextConfig",
+    "Moondream3VisionConfig",
+    "Moondream3RegionConfig",
+    "Moondream3PreTrainedModel",
+    "Moondream3Model",
+    "Moondream3TextModel",
+    "Moondream3VisionModel",
+    "Moondream3ForConditionalGeneration",
+]

modeling_moondream3_fusedmoe.py

@@ -0,0 +1,301 @@
+# coding=utf-8
+# Copyright 2025 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.
+
+from typing import Callable, Optional, Tuple, Union
+
+import numpy as np
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from PIL import Image
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.processing_utils import Unpack
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel, load_state_dict
+from transformers.generation import GenerationMixin
+from transformers.utils import logging, TransformersKwargs
+
+from .moondream3_moe_fused.moe_fused_linear import MoeFusedLinear
+from .moondream3_moe_fused.kernels.indexing import get_expert_counts_and_idx
+from .configuration_moondream3 import Moondream3Config, Moondream3TextConfig, Moondream3VisionConfig, Moondream3RegionConfig
+
+from . import modeling_moondream3
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "Moondream3Config"
+
+class Moondream3FusedSparseMoeBlock(nn.Module):
+    def __init__(self, config: Moondream3TextConfig) -> None:
+        super().__init__()
+        self.num_experts = config.num_experts
+        self.num_selected = config.num_experts_per_tok
+        self.hidden_size = config.hidden_size
+        self.moe_intermediate_size = config.moe_intermediate_size
+
+        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
+        self.gate_proj = MoeFusedLinear(self.hidden_size, self.moe_intermediate_size, config.num_experts)
+        self.up_proj = MoeFusedLinear(self.hidden_size, self.moe_intermediate_size, config.num_experts)
+        self.down_proj = MoeFusedLinear(self.moe_intermediate_size, self.hidden_size, config.num_experts)
+
+    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        M = batch_size * sequence_length
+
+        hidden_states = hidden_states.view(M, hidden_dim)
+        # router_logits: (M, num_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float32)
+        # routing_weights, selected_experts: (M, num_selected)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.num_selected, dim=-1)
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        hidden_states = hidden_states.unsqueeze(1).expand(M, self.num_selected, hidden_dim)
+        # hidden_states must be contiguous
+        hidden_states = hidden_states.reshape(M * self.num_selected, hidden_dim)
+        selected_experts = selected_experts.view(M * self.num_selected)
+
+        # Sort selected_experts and hidden_states for better memory coalescence of weight
+        # It's possible to fuse a sort and a MoeFusedLinear layer, but for now we separate them for clarity
+        m_sizes, sort_idx, inv_sort_idx = get_expert_counts_and_idx(selected_experts, self.num_experts)
+        hidden_states = hidden_states[sort_idx]
+
+        # It's possible to fuse gate_h and up_h, but this affects the shape of LoRA
+        gate_h = self.gate_proj(hidden_states, m_sizes)
+        up_h = self.up_proj(hidden_states, m_sizes)
+        hidden_states = F.gelu(up_h) * (gate_h + 1)
+        del gate_h, up_h
+        hidden_states = self.down_proj(hidden_states, m_sizes)
+
+        hidden_states = hidden_states[inv_sort_idx]
+
+        hidden_states = hidden_states.view(M, self.num_selected, hidden_dim)
+        hidden_states = torch.einsum("beo,be->bo", hidden_states, routing_weights)
+
+        hidden_states = hidden_states.view(batch_size, sequence_length, hidden_dim)
+        return hidden_states, router_logits
+
+modeling_moondream3.Moondream3SparseMoeBlock = Moondream3FusedSparseMoeBlock
+from .modeling_moondream3 import Moondream3Config, Moondream3TextConfig, Moondream3VisionConfig, Moondream3RegionConfig, Moondream3PreTrainedModel, Moondream3Model, Moondream3TextModel, Moondream3VisionModel, Moondream3ForConditionalGeneration
+
+
+class Moondream3ForConditionalGeneration(Moondream3PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: Moondream3Config):
+        super().__init__(config)
+        self.model = Moondream3Model(config)
+        self.vocab_size = config.text_config.vocab_size
+        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=True)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.text_model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.text_model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model.text_model = decoder
+
+    def get_decoder(self):
+        return self.model.text_model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        pixel_values: torch.FloatTensor = None,
+        tiling: 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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: int = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        # Get hidden states from the base model (it already builds the multimodal prefix)
+        model_outputs = self.model(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            tiling=tiling,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=None,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+        )
+
+        hidden_states = model_outputs.last_hidden_state  # [B, T, D]
+
+        # Compute logits; only keep the tail if requested
+        if isinstance(logits_to_keep, int) and logits_to_keep > 0:
+            hs = hidden_states[:, -logits_to_keep:, :]
+        elif isinstance(logits_to_keep, slice):
+            hs = hidden_states[:, logits_to_keep, :]
+        else:
+            hs = hidden_states
+
+        logits = self.lm_head(hs)  # [B, T', V]
+
+        loss = None
+        if labels is not None:
+            # Shift if your training uses standard LM convention; here we assume labels aligned with hs
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.vocab_size)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=getattr(model_outputs, "past_key_values", None),
+            hidden_states=getattr(model_outputs, "hidden_states", None),
+            attentions=getattr(model_outputs, "attentions", None),
+        )
+
+    @classmethod
+    def _load_pretrained_model(
+        cls,
+        model: "PreTrainedModel",
+        state_dict: Optional[dict],
+        checkpoint_files: Optional[list[str]],
+        pretrained_model_name_or_path,
+        weights_only: bool = True,
+        **kwargs,
+    ):
+        if checkpoint_files is not None:
+            state_dict = {}
+            for file in checkpoint_files:
+                sd = load_state_dict(file, map_location="cpu", weights_only=weights_only)
+                for key, value in sd.items():
+                    state_dict[key] = value
+
+            from collections import defaultdict
+
+            moe_layer_experts = defaultdict(set)
+
+            for key in state_dict.keys():
+                if key.startswith("model.text_model.layers."):
+                    parts = key.split(".")
+                    # Expected: model.text_model.layers.{layer}.mlp.experts.{expert_id}.down_proj.weight
+                    if len(parts) > 6 and parts[5] == "experts" and parts[3].isdigit() and parts[6].isdigit():
+                        layer_idx = int(parts[3])
+                        expert_idx = int(parts[6])
+                        moe_layer_experts[layer_idx].add(expert_idx)
+
+            moe_layers = {layer: len(experts) for layer, experts in moe_layer_experts.items()}
+            for layer_idx, num_experts in moe_layers.items():
+                state_dict[f"model.text_model.layers.{layer_idx}.mlp.down_proj.weight"] = torch.stack(
+                    [
+                        state_dict[f"model.text_model.layers.{layer_idx}.mlp.experts.{i}.down_proj.weight"] for i in range(num_experts)
+                    ]
+                )
+                for i in range(num_experts):
+                    del state_dict[f"model.text_model.layers.{layer_idx}.mlp.experts.{i}.down_proj.weight"]
+
+                state_dict[f"model.text_model.layers.{layer_idx}.mlp.up_proj.weight"] = torch.stack(
+                    [
+                        state_dict[f"model.text_model.layers.{layer_idx}.mlp.experts.{i}.up_proj.weight"] for i in range(num_experts)
+                    ]
+                )
+                for i in range(num_experts):
+                    del state_dict[f"model.text_model.layers.{layer_idx}.mlp.experts.{i}.up_proj.weight"]
+
+                state_dict[f"model.text_model.layers.{layer_idx}.mlp.gate_proj.weight"] = torch.stack(
+                    [
+                        state_dict[f"model.text_model.layers.{layer_idx}.mlp.experts.{i}.gate_proj.weight"] for i in range(num_experts)
+                    ]
+                )
+                for i in range(num_experts):
+                    del state_dict[f"model.text_model.layers.{layer_idx}.mlp.experts.{i}.gate_proj.weight"]
+            checkpoint_files = None
+
+        model, missing_keys, unexpected_keys, mismatched_keys, disk_offload_index, error_msgs = super()._load_pretrained_model(
+            model,
+            state_dict,
+            checkpoint_files,
+            pretrained_model_name_or_path,
+            **kwargs,
+        )
+        return model, missing_keys, unexpected_keys, mismatched_keys, disk_offload_index, error_msgs
+
+    def _fix_state_dict_keys_on_save(self, state_dict: dict):
+        for layer_idx in range(self.config.text_config.moe_start_layer, self.config.text_config.num_hidden_layers):
+            layer_key = f"model.text_model.layers.{layer_idx}"
+            tensor = state_dict.pop(f"{layer_key}.mlp.down_proj.weight").cpu()
+            for i, t in enumerate(torch.unbind(tensor)):
+                base_key = f"{layer_key}.mlp.experts.{i}"
+                state_dict[f"{base_key}.down_proj.weight"] = t.contiguous()
+
+            tensor = state_dict.pop(f"{layer_key}.mlp.up_proj.weight").cpu()
+            for i, t in enumerate(torch.unbind(tensor)):
+                base_key = f"{layer_key}.mlp.experts.{i}"
+                state_dict[f"{base_key}.up_proj.weight"] = t.contiguous()
+
+            tensor = state_dict.pop(f"{layer_key}.mlp.gate_proj.weight").cpu()
+            for i, t in enumerate(torch.unbind(tensor)):
+                base_key = f"{layer_key}.mlp.experts.{i}"
+                state_dict[f"{base_key}.gate_proj.weight"] = t.contiguous()
+        return state_dict
+
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+
+__all__ = [
+    "Moondream3Config",
+    "Moondream3TextConfig",
+    "Moondream3VisionConfig",
+    "Moondream3RegionConfig",
+    "Moondream3PreTrainedModel",
+    "Moondream3Model",
+    "Moondream3TextModel",
+    "Moondream3VisionModel",
+    "Moondream3ForConditionalGeneration",
+]

moondream3_moe_fused/functional.py

@@ -0,0 +1,79 @@
+import torch
+
+from .grouped_gemm.interface import grouped_gemm
+
+
+# output[b, o] = sum_i weight[selected_experts[b], o, i] * input[b, i]
+def _moe_fused_linear_naive_fwd(
+    input: torch.Tensor, weight: torch.Tensor, selected_experts: torch.Tensor
+) -> torch.Tensor:
+    """
+    Computes a MoE linear operation.
+
+    The operation is defined as:
+    `output[b, o] = sum_i weight[selected_experts[b], o, i] * input[b, i]`
+
+    Args:
+        input (`torch.FloatTensor`): input tensor of shape `(batch_size, in_features)`.
+        weight (`torch.FloatTensor`): weight tensor of shape `(num_experts, out_features, in_features)`.
+        selected_experts (`torch.LongTensor`): tensor of selected expert indices in shape `(batch_size,)`.
+            Each element is in the range `[0, num_experts)`.
+
+    Returns:
+        output (`torch.FloatTensor`): output tensor of shape `(batch_size, out_features)`.
+    """
+    batch_size, in_features = input.shape
+    num_experts, out_features, _ = weight.shape
+
+    output = torch.empty(batch_size, out_features, device=input.device, dtype=input.dtype)
+    for b in range(batch_size):
+        _weight = weight[selected_experts[b], :, :]
+        _input = input[b, :]
+        output[b, :] = _weight @ _input
+    return output
+
+
+# grad_input[b, i] = sum_o weight[selected_experts[b], o, i] * grad_output[b, o]
+def _moe_fused_linear_naive_bwd_input(
+    grad_output: torch.Tensor, input: torch.Tensor, weight: torch.Tensor, selected_experts: torch.Tensor
+) -> torch.Tensor:
+    batch_size, in_features = input.shape
+    num_experts, out_features, _ = weight.shape
+
+    grad_input = torch.empty_like(input)
+    for b in range(batch_size):
+        _weight = weight[selected_experts[b], :, :]
+        _grad_output = grad_output[b, :]
+        grad_input[b, :] = _grad_output @ _weight
+    return grad_input
+
+
+# grad_weight[e, o, i] = sum_b if(selected_experts[b] == e) grad_output[b, o] * input[b, i]
+def _moe_fused_linear_naive_bwd_weight(
+    grad_output: torch.Tensor, input: torch.Tensor, weight: torch.Tensor, selected_experts: torch.Tensor
+) -> torch.Tensor:
+    batch_size, in_features = input.shape
+    num_experts, out_features, _ = weight.shape
+
+    grad_weight = torch.zeros_like(weight)
+    for b in range(batch_size):
+        grad_weight[selected_experts[b], :, :] += grad_output[b, :, None] * input[b, None, :]
+    return grad_weight
+
+
+def moe_fused_linear(input: torch.Tensor, weight: torch.Tensor, m_sizes: torch.Tensor) -> torch.Tensor:
+    """
+    Computes a MoE linear operation using grouped GEMM.
+
+    The operation is defined as:
+    `output[b, o] = sum_i weight[selected_experts[b], o, i] * input[b, i]`
+
+    Args:
+        input (`torch.FloatTensor`): input tensor of shape `(batch_size, in_features)`.
+        weight (`torch.FloatTensor`): weight tensor of shape `(num_experts, out_features, in_features)`.
+        m_sizes (`torch.LongTensor`): counts of selected experts in shape `(num_experts)`. Should sum to `batch_size`.
+
+    Returns:
+        output (`torch.FloatTensor`): output tensor of shape `(batch_size, out_features)`.
+    """
+    return grouped_gemm(input, weight, m_sizes)

moondream3_moe_fused/grouped_gemm/LICENSE

@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published
+    by the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<https://www.gnu.org/licenses/>.

moondream3_moe_fused/grouped_gemm/autotuning.py

@@ -0,0 +1,111 @@
+import os
+from itertools import product
+from typing import Any
+
+import torch
+import triton
+
+
+DEFAULT_M_BLOCK_SIZES = [16, 32, 64, 128, 256]
+DEFAULT_N_BLOCK_SIZES = [16, 32, 64, 128, 256]
+DEFAULT_K_BLOCK_SIZES = [16, 32, 64, 128, 256]
+DEFAULT_NUM_WARPS = [4, 8]
+DEFAULT_NUM_STAGES = [3, 4, 5, 6]
+
+
+def get_num_sms() -> int:
+    return torch.cuda.get_device_properties("cuda").multi_processor_count
+
+
+def get_autotune_configs() -> list[triton.Config]:
+    return [
+        triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_K': 32, 'BLOCK_SIZE_M': 64}, num_stages=2, num_warps=4),
+    ]
+
+    configs = []
+    for m, n, k, w, s in product(
+        DEFAULT_M_BLOCK_SIZES,
+        DEFAULT_N_BLOCK_SIZES,
+        DEFAULT_K_BLOCK_SIZES,
+        DEFAULT_NUM_WARPS,
+        DEFAULT_NUM_STAGES,
+    ):
+        configs.append(
+            triton.Config({"BLOCK_SIZE_M": m, "BLOCK_SIZE_N": n, "BLOCK_SIZE_K": k}, num_warps=w, num_stages=s)
+        )
+    return configs
+
+
+def _get_device_properties() -> dict[str, Any]:
+    return triton.runtime.driver.active.utils.get_device_properties(torch.cuda.current_device())
+
+
+def _exceeds_smem_capacity(
+    num_stages: int,
+    BLOCK_SIZE_M: int,
+    BLOCK_SIZE_N: int,
+    BLOCK_SIZE_K: int,
+    dtype: torch.dtype,
+    smem_size: int,
+    slack: int = 0,
+) -> bool:
+    return (
+        num_stages * BLOCK_SIZE_K * (BLOCK_SIZE_M + BLOCK_SIZE_N) + BLOCK_SIZE_M * BLOCK_SIZE_N
+    ) * dtype.itemsize > smem_size + slack
+
+
+def _common_prune_criteria(config: triton.Config, kwargs: dict[str, Any]) -> bool:
+    return False
+    num_stages = config.num_stages
+    BLOCK_SIZE_M = config.kwargs["BLOCK_SIZE_M"]
+    BLOCK_SIZE_N = config.kwargs["BLOCK_SIZE_N"]
+    BLOCK_SIZE_K = config.kwargs["BLOCK_SIZE_K"]
+    dtype = kwargs["x_ptr"].dtype
+    device_properties = _get_device_properties()
+    smem_size = device_properties["max_shared_mem"]
+    if _exceeds_smem_capacity(num_stages, BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K, dtype, smem_size):
+        return True
+
+    M = kwargs["M"]
+    N = kwargs["N"]
+    K = kwargs["K"]
+    num_experts = kwargs["NUM_EXPERTS"]
+    tokens_per_expert = M // num_experts
+    max_block_size_M = max(tokens_per_expert * 2, DEFAULT_M_BLOCK_SIZES[0])
+    max_block_size_N = max(N, DEFAULT_N_BLOCK_SIZES[0])
+    max_block_size_K = max(K, DEFAULT_K_BLOCK_SIZES[0])
+    if BLOCK_SIZE_M >= max_block_size_M:
+        return True
+    if BLOCK_SIZE_N >= max_block_size_N:
+        return True
+    if BLOCK_SIZE_K >= max_block_size_K:
+        return True
+
+    min_block_size_M = min(triton.next_power_of_2(tokens_per_expert // 2 + 1), 64)
+    min_block_size_N = min(triton.next_power_of_2(N // 2 + 1), 64)
+    min_block_size_K = min(triton.next_power_of_2(K // 2 + 1), 64)
+    if BLOCK_SIZE_M * BLOCK_SIZE_N <= min_block_size_M * min_block_size_N:
+        return True
+    if BLOCK_SIZE_M * BLOCK_SIZE_K <= min_block_size_M * min_block_size_K:
+        return True
+    if BLOCK_SIZE_N * BLOCK_SIZE_K <= min_block_size_N * min_block_size_K:
+        return True
+
+    return False
+
+
+def prune_configs(configs: list[triton.Config], args, **kwargs) -> list[triton.Config]:
+    pruned_configs = []
+    for config in configs:
+        if _common_prune_criteria(config, args):
+            continue
+        pruned_configs.append(config)
+    return pruned_configs
+
+
+# We need to autotune on batch size only when benchmarking with a large range of batch sizes
+def get_autotune_keys() -> list[str]:
+    if os.getenv("AUTOTUNE_BATCH_SIZE", "0") == "1":
+        return ["M", "N", "K", "NUM_EXPERTS"]
+    else:
+        return ["N", "K", "NUM_EXPERTS"]

moondream3_moe_fused/grouped_gemm/backward_dw.py

@@ -0,0 +1,135 @@
+# w[e, n, k] = sum_m if(s[m] == e) y[m, n] * x[m, k]
+
+import torch
+import triton
+import triton.language as tl
+
+from .autotuning import (
+    get_autotune_configs,
+    get_autotune_keys,
+    get_num_sms,
+    prune_configs,
+)
+from .forward import is_int_tensor
+
+
+@triton.autotune(
+    configs=get_autotune_configs(),
+    prune_configs_by={"early_config_prune": prune_configs},
+    key=get_autotune_keys(),
+)
+@triton.jit
+def _grouped_gemm_backward_dw_kernel(
+    # Pointers
+    x_ptr,
+    y_ptr,
+    m_sizes_ptr,
+    w_ptr,
+    # Dimensions
+    M: int,
+    N: tl.constexpr,
+    K: tl.constexpr,
+    NUM_EXPERTS: tl.constexpr,
+    NUM_SMS: tl.constexpr,
+    # Strides
+    stride_xm: tl.constexpr,
+    stride_xk: tl.constexpr,
+    stride_ym: tl.constexpr,
+    stride_yn: tl.constexpr,
+    stride_we: tl.constexpr,
+    stride_wn: tl.constexpr,
+    stride_wk: tl.constexpr,
+    # Metadata
+    BLOCK_SIZE_M: tl.constexpr = 64,
+    BLOCK_SIZE_N: tl.constexpr = 64,
+    BLOCK_SIZE_K: tl.constexpr = 64,
+) -> None:
+    tidx = tl.program_id(0)
+
+    # Output tiles per expert, since each expert weight matrix is [N, K]
+    num_n_tiles = tl.cdiv(N, BLOCK_SIZE_N)
+    num_k_tiles = tl.cdiv(K, BLOCK_SIZE_K)
+    num_tiles_per_expert = num_n_tiles * num_k_tiles
+
+    for tile_idx in range(tidx, num_tiles_per_expert, NUM_SMS):
+        # Output tile index
+        tile_n_idx = tile_idx % num_n_tiles
+        tile_k_idx = tile_idx // num_n_tiles
+
+        offs_n = tile_n_idx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+        offs_k = tile_k_idx * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
+        mask_n = offs_n < N
+        mask_k = offs_k < K
+
+        m_end = 0
+        for expert_idx in range(NUM_EXPERTS):
+            m_start = m_end
+            m_size = tl.load(m_sizes_ptr + expert_idx).to(tl.int32)
+            m_end = m_start + m_size
+            if m_size > 0:
+                offs_m = m_start + tl.arange(0, BLOCK_SIZE_M)
+
+                x_ptrs = x_ptr + stride_xm * offs_m[:, None] + stride_xk * offs_k[None, :]
+                y_ptrs = y_ptr + stride_ym * offs_m[:, None] + stride_yn * offs_n[None, :]
+
+                accumulator = tl.zeros((BLOCK_SIZE_N, BLOCK_SIZE_K), dtype=tl.float32)
+                for _ in range(tl.cdiv(m_size, BLOCK_SIZE_M)):
+                    mask_m = offs_m < m_end
+                    x = tl.load(x_ptrs, mask=mask_m[:, None] & mask_k[None, :])
+                    y = tl.load(y_ptrs, mask=mask_m[:, None] & mask_n[None, :])
+
+                    accumulator += tl.dot(y.T, x)
+
+                    offs_m += BLOCK_SIZE_M
+                    x_ptrs += stride_xm * BLOCK_SIZE_M
+                    y_ptrs += stride_ym * BLOCK_SIZE_M
+                w = accumulator.to(w_ptr.dtype.element_ty)
+
+                w_ptrs = w_ptr + stride_we * expert_idx + stride_wn * offs_n[:, None] + stride_wk * offs_k[None, :]
+                tl.store(w_ptrs, w, mask=mask_n[:, None] & mask_k[None, :])
+
+
+def grouped_gemm_backward_dw(
+    x: torch.Tensor, y: torch.Tensor, m_sizes: torch.Tensor, dtype: torch.dtype
+) -> torch.Tensor:
+    assert x.is_cuda
+    assert y.device == x.device
+    assert m_sizes.device == x.device
+    assert is_int_tensor(m_sizes)
+    assert x.is_contiguous()
+    assert y.is_contiguous()
+    assert m_sizes.is_contiguous()
+    assert x.ndim == 2
+    assert y.ndim == 2
+    assert m_sizes.ndim == 1
+    M, K = x.shape
+    _, N = y.shape
+    assert y.shape[0] == M
+    E = m_sizes.numel()
+
+    w = torch.zeros((E, N, K), device=x.device, dtype=dtype)
+    NUM_SMS = get_num_sms()
+    grid = lambda META: (NUM_SMS,)
+    with torch.cuda.device(x.device):
+        _grouped_gemm_backward_dw_kernel[grid](
+            # Pointers
+            x,
+            y,
+            m_sizes,
+            w,
+            # Dimensions
+            M,
+            N,
+            K,
+            E,
+            NUM_SMS,
+            # Strides
+            x.stride(0),
+            x.stride(1),
+            y.stride(0),
+            y.stride(1),
+            w.stride(0),
+            w.stride(1),
+            w.stride(2),
+        )
+    return w

moondream3_moe_fused/grouped_gemm/forward.py

@@ -0,0 +1,164 @@
+# y[m, n] = sum_k w[s[m], n, k] * x[m, k]
+
+from typing import Optional
+
+import torch
+import triton
+import triton.language as tl
+
+from .autotuning import (
+    get_autotune_configs,
+    get_autotune_keys,
+    get_num_sms,
+    prune_configs,
+)
+
+
+@triton.autotune(
+    configs=get_autotune_configs(),
+    # configs=[
+    #     triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_K': 32, 'BLOCK_SIZE_M': 128}, num_stages=4, num_warps=4),
+    # ],
+    prune_configs_by={"early_config_prune": prune_configs},
+    key=get_autotune_keys(),
+)
+@triton.jit
+def _grouped_gemm_forward_kernel(
+    # Pointers
+    x_ptr,
+    w_ptr,
+    m_sizes_ptr,
+    y_ptr,
+    # Dimensions
+    M: int,
+    N: tl.constexpr,
+    K: tl.constexpr,
+    NUM_EXPERTS: tl.constexpr,
+    NUM_SMS: tl.constexpr,
+    # Strides
+    stride_xm: tl.constexpr,
+    stride_xk: tl.constexpr,
+    stride_we: tl.constexpr,
+    stride_wn: tl.constexpr,
+    stride_wk: tl.constexpr,
+    stride_ym: tl.constexpr,
+    stride_yn: tl.constexpr,
+    # Metadata
+    BLOCK_SIZE_M: tl.constexpr = 64,
+    BLOCK_SIZE_N: tl.constexpr = 64,
+    BLOCK_SIZE_K: tl.constexpr = 64,
+) -> None:
+    tidx = tl.program_id(0)
+    m_end = 0
+    processed_tiles = 0
+    for expert_idx in range(NUM_EXPERTS):
+        m_start = m_end
+        m_size = tl.load(m_sizes_ptr + expert_idx).to(tl.int32)
+        m_end = m_start + m_size
+        if m_size > 0:
+            # tiles for this group's GEMM
+            num_m_tiles = tl.cdiv(m_size, BLOCK_SIZE_M)
+            num_n_tiles = tl.cdiv(N, BLOCK_SIZE_N)
+            num_tiles_per_expert = num_m_tiles * num_n_tiles
+
+            # Lower bound and upper bound are defined relative to the total tiles processed so far
+            # This ensures that we are only processing tiles for the current expert group AND
+            # we never exceed the total number of tiles for all expert groups
+            while tidx >= processed_tiles and tidx < processed_tiles + num_tiles_per_expert:
+                tile_idx = tidx - processed_tiles
+
+                # Output tile for this thread block for this expert group
+                # TODO: Check if L2 cache re-use for this order is optimal
+                tile_m_idx = tile_idx % num_m_tiles
+                tile_n_idx = tile_idx // num_m_tiles
+
+                offs_k = tl.arange(0, BLOCK_SIZE_K)
+
+                offs_m = m_start + tile_m_idx * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+                x_ptrs = x_ptr + stride_xm * offs_m[:, None] + stride_xk * offs_k[None, :]
+                mask_m = offs_m < m_end
+
+                offs_n = tile_n_idx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+                w_ptrs = w_ptr + stride_we * expert_idx + stride_wn * offs_n[:, None] + stride_wk * offs_k[None, :]
+                mask_n = offs_n < N
+
+                accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+                # GEMM main loop
+                for _ in range(tl.cdiv(K, BLOCK_SIZE_K)):
+                    mask_k = offs_k < K
+                    x = tl.load(x_ptrs, mask=mask_m[:, None] & mask_k[None, :])
+                    w = tl.load(w_ptrs, mask=mask_n[:, None] & mask_k[None, :])
+
+                    accumulator += tl.dot(x.to(w.dtype), w.T)
+
+                    offs_k += BLOCK_SIZE_K
+                    x_ptrs += stride_xk * BLOCK_SIZE_K
+                    w_ptrs += stride_wk * BLOCK_SIZE_K
+                y = accumulator.to(y_ptr.dtype.element_ty)
+
+                y_ptrs = y_ptr + stride_ym * offs_m[:, None] + stride_yn * offs_n[None, :]
+                tl.store(y_ptrs, y, mask=mask_m[:, None] & mask_n[None, :])
+
+                # Move to the next tile within this expert group
+                tidx += NUM_SMS
+
+            # Update the total tiles count for the next expert group
+            processed_tiles += num_tiles_per_expert
+
+
+def is_int_tensor(x: torch.Tensor) -> bool:
+    return x.dtype in {
+        torch.uint8,
+        torch.int8,
+        torch.int16,
+        torch.int32,
+        torch.int64,
+    }
+
+
+def grouped_gemm_forward(
+    x: torch.Tensor, w: torch.Tensor, m_sizes: torch.Tensor, dtype: Optional[torch.dtype] = None
+) -> torch.Tensor:
+    assert x.is_cuda
+    assert w.device == x.device
+    assert m_sizes.device == x.device
+    assert is_int_tensor(m_sizes)
+    assert x.is_contiguous()
+    assert w.is_contiguous()
+    assert m_sizes.is_contiguous()
+    assert x.ndim == 2
+    assert w.ndim == 3
+    assert m_sizes.ndim == 1
+    M, K = x.shape
+    E, N, _ = w.shape
+    assert w.shape[2] == K
+    assert m_sizes.numel() == E
+
+    if dtype is None:
+        dtype = x.dtype
+    y = torch.empty((M, N), device=x.device, dtype=dtype)
+    NUM_SMS = get_num_sms()
+    grid = lambda META: (NUM_SMS,)
+    with torch.cuda.device(x.device):
+        _grouped_gemm_forward_kernel[grid](
+            # Pointers
+            x,
+            w,
+            m_sizes,
+            y,
+            # Dimensions
+            M,
+            N,
+            K,
+            E,
+            NUM_SMS,
+            # Strides
+            x.stride(0),
+            x.stride(1),
+            w.stride(0),
+            w.stride(1),
+            w.stride(2),
+            y.stride(0),
+            y.stride(1),
+        )
+    return y

moondream3_moe_fused/grouped_gemm/forward_4bit.py

@@ -0,0 +1,209 @@
+# y[m, n] = sum_k w[s[m], n, k] * x[m, k]
+# Currently this is slower than the unfused dequant + linear when M is large
+
+from typing import Optional
+
+import torch
+import triton
+import triton.language as tl
+from bitsandbytes.functional import QuantState, dequantize_blockwise
+
+from .autotuning import (
+    get_autotune_configs,
+    get_autotune_keys,
+    get_num_sms,
+    prune_configs,
+)
+
+
+@triton.autotune(
+    configs=get_autotune_configs(),
+    prune_configs_by={"early_config_prune": prune_configs},
+    key=get_autotune_keys(),
+)
+@triton.jit
+def _grouped_gemm_forward_4bit_kernel(
+    # Pointers
+    x_ptr,
+    w_quant_ptr,
+    w_code_ptr,
+    w_absmax_ptr,
+    w_blocksize: tl.constexpr,
+    m_sizes_ptr,
+    y_ptr,
+    # Dimensions
+    M: int,
+    N: tl.constexpr,
+    K: tl.constexpr,
+    NUM_EXPERTS: tl.constexpr,
+    NUM_SMS: tl.constexpr,
+    # Strides
+    stride_xm: tl.constexpr,
+    stride_xk: tl.constexpr,
+    stride_we: tl.constexpr,
+    stride_wn: tl.constexpr,
+    stride_wk: tl.constexpr,
+    stride_ym: tl.constexpr,
+    stride_yn: tl.constexpr,
+    # Metadata
+    BLOCK_SIZE_M: tl.constexpr = 64,
+    BLOCK_SIZE_N: tl.constexpr = 64,
+    BLOCK_SIZE_K: tl.constexpr = 64,
+) -> None:
+    tidx = tl.program_id(0)
+    m_end = 0
+    processed_tiles = 0
+    for expert_idx in range(NUM_EXPERTS):
+        m_start = m_end
+        m_size = tl.load(m_sizes_ptr + expert_idx).to(tl.int32)
+        m_end = m_start + m_size
+        if m_size > 0:
+            # tiles for this group's GEMM
+            num_m_tiles = tl.cdiv(m_size, BLOCK_SIZE_M)
+            num_n_tiles = tl.cdiv(N, BLOCK_SIZE_N)
+            num_tiles_per_expert = num_m_tiles * num_n_tiles
+
+            # Lower bound and upper bound are defined relative to the total tiles processed so far
+            # This ensures that we are only processing tiles for the current expert group AND
+            # we never exceed the total number of tiles for all expert groups
+            while tidx >= processed_tiles and tidx < processed_tiles + num_tiles_per_expert:
+                tile_idx = tidx - processed_tiles
+
+                # Output tile for this thread block for this expert group
+                # TODO: Check if L2 cache re-use for this order is optimal
+                tile_m_idx = tile_idx % num_m_tiles
+                tile_n_idx = tile_idx // num_m_tiles
+
+                offs_k = tl.arange(0, BLOCK_SIZE_K)
+                offs_k_d2 = tl.arange(0, BLOCK_SIZE_K // 2)
+
+                offs_m = m_start + tile_m_idx * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+                x_ptrs = x_ptr + stride_xm * offs_m[:, None] + stride_xk * offs_k[None, :]
+                mask_m = offs_m < m_end
+
+                offs_n = tile_n_idx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+                w_offs = stride_we * expert_idx + stride_wn * offs_n[:, None] + stride_wk * offs_k[None, :]
+                w_offs_d2 = (
+                    stride_we // 2 * expert_idx + stride_wn // 2 * offs_n[:, None] + stride_wk * offs_k_d2[None, :]
+                )
+                mask_n = offs_n < N
+
+                accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+                # GEMM main loop
+                for _ in range(tl.cdiv(K, BLOCK_SIZE_K)):
+                    mask_k = offs_k < K
+                    mask_k_d2 = offs_k_d2 < K // 2
+                    x = tl.load(x_ptrs, mask=mask_m[:, None] & mask_k[None, :])
+
+                    # Dequantize
+                    w_code_offs = tl.load(w_quant_ptr + w_offs_d2, mask=mask_n[:, None] & mask_k_d2[None, :])
+                    w_code_offs = tl.interleave(w_code_offs // 16, w_code_offs % 16)
+                    w_code = tl.load(w_code_ptr + w_code_offs, mask=mask_n[:, None] & mask_k[None, :])
+                    w_absmax = tl.load(w_absmax_ptr + w_offs // w_blocksize)
+                    # w_quant_state.dtype is ignored, and w is always cast to x.dtype
+                    w = (w_code * w_absmax).to(x.dtype)
+
+                    accumulator += tl.dot(x, w.T)
+
+                    offs_k += BLOCK_SIZE_K
+                    offs_k_d2 += BLOCK_SIZE_K // 2
+                    x_ptrs += stride_xk * BLOCK_SIZE_K
+                    w_offs += stride_wk * BLOCK_SIZE_K
+                    w_offs_d2 += stride_wk * BLOCK_SIZE_K // 2
+                y = accumulator.to(y_ptr.dtype.element_ty)
+
+                y_ptrs = y_ptr + stride_ym * offs_m[:, None] + stride_yn * offs_n[None, :]
+                tl.store(y_ptrs, y, mask=mask_m[:, None] & mask_n[None, :])
+
+                # Move to the next tile within this expert group
+                tidx += NUM_SMS
+
+            # Update the total tiles count for the next expert group
+            processed_tiles += num_tiles_per_expert
+
+
+def is_int_tensor(x: torch.Tensor) -> bool:
+    return x.dtype in {
+        torch.uint8,
+        torch.int8,
+        torch.int16,
+        torch.int32,
+        torch.int64,
+    }
+
+
+def grouped_gemm_forward_4bit(
+    x: torch.Tensor,
+    w_quant: torch.Tensor,
+    w_quant_state: QuantState,
+    m_sizes: torch.Tensor,
+    dtype: Optional[torch.dtype] = None,
+) -> torch.Tensor:
+    assert w_quant_state.quant_type == "nf4"
+    assert w_quant_state.blocksize == triton.next_power_of_2(w_quant_state.blocksize)
+
+    # code and absmax should be float32. After computing code * absmax, w may be cast to bfloat16
+    w_code = w_quant_state.code
+    assert w_code.dtype == torch.float32
+
+    w_absmax = w_quant_state.absmax
+    if w_quant_state.nested:
+        w_absmax = dequantize_blockwise(w_absmax, w_quant_state.state2)
+        w_absmax += w_quant_state.offset
+    assert w_absmax.dtype == torch.float32
+
+    assert x.is_cuda
+    assert w_quant.device == x.device
+    assert w_code.device == x.device
+    assert w_absmax.device == x.device
+    assert m_sizes.device == x.device
+    assert is_int_tensor(w_quant)
+    assert is_int_tensor(m_sizes)
+    assert x.is_contiguous()
+    assert w_quant.is_contiguous()
+    assert w_code.is_contiguous()
+    assert w_absmax.is_contiguous()
+    assert m_sizes.is_contiguous()
+    assert x.ndim == 2
+    assert len(w_quant_state.shape) == 3
+    assert m_sizes.ndim == 1
+    M, K = x.shape
+    E, N, _ = w_quant_state.shape
+    assert w_quant_state.shape[2] == K
+    assert K % 2 == 0
+    assert E * N * K % w_quant_state.blocksize == 0
+    assert w_quant.numel() == E * N * K // 2
+    assert w_absmax.numel() == E * N * K // w_quant_state.blocksize
+    assert m_sizes.numel() == E
+
+    if dtype is None:
+        dtype = x.dtype
+    y = torch.empty((M, N), device=x.device, dtype=dtype)
+    NUM_SMS = get_num_sms()
+    grid = lambda META: (NUM_SMS,)
+    with torch.cuda.device(x.device):
+        _grouped_gemm_forward_4bit_kernel[grid](
+            # Pointers
+            x,
+            w_quant,
+            w_code,
+            w_absmax,
+            w_quant_state.blocksize,
+            m_sizes,
+            y,
+            # Dimensions
+            M,
+            N,
+            K,
+            E,
+            NUM_SMS,
+            # Strides
+            x.stride(0),
+            x.stride(1),
+            N * K,
+            K,
+            1,
+            y.stride(0),
+            y.stride(1),
+        )
+    return y

moondream3_moe_fused/grouped_gemm/forward_transposed.py

@@ -0,0 +1,152 @@
+# y[m, n] = sum_k w[s[m], k, n] * x[m, k]
+
+from typing import Optional
+
+import torch
+import triton
+import triton.language as tl
+
+from .autotuning import (
+    get_autotune_configs,
+    get_autotune_keys,
+    get_num_sms,
+    prune_configs,
+)
+from .forward import is_int_tensor
+
+
+@triton.autotune(
+    configs=get_autotune_configs(),
+    prune_configs_by={"early_config_prune": prune_configs},
+    key=get_autotune_keys(),
+)
+@triton.jit
+def _grouped_gemm_forward_transposed_kernel(
+    # Pointers
+    x_ptr,
+    w_ptr,
+    m_sizes_ptr,
+    y_ptr,
+    # Dimensions
+    M: int,
+    N: tl.constexpr,
+    K: tl.constexpr,
+    NUM_EXPERTS: tl.constexpr,
+    NUM_SMS: tl.constexpr,
+    # Strides
+    stride_xm: tl.constexpr,
+    stride_xk: tl.constexpr,
+    stride_we: tl.constexpr,
+    stride_wk: tl.constexpr,
+    stride_wn: tl.constexpr,
+    stride_ym: tl.constexpr,
+    stride_yn: tl.constexpr,
+    # Metadata
+    BLOCK_SIZE_M: tl.constexpr = 64,
+    BLOCK_SIZE_N: tl.constexpr = 64,
+    BLOCK_SIZE_K: tl.constexpr = 64,
+) -> None:
+    tidx = tl.program_id(0)
+    m_end = 0
+    processed_tiles = 0
+    for expert_idx in range(NUM_EXPERTS):
+        m_start = m_end
+        m_size = tl.load(m_sizes_ptr + expert_idx).to(tl.int32)
+        m_end = m_start + m_size
+        if m_size > 0:
+            # tiles for this group's GEMM
+            num_m_tiles = tl.cdiv(m_size, BLOCK_SIZE_M)
+            num_n_tiles = tl.cdiv(N, BLOCK_SIZE_N)
+            num_tiles_per_expert = num_m_tiles * num_n_tiles
+
+            # Lower bound and upper bound are defined relative to the total tiles processed so far
+            # This ensures that we are only processing tiles for the current expert group AND
+            # we never exceed the total number of tiles for all expert groups
+            while tidx >= processed_tiles and tidx < processed_tiles + num_tiles_per_expert:
+                tile_idx = tidx - processed_tiles
+
+                # Output tile for this thread block for this expert group
+                # TODO: Check if L2 cache re-use for this order is optimal
+                tile_m_idx = tile_idx % num_m_tiles
+                tile_n_idx = tile_idx // num_m_tiles
+
+                offs_k = tl.arange(0, BLOCK_SIZE_K)
+
+                offs_m = m_start + tile_m_idx * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+                x_ptrs = x_ptr + stride_xm * offs_m[:, None] + stride_xk * offs_k[None, :]
+                mask_m = offs_m < m_start + m_size
+
+                offs_n = tile_n_idx * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+                w_ptrs = w_ptr + stride_we * expert_idx + stride_wn * offs_n[:, None] + stride_wk * offs_k[None, :]
+                mask_n = offs_n < N
+
+                accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+                # GEMM main loop
+                for _ in range(tl.cdiv(K, BLOCK_SIZE_K)):
+                    mask_k = offs_k < K
+                    x = tl.load(x_ptrs, mask=mask_m[:, None] & mask_k[None, :])
+                    w = tl.load(w_ptrs, mask=mask_n[:, None] & mask_k[None, :])
+
+                    accumulator += tl.dot(x.to(w.dtype), w.T)
+
+                    offs_k += BLOCK_SIZE_K
+                    x_ptrs += stride_xk * BLOCK_SIZE_K
+                    w_ptrs += stride_wk * BLOCK_SIZE_K
+                y = accumulator.to(y_ptr.dtype.element_ty)
+
+                y_ptrs = y_ptr + stride_ym * offs_m[:, None] + stride_yn * offs_n[None, :]
+                tl.store(y_ptrs, y, mask=mask_m[:, None] & mask_n[None, :])
+
+                # Move to the next tile within this expert group
+                tidx += NUM_SMS
+
+            # Update the total tiles count for the next expert group
+            processed_tiles += num_tiles_per_expert
+
+
+def grouped_gemm_forward_transposed(
+    x: torch.Tensor, w: torch.Tensor, m_sizes: torch.Tensor, dtype: Optional[torch.dtype] = None
+) -> torch.Tensor:
+    assert x.is_cuda
+    assert w.device == x.device
+    assert m_sizes.device == x.device
+    assert is_int_tensor(m_sizes)
+    assert x.is_contiguous()
+    assert w.is_contiguous()
+    assert m_sizes.is_contiguous()
+    assert x.ndim == 2
+    assert w.ndim == 3
+    assert m_sizes.ndim == 1
+    M, K = x.shape
+    E, _, N = w.shape
+    assert w.shape[1] == K
+    assert m_sizes.numel() == E
+
+    if dtype is None:
+        dtype = x.dtype
+    y = torch.empty((M, N), device=x.device, dtype=dtype)
+    NUM_SMS = get_num_sms()
+    grid = lambda META: (NUM_SMS,)
+    with torch.cuda.device(x.device):
+        _grouped_gemm_forward_transposed_kernel[grid](
+            # Pointers
+            x,
+            w,
+            m_sizes,
+            y,
+            # Dimensions
+            M,
+            N,
+            K,
+            E,
+            NUM_SMS,
+            # Strides
+            x.stride(0),
+            x.stride(1),
+            w.stride(0),
+            w.stride(1),
+            w.stride(2),
+            y.stride(0),
+            y.stride(1),
+        )
+    return y

moondream3_moe_fused/grouped_gemm/interface.py

@@ -0,0 +1,31 @@
+import torch
+
+from .backward_dw import grouped_gemm_backward_dw
+from .forward import grouped_gemm_forward
+from .forward_transposed import grouped_gemm_forward_transposed
+
+
+class GroupedGemm(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x, w, m_sizes):
+        ctx.save_for_backward(x, w, m_sizes)
+        return grouped_gemm_forward(x, w, m_sizes)
+
+    @staticmethod
+    def backward(ctx, dy):
+        x, w, m_sizes = ctx.saved_tensors
+
+        if x.requires_grad:
+            dx = grouped_gemm_forward_transposed(dy, w, m_sizes, x.dtype)
+        else:
+            dx = None
+
+        if w.requires_grad:
+            dw = grouped_gemm_backward_dw(x, dy, m_sizes, w.dtype)
+        else:
+            dw = None
+
+        return dx, dw, None
+
+
+grouped_gemm = GroupedGemm.apply

moondream3_moe_fused/kernels/indexing.py

@@ -0,0 +1,65 @@
+from functools import partial
+
+import torch
+
+
+# Assume s is sorted
+@partial(torch.compile, fullgraph=True, mode="max-autotune-no-cudagraphs")
+@torch.no_grad()
+def get_batch_begins_ends(s: torch.Tensor, E: int) -> torch.Tensor:
+    arange = torch.arange(E, device=s.device, dtype=s.dtype)
+    s_begins = (arange[:, None] > s[None, :]).sum(dim=1, dtype=torch.int32)
+    s_ends = (arange[:, None] >= s[None, :]).sum(dim=1, dtype=torch.int32)
+    s_begins_ends = torch.stack([s_begins, s_ends], dim=1)
+    return s_begins_ends
+
+
+# Faster than torch.histc when each element of s is an int in [0, E)
+@partial(torch.compile, fullgraph=True, mode="max-autotune-no-cudagraphs")
+@torch.no_grad()
+def get_expert_counts(s: torch.Tensor, E: int) -> torch.Tensor:
+    arange = torch.arange(E, device=s.device, dtype=s.dtype)
+    counts = (arange[:, None] == s[None, :]).sum(dim=1, dtype=torch.int32)
+    return counts
+
+
+# Faster than torch.sort when each element of s is an int in [0, E)
+@partial(torch.compile, fullgraph=True, mode="max-autotune-no-cudagraphs")
+@torch.no_grad()
+def sort_experts(s: torch.Tensor, E: int) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    E_arange = torch.arange(E, device=s.device, dtype=s.dtype)
+    compare = E_arange[:, None] == s[None, :]
+    counts = compare.sum(dim=1, dtype=torch.int32)
+    s_sorted = torch.repeat_interleave(counts, output_size=s.numel())  # int32
+
+    s_arange = torch.arange(s.numel(), device=s.device, dtype=s.dtype)
+    ranks_in_bin = compare.cumsum(dim=1, dtype=torch.int32)
+    ranks_in_bin = ranks_in_bin[s, s_arange]
+    offsets = counts.cumsum(dim=0, dtype=torch.int32) - counts
+    idx = ranks_in_bin + offsets[s] - 1  # int32
+
+    inv_idx = torch.empty_like(idx)  # int32
+    inv_idx[idx] = s_arange.to(inv_idx.dtype)
+
+    # The above definition of idx is the opposite of torch.sort
+    return s_sorted, inv_idx, idx
+
+
+@partial(torch.compile, fullgraph=True, mode="max-autotune-no-cudagraphs")
+@torch.no_grad()
+def get_expert_counts_and_idx(s: torch.Tensor, E: int) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    E_arange = torch.arange(E, device=s.device, dtype=s.dtype)
+    compare = E_arange[:, None] == s[None, :]
+    counts = compare.sum(dim=1, dtype=torch.int32)
+
+    s_arange = torch.arange(s.numel(), device=s.device, dtype=s.dtype)
+    ranks_in_bin = compare.cumsum(dim=1, dtype=torch.int32)
+    ranks_in_bin = ranks_in_bin[s, s_arange]
+    offsets = counts.cumsum(dim=0, dtype=torch.int32) - counts
+    idx = ranks_in_bin + offsets[s] - 1  # int32
+
+    inv_idx = torch.empty_like(idx)  # int32
+    inv_idx[idx] = s_arange.to(inv_idx.dtype)
+
+    # The above definition of idx is the opposite of torch.sort
+    return counts, inv_idx, idx

moondream3_moe_fused/lora.py

@@ -0,0 +1,143 @@
+# Modified from https://github.com/huggingface/peft/blob/e34852f7b67d51ba7ef871051b1236e9558c650e/src/peft/tuners/lora/layer.py#L585
+
+import functools
+import math
+from typing import Optional, Union
+
+import torch
+from peft import LoraConfig
+from peft.tuners.lora.layer import LoraLayer
+from torch import nn
+
+from .moe_fused_linear import MoeFusedLinear, moe_fused_kaiming_uniform_
+
+
+class LoraMoeFusedLinear(nn.Module, LoraLayer):
+    def __init__(
+        self,
+        base_layer: MoeFusedLinear,
+        adapter_name: str,
+        r: int = 0,
+        lora_alpha: int = 1,
+        lora_dropout: float = 0.0,
+        init_lora_weights: Union[bool, str] = True,
+        use_rslora: bool = False,
+        use_dora: bool = False,
+        lora_bias: bool = False,
+        **kwargs,
+    ) -> None:
+        if init_lora_weights not in {True, False}:
+            raise NotImplementedError
+        if use_dora:
+            raise NotImplementedError
+        if lora_bias:
+            raise NotImplementedError
+
+        super().__init__()
+        LoraLayer.__init__(self, base_layer, **kwargs)
+        self.num_experts = base_layer.num_experts
+        self._active_adapter = adapter_name
+
+        self.update_layer(
+            adapter_name,
+            r,
+            lora_alpha=lora_alpha,
+            lora_dropout=lora_dropout,
+            init_lora_weights=init_lora_weights,
+            use_rslora=use_rslora,
+        )
+
+    def update_layer(
+        self,
+        adapter_name: str,
+        r: int,
+        lora_alpha: int,
+        lora_dropout: float,
+        init_lora_weights: Union[bool, str],
+        use_rslora: bool,
+    ) -> None:
+        # This code works for linear layers, override for other layer types
+        if r <= 0:
+            raise ValueError(f"`r` should be a positive integer value but the value passed is {r}")
+
+        self.r[adapter_name] = r
+        self.lora_alpha[adapter_name] = lora_alpha
+        if lora_dropout > 0.0:
+            lora_dropout_layer = nn.Dropout(p=lora_dropout)
+        else:
+            lora_dropout_layer = nn.Identity()
+        self.lora_dropout.update(nn.ModuleDict({adapter_name: lora_dropout_layer}))
+
+        # Actual trainable parameters
+        self.lora_A[adapter_name] = MoeFusedLinear(self.in_features, r, self.num_experts)
+        self.lora_B[adapter_name] = MoeFusedLinear(r, self.out_features, self.num_experts)
+
+        if use_rslora:
+            self.scaling[adapter_name] = lora_alpha / math.sqrt(r)
+        else:
+            self.scaling[adapter_name] = lora_alpha / r
+
+        if init_lora_weights is True:
+            self.reset_lora_parameters(adapter_name, init_lora_weights)
+
+        # call this before dora_init
+        self._move_adapter_to_device_of_base_layer(adapter_name)
+
+        self.set_adapter(self.active_adapters)
+
+    def reset_lora_parameters(self, adapter_name: str, init_lora_weights: Union[bool, str]) -> None:
+        if init_lora_weights is False:
+            return
+
+        if adapter_name in self.lora_A:
+            if init_lora_weights is True:
+                # initialize A the same way as the default for nn.Linear and B to zero
+                moe_fused_kaiming_uniform_(self.lora_A[adapter_name].weight)
+                nn.init.zeros_(self.lora_B[adapter_name].weight)
+            else:
+                raise ValueError(f"Unknown initialization {init_lora_weights=}")
+
+    def merge(self, safe_merge: bool = False, adapter_names: Optional[list[str]] = None) -> None:
+        raise NotImplementedError
+
+    def unmerge(self) -> None:
+        raise NotImplementedError
+
+    def forward(self, x: torch.Tensor, m_sizes: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        self._check_forward_args(x, *args, **kwargs)
+        adapter_names = kwargs.pop("adapter_names", None)
+
+        if self.disable_adapters:
+            if self.merged:
+                self.unmerge()
+            result = self.base_layer(x, m_sizes, *args, **kwargs)
+        elif adapter_names is not None:
+            raise NotImplementedError
+            # result = self._mixed_batch_forward(x, *args, adapter_names=adapter_names, **kwargs)
+            # In _mixed_batch_forward, we need to change `lora_B(lora_A(dropout(x)))`
+            # to `lora_B(lora_A(dropout(x), m_sizes), m_sizes)`
+        elif self.merged:
+            result = self.base_layer(x, m_sizes, *args, **kwargs)
+        else:
+            result = self.base_layer(x, m_sizes, *args, **kwargs)
+            torch_result_dtype = result.dtype
+
+            lora_A_keys = self.lora_A.keys()
+            for active_adapter in self.active_adapters:
+                if active_adapter not in lora_A_keys:
+                    continue
+
+                lora_A = self.lora_A[active_adapter]
+                lora_B = self.lora_B[active_adapter]
+                dropout = self.lora_dropout[active_adapter]
+                scaling = self.scaling[active_adapter]
+                x = self._cast_input_dtype(x, lora_A.weight.dtype)
+                result = result + lora_B(lora_A(dropout(x), m_sizes), m_sizes) * scaling
+
+            result = result.to(torch_result_dtype)
+
+        return result
+
+    def __repr__(self) -> str:
+        rep = super().__repr__()
+        return "lora." + rep

moondream3_moe_fused/moe_fused_linear.py

@@ -0,0 +1,44 @@
+import torch
+from torch import nn
+import math
+from typing import Optional
+from .functional import moe_fused_linear
+
+def moe_fused_kaiming_uniform_(weight: torch.Tensor) -> None:
+    # Kaiming uniform on in_features
+    # Although Qwen's default activation is silu, we set the gain `a = sqrt(5)` following the original Linear
+    in_features = weight.shape[-1]
+    bound = math.sqrt(3 * 5 / in_features)
+    nn.init.uniform_(weight, -bound, bound)
+
+class MoeFusedLinear(nn.Module):
+    __constants__ = ["in_features", "out_features", "num_experts"]
+    in_features: int
+    out_features: int
+    num_experts: int
+    weight: torch.Tensor
+
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int,
+        num_experts: int,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.num_experts = num_experts
+        self.weight = nn.Parameter(torch.empty((num_experts, out_features, in_features), **factory_kwargs))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        moe_fused_kaiming_uniform_(self.weight)
+
+    def forward(self, input: torch.Tensor, m_sizes: torch.Tensor) -> torch.Tensor:
+        return moe_fused_linear(input, self.weight, m_sizes)
+
+    def extra_repr(self) -> str:
+        return f"in_features={self.in_features}, out_features={self.out_features}, num_experts={self.num_experts}"

moondream3_moe_fused/quantize/layer.py

@@ -0,0 +1,96 @@
+# Modified from https://github.com/bitsandbytes-foundation/bitsandbytes/blob/888788d75db8ff8e8888838307119f98d1235c24/bitsandbytes/nn/modules.py#L377
+# TODO: support IPEX
+
+import warnings
+from typing import Any, Optional
+
+import torch
+from bitsandbytes.functional import dequantize_4bit
+from bitsandbytes.nn.modules import Params4bit, fix_4bit_weight_quant_state_from_module
+from torch import nn
+
+from ..functional import moe_fused_linear
+from ..moe_fused_linear import MoeFusedLinear
+
+
+# TODO: Fuse this
+def moe_fused_linear_4bit(input: torch.Tensor, weight: Params4bit, m_sizes: torch.Tensor) -> torch.Tensor:
+    assert not weight.requires_grad
+    # Cast weight to input.dtype
+    # The grouped GEMM kernels use float32 accumulator
+    weight = dequantize_4bit(weight, weight.quant_state).to(input.dtype)
+    return moe_fused_linear(input, weight, m_sizes)
+
+
+class MoeFusedLinear4bit(MoeFusedLinear):
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int,
+        num_experts: int,
+        *,
+        weight: Optional[nn.Parameter] = None,  # Used for initializing from a non-quantized module
+        compute_dtype: Optional[torch.dtype] = None,
+        compress_statistics: bool = True,
+        quant_type: str = "fp4",
+        quant_storage: torch.dtype = torch.uint8,
+        device: Optional[torch.device] = None,
+    ) -> None:
+        super().__init__(in_features, out_features, num_experts, device=device)
+        self.weight = Params4bit(
+            self.weight,
+            requires_grad=False,
+            compress_statistics=compress_statistics,
+            quant_type=quant_type,
+            quant_storage=quant_storage,
+            module=self,
+        )
+        # self.persistent_buffers = []  # TODO consider as way to save quant state
+        self.compute_dtype = compute_dtype
+        self.compute_type_is_set = compute_dtype is not None
+        self.quant_state = None
+        self.quant_storage = quant_storage
+
+    def set_compute_type(self, x: torch.Tensor) -> None:
+        if x.dtype in [torch.float32, torch.bfloat16]:
+            # the input is in a dtype that is safe to compute in, we switch
+            # to this type for speed and stability
+            self.compute_dtype = x.dtype
+        elif x.dtype == torch.float16:
+            # we take the compoute dtype passed into the layer
+            if self.compute_dtype in [None, torch.float32] and (x.numel() == x.shape[-1]):
+                # single batch inference with input torch.float16 and compute_dtype float32 -> slow inference when it could be fast
+                # warn the user about this
+                warnings.warn(
+                    "Input type into Linear4bit is torch.float16, but bnb_4bit_compute_dtype=torch.float32 (default). "
+                    "This will lead to slow inference.",
+                )
+                warnings.filterwarnings("ignore", message=".*inference.")
+            if self.compute_dtype in [None, torch.float32] and (x.numel() != x.shape[-1]):
+                warnings.warn(
+                    "Input type into Linear4bit is torch.float16, but bnb_4bit_compute_dtype=torch.float32 (default). "
+                    "This will lead to slow inference or training speed.",
+                )
+                warnings.filterwarnings("ignore", message=".*inference or training")
+
+    def _save_to_state_dict(self, destination: dict[str, Any], prefix: str, keep_vars: bool) -> None:
+        super()._save_to_state_dict(destination, prefix, keep_vars)
+
+        if getattr(self.weight, "quant_state", None) is not None:
+            for k, v in self.weight.quant_state.as_dict(packed=True).items():
+                destination[prefix + "weight." + k] = v if keep_vars else v.detach()
+
+    def forward(self, x: torch.Tensor, m_sizes: torch.Tensor) -> torch.Tensor:
+        fix_4bit_weight_quant_state_from_module(self)
+
+        if not self.compute_type_is_set:
+            self.set_compute_type(x)
+            self.compute_type_is_set = True
+
+        inp_dtype = x.dtype
+        if self.compute_dtype is not None:
+            x = x.to(self.compute_dtype)
+
+        x = moe_fused_linear_4bit(x, self.weight, m_sizes)
+        x = x.to(inp_dtype)
+        return x

moondream3_moe_fused/quantize/quantizer.py

@@ -0,0 +1,133 @@
+from inspect import signature
+from typing import Any, Optional, Union
+
+from accelerate import init_empty_weights
+from bitsandbytes.nn import Linear4bit
+from torch import nn
+from transformers import BitsAndBytesConfig
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import Conv1D
+from transformers.quantizers.quantizer_bnb_4bit import Bnb4BitHfQuantizer
+from transformers.utils import logging
+
+from ..modular_qwen3_moe_fused import MoeFusedLinear
+from .layer import MoeFusedLinear4bit
+
+
+logger = logging.get_logger(__name__)
+
+
+# Modified from https://github.com/huggingface/transformers/blob/508a7040556dc6b45f09174c662a9632284b2445/src/transformers/integrations/bitsandbytes.py#L150
+def _replace_with_bnb_moe_fused_linear(
+    model: nn.Module,
+    modules_to_not_convert: list[str],
+    current_key_name: list[str],
+    quantization_config: BitsAndBytesConfig,
+    has_been_replaced: bool,
+) -> bool:
+    for name, module in model.named_children():
+        current_key_name.append(name)
+
+        if isinstance(module, (nn.Linear, Conv1D, MoeFusedLinear)) and name not in modules_to_not_convert:
+            # Check if the current key is not in the `modules_to_not_convert`
+            current_key_name_str = ".".join(current_key_name)
+            if not any(
+                (key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert
+            ):
+                num_experts = None
+                if isinstance(module, MoeFusedLinear):
+                    in_features = module.in_features
+                    out_features = module.out_features
+                    num_experts = module.num_experts
+                elif isinstance(module, Conv1D):
+                    in_features, out_features = module.weight.shape
+                else:
+                    in_features = module.in_features
+                    out_features = module.out_features
+
+                if isinstance(module, MoeFusedLinear):
+                    model._modules[name] = MoeFusedLinear4bit(
+                        in_features,
+                        out_features,
+                        num_experts,
+                        compute_dtype=quantization_config.bnb_4bit_compute_dtype,
+                        compress_statistics=quantization_config.bnb_4bit_use_double_quant,
+                        quant_type=quantization_config.bnb_4bit_quant_type,
+                        quant_storage=quantization_config.bnb_4bit_quant_storage,
+                    )
+                else:
+                    extra_kwargs = (
+                        {"quant_storage": quantization_config.bnb_4bit_quant_storage}
+                        if "quant_storage" in list(signature(Linear4bit).parameters)
+                        else {}
+                    )
+                    model._modules[name] = Linear4bit(
+                        in_features,
+                        out_features,
+                        module.bias is not None,
+                        quantization_config.bnb_4bit_compute_dtype,
+                        compress_statistics=quantization_config.bnb_4bit_use_double_quant,
+                        quant_type=quantization_config.bnb_4bit_quant_type,
+                        **extra_kwargs,
+                    )
+
+                has_been_replaced = True
+                # Store the module class in case we need to transpose the weight later
+                model._modules[name].source_cls = type(module)
+                # Force requires grad to False to avoid unexpected errors
+                model._modules[name].requires_grad_(False)
+
+        if len(list(module.children())) > 0:
+            has_been_replaced = _replace_with_bnb_moe_fused_linear(
+                module, modules_to_not_convert, current_key_name, quantization_config, has_been_replaced
+            )
+
+        # Remove the last key for recursion
+        current_key_name.pop(-1)
+
+    return has_been_replaced
+
+
+# model is modified in place
+def replace_with_bnb_moe_fused_linear(
+    model: nn.Module, modules_to_not_convert: Optional[list[str]], quantization_config: BitsAndBytesConfig
+) -> None:
+    modules_to_not_convert = ["lm_head"] if modules_to_not_convert is None else modules_to_not_convert
+    with init_empty_weights():
+        has_been_replaced = _replace_with_bnb_moe_fused_linear(
+            model, modules_to_not_convert, [], quantization_config, False
+        )
+
+    if not has_been_replaced:
+        logger.warning(
+            "You are loading your model in 8bit or 4bit but no linear modules were found in your model."
+            " Please double check your model architecture, or submit an issue on github if you think this is"
+            " a bug."
+        )
+
+
+def _process_model_before_weight_loading(
+    self: Bnb4BitHfQuantizer,
+    model: PreTrainedModel,
+    device_map: Union[str, dict[str, Any]],
+    keep_in_fp32_modules: Optional[list[str]] = None,
+    **kwargs,
+) -> None:
+    self.modules_to_not_convert = self.get_modules_to_not_convert(
+        model, self.quantization_config.llm_int8_skip_modules, keep_in_fp32_modules
+    )
+
+    # Extend `self.modules_to_not_convert` to keys that are supposed to be offloaded to `cpu` or `disk`
+    if isinstance(device_map, dict) and len(device_map) > 1:
+        keys_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]]
+        self.modules_to_not_convert.extend(keys_on_cpu)
+
+    replace_with_bnb_moe_fused_linear(
+        model, modules_to_not_convert=self.modules_to_not_convert, quantization_config=self.quantization_config
+    )
+
+    model.config.quantization_config = self.quantization_config
+
+
+def patch_bnb_quantizer() -> None:
+    Bnb4BitHfQuantizer._process_model_before_weight_loading = _process_model_before_weight_loading

preprocessor_config.json

@@ -0,0 +1,10 @@
+{
+  "auto_map": {
+    "AutoImageProcessor": "image_processing_moondream3.Moondream3ImageProcessor",
+    "AutoProcessor": "processing_moondream3.Moondream3Processor"
+  },
+  "image_processor_type": "Moondream3ImageProcessor",
+  "max_crops": 12,
+  "overlap_margin": 4,
+  "processor_class": "Moondream3Processor"
+}

processing_moondream3.py

@@ -0,0 +1,256 @@
+# coding=utf-8
+# Copyright 2024 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.
+"""
+Processor class for Moondream3.
+"""
+
+from typing import Optional, Union
+
+import numpy as np
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput, is_valid_image
+from transformers.processing_utils import (
+    MultiModalData,
+    ProcessingKwargs,
+    ProcessorMixin,
+    Unpack,
+)
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+from transformers.utils import is_vision_available, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Moondream3ProcessorKwargs(ProcessingKwargs, total=False):
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+            "return_token_type_ids": False
+        },
+        "common_kwargs": {
+            "return_tensors": "pt",
+        },
+    }
+
+def _rotate_right_array(x, k: int):
+    """
+    Rotate a 1D or 2D structure k steps to the right along the last axis.
+    Supports: list, numpy.ndarray, torch.Tensor.
+    Works even if numpy or torch are not installed.
+    Raises TypeError for unsupported input types.
+    """
+    # optional imports
+    try:
+        import numpy as np
+    except ImportError:
+        np = None
+
+    try:
+        import torch
+    except ImportError:
+        torch = None
+
+    # torch.Tensor
+    if torch is not None and isinstance(x, torch.Tensor):
+        if x.size(-1) == 0:
+            return x
+        return torch.roll(x, shifts=k % x.size(-1), dims=-1)
+
+    # numpy.ndarray
+    if np is not None and isinstance(x, np.ndarray):
+        if x.shape[-1] == 0:
+            return x
+        return np.roll(x, k % x.shape[-1], axis=-1)
+
+    # python list (1D or 2D)
+    if isinstance(x, list):
+        if not x:  # empty list
+            return x
+        # 2D (batch, seq)
+        if isinstance(x[0], list):
+            out = []
+            for row in x:
+                if not row:
+                    out.append(row)
+                    continue
+                shift = k % len(row)
+                out.append(row[-shift:] + row[:-shift] if shift else row[:])
+            return out
+        # 1D
+        shift = k % len(x)
+        return x[-shift:] + x[:-shift] if shift else x[:]
+
+    # unsupported type
+    raise TypeError(
+        f"Unsupported type {type(x).__name__} for rotation. "
+        f"Expected list, numpy.ndarray, or torch.Tensor. "
+        f"(numpy or torch are optional dependencies)"
+    )
+
+
+# Copied from transformers.models.idefics2.processing_idefics2.is_url
+def is_url(val) -> bool:
+    return isinstance(val, str) and val.startswith("http")
+
+
+# Copied from transformers.models.idefics2.processing_idefics2.is_image_or_image_url
+def is_image_or_image_url(elem):
+    return is_url(elem) or is_valid_image(elem)
+
+
+class Moondream3Processor(ProcessorMixin):
+    r"""
+    Constructs a Moondream3 processor which wraps a Moondream3 image processor and a Moondream3 tokenizer into a single processor.
+
+    [`Moondream3Processor`] offers all the functionalities of [`CLIPImageProcessor`] and [`LlamaTokenizerFast`]. See the
+    [`~Moondream3Processor.__call__`] and [`~Moondream3Processor.decode`] for more information.
+
+    Args:
+        image_processor ([`Moondream3ImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`LlamaTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        patch_size (`int`, *optional*, defaults to 16):
+            Patch size from the vision tower.
+        spatial_merge_size (`int`, *optional*, defaults to 1):
+            The downsampling factor for the spatial merge operation.
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+        image_token (`str`, *optional*, defaults to `"[IMG]"`):
+            Special token used to denote image location.
+        image_break_token (`str`, *optional*, defaults to `"[IMG_BREAK]"`):
+            Special token used to denote the end of a line of pixels in an image.
+        image_end_token (`str`, *optional*, defaults to `"[IMG_END]"`):
+            Special token used to denote the end of an image input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        chat_template=None,
+        image_token_id=0,
+        **kwargs,
+    ):
+        self.image_token_id = image_token_id
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+
+    def __call__(
+        self,
+        images: Optional[ImageInput] = None,
+        text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,
+        **kwargs: Unpack[Moondream3ProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to LlamaTokenizerFast's [`~LlamaTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwargs` arguments to
+        CLIPImageProcessor's [`~CLIPImageProcessor.__call__`] if `images` is not `None`. Please refer to the docstring
+        of the above two methods for more information.
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `list[str]`, `list[list[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+            `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+
+        output_kwargs = self._merge_kwargs(
+            Moondream3ProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+
+        if images is not None:
+            image_inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
+        else:
+            image_inputs = {}
+
+        if isinstance(text, str):
+            text = [text]
+        elif not isinstance(text, list) and not isinstance(text[0], str):
+            raise TypeError("Invalid input text. Please provide a string, or a list of strings")
+
+        # try to expand inputs in processing if we have the necessary parts
+        prompt_strings = text
+
+        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
+        text_inputs = self.tokenizer(prompt_strings, **output_kwargs["text_kwargs"], return_tensors=None)
+        if "input_ids" in text_inputs:
+            # prepend 1 bos_token_id and 729 image_token_id to the text_inputs
+            for i in range(len(text_inputs["input_ids"])):
+                prepended_tokens = [self.tokenizer.bos_token_id] + [self.image_token_id] * 729
+                text_inputs["input_ids"][i] = prepended_tokens + text_inputs["input_ids"][i]
+        if "attention_mask" in text_inputs:
+            # attend to the 730 prepended tokens
+            for i in range(len(text_inputs["attention_mask"])):
+                prepended_mask = [1] * 730
+                text_inputs["attention_mask"][i] = prepended_mask + text_inputs["attention_mask"][i]
+
+        return BatchFeature(data={**text_inputs, **image_inputs}, tensor_type=return_tensors)
+
+    def apply_chat_template(
+        self,
+        conversation: Union[list[dict[str, str]], list[list[dict[str, str]]]],
+        chat_template: Optional[str] = None,
+        **kwargs,
+    ) -> str:
+            # Call the original behavior first
+        out = super().apply_chat_template(
+            conversation=conversation,
+            chat_template=chat_template,
+            **kwargs,
+        )
+
+        # Only post-process when:
+        # - user requested assistant mask
+        # - output is a dict (tokenized + return_dict=True path)
+        if isinstance(out, BatchFeature) and kwargs.get("return_assistant_tokens_mask", False):
+            if "assistant_masks" in out and out["assistant_masks"] is not None:
+                out["assistant_masks"] = _rotate_right_array(out["assistant_masks"], 730)
+
+        return out
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return tokenizer_input_names + image_processor_input_names + ["image_sizes"]
+
+
+__all__ = ["Moondream3Processor"]

processor_config.json

@@ -0,0 +1,7 @@
+{
+  "auto_map": {
+    "AutoProcessor": "processing_moondream3.Moondream3Processor"
+  },
+  "image_token_id": 0,
+  "processor_class": "Moondream3Processor"
+}

special_tokens_map.json

@@ -0,0 +1,10 @@
+{
+  "bos_token": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+  }
+}

tokenizer_config.json

@@ -0,0 +1,180 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<|md_reserved_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "<|md_reserved_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<|md_reserved_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "<|md_reserved_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "5": {
+      "content": "<|md_reserved_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "<|md_reserved_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "7": {
+      "content": "<|md_reserved_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "8": {
+      "content": "<|md_reserved_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "9": {
+      "content": "<|md_reserved_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "10": {
+      "content": "<|md_reserved_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "11": {
+      "content": "<|md_reserved_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "12": {
+      "content": "<|md_reserved_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "13": {
+      "content": "<|md_reserved_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "14": {
+      "content": "<|md_reserved_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "15": {
+      "content": "<|md_reserved_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "16": {
+      "content": "<|md_reserved_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "17": {
+      "content": "<|md_reserved_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "18": {
+      "content": "<|md_reserved_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "19": {
+      "content": "<|md_reserved_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "20": {
+      "content": "<|md_reserved_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "extra_special_tokens": {},
+  "model_max_length": 1000000000000000019884624838656,
+  "processor_class": "Moondream3Processor",
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}