Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,652 @@
+---
+library_name: transformers
+pipeline_tag: image-text-to-text
+inference: true
+widget:
+  - text: Hello!
+    example_title: Hello world
+    group: Python
+base_model:
+- google/gemma-3n-E4B-it
+---
+
+This tiny model is for debugging. It is randomly initialized with the config adapted from [google/gemma-3n-E4B-it](https://huggingface.co/google/gemma-3n-E4B-it).
+
+### Example usage:
+
+```python
+import torch
+
+from transformers import pipeline
+
+model_id = "tiny-random/gemma-3n"
+pipe = pipeline(
+    task="image-text-to-text",
+    model=model_id,
+    device=0,
+    torch_dtype=torch.bfloat16
+)
+
+# temporary patch for audio tower
+from accelerate.hooks import ModelHook, add_hook_to_module
+
+class EnsureDtype(ModelHook):
+    def pre_forward(self, module, *args, **kwargs):
+        args = list(args)
+        args[0] = args[0].to(module.dtype)
+        return super().pre_forward(module, *args, **kwargs)
+add_hook_to_module(pipe.model.audio_tower, EnsureDtype())
+
+messages = [
+    {
+        "role": "system",
+        "content": [
+            {"type": "text", "text": "You are a helpful assistant."}
+        ]
+    },
+    {
+        "role": "user",
+        "content": [
+            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"},
+            # audio is buggy for now: bf16 x fp32
+            {"type": "audio", "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Audio/glass-breaking-151256.mp3"},
+            {"type": "text", "text": "Which image is cuter?"},
+        ]
+    },
+]
+result = pipe(messages, min_new_tokens=512, max_new_tokens=512, do_sample=True)
+print(result)
+```
+
+### Codes to create this repo:
+
+```python
+import json
+from pathlib import Path
+
+import torch
+
+import accelerate
+from huggingface_hub import file_exists, hf_hub_download
+from timm.models.mobilenetv5 import decode_arch_def
+from transformers import (
+    AutoConfig,
+    AutoModelForCausalLM,
+    AutoProcessor,
+    AutoTokenizer,
+    Gemma3nForConditionalGeneration,
+    GenerationConfig,
+    set_seed,
+)
+
+source_model_id = "google/gemma-3n-E4B-it"
+save_folder = "/tmp/tiny-random/gemma-3n"
+
+processor = AutoProcessor.from_pretrained(source_model_id)
+processor.save_pretrained(save_folder)
+
+with open(hf_hub_download(source_model_id, filename='config.json', repo_type='model'), 'r', encoding='utf-8') as f:
+    config_json = json.load(f)
+
+config_json['audio_config'].update({
+    "conf_num_attention_heads": 2,
+    "conf_num_hidden_layers": 2,
+    "hidden_size": 64,
+})
+config_json['text_config'].update({
+    "activation_sparsity_pattern": [0.95, 0.95, 0.0, 0.0],
+    "head_dim": 32,  # required by vllm
+    "hidden_size": 32,
+    "hidden_size_per_layer_input": 2,
+    "intermediate_size": 64,
+    "laurel_rank": 8,
+    "layer_types": ['sliding_attention', 'full_attention', 'sliding_attention', 'full_attention'],
+    "num_attention_heads": 1,
+    "num_hidden_layers": 4,
+    "num_key_value_heads": 1,
+    "num_kv_shared_layers": 2,
+    "sliding_window": 512,
+})
+block_args = decode_arch_def(
+    [
+        # Stage 0: 128x128 in
+        [
+            'er_r1_k3_s2_e4_c32',
+            'er_r1_k3_s1_e4_c32',
+        ],
+        # Stage 1: 256x256 in
+        [
+            'uir_r1_a3_k5_s2_e6_c32',
+            'uir_r1_a5_k0_s1_e4_c32',
+            'uir_r1_a3_k0_s1_e4_c32',
+        ],
+        # Stage 2: 640x640 in
+        [
+            "uir_r1_a5_k5_s2_e6_c32",
+            "uir_r1_a0_k0_s1_e1_c32",
+            "mqa_r1_k3_h2_v2_s1_d64_c32",
+            "uir_r1_a0_k0_s1_e2_c32",
+        ],
+        # Stage 3: 1280x1280 in
+        [
+            "uir_r1_a5_k5_s2_e6_c32",
+            "mqa_r1_k3_h2_s1_d64_c32",
+            "uir_r1_a0_k0_s1_e2_c32",
+        ],
+    ]
+)
+config_json['vision_config'].update({
+    "hidden_size": 2048,  # hard-coded in timm
+    "model_args": {
+        "block_args": block_args,
+    }
+})
+
+with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
+    json.dump(config_json, f, indent=2)
+
+config = AutoConfig.from_pretrained(
+    save_folder,
+    trust_remote_code=True,
+)
+print(config)
+
+torch.set_default_dtype(torch.bfloat16)
+model = Gemma3nForConditionalGeneration(config)
+torch.set_default_dtype(torch.float32)
+if file_exists(filename="generation_config.json", repo_id=source_model_id, repo_type='model'):
+    model.generation_config = GenerationConfig.from_pretrained(
+        source_model_id, trust_remote_code=True,
+    )
+set_seed(42)
+model = model.cpu()
+all_numels = 0
+for name, p in sorted(model.named_parameters()):
+    all_numels += p.numel()
+with torch.no_grad():
+    for name, p in sorted(model.named_parameters()):
+        torch.nn.init.normal_(p, 0, 0.2)
+        print(name, p.shape, f'{p.numel() / all_numels * 100: .4f}%')
+model.save_pretrained(save_folder)
+```
+
+### Printing the model:
+
+```text
+Gemma3nForConditionalGeneration(
+  (model): Gemma3nModel(
+    (vision_tower): TimmWrapperModel(
+      (timm_model): MobileNetV5Encoder(
+        (conv_stem): ConvNormAct(
+          (conv): Conv2dSame(3, 64, kernel_size=(3, 3), stride=(2, 2), bias=False)
+          (bn): RmsNormAct2d(
+            (drop): Identity()
+            (act): GELU(approximate='none')
+          )
+        )
+        (blocks): Sequential(
+          (0): Sequential(
+            (0): EdgeResidual(
+              (conv_exp): Conv2dSame(64, 256, kernel_size=(3, 3), stride=(2, 2), bias=False)
+              (bn1): RmsNormAct2d(
+                (drop): Identity()
+                (act): GELU(approximate='none')
+              )
+              (aa): Identity()
+              (se): Identity()
+              (conv_pwl): Conv2d(256, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+              (bn2): RmsNormAct2d(
+                (drop): Identity()
+                (act): Identity()
+              )
+              (drop_path): Identity()
+            )
+            (1): EdgeResidual(
+              (conv_exp): Conv2d(32, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
+              (bn1): RmsNormAct2d(
+                (drop): Identity()
+                (act): GELU(approximate='none')
+              )
+              (aa): Identity()
+              (se): Identity()
+              (conv_pwl): Conv2d(128, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+              (bn2): RmsNormAct2d(
+                (drop): Identity()
+                (act): Identity()
+              )
+              (drop_path): Identity()
+            )
+          )
+          (1): Sequential(
+            (0): UniversalInvertedResidual(
+              (dw_start): ConvNormAct(
+                (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (pw_exp): ConvNormAct(
+                (conv): Conv2d(32, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (dw_mid): ConvNormAct(
+                (conv): Conv2dSame(192, 192, kernel_size=(5, 5), stride=(2, 2), groups=192, bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (se): Identity()
+              (pw_proj): ConvNormAct(
+                (conv): Conv2d(192, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (dw_end): Identity()
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+            (1): UniversalInvertedResidual(
+              (dw_start): ConvNormAct(
+                (conv): Conv2d(32, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=32, bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (pw_exp): ConvNormAct(
+                (conv): Conv2d(32, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (dw_mid): Identity()
+              (se): Identity()
+              (pw_proj): ConvNormAct(
+                (conv): Conv2d(128, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (dw_end): Identity()
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+            (2): UniversalInvertedResidual(
+              (dw_start): ConvNormAct(
+                (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (pw_exp): ConvNormAct(
+                (conv): Conv2d(32, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (dw_mid): Identity()
+              (se): Identity()
+              (pw_proj): ConvNormAct(
+                (conv): Conv2d(128, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (dw_end): Identity()
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+          )
+          (2): Sequential(
+            (0): UniversalInvertedResidual(
+              (dw_start): ConvNormAct(
+                (conv): Conv2d(32, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=32, bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (pw_exp): ConvNormAct(
+                (conv): Conv2d(32, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (dw_mid): ConvNormAct(
+                (conv): Conv2dSame(192, 192, kernel_size=(5, 5), stride=(2, 2), groups=192, bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (se): Identity()
+              (pw_proj): ConvNormAct(
+                (conv): Conv2d(192, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (dw_end): Identity()
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+            (1): UniversalInvertedResidual(
+              (dw_start): Identity()
+              (pw_exp): ConvNormAct(
+                (conv): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (dw_mid): Identity()
+              (se): Identity()
+              (pw_proj): ConvNormAct(
+                (conv): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (dw_end): Identity()
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+            (2): MobileAttention(
+              (norm): RmsNormAct2d(
+                (drop): Identity()
+                (act): Identity()
+              )
+              (attn): MultiQueryAttention2d(
+                (query): Sequential(
+                  (proj): Conv2d(32, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                )
+                (key): Sequential(
+                  (down_conv): Conv2dSame(32, 32, kernel_size=(3, 3), stride=(2, 2), groups=32, bias=False)
+                  (norm): RmsNorm2d()
+                  (proj): Conv2d(32, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                )
+                (value): Sequential(
+                  (down_conv): Conv2dSame(32, 32, kernel_size=(3, 3), stride=(2, 2), groups=32, bias=False)
+                  (norm): RmsNorm2d()
+                  (proj): Conv2d(32, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                )
+                (attn_drop): Dropout(p=0.0, inplace=False)
+                (output): Sequential(
+                  (proj): Conv2d(128, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                  (drop): Dropout(p=0.0, inplace=False)
+                )
+              )
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+            (3): UniversalInvertedResidual(
+              (dw_start): Identity()
+              (pw_exp): ConvNormAct(
+                (conv): Conv2d(32, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (dw_mid): Identity()
+              (se): Identity()
+              (pw_proj): ConvNormAct(
+                (conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (dw_end): Identity()
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+          )
+          (3): Sequential(
+            (0): UniversalInvertedResidual(
+              (dw_start): ConvNormAct(
+                (conv): Conv2d(32, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=32, bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (pw_exp): ConvNormAct(
+                (conv): Conv2d(32, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (dw_mid): ConvNormAct(
+                (conv): Conv2dSame(192, 192, kernel_size=(5, 5), stride=(2, 2), groups=192, bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (se): Identity()
+              (pw_proj): ConvNormAct(
+                (conv): Conv2d(192, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (dw_end): Identity()
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+            (1): MobileAttention(
+              (norm): RmsNormAct2d(
+                (drop): Identity()
+                (act): Identity()
+              )
+              (attn): MultiQueryAttention2d(
+                (query): Sequential(
+                  (proj): Conv2d(32, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                )
+                (key): Sequential(
+                  (proj): Conv2d(32, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                )
+                (value): Sequential(
+                  (proj): Conv2d(32, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                )
+                (attn_drop): Dropout(p=0.0, inplace=False)
+                (output): Sequential(
+                  (proj): Conv2d(128, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                  (drop): Dropout(p=0.0, inplace=False)
+                )
+              )
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+            (2): UniversalInvertedResidual(
+              (dw_start): Identity()
+              (pw_exp): ConvNormAct(
+                (conv): Conv2d(32, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): GELU(approximate='none')
+                )
+              )
+              (dw_mid): Identity()
+              (se): Identity()
+              (pw_proj): ConvNormAct(
+                (conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
+                (bn): RmsNormAct2d(
+                  (drop): Identity()
+                  (act): Identity()
+                )
+              )
+              (dw_end): Identity()
+              (layer_scale): LayerScale2d()
+              (drop_path): Identity()
+            )
+          )
+        )
+        (msfa): MobileNetV5MultiScaleFusionAdapter(
+          (ffn): UniversalInvertedResidual(
+            (dw_start): Identity()
+            (pw_exp): ConvNormAct(
+              (conv): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
+              (bn): RmsNormAct2d(
+                (drop): Identity()
+                (act): GELU(approximate='none')
+              )
+            )
+            (dw_mid): Identity()
+            (se): Identity()
+            (pw_proj): ConvNormAct(
+              (conv): Conv2d(128, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)
+              (bn): RmsNormAct2d(
+                (drop): Identity()
+                (act): Identity()
+              )
+            )
+            (dw_end): Identity()
+            (layer_scale): Identity()
+            (drop_path): Identity()
+          )
+          (norm): RmsNorm2d()
+        )
+      )
+    )
+    (language_model): Gemma3nTextModel(
+      (embed_tokens): Gemma3nTextScaledWordEmbedding(262400, 32, padding_idx=0)
+      (layers): ModuleList(
+        (0-3): 4 x Gemma3nTextDecoderLayer(
+          (self_attn): Gemma3nTextAttention(
+            (q_proj): Linear(in_features=32, out_features=32, bias=False)
+            (k_proj): Linear(in_features=32, out_features=32, bias=False)
+            (v_proj): Linear(in_features=32, out_features=32, bias=False)
+            (o_proj): Linear(in_features=32, out_features=32, bias=False)
+            (q_norm): Gemma3nRMSNorm((32,), eps=1e-06)
+            (k_norm): Gemma3nRMSNorm((32,), eps=1e-06)
+            (v_norm): Gemma3nRMSNorm((), eps=1e-06)
+          )
+          (mlp): Gemma3nTextMLP(
+            (gate_proj): Linear(in_features=32, out_features=64, bias=False)
+            (up_proj): Linear(in_features=32, out_features=64, bias=False)
+            (down_proj): Linear(in_features=64, out_features=32, bias=False)
+            (act_fn): PytorchGELUTanh()
+          )
+          (input_layernorm): Gemma3nRMSNorm((32,), eps=1e-06)
+          (post_attention_layernorm): Gemma3nRMSNorm((32,), eps=1e-06)
+          (pre_feedforward_layernorm): Gemma3nRMSNorm((32,), eps=1e-06)
+          (post_feedforward_layernorm): Gemma3nRMSNorm((32,), eps=1e-06)
+          (act_fn): PytorchGELUTanh()
+          (altup): Gemma3nTextAltUp(
+            (correction_coefs): Linear(in_features=4, out_features=4, bias=False)
+            (prediction_coefs): Linear(in_features=4, out_features=16, bias=False)
+            (modality_router): Linear(in_features=32, out_features=4, bias=False)
+            (router_norm): Gemma3nRMSNorm((32,), eps=1e-06)
+          )
+          (laurel): Gemma3nTextLaurelBlock(
+            (linear_left): Linear(in_features=32, out_features=8, bias=False)
+            (linear_right): Linear(in_features=8, out_features=32, bias=False)
+            (post_laurel_norm): Gemma3nRMSNorm((32,), eps=1e-06)
+          )
+          (per_layer_input_gate): Linear(in_features=32, out_features=2, bias=False)
+          (per_layer_projection): Linear(in_features=2, out_features=32, bias=False)
+          (post_per_layer_input_norm): Gemma3nRMSNorm((32,), eps=1e-06)
+        )
+      )
+      (norm): Gemma3nRMSNorm((32,), eps=1e-06)
+      (rotary_emb): Gemma3nTextRotaryEmbedding()
+      (rotary_emb_local): Gemma3nTextRotaryEmbedding()
+      (embed_tokens_per_layer): Gemma3nTextScaledWordEmbedding(262144, 8, padding_idx=0)
+      (per_layer_model_projection): Linear(in_features=32, out_features=8, bias=False)
+      (per_layer_projection_norm): Gemma3nRMSNorm((2,), eps=1e-06)
+      (altup_projections): ModuleList(
+        (0-2): 3 x Linear(in_features=32, out_features=32, bias=False)
+      )
+      (altup_unembed_projections): ModuleList(
+        (0-2): 3 x Linear(in_features=32, out_features=32, bias=False)
+      )
+    )
+    (audio_tower): Gemma3nAudioEncoder(
+      (subsample_conv_projection): Gemma3nAudioSubSampleConvProjection(
+        (conv_0): Gemma3nAudioSSCPConvBlock(
+          (conv): Conv2d(1, 128, kernel_size=(3, 3), stride=(2, 2), bias=False)
+          (norm): Gemma3nAudioCumulativeGroupNorm()
+          (activation): ReLU()
+        )
+        (conv_1): Gemma3nAudioSSCPConvBlock(
+          (conv): Conv2d(128, 32, kernel_size=(3, 3), stride=(2, 2), bias=False)
+          (norm): Gemma3nAudioCumulativeGroupNorm()
+          (activation): ReLU()
+        )
+        (input_proj_linear): Linear(in_features=1024, out_features=64, bias=False)
+      )
+      (conformer): ModuleList(
+        (0-1): 2 x Gemma3nAudioConformerBlock(
+          (ffw_layer_start): Gemma3nAudioConformerFeedForward(
+            (pre_layer_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+            (ffw_layer_1): Linear(in_features=64, out_features=256, bias=False)
+            (ffw_layer_2): Linear(in_features=256, out_features=64, bias=False)
+            (post_layer_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+          )
+          (attention): Gemma3nAudioConformerAttention(
+            (pre_attn_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+            (attn): Gemma3nAudioAttention(
+              (relative_position_embedding): Gemma3nAudioRelativePositionEmbedding(
+                (pos_proj): Linear(in_features=64, out_features=64, bias=False)
+              )
+              (q_proj): Linear(in_features=64, out_features=64, bias=False)
+              (k_proj): Linear(in_features=64, out_features=64, bias=False)
+              (v_proj): Linear(in_features=64, out_features=64, bias=False)
+            )
+            (post): Linear(in_features=64, out_features=64, bias=False)
+            (post_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+          )
+          (lconv1d): Gemma3nAudioConformerLightConv1d(
+            (pre_layer_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+            (linear_start): Linear(in_features=64, out_features=128, bias=False)
+            (depthwise_conv1d): Conv1d(64, 64, kernel_size=(5,), stride=(1,), groups=64, bias=False)
+            (conv_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+            (linear_end): Linear(in_features=64, out_features=64, bias=False)
+          )
+          (ffw_layer_end): Gemma3nAudioConformerFeedForward(
+            (pre_layer_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+            (ffw_layer_1): Linear(in_features=64, out_features=256, bias=False)
+            (ffw_layer_2): Linear(in_features=256, out_features=64, bias=False)
+            (post_layer_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+          )
+          (norm): Gemma3nRMSNorm((64,), eps=1e-06)
+        )
+      )
+    )
+    (embed_vision): Gemma3nMultimodalEmbedder(
+      (embedding): Embedding(128, 2048)
+      (hard_embedding_norm): Gemma3nRMSNorm((2048,), eps=1e-06)
+      (soft_embedding_norm): Gemma3nRMSNorm((2048,), eps=1e-06)
+      (embedding_projection): Linear(in_features=2048, out_features=32, bias=False)
+      (embedding_post_projection_norm): Gemma3nRMSNorm((), eps=1e-06)
+    )
+    (embed_audio): Gemma3nMultimodalEmbedder(
+      (embedding): Embedding(128, 64)
+      (hard_embedding_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+      (soft_embedding_norm): Gemma3nRMSNorm((64,), eps=1e-06)
+      (embedding_projection): Linear(in_features=64, out_features=32, bias=False)
+      (embedding_post_projection_norm): Gemma3nRMSNorm((), eps=1e-06)
+    )
+  )
+  (lm_head): Linear(in_features=32, out_features=262400, bias=False)
+)
+```

chat_template.jinja

@@ -0,0 +1,49 @@
+{{ bos_token }}
+{%- if messages[0]['role'] == 'system' -%}
+    {%- if messages[0]['content'] is string -%}
+        {%- set first_user_prefix = messages[0]['content'] + '
+
+' -%}
+    {%- else -%}
+        {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
+
+' -%}
+    {%- endif -%}
+    {%- set loop_messages = messages[1:] -%}
+{%- else -%}
+    {%- set first_user_prefix = "" -%}
+    {%- set loop_messages = messages -%}
+{%- endif -%}
+{%- for message in loop_messages -%}
+    {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
+        {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
+    {%- endif -%}
+    {%- if (message['role'] == 'assistant') -%}
+        {%- set role = "model" -%}
+    {%- else -%}
+        {%- set role = message['role'] -%}
+    {%- endif -%}
+    {{ '<start_of_turn>' + role + '
+' + (first_user_prefix if loop.first else "") }}
+    {%- if message['content'] is string -%}
+        {{ message['content'] | trim }}
+    {%- elif message['content'] is iterable -%}
+        {%- for item in message['content'] -%}
+            {%- if item['type'] == 'audio' -%}
+                {{ '<audio_soft_token>' }}
+            {%- elif item['type'] == 'image' -%}
+                {{ '<image_soft_token>' }}
+            {%- elif item['type'] == 'text' -%}
+                {{ item['text'] | trim }}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- else -%}
+        {{ raise_exception("Invalid content type") }}
+    {%- endif -%}
+    {{ '<end_of_turn>
+' }}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    {{'<start_of_turn>model
+'}}
+{%- endif -%}

config.json

@@ -0,0 +1,289 @@
+{
+  "architectures": [
+    "Gemma3nForConditionalGeneration"
+  ],
+  "audio_config": {
+    "conf_attention_chunk_size": 12,
+    "conf_attention_context_left": 13,
+    "conf_attention_context_right": 0,
+    "conf_attention_logit_cap": 50.0,
+    "conf_conv_kernel_size": 5,
+    "conf_num_attention_heads": 2,
+    "conf_num_hidden_layers": 2,
+    "conf_positional_bias_size": 256,
+    "conf_reduction_factor": 4,
+    "conf_residual_weight": 0.5,
+    "gradient_clipping": 10000000000.0,
+    "hidden_size": 64,
+    "input_feat_size": 128,
+    "model_type": "gemma3n_audio",
+    "rms_norm_eps": 1e-06,
+    "sscp_conv_channel_size": [
+      128,
+      32
+    ],
+    "sscp_conv_eps": 0.001,
+    "sscp_conv_group_norm_eps": 0.001,
+    "sscp_conv_kernel_size": [
+      [
+        3,
+        3
+      ],
+      [
+        3,
+        3
+      ]
+    ],
+    "sscp_conv_stride_size": [
+      [
+        2,
+        2
+      ],
+      [
+        2,
+        2
+      ]
+    ],
+    "torch_dtype": "bfloat16",
+    "vocab_offset": 262272,
+    "vocab_size": 128
+  },
+  "audio_soft_tokens_per_image": 188,
+  "audio_token_id": 262273,
+  "boa_token_id": 256000,
+  "boi_token_id": 255999,
+  "eoa_token_id": 262272,
+  "eoi_token_id": 262144,
+  "eos_token_id": [
+    1,
+    106
+  ],
+  "image_token_id": 262145,
+  "initializer_range": 0.02,
+  "model_type": "gemma3n",
+  "text_config": {
+    "activation_sparsity_pattern": [
+      0.95,
+      0.95,
+      0.0,
+      0.0
+    ],
+    "altup_active_idx": 0,
+    "altup_coef_clip": 120.0,
+    "altup_correct_scale": true,
+    "altup_lr_multiplier": 1.0,
+    "altup_num_inputs": 4,
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "final_logit_softcapping": 30.0,
+    "head_dim": 32,
+    "hidden_activation": "gelu_pytorch_tanh",
+    "hidden_size": 32,
+    "hidden_size_per_layer_input": 2,
+    "initializer_range": 0.02,
+    "intermediate_size": [
+      64,
+      64,
+      64,
+      64
+    ],
+    "laurel_rank": 8,
+    "layer_types": [
+      "sliding_attention",
+      "full_attention",
+      "sliding_attention",
+      "full_attention"
+    ],
+    "max_position_embeddings": 32768,
+    "model_type": "gemma3n_text",
+    "num_attention_heads": 1,
+    "num_hidden_layers": 4,
+    "num_key_value_heads": 1,
+    "num_kv_shared_layers": 2,
+    "query_pre_attn_scalar": 256,
+    "rms_norm_eps": 1e-06,
+    "rope_local_base_freq": 10000.0,
+    "rope_scaling": null,
+    "rope_theta": 1000000.0,
+    "sliding_window": 512,
+    "torch_dtype": "bfloat16",
+    "use_cache": true,
+    "vocab_size": 262400,
+    "vocab_size_per_layer_input": 262144
+  },
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.54.0.dev0",
+  "vision_config": {
+    "architecture": "mobilenetv5_300m_enc",
+    "do_pooling": true,
+    "hidden_size": 2048,
+    "initializer_range": 0.02,
+    "label_names": [
+      "LABEL_0",
+      "LABEL_1"
+    ],
+    "model_args": {
+      "block_args": [
+        [
+          {
+            "act_layer": null,
+            "block_type": "er",
+            "exp_kernel_size": 3,
+            "exp_ratio": 4.0,
+            "force_in_chs": 0,
+            "noskip": false,
+            "out_chs": 32,
+            "pw_kernel_size": 1,
+            "se_ratio": 0.0,
+            "stride": 2
+          },
+          {
+            "act_layer": null,
+            "block_type": "er",
+            "exp_kernel_size": 3,
+            "exp_ratio": 4.0,
+            "force_in_chs": 0,
+            "noskip": false,
+            "out_chs": 32,
+            "pw_kernel_size": 1,
+            "se_ratio": 0.0,
+            "stride": 1
+          }
+        ],
+        [
+          {
+            "act_layer": null,
+            "block_type": "uir",
+            "dw_kernel_size_end": 0,
+            "dw_kernel_size_mid": 5,
+            "dw_kernel_size_start": 3,
+            "exp_ratio": 6.0,
+            "noskip": false,
+            "out_chs": 32,
+            "se_ratio": 0.0,
+            "stride": 2
+          },
+          {
+            "act_layer": null,
+            "block_type": "uir",
+            "dw_kernel_size_end": 0,
+            "dw_kernel_size_mid": 0,
+            "dw_kernel_size_start": 5,
+            "exp_ratio": 4.0,
+            "noskip": false,
+            "out_chs": 32,
+            "se_ratio": 0.0,
+            "stride": 1
+          },
+          {
+            "act_layer": null,
+            "block_type": "uir",
+            "dw_kernel_size_end": 0,
+            "dw_kernel_size_mid": 0,
+            "dw_kernel_size_start": 3,
+            "exp_ratio": 4.0,
+            "noskip": false,
+            "out_chs": 32,
+            "se_ratio": 0.0,
+            "stride": 1
+          }
+        ],
+        [
+          {
+            "act_layer": null,
+            "block_type": "uir",
+            "dw_kernel_size_end": 0,
+            "dw_kernel_size_mid": 5,
+            "dw_kernel_size_start": 5,
+            "exp_ratio": 6.0,
+            "noskip": false,
+            "out_chs": 32,
+            "se_ratio": 0.0,
+            "stride": 2
+          },
+          {
+            "act_layer": null,
+            "block_type": "uir",
+            "dw_kernel_size_end": 0,
+            "dw_kernel_size_mid": 0,
+            "dw_kernel_size_start": 0,
+            "exp_ratio": 1.0,
+            "noskip": false,
+            "out_chs": 32,
+            "se_ratio": 0.0,
+            "stride": 1
+          },
+          {
+            "act_layer": null,
+            "block_type": "mqa",
+            "dw_kernel_size": 3,
+            "key_dim": 64,
+            "kv_stride": 2,
+            "noskip": false,
+            "num_heads": 2,
+            "out_chs": 32,
+            "stride": 1,
+            "value_dim": 64
+          },
+          {
+            "act_layer": null,
+            "block_type": "uir",
+            "dw_kernel_size_end": 0,
+            "dw_kernel_size_mid": 0,
+            "dw_kernel_size_start": 0,
+            "exp_ratio": 2.0,
+            "noskip": false,
+            "out_chs": 32,
+            "se_ratio": 0.0,
+            "stride": 1
+          }
+        ],
+        [
+          {
+            "act_layer": null,
+            "block_type": "uir",
+            "dw_kernel_size_end": 0,
+            "dw_kernel_size_mid": 5,
+            "dw_kernel_size_start": 5,
+            "exp_ratio": 6.0,
+            "noskip": false,
+            "out_chs": 32,
+            "se_ratio": 0.0,
+            "stride": 2
+          },
+          {
+            "act_layer": null,
+            "block_type": "mqa",
+            "dw_kernel_size": 3,
+            "key_dim": 64,
+            "kv_stride": 1,
+            "noskip": false,
+            "num_heads": 2,
+            "out_chs": 32,
+            "stride": 1,
+            "value_dim": 64
+          },
+          {
+            "act_layer": null,
+            "block_type": "uir",
+            "dw_kernel_size_end": 0,
+            "dw_kernel_size_mid": 0,
+            "dw_kernel_size_start": 0,
+            "exp_ratio": 2.0,
+            "noskip": false,
+            "out_chs": 32,
+            "se_ratio": 0.0,
+            "stride": 1
+          }
+        ]
+      ]
+    },
+    "model_type": "gemma3n_vision",
+    "num_classes": 2,
+    "rms_norm_eps": 1e-06,
+    "torch_dtype": "bfloat16",
+    "vocab_offset": 262144,
+    "vocab_size": 128
+  },
+  "vision_soft_tokens_per_image": 256
+}

generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 2,
+  "cache_implementation": "hybrid",
+  "do_sample": true,
+  "eos_token_id": [
+    1,
+    106
+  ],
+  "pad_token_id": 0,
+  "top_k": 64,
+  "top_p": 0.95,
+  "transformers_version": "4.54.0.dev0",
+  "trust_remote_code": true
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07cf4ccabeb400baa5b96ad6df316bb7ff86578f9450bf5b1e2a226e15c448f5
+size 23576116

preprocessor_config.json

@@ -0,0 +1,51 @@
+{
+  "crop_size": null,
+  "data_format": "channels_first",
+  "default_to_square": false,
+  "device": null,
+  "disable_grouping": null,
+  "dither": 0.0,
+  "do_center_crop": null,
+  "do_convert_rgb": null,
+  "do_normalize": false,
+  "do_rescale": true,
+  "do_resize": true,
+  "feature_size": 128,
+  "fft_length": 1024,
+  "fft_overdrive": true,
+  "frame_length": 512,
+  "hop_length": 160,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_processor_type": "SiglipImageProcessorFast",
+  "image_seq_length": 256,
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "input_data_format": null,
+  "input_scale_factor": 1.0,
+  "max_frequency": 7600.0,
+  "mel_floor": 1e-05,
+  "min_frequency": 125.0,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "per_bin_mean": null,
+  "per_bin_stddev": null,
+  "preemphasis": 0.97,
+  "preemphasis_htk_flavor": true,
+  "processor_class": "Gemma3nProcessor",
+  "resample": 2,
+  "rescale_factor": 0.00392156862745098,
+  "return_attention_mask": false,
+  "return_tensors": null,
+  "sampling_rate": 16000,
+  "size": {
+    "height": 768,
+    "width": 768
+  }
+}

processor_config.json

@@ -0,0 +1,5 @@
+{
+  "audio_seq_length": 188,
+  "image_seq_length": 256,
+  "processor_class": "Gemma3nProcessor"
+}

special_tokens_map.json

@@ -0,0 +1,36 @@
+{
+  "audio_token": "<audio_soft_token>",
+  "boa_token": "<start_of_audio>",
+  "boi_token": "<start_of_image>",
+  "bos_token": {
+    "content": "<bos>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eoa_token": "<end_of_audio>",
+  "eoi_token": "<end_of_image>",
+  "eos_token": {
+    "content": "<eos>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "image_token": "<image_soft_token>",
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6c35ee648c07754b44cd9e371c75d4caa05c4504910b7ad29b1847ee9d8ba5d
+size 33442553

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea5f0cc48abfbfc04d14562270a32e02149a3e7035f368cc5a462786f4a59961
+size 4696020