google

/

gemma-3n-E4B-it

like 919

Follow

Google 64.3k

Hi @EsatKara , Gemma-3n-E4Bmodels requires Transformers version 4.53.0. Please update your Transformers library to at least version 4.53.0 to use Gemma-3n-E4B and let us know if you still face the same issue. Thank you

Hi @EsatKara , Gemma-3n-E4Bmodels requires Transformers version 4.53.0. Please update your Transformers library to at least version 4.53.0 to use Gemma-3n-E4B and let us know if you still face the same issue. Thank you

I upgraded Transformers to 4.53.0 with vllm of 0.9.1, but encountered the same issue

Hi @EsatKara I’m using the same environment as you. I suspect the root cause is that the config.json does not explicitly define top-level keys like vocab_size, num_hidden_layers, num_attention_heads, etc. Instead, these values are nested under the "audio_config" field, like so:

json
"audio_config": {
"vocab_size": 128,
"hidden_size": 1536,
"conf_num_attention_heads": 8,
"conf_num_hidden_layers": 12,
...
}
In my case, I manually copied relevant values (like vocab_size, num_attention_heads, num_hidden_layers, etc.) from audio_config and added them as top-level fields in config.json. After that, VLLM was able to launch successfully.

However, I’m not sure if this workaround would affect model performance or correctness — since this is not an officially supported or standard method. I also suspect this issue could be related to VLLM not fully supporting models with nested configs, or possibly a version mismatch.

@EsatKara vllm 0.9.1 does not support gemma 3n yet, see https://github.com/vllm-project/vllm/pull/20134 the PR only got merged 4 days ago. You need to wait for new release or build from source.

HI, @Lqqs
could you please share the config file.the json that you edited

HI @kloblok
Here is my JSON file. I hope it will be helpful to you.
I am currently using transformers version 4.53.0 and vllm version 0.7.4. With the following config.json, I am now able to run vLLM successfully.

{
"architectures": [
"Gemma3nForConditionalGeneration"
],
"hidden_activation": "gelu_pytorch_tanh",
"hidden_size": 2048,
"hidden_size_per_layer_input": 256,
"initializer_range": 0.02,
"intermediate_size": 16384,
"laurel_rank": 64,
"vocab_size": 262400,
"head_dim": 256,
"base_model_tp_plan": {},
"max_position_embeddings": 32768,
"model_type": "gemma3n_text",
"num_attention_heads": 8,
"num_hidden_layers": 35,
"num_key_value_heads": 2,
"num_kv_shared_layers": 15,
"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,
"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": 8,
"conf_num_hidden_layers": 12,
"conf_positional_bias_size": 256,
"conf_reduction_factor": 4,
"conf_residual_weight": 0.5,
"gradient_clipping": 10000000000.0,
"hidden_size": 1536,
"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_kernel_size": [
[
3,
3
],
[
3,
3
]
],
"sscp_conv_stride_size": [
[
2,
2
],
[
2,
2
]
],
"torch_dtype": "bfloat16",
"vocab_offset": 262273,
"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,
0.95,
0.95,
0.95,
0.95,
0.95,
0.95,
0.95,
0.95,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
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": 256,
"hidden_activation": "gelu_pytorch_tanh",
"hidden_size": 2048,
"hidden_size_per_layer_input": 256,
"initializer_range": 0.02,
"intermediate_size": 16384,
"laurel_rank": 64,
"layer_types": [
"sliding_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention"
],
"max_position_embeddings": 32768,
"model_type": "gemma3n_text",
"num_attention_heads": 8,
"num_hidden_layers": 35,
"num_key_value_heads": 2,
"num_kv_shared_layers": 15,
"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": "4.53.0.dev0",
"vision_config": {
"architecture": "mobilenetv5_300m_enc",
"do_pooling": true,
"hidden_size": 2048,
"initializer_range": 0.02,
"label_names": [
"LABEL_0",
"LABEL_1"
],
"model_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
}

Additionally, using fp32 precision during vLLM deployment may lead to runtime errors.DEBUG 07-01 10:59:16 client.py:171] Heartbeat successful. Sampling probabilities contain NaN or Inf values. So far, the model runs more stably with bfloat16 (bf16) precision.