import torch
from transformers import AutoProcessor, VoxtralForConditionalGeneration

model_id = "yujiepan/voxtral-tiny-random"

device = "cuda"
processor = AutoProcessor.from_pretrained(model_id)
model = VoxtralForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map=device)

conversation = [
    {
        "role": "user",
        "content": [
            {
                "type": "audio",
                "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/mary_had_lamb.mp3",
            },
            {
                "type": "audio",
                "path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/winning_call.mp3",
            },
            {"type": "text", "text": "What sport and what nursery rhyme are referenced?"},
        ],
    }
]

inputs = processor.apply_chat_template(conversation)
inputs = inputs.to(device, dtype=torch.bfloat16)

outputs = model.generate(**inputs, max_new_tokens=32)
decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)

print("\nGenerated response:")
print("=" * 80)
print(decoded_outputs[0])
print("=" * 80)

Codes to create this repo:

python

import json
from pathlib import Path

import accelerate
import torch
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModel,
    AutoModelForCausalLM,
    AutoProcessor,
    GenerationConfig,
    set_seed,
)

source_model_id = "mistralai/Voxtral-Small-24B-2507"
save_folder = "/tmp/yujiepan/voxtral-tiny-random"

processor = AutoProcessor.from_pretrained(source_model_id, trust_remote_code=True)
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(
    {
        "head_dim": 32,
        "hidden_size": 64,
        "intermediate_size": 256,
        "num_attention_heads": 2,
        "num_key_value_heads": 2,
        "num_hidden_layers": 2,
    }
)
config_json['hidden_size'] = 64
config_json['text_config'].update(
    {
        "head_dim": 32,
        "hidden_size": 64,
        "intermediate_size": 128,
        "num_attention_heads": 2,
        "num_key_value_heads": 1,
        "num_hidden_layers": 2,
        'tie_word_embeddings': True,
    }
)
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 = AutoModel.from_config(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()  # cpu is more stable for random initialization across machines
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.2)
        print(name, p.shape)
model.save_pretrained(save_folder)
print(model)

Printing the model:

text

VoxtralForConditionalGeneration(
  (audio_tower): VoxtralEncoder(
    (conv1): Conv1d(128, 64, kernel_size=(3,), stride=(1,), padding=(1,))
    (conv2): Conv1d(64, 64, kernel_size=(3,), stride=(2,), padding=(1,))
    (embed_positions): Embedding(1500, 64)
    (layers): ModuleList(
      (0-1): 2 x VoxtralEncoderLayer(
        (self_attn): VoxtralAttention(
          (k_proj): Linear(in_features=64, out_features=64, bias=False)
          (v_proj): Linear(in_features=64, out_features=64, bias=True)
          (q_proj): Linear(in_features=64, out_features=64, bias=True)
          (out_proj): Linear(in_features=64, out_features=64, bias=True)
        )
        (self_attn_layer_norm): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
        (activation_fn): GELUActivation()
        (fc1): Linear(in_features=64, out_features=256, bias=True)
        (fc2): Linear(in_features=256, out_features=64, bias=True)
        (final_layer_norm): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
      )
    )
    (layer_norm): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
    (avg_pooler): AvgPool1d(kernel_size=(2,), stride=(2,), padding=(0,))
  )
  (language_model): LlamaForCausalLM(
    (model): LlamaModel(
      (embed_tokens): Embedding(131072, 64)
      (layers): ModuleList(
        (0-1): 2 x LlamaDecoderLayer(
          (self_attn): LlamaAttention(
            (q_proj): Linear(in_features=64, out_features=64, bias=False)
            (k_proj): Linear(in_features=64, out_features=32, bias=False)
            (v_proj): Linear(in_features=64, out_features=32, bias=False)
            (o_proj): Linear(in_features=64, out_features=64, bias=False)
          )
          (mlp): LlamaMLP(
            (gate_proj): Linear(in_features=64, out_features=128, bias=False)
            (up_proj): Linear(in_features=64, out_features=128, bias=False)
            (down_proj): Linear(in_features=128, out_features=64, bias=False)
            (act_fn): SiLU()
          )
          (input_layernorm): LlamaRMSNorm((64,), eps=1e-05)
          (post_attention_layernorm): LlamaRMSNorm((64,), eps=1e-05)
        )
      )
      (norm): LlamaRMSNorm((64,), eps=1e-05)
      (rotary_emb): LlamaRotaryEmbedding()
    )
    (lm_head): Linear(in_features=64, out_features=131072, bias=False)
  )
  (multi_modal_projector): VoxtralMultiModalProjector(
    (linear_1): Linear(in_features=256, out_features=64, bias=False)
    (act): GELUActivation()
    (linear_2): Linear(in_features=64, out_features=64, bias=False)
  )
)