devstral-2

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from mistralai/Devstral-2-123B-Instruct-2512.

Example usage:

import torch
from transformers import Ministral3ForCausalLM, MistralCommonBackend

# Load model and tokenizer
model_id = "tiny-random/devstral-2"
model = Ministral3ForCausalLM.from_pretrained(
    model_id,
    device_map="cuda",
    torch_dtype="bfloat16",
    trust_remote_code=True,
)
tokenizer = MistralCommonBackend.from_pretrained(model_id)
messages = [
    {
        "role": "user",
        "content": "Hi",
    },
]

tokenized = tokenizer.apply_chat_template(
    messages, return_tensors="pt", return_dict=True)
output = model.generate(
    **tokenized.to("cuda"),
    max_new_tokens=32,
)[0]
decoded_output = tokenizer.decode(output[len(tokenized["input_ids"][0]):])
print(decoded_output)

Codes to create this repo:

import json
from pathlib import Path

import accelerate
import torch
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    GenerationConfig,
    Ministral3ForCausalLM,
    MistralCommonBackend,
    set_seed,
)

source_model_id = "mistralai/Devstral-2-123B-Instruct-2512"
save_folder = "/tmp/tiny-random/devstral-2"

processor = AutoProcessor.from_pretrained(
    source_model_id, trust_remote_code=True)
processor.save_pretrained(save_folder)
processor = MistralCommonBackend.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.update({
    "head_dim": 32,
    "hidden_size": 8,
    "intermediate_size": 64,
    "num_attention_heads": 8,
    "num_hidden_layers": 2,
    "num_key_value_heads": 4,
    "tie_word_embeddings": True,
})
del config_json['quantization_config']
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 = Ministral3ForCausalLM(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,
    )
    model.generation_config.do_sample = True
    print(model.generation_config)
model = model.cpu()
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.1)
        print(name, p.shape)
model.save_pretrained(save_folder)
print(model)

Printing the model:

Ministral3ForCausalLM(
  (model): Ministral3Model(
    (embed_tokens): Embedding(131072, 8, padding_idx=11)
    (layers): ModuleList(
      (0-1): 2 x Ministral3DecoderLayer(
        (self_attn): Ministral3Attention(
          (q_proj): Linear(in_features=8, out_features=256, bias=False)
          (k_proj): Linear(in_features=8, out_features=128, bias=False)
          (v_proj): Linear(in_features=8, out_features=128, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
        )
        (mlp): Ministral3MLP(
          (gate_proj): Linear(in_features=8, out_features=64, bias=False)
          (up_proj): Linear(in_features=8, out_features=64, bias=False)
          (down_proj): Linear(in_features=64, out_features=8, bias=False)
          (act_fn): SiLUActivation()
        )
        (input_layernorm): Ministral3RMSNorm((8,), eps=1e-05)
        (post_attention_layernorm): Ministral3RMSNorm((8,), eps=1e-05)
      )
    )
    (norm): Ministral3RMSNorm((8,), eps=1e-05)
    (rotary_emb): Ministral3RotaryEmbedding()
  )
  (lm_head): Linear(in_features=8, out_features=131072, bias=False)
)