You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

199 lines
9.5 KiB

# Copyright 2025 Mistral AI and 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.
"""Ministral model configuration"""
from ...configuration_utils import PreTrainedConfig
from ...modeling_rope_utils import RopeParameters
from ...utils import logging
logger = logging.get_logger(__name__)
class Ministral3Config(PreTrainedConfig):
r"""
This is the configuration class to store the configuration of a [`Ministral3Model`]. It is used to instantiate an
Mistral model according to the specified arguments, defining the model architecture. Instantiating a configuration
with the defaults will yield a similar configuration to that of the mistralai/Ministral-3-8B-Base-2512, mistralai/Ministral-3-8B-Instruct-2512 or mistralai/Ministral-3-8B-Reasoning-2512.
[mistralai/Ministral-3-8B-Base-2512](https://huggingface.co/mistralai/Ministral-3-8B-Base-2512)
[mistralai/Ministral-3-8B-Instruct-2512](https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512)
[mistralai/Ministral-3-8B-Reasoning-2512](https://huggingface.co/mistralai/Ministral-3-8B-Reasoning-2512)
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 (`Optional`, *optional*, defaults to 131072):
Vocabulary size of the Ministral3 model. Defines the number of different tokens that can be represented by
the `inputs_ids` passed when calling [`Ministral3Model`].
hidden_size (`Optional`, *optional*, defaults to 4096):
Dimensionality of the embeddings and hidden states.
intermediate_size (`Optional`, *optional*, defaults to 14336):
Dimensionality of the intermediate (feed-forward) layer.
num_hidden_layers (`Optional`, *optional*, defaults to 34):
Number of hidden layers in the Transformer decoder.
num_attention_heads (`Optional`, *optional*, defaults to 32):
Number of attention heads for each attention layer in the Transformer decoder.
num_key_value_heads (`Optional`, *optional*, defaults to 8):
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA); if
`num_key_value_heads=1`, the model will use Multi Query Attention (MQA); otherwise GQA is used.
head_dim (`Optional`, *optional*, defaults to 128):
The attention head dimension. If not specified, will default to `hidden_size // num_attention_heads`.
hidden_act (`Optional`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the decoder.
max_position_embeddings (`Optional`, *optional*, defaults to 262144):
The maximum sequence length that this model might ever be used with.
initializer_range (`Optional`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
rms_norm_eps (`Optional`, *optional*, defaults to 1e-05):
The epsilon used by the rms normalization layers.
use_cache (`Optional`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
pad_token_id (`Optional`, *optional*, defaults to 11):
The id of the padding token.
bos_token_id (`Optional`, *optional*, defaults to 1):
The id of the "beginning-of-sequence" token.
eos_token_id (`Optional`, *optional*, defaults to 2):
The id of the "end-of-sequence" token.
tie_word_embeddings (`Optional`, *optional*, defaults to `False`):
Whether the model's input and output word embeddings should be tied.
rope_parameters (`Union`, *optional*, defaults to `{'type': 'yarn', 'rope_theta': 1000000.0, 'factor': 16.0, 'original_max_position_embeddings': 16384, 'beta_fast': 32.0, 'beta_slow': 1.0, 'mscale_all_dim': 1.0, 'mscale': 1.0, 'llama_4_scaling_beta': 0.1}`):
Dictionary containing the configuration parameters for the RoPE embeddings, including optional Yarn scaling
settings such as `factor`, `original_max_position_embeddings`, `mscale`, and `llama_4_scaling_beta`.
sliding_window (`Optional`, *optional*):
Sliding window attention window size. If `None`, full attention is used.
attention_dropout (`Optional`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
Example:
```python
>>> from transformers import Ministral3Config, Ministral3ForCausalLM, Mistral3Config, Mistral3ForConditionalGeneration, PixtralVisionConfig
>>> # Initializing a Pixtral-vision config
>>> vision_config = PixtralVisionConfig()
>>> # Initializing a Ministral3 config
>>> text_config = Ministral3Config()
>>> # Initializing a Mistral3 configuration
>>> configuration = Mistral3Config(vision_config, text_config)
>>> # Initializing a model from the Ministral3 configuration
>>> text_model = Ministral3ForCausalLM(text_config)
>>> # Initializing a model from the Mistral3 configuration
>>> model = Mistral3ForConditionalGeneration(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "ministral3"
keys_to_ignore_at_inference = ["past_key_values"]
# Default tensor parallel plan for base model `MistralModel`
base_model_tp_plan = {
"layers.*.self_attn.q_proj": "colwise",
"layers.*.self_attn.k_proj": "colwise",
"layers.*.self_attn.v_proj": "colwise",
"layers.*.self_attn.o_proj": "rowwise",
"layers.*.mlp.gate_proj": "colwise",
"layers.*.mlp.up_proj": "colwise",
"layers.*.mlp.down_proj": "rowwise",
}
base_model_pp_plan = {
"embed_tokens": (["input_ids"], ["inputs_embeds"]),
"layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
"norm": (["hidden_states"], ["hidden_states"]),
}
def __init__(
self,
vocab_size: int | None = 131072,
hidden_size: int | None = 4096,
intermediate_size: int | None = 14336,
num_hidden_layers: int | None = 34,
num_attention_heads: int | None = 32,
num_key_value_heads: int | None = 8,
head_dim: int | None = 128,
hidden_act: str | None = "silu",
max_position_embeddings: int | None = 262144,
initializer_range: float | None = 0.02,
rms_norm_eps: float | None = 1e-5,
use_cache: bool | None = True,
pad_token_id: int | None = 11,
bos_token_id: int | None = 1,
eos_token_id: int | None = 2,
tie_word_embeddings: bool | None = False,
rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
sliding_window: int | None = None,
attention_dropout: float | None = 0.0,
**kwargs,
):
if rope_parameters is None:
rope_parameters = {
"type": "yarn",
"rope_theta": 1000000.0,
"factor": 16.0,
"original_max_position_embeddings": 16384,
"max_position_embeddings": max_position_embeddings,
"beta_fast": 32.0,
"beta_slow": 1.0,
"mscale_all_dim": 1.0,
"mscale": 1.0,
"llama_4_scaling_beta": 0.1,
}
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.sliding_window = sliding_window
self.head_dim = head_dim if head_dim is not None else hidden_size // num_attention_heads
# for backward compatibility
if num_key_value_heads is None:
num_key_value_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_dropout = attention_dropout
if "layer_types" in kwargs:
logger.warning_once(
"Detected Mistral model with layer_types. Consider using AutoModel or Ministral classes instead to enable alternating attention compatibility."
)
self.rope_parameters = rope_parameters
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.eos_token_id = eos_token_id
self.tie_word_embeddings = tie_word_embeddings
super().__init__(
ignore_keys_at_rope_validation={"llama_4_scaling_beta", "max_position_embeddings"},
**kwargs,
)
__all__ = ["Ministral3Config"]