Hugging Face's logo

Transformers documentation

DeepSeek-V3.2

Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

This model was published in HF papers on 2025-12-02 and contributed to Hugging Face Transformers on 2026-06-10.

PyTorch FlashAttention SDPA FP8

DeepSeek-V3.2

Overview

DeepSeek-V3.2-Exp is an experimental release from DeepSeek-AI that introduces DeepSeek Sparse Attention (DSA), a trainable, fine-grained sparse attention mechanism designed to improve training and inference efficiency in long-context scenarios. It is built directly on top of DeepSeek-V3.1-Terminus: the model keeps the same 685B-parameter Mixture-of-Experts (MoE) backbone and Multi-head Latent Attention (MLA), and is obtained through continued training that adds the sparse-attention indexer while deliberately aligning the training distribution with V3.1-Terminus so the two models can be compared head-to-head.

The work was later extended in the DeepSeek-V3.2 technical report, DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models, which pairs DSA with a scalable reinforcement-learning framework and reports gold-medal level results on competition math (IMO) and competitive programming (IOI) benchmarks.

The abstract from the DeepSeek-V3.2-Exp release is the following:

We introduce DeepSeek-V3.2-Exp, an experimental version of our model that incorporates DeepSeek Sparse Attention (DSA) to explore and validate optimizations for training and inference efficiency in long-context scenarios. DeepSeek Sparse Attention achieves fine-grained sparse attention for the first time with minimal impact on model output quality. Built upon DeepSeek-V3.1-Terminus, DeepSeek-V3.2-Exp delivers substantially improved efficiency in both training and inference, especially in long-context settings, while maintaining virtually identical benchmark performance.

DeepSeek Sparse Attention (DSA)

DSA reduces the quadratic cost of attention over long sequences by attending only to a selected subset of past tokens. It has two components:

  1. Lightning indexer. A lightweight, low-head-count scoring module computes an index score between each query and every preceding key. In the reference implementation it runs in FP8 with a Hadamard (rotate_activation) transform; because the transform is orthogonal (Hq·Hk = q·k) and FP8 is only a precision optimization, the transformers port computes the same scores directly in bf16/fp32, keeping the indexer cheap relative to the main attention.
  2. Fine-grained token selection. For each query the indexer keeps the top-index_topk (2048 by default) tokens, and main MLA attention is then computed only over those tokens via an additive mask. This turns the per-query attention cost from O(L) to O(index_topk) for long sequences when using flash_mla, which is not supported yet 😉.

The indexer keeps its own small per-token key cache (single-head, index_head_dim) alongside the main K/V cache. In transformers this lives on a dedicated cache layer — DynamicIndexedLayer for growing caches and StaticIndexedLayer for static / torch.compile caches — and is updated through past_key_values.update_indexer().

In DeepSeek-V3.2 every layer runs its own indexer — there is no cross-layer top-k sharing.

The MLA query LoRA path (q_lora_rank) is required. The indexer scores queries from the low-rank query latent q_a_layernorm(q_a_proj(x)) (its wq_b projection is sized by q_lora_rank), so the model always uses the LoRA query path and q_lora_rank must be set — the released checkpoint uses 1536. The optional non-LoRA q_proj path that DeepSeek-V3 exposes for q_lora_rank=None is not supported here: without the query latent there is nothing for the indexer to consume.

Usage examples

DeepSeek-V3.2-Exp is distributed as an FP8 checkpoint. The indexer projections are kept out of FP8 quantization, since the checkpoint stores them in bf16/fp32:

from transformers import FineGrainedFP8Config, AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "deepseek-ai/DeepSeek-V3.2-Exp"
quantization_config = FineGrainedFP8Config(
    modules_to_not_convert=["model.layers.*.mlp.gate.*", "*.self_attn.indexer.weights_proj.*"],
    weight_block_size=(128, 128),
)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
    quantization_config=quantization_config,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

inputs = tokenizer("What are we having for dinner?", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

The original code can be found here.

DeepseekV32Config

class transformers.DeepseekV32Config

< >

( transformers_version: str | None = None architectures: list[str] | None = None output_hidden_states: bool | None = False return_dict: bool | None = True dtype: typing.Union[str, ForwardRef('torch.dtype'), NoneType] = None chunk_size_feed_forward: int = 0 is_encoder_decoder: bool = False id2label: dict[int, str] | dict[str, str] | None = None label2id: dict[str, int] | dict[str, str] | None = None problem_type: typing.Optional[typing.Literal['regression', 'single_label_classification', 'multi_label_classification']] = None vocab_size: int = 129280 hidden_size: int = 7168 intermediate_size: int = 18432 moe_intermediate_size: int = 2048 num_hidden_layers: int = 61 num_attention_heads: int = 128 num_key_value_heads: int = 128 n_shared_experts: int = 1 n_routed_experts: int = 256 routed_scaling_factor: float = 2.5 kv_lora_rank: int = 512 q_lora_rank: int = 1536 qk_rope_head_dim: int = 64 v_head_dim: int = 128 qk_nope_head_dim: int = 128 n_group: int = 8 topk_group: int = 4 num_experts_per_tok: int = 8 norm_topk_prob: bool = True hidden_act: str = 'silu' max_position_embeddings: int = 163840 initializer_range: float = 0.02 rms_norm_eps: float = 1e-06 use_cache: bool = True pad_token_id: int | None = None bos_token_id: int | None = 0 eos_token_id: int | list[int] | None = 1 tie_word_embeddings: bool = False rope_parameters: dict | None = None mlp_layer_types: list[str] | None = None attention_bias: bool = False attention_dropout: float | int = 0.0 index_topk: int = 2048 index_head_dim: int = 128 index_n_heads: int = 64 mlp_bias: bool = False num_experts: int = 256 head_dim: int = 64 first_k_dense_replace: int = 3 layer_types: list[str] | None = None )

Parameters

  • vocab_size (int, optional, defaults to 129280) — Vocabulary size of the model. Defines the number of different tokens that can be represented by the input_ids.
  • hidden_size (int, optional, defaults to 7168) — Dimension of the hidden representations.
  • intermediate_size (int, optional, defaults to 18432) — Dimension of the MLP representations.
  • moe_intermediate_size (int, optional, defaults to 2048) — Intermediate size of the routed expert MLPs.
  • num_hidden_layers (int, optional, defaults to 61) — Number of hidden layers in the Transformer decoder.
  • num_attention_heads (int, optional, defaults to 128) — Number of attention heads for each attention layer in the Transformer decoder.
  • num_key_value_heads (int, optional, defaults to 128) — 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. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details, check out this paper. If it is not specified, will default to num_attention_heads.
  • n_shared_experts (int, optional, defaults to 1) — Number of shared experts.
  • n_routed_experts (int, optional, defaults to 256) — Number of routed experts.
  • routed_scaling_factor (float, optional, defaults to 2.5) — Scaling factor or routed experts.
  • kv_lora_rank (int, optional, defaults to 512) — Rank of the LoRA matrices for key and value projections.
  • q_lora_rank (int, optional, defaults to 1536) — Rank of the LoRA matrices for query projections.
  • qk_rope_head_dim (int, optional, defaults to 64) — Dimension of the query/key heads that use rotary position embeddings.
  • v_head_dim (int, optional, defaults to 128) — Dimension of the value heads.
  • qk_nope_head_dim (int, optional, defaults to 128) — Dimension of the query/key heads that don’t use rotary position embeddings.
  • n_group (int, optional, defaults to 1) — Number of groups for routed experts.
  • topk_group (int, optional, defaults to 4) — Number of selected groups for each token (for each token, ensuring the selected experts is only within topk_group groups).
  • num_experts_per_tok (int, optional, defaults to 8) — Number of experts to route each token to. This is the top-k value for the token-choice routing.
  • norm_topk_prob (bool, optional, defaults to True) — Whether to normalize the weights of the routed experts.
  • hidden_act (str, optional, defaults to silu) — The non-linear activation function (function or string) in the decoder. For example, "gelu", "relu", "silu", etc.
  • max_position_embeddings (int, optional, defaults to 163840) — The maximum sequence length that this model might ever be used with.
  • initializer_range (float, optional, defaults to 0.02) — The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
  • rms_norm_eps (float, optional, defaults to 1e-06) — The epsilon used by the rms normalization layers.
  • use_cache (bool, 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 or when the model is a decoder-only generative model.
  • pad_token_id (int, optional) — Token id used for padding in the vocabulary.
  • bos_token_id (int, optional, defaults to 0) — Token id used for beginning-of-stream in the vocabulary.
  • eos_token_id (Union[int, list[int]], optional, defaults to 1) — Token id used for end-of-stream in the vocabulary.
  • tie_word_embeddings (bool, optional, defaults to False) — Whether to tie weight embeddings according to model’s tied_weights_keys mapping.
  • rope_parameters (dict, optional) — Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain a value for rope_theta and optionally parameters used for scaling in case you want to use RoPE with longer max_position_embeddings.
  • mlp_layer_types (list, optional) — MLP type pattern for each layer ("dense" or "sparse"). Defaults to 3 dense + rest sparse.
  • attention_bias (bool, optional, defaults to False) — Whether to use a bias in the query, key, value and output projection layers during self-attention.
  • attention_dropout (Union[float, int], optional, defaults to 0.0) — The dropout ratio for the attention probabilities.
  • index_topk (int, optional, defaults to 2048) — Number of top tokens selected by the indexer for sparse attention.
  • index_head_dim (int, optional, defaults to 128) — Head dimension for the indexer projections (DSA).
  • index_n_heads (int, optional, defaults to 64) — Number of heads for the indexer projections (DSA).
  • mlp_bias (bool, optional, defaults to False) — Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
  • num_experts (int, optional, defaults to 256) — Number of routed experts in MoE layers.
  • head_dim (int, optional, defaults to 64) — The attention head dimension. If None, it will default to hidden_size // num_attention_heads
  • first_k_dense_replace (int, optional, defaults to 3) — Number of leading layers that use a dense MLP; the rest use the MoE block.
  • layer_types (list[str], optional) — A list that explicitly maps each layer index with its layer type. If not provided, it will be automatically generated based on config values.

This is the configuration class to store the configuration of a DeepseekV32Model. It is used to instantiate a Deepseek V32 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 deepseek-ai/DeepSeek-V3.2-Exp

Configuration objects inherit from PreTrainedConfig and can be used to control the model outputs. Read the documentation from PreTrainedConfig for more information.

>>> from transformers import DeepseekV32Config, DeepseekV32Model

>>> # Initializing a DeepSeek-V3.2 configuration
>>> configuration = DeepseekV32Config()

>>> # Initializing a model from the configuration
>>> model = DeepseekV32Model(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config

DeepseekV32PreTrainedModel

class transformers.DeepseekV32PreTrainedModel

< >

( config: PreTrainedConfig *inputs **kwargs )

Parameters

  • config (PreTrainedConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

_forward_unimplemented

< >

( *input: typing.Any )

Define the computation performed at every call.

Should be overridden by all subclasses.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

DeepseekV32Model

class transformers.DeepseekV32Model

< >

( config: DeepseekV32Config )

Parameters

  • config (DeepseekV32Config) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The bare Deepseek V32 Model outputting raw hidden-states without any specific head on top.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

< >

( input_ids: torch.LongTensor | None = None attention_mask: torch.Tensor | None = None position_ids: torch.LongTensor | None = None past_key_values: transformers.cache_utils.Cache | None = None inputs_embeds: torch.FloatTensor | None = None use_cache: bool | None = None **kwargs: typing_extensions.Unpack[transformers.utils.generic.TransformersKwargs] ) → BaseModelOutputWithPast or tuple(torch.FloatTensor)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • attention_mask (torch.Tensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

    • 1 for tokens that are not masked,
    • 0 for tokens that are masked.

    What are attention masks?

  • position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.n_positions - 1].

    What are position IDs?

  • past_key_values (~cache_utils.Cache, optional) — Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the past_key_values returned by the model at a previous stage of decoding, when use_cache=True or config.use_cache=True.

    Only Cache instance is allowed as input, see our kv cache guide. If no past_key_values are passed, DynamicCache will be initialized by default.

    The model will output the same cache format that is fed as input.

    If past_key_values are used, the user is expected to input only unprocessed input_ids (those that don’t have their past key value states given to this model) of shape (batch_size, unprocessed_length) instead of all input_ids of shape (batch_size, sequence_length).

  • inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing input_ids you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert input_ids indices into associated vectors than the model’s internal embedding lookup matrix.
  • use_cache (bool, optional) — If set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).

Returns

BaseModelOutputWithPast or tuple(torch.FloatTensor)

A BaseModelOutputWithPast or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (DeepseekV32Config) and inputs.

The DeepseekV32Model forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

  • last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the model.

    If past_key_values is used only the last hidden-state of the sequences of shape (batch_size, 1, hidden_size) is output.

  • past_key_values (Cache, optional, returned when use_cache=True is passed or when config.use_cache=True) — It is a Cache instance. For more details, see our kv cache guide.

    Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if config.is_encoder_decoder=True in the cross-attention blocks) that can be used (see past_key_values input) to speed up sequential decoding.

  • hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

    Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.

  • attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

    Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

DeepseekV32ForCausalLM

class transformers.DeepseekV32ForCausalLM

< >

( config )

Parameters

  • config (DeepseekV32ForCausalLM) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The Deepseek V32 Model for causal language modeling.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

< >

( input_ids: torch.LongTensor | None = None attention_mask: torch.Tensor | None = None position_ids: torch.LongTensor | None = None past_key_values: transformers.cache_utils.Cache | None = None inputs_embeds: torch.FloatTensor | None = None labels: torch.LongTensor | None = None use_cache: bool | None = None logits_to_keep: int | torch.Tensor = 0 **kwargs: typing_extensions.Unpack[transformers.utils.generic.TransformersKwargs] ) → CausalLMOutputWithPast or tuple(torch.FloatTensor)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • attention_mask (torch.Tensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

    • 1 for tokens that are not masked,
    • 0 for tokens that are masked.

    What are attention masks?

  • position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.n_positions - 1].

    What are position IDs?

  • past_key_values (~cache_utils.Cache, optional) — Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the past_key_values returned by the model at a previous stage of decoding, when use_cache=True or config.use_cache=True.

    Only Cache instance is allowed as input, see our kv cache guide. If no past_key_values are passed, DynamicCache will be initialized by default.

    The model will output the same cache format that is fed as input.

    If past_key_values are used, the user is expected to input only unprocessed input_ids (those that don’t have their past key value states given to this model) of shape (batch_size, unprocessed_length) instead of all input_ids of shape (batch_size, sequence_length).

  • inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing input_ids you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert input_ids indices into associated vectors than the model’s internal embedding lookup matrix.
  • labels (torch.LongTensor of shape (batch_size, sequence_length), optional) — Labels for computing the masked language modeling loss. Indices should either be in [0, ..., config.vocab_size] or -100 (see input_ids docstring). Tokens with indices set to -100 are ignored (masked), the loss is only computed for the tokens with labels in [0, ..., config.vocab_size].
  • use_cache (bool, optional) — If set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).
  • logits_to_keep (Union[int, torch.Tensor], optional, defaults to 0) — If an int, compute logits for the last logits_to_keep tokens. If 0, calculate logits for all input_ids (special case). Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences or large vocabulary size. If a torch.Tensor, must be 1D corresponding to the indices to keep in the sequence length dimension. This is useful when using packed tensor format (single dimension for batch and sequence length).

Returns

CausalLMOutputWithPast or tuple(torch.FloatTensor)

A CausalLMOutputWithPast or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (DeepseekV32Config) and inputs.

The DeepseekV32ForCausalLM forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

  • loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Language modeling loss (for next-token prediction).

  • logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).

  • past_key_values (Cache, optional, returned when use_cache=True is passed or when config.use_cache=True) — It is a Cache instance. For more details, see our kv cache guide.

    Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see past_key_values input) to speed up sequential decoding.

  • hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

    Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.

  • attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

    Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Example:

>>> from transformers import AutoTokenizer, DeepseekV32ForCausalLM

>>> model = DeepseekV32ForCausalLM.from_pretrained("meta-deepseek_v32/DeepseekV32-2-7b-hf")
>>> tokenizer = AutoTokenizer.from_pretrained("meta-deepseek_v32/DeepseekV32-2-7b-hf")

>>> prompt = "Hey, are you conscious? Can you talk to me?"
>>> inputs = tokenizer(prompt, return_tensors="pt")

>>> # Generate
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
Update on GitHub

←DeepSeek-V3 DeepSeek-V4→