# Copyright 2024 The Qwen team, Alibaba Group 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. """Qwen3 model configuration""" from transformers.configuration_utils import PretrainedConfig, layer_type_validation from transformers.modeling_rope_utils import rope_config_validation from transformers.utils import logging logger = logging.get_logger(__name__) class ActionModelConfig(PretrainedConfig): r""" This is the configuration class to store the configuration of a [`Qwen3Model`]. It is used to instantiate a Qwen3 model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of Qwen3-8B [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B). 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 (`int`, *optional*, defaults to 151936): Vocabulary size of the Qwen3 model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`Qwen3Model`] hidden_size (`int`, *optional*, defaults to 4096): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 22016): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 32): Number of hidden layers in the Transformer encoder. num_attention_heads (`int`, *optional*, defaults to 32): Number of attention heads for each attention layer in the Transformer encoder. num_key_value_heads (`int`, *optional*, defaults to 32): 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](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`. head_dim (`int`, *optional*, defaults to 128): The attention head dimension. hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): The non-linear activation function (function or string) in the decoder. max_position_embeddings (`int`, *optional*, defaults to 32768): 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`. tie_word_embeddings (`bool`, *optional*, defaults to `False`): Whether the model's input and output word embeddings should be tied. rope_theta (`float`, *optional*, defaults to 10000.0): The base period of the RoPE embeddings. rope_scaling (`Dict`, *optional*): Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value accordingly. Expected contents: `rope_type` (`str`): The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope', 'llama3'], with 'default' being the original RoPE implementation. `factor` (`float`, *optional*): Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In most scaling types, a `factor` of x will enable the model to handle sequences of length x * original maximum pre-trained length. `original_max_position_embeddings` (`int`, *optional*): Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during pretraining. `attention_factor` (`float`, *optional*): Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention computation. If unspecified, it defaults to value recommended by the implementation, using the `factor` field to infer the suggested value. `beta_fast` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear ramp function. If unspecified, it defaults to 32. `beta_slow` (`float`, *optional*): Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear ramp function. If unspecified, it defaults to 1. `short_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to short contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `long_factor` (`list[float]`, *optional*): Only used with 'longrope'. The scaling factor to be applied to long contexts (< `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2 `low_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE `high_freq_factor` (`float`, *optional*): Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during self-attention. use_sliding_window (`bool`, *optional*, defaults to `False`): Whether to use sliding window attention. sliding_window (`int`, *optional*, defaults to 4096): Sliding window attention (SWA) window size. If not specified, will default to `4096`. max_window_layers (`int`, *optional*, defaults to 28): The number of layers using full attention. The first `max_window_layers` layers will use full attention, while any additional layer afterwards will use SWA (Sliding Window Attention). layer_types (`list`, *optional*): Attention pattern for each layer. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. ```python >>> from transformers import Qwen3Model, Qwen3Config >>> # Initializing a Qwen3 style configuration >>> configuration = Qwen3Config() >>> # Initializing a model from the Qwen3-8B style configuration >>> model = Qwen3Model(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "qwen3" keys_to_ignore_at_inference = ["past_key_values"] # Default tensor parallel plan for base model `Qwen3` base_model_tp_plan = { "action_encoder.*.self_attn.q_proj": "colwise", "action_encoder.*.self_attn.k_proj": "colwise", "action_encoder.*.self_attn.v_proj": "colwise", "action_encoder.*.self_attn.o_proj": "rowwise", "action_encoder.*.mlp.gate_proj": "colwise", "action_encoder.*.mlp.up_proj": "colwise", "action_encoder.*.mlp.down_proj": "rowwise", "action_decoder.*.self_attn.q_proj": "colwise", "action_decoder.*.self_attn.k_proj": "colwise", "action_decoder.*.self_attn.v_proj": "colwise", "action_decoder.*.self_attn.o_proj": "rowwise", "action_decoder.*.mlp.gate_proj": "colwise", "action_decoder.*.mlp.up_proj": "colwise", "action_decoder.*.mlp.down_proj": "rowwise", } base_model_pp_plan = { "action_encoder": (["hidden_states", "attention_mask"], ["hidden_states"]), "action_decoder": (["hidden_states", "attention_mask"], ["hidden_states"]), "norm": (["hidden_states"], ["hidden_states"]), } def __init__( self, action_size=64, state_size=96, hidden_size=1024, intermediate_size=3072, dataset_vocab_size=256, # 假设有100个数据集 num_data_tokens=8, mask_ratio=0.25, # Action Mask 比例 # Mask ratio sampling for DAE masking: # - "fixed": use `mask_ratio` for all trajectories # - "uniform_per_traj": sample per trajectory in [mask_ratio_min, mask_ratio_max] mask_ratio_mode="uniform_per_traj", mask_ratio_min=0.25, mask_ratio_max=0.75, # Loss mode: whether to add reconstruction loss for masked-action view (two-view training). # - False: only current action reconstruction (single view). # - True: current action recon + masked action recon (two views, two recon losses). use_masked_action_recon=False, # Optional contrastive loss on action embedding (InfoNCE). When True, adds contrastive # between clean and masked embeddings; typically used together with use_masked_action_recon. use_contrastive_loss=False, # Optional domain-adversarial loss (GRL + MLP): predict dataset_id from embedding, then reverse # gradient so encoder learns domain-invariant features. Use with multi-domain data. use_domain_adversarial=False, domain_adversarial_weight=0.1, domain_adversarial_lambda=1.0, domain_adversarial_mlp_hidden=None, contrastive_temperature=0.07, contrastive_weight=0.1, contrastive_use_proj=False, contrastive_proj_dim=256, contrastive_use_distributed=True, state_drop_prob=0.5, # State Dropout 比例 min_action_len=5, # 最小切片长度 num_encoder_layers=28, num_decoder_layers=28, num_attention_heads=16, num_key_value_heads=8, head_dim=128, hidden_act="silu", max_position_embeddings=2048, max_action_chunk_size=256, initializer_range=0.02, rms_norm_eps=1e-6, use_cache=True, tie_word_embeddings=False, rope_theta=10000.0, rope_scaling=None, attention_bias=False, use_sliding_window=False, sliding_window=4096, max_window_layers=28, # layer_types=None, attention_dropout=0.0, use_vae_reparameterization=False, # ---- Qwen3 pretrained init (optional) ---- # Can be a HuggingFace model id (e.g. "Qwen/Qwen3-0.6B") or a local checkpoint folder. qwen3_pretrained_name_or_path="Qwen/Qwen3-0.6B", # Copy transformer block weights into action_encoder / action_decoder. qwen3_init_action_encoder=True, qwen3_init_action_decoder=True, # Copy final RMSNorm weights if shape matches. qwen3_init_norm=True, # Which source layer index maps to target layer 0. qwen3_encoder_layer_offset=0, qwen3_decoder_layer_offset=0, **kwargs, ): self.action_size = action_size self.state_size = state_size self.max_position_embeddings = max_position_embeddings self.max_action_chunk_size = max_action_chunk_size self.hidden_size = hidden_size self.intermediate_size = intermediate_size self.dataset_vocab_size = dataset_vocab_size self.num_data_tokens = num_data_tokens self.mask_ratio = mask_ratio self.mask_ratio_mode = mask_ratio_mode self.mask_ratio_min = mask_ratio_min self.mask_ratio_max = mask_ratio_max self.use_masked_action_recon = use_masked_action_recon self.use_contrastive_loss = use_contrastive_loss self.use_domain_adversarial = use_domain_adversarial self.domain_adversarial_weight = domain_adversarial_weight self.domain_adversarial_lambda = domain_adversarial_lambda self.domain_adversarial_mlp_hidden = ( domain_adversarial_mlp_hidden if domain_adversarial_mlp_hidden is not None else hidden_size ) self.contrastive_temperature = contrastive_temperature self.contrastive_weight = contrastive_weight self.contrastive_use_proj = contrastive_use_proj self.contrastive_proj_dim = contrastive_proj_dim self.contrastive_use_distributed = contrastive_use_distributed self.state_drop_prob = state_drop_prob self.min_action_len = min_action_len self.num_encoder_layers = num_encoder_layers self.num_decoder_layers = num_decoder_layers self.num_attention_heads = num_attention_heads self.use_sliding_window = use_sliding_window self.sliding_window = sliding_window if self.use_sliding_window else None self.max_window_layers = max_window_layers self.use_vae_reparameterization = use_vae_reparameterization print(f"use_vae_reparameterization? {use_vae_reparameterization}") # Qwen3 pretrained init (optional) self.qwen3_pretrained_name_or_path = qwen3_pretrained_name_or_path self.qwen3_init_action_encoder = qwen3_init_action_encoder self.qwen3_init_action_decoder = qwen3_init_action_decoder self.qwen3_init_norm = qwen3_init_norm self.qwen3_encoder_layer_offset = qwen3_encoder_layer_offset self.qwen3_decoder_layer_offset = qwen3_decoder_layer_offset # 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.head_dim = head_dim self.hidden_act = hidden_act self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.rope_theta = rope_theta self.rope_scaling = rope_scaling self.attention_bias = attention_bias self.attention_dropout = attention_dropout # Validate the correctness of rotary position embeddings parameters # BC: if there is a 'type' field, move it to 'rope_type'. if self.rope_scaling is not None and "type" in self.rope_scaling: self.rope_scaling["rope_type"] = self.rope_scaling["type"] rope_config_validation(self) # self.layer_types = layer_types # if self.layer_types is None: # self.layer_types = [ # "sliding_attention" # if self.sliding_window is not None and i >= self.max_window_layers # else "full_attention" # for i in range(self.num_hidden_layers) # ] # layer_type_validation(self.layer_types, self.num_hidden_layers) super().__init__( tie_word_embeddings=tie_word_embeddings, **kwargs, ) __all__ = ["ActionModelConfig"]