Qwen2MoE
Overview
Qwen2MoE is the new model series of large language models from the Qwen team. Previously, we released the Qwen series, including Qwen-72B, Qwen-1.8B, Qwen-VL, Qwen-Audio, etc.
Model Details
Qwen2MoE is a language model series including decoder language models of different model sizes. For each size, we release the base language model and the aligned chat model. Qwen2MoE has the following architectural choices:
- Qwen2MoE is based on the Transformer architecture with SwiGLU activation, attention QKV bias, group query attention, mixture of sliding window attention and full attention, etc. Additionally, we have an improved tokenizer adaptive to multiple natural languages and codes.
- Qwen2MoE employs Mixture of Experts (MoE) architecture, where the models are upcycled from dense language models. For instance,
Qwen1.5-MoE-A2.7Bis upcycled fromQwen-1.8B. It has 14.3B parameters in total and 2.7B activated parameters during runtime, while it achieves comparable performance withQwen1.5-7B, with only 25% of the training resources.
For more details refer to the release blog post.
Usage tips
Qwen1.5-MoE-A2.7B and Qwen1.5-MoE-A2.7B-Chat can be found on the Huggingface Hub
In the following, we demonstrate how to use Qwen1.5-MoE-A2.7B-Chat for the inference. Note that we have used the ChatML format for dialog, in this demo we show how to leverage apply_chat_template for this purpose.
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> device = "cuda" # the device to load the model onto
>>> model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-MoE-A2.7B-Chat", device_map="auto")
>>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-MoE-A2.7B-Chat")
>>> prompt = "Give me a short introduction to large language model."
>>> messages = [{"role": "user", "content": prompt}]
>>> text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
>>> model_inputs = tokenizer([text], return_tensors="pt").to(device)
>>> generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512, do_sample=True)
>>> generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
>>> response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]Qwen2MoeConfig
class transformers.Qwen2MoeConfig
< source >( vocab_size = 151936 hidden_size = 2048 intermediate_size = 5632 num_hidden_layers = 24 num_attention_heads = 16 num_key_value_heads = 16 hidden_act = 'silu' max_position_embeddings = 32768 initializer_range = 0.02 rms_norm_eps = 1e-06 use_cache = True tie_word_embeddings = False rope_theta = 10000.0 use_sliding_window = False sliding_window = 4096 max_window_layers = 28 attention_dropout = 0.0 decoder_sparse_step = 1 moe_intermediate_size = 1408 shared_expert_intermediate_size = 5632 num_experts_per_tok = 4 num_experts = 60 norm_topk_prob = False output_router_logits = False router_aux_loss_coef = 0.001 **kwargs )
Parameters
- vocab_size (
int, optional, defaults to 151936) — Vocabulary size of the Qwen2MoE model. Defines the number of different tokens that can be represented by theinputs_idspassed when calling Qwen2MoeModel - hidden_size (
int, optional, defaults to 2048) — Dimension of the hidden representations. - intermediate_size (
int, optional, defaults to 5632) — Dimension of the MLP representations. - num_hidden_layers (
int, optional, defaults to 24) — Number of hidden layers in the Transformer encoder. - num_attention_heads (
int, optional, defaults to 16) — Number of attention heads for each attention layer in the Transformer encoder. - num_key_value_heads (
int, optional, defaults to 16) — This is the number of key_value heads that should be used to implement Grouped Query Attention. Ifnum_key_value_heads=num_attention_heads, the model will use Multi Head Attention (MHA), ifnum_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 checkout [this paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to32`. - hidden_act (
strorfunction, 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 toTrue) — Whether or not the model should return the last key/values attentions (not used by all models). Only relevant ifconfig.is_decoder=True. - tie_word_embeddings (
bool, optional, defaults toFalse) — 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. - use_sliding_window (
bool, optional, defaults toFalse) — Whether to use sliding window attention. - sliding_window (
int, optional, defaults to 4096) — Sliding window attention (SWA) window size. If not specified, will default to4096. - max_window_layers (
int, optional, defaults to 28) — The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention. - attention_dropout (
float, optional, defaults to 0.0) — The dropout ratio for the attention probabilities. - decoder_sparse_step (
int, optional, defaults to 1) — The frequency of the MoE layer. - moe_intermediate_size (
int, optional, defaults to 1408) — Intermediate size of the routed expert. - shared_expert_intermediate_size (
int, optional, defaults to 5632) — Intermediate size of the shared expert. - num_experts_per_tok (
int, optional, defaults to 4) — Number of selected experts. - num_experts (
int, optional, defaults to 60) — Number of routed experts. - norm_topk_prob (
bool, optional, defaults toFalse) — Whether to normalize the topk probabilities. - output_router_logits (
bool, optional, defaults toFalse) — Whether or not the router logits should be returned by the model. Enabeling this will also allow the model to output the auxiliary loss, including load balancing loss and router z-loss. - router_aux_loss_coef (
float, optional, defaults to 0.001) — The aux loss factor for the total loss.
This is the configuration class to store the configuration of a Qwen2MoeModel. It is used to instantiate a Qwen2MoE model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of Qwen1.5-MoE-A2.7B” Qwen/Qwen1.5-MoE-A2.7B”.
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 Qwen2MoeModel, Qwen2MoeConfig
>>> # Initializing a Qwen2MoE style configuration
>>> configuration = Qwen2MoeConfig()
>>> # Initializing a model from the Qwen1.5-MoE-A2.7B" style configuration
>>> model = Qwen2MoeModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.configQwen2MoeModel
class transformers.Qwen2MoeModel
< source >( config: Qwen2MoeConfig )
Parameters
- config (Qwen2MoeConfig) — 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. config — Qwen2MoeConfig
The bare Qwen2MoE 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.
Transformer decoder consisting of config.num_hidden_layers layers. Each layer is a Qwen2MoeDecoderLayer
forward
< source >( input_ids: LongTensor = None attention_mask: Optional = None position_ids: Optional = None past_key_values: Optional = None inputs_embeds: Optional = None use_cache: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None output_router_logits: Optional = None return_dict: Optional = None )
Parameters
- input_ids (
torch.LongTensorof shape(batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
- attention_mask (
torch.Tensorof 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.
Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If
past_key_valuesis used, optionally only the lastdecoder_input_idshave to be input (seepast_key_values).If you want to change padding behavior, you should read
modeling_opt._prepare_decoder_attention_maskand modify to your needs. See diagram 1 in the paper for more information on the default strategy.- 1 indicates the head is not masked,
- 0 indicates the head is masked.
- position_ids (
torch.LongTensorof 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]. - past_key_values (
Cacheortuple(tuple(torch.FloatTensor)), 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 thepast_key_valuesreturned by the model at a previous stage of decoding, whenuse_cache=Trueorconfig.use_cache=True.Two formats are allowed:
- a Cache instance;
- Tuple of
tuple(torch.FloatTensor)of lengthconfig.n_layers, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)). This is also known as the legacy cache format.
The model will output the same cache format that is fed as input. If no
past_key_valuesare passed, the legacy cache format will be returned.If
past_key_valuesare used, the user can optionally input only the lastinput_ids(those that don’t have their past key value states given to this model) of shape(batch_size, 1)instead of allinput_idsof shape(batch_size, sequence_length). - inputs_embeds (
torch.FloatTensorof shape(batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passinginput_idsyou can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_idsindices into associated vectors than the model’s internal embedding lookup matrix. - use_cache (
bool, optional) — If set toTrue,past_key_valueskey value states are returned and can be used to speed up decoding (seepast_key_values). - output_attentions (
bool, optional) — Whether or not to return the attentions tensors of all attention layers. Seeattentionsunder returned tensors for more detail. - output_hidden_states (
bool, optional) — Whether or not to return the hidden states of all layers. Seehidden_statesunder returned tensors for more detail. - output_router_logits (
bool, optional) — Whether or not to return the logits of all the routers. They are useful for computing the router loss, and should not be returned during inference. - return_dict (
bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.
The Qwen2MoeModel 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.
Qwen2MoeForCausalLM
forward
< source >( input_ids: LongTensor = None attention_mask: Optional = None position_ids: Optional = None past_key_values: Optional = None inputs_embeds: Optional = None labels: Optional = None use_cache: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None output_router_logits: Optional = None return_dict: Optional = None ) → transformers.modeling_outputs.MoeCausalLMOutputWithPast or tuple(torch.FloatTensor)
Parameters
- input_ids (
torch.LongTensorof shape(batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
- attention_mask (
torch.Tensorof 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.
Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If
past_key_valuesis used, optionally only the lastdecoder_input_idshave to be input (seepast_key_values).If you want to change padding behavior, you should read
modeling_opt._prepare_decoder_attention_maskand modify to your needs. See diagram 1 in the paper for more information on the default strategy.- 1 indicates the head is not masked,
- 0 indicates the head is masked.
- position_ids (
torch.LongTensorof 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]. - past_key_values (
Cacheortuple(tuple(torch.FloatTensor)), 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 thepast_key_valuesreturned by the model at a previous stage of decoding, whenuse_cache=Trueorconfig.use_cache=True.Two formats are allowed:
- a Cache instance;
- Tuple of
tuple(torch.FloatTensor)of lengthconfig.n_layers, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)). This is also known as the legacy cache format.
The model will output the same cache format that is fed as input. If no
past_key_valuesare passed, the legacy cache format will be returned.If
past_key_valuesare used, the user can optionally input only the lastinput_ids(those that don’t have their past key value states given to this model) of shape(batch_size, 1)instead of allinput_idsof shape(batch_size, sequence_length). - inputs_embeds (
torch.FloatTensorof shape(batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passinginput_idsyou can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_idsindices into associated vectors than the model’s internal embedding lookup matrix. - use_cache (
bool, optional) — If set toTrue,past_key_valueskey value states are returned and can be used to speed up decoding (seepast_key_values). - output_attentions (
bool, optional) — Whether or not to return the attentions tensors of all attention layers. Seeattentionsunder returned tensors for more detail. - output_hidden_states (
bool, optional) — Whether or not to return the hidden states of all layers. Seehidden_statesunder returned tensors for more detail. - output_router_logits (
bool, optional) — Whether or not to return the logits of all the routers. They are useful for computing the router loss, and should not be returned during inference. - return_dict (
bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.Args — labels (
torch.LongTensorof 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 (seeinput_idsdocstring). Tokens with indices set to-100are ignored (masked), the loss is only computed for the tokens with labels in[0, ..., config.vocab_size].
Returns
transformers.modeling_outputs.MoeCausalLMOutputWithPast or tuple(torch.FloatTensor)
A transformers.modeling_outputs.MoeCausalLMOutputWithPast 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 (Qwen2MoeConfig) and inputs.
-
loss (
torch.FloatTensorof shape(1,), optional, returned whenlabelsis provided) — Language modeling loss (for next-token prediction). -
logits (
torch.FloatTensorof shape(batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). -
aux_loss (
torch.FloatTensor, optional, returned whenlabelsis provided) — aux_loss for the sparse modules. -
router_logits (
tuple(torch.FloatTensor), optional, returned whenoutput_router_probs=Trueandconfig.add_router_probs=Trueis passed or whenconfig.output_router_probs=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, sequence_length, num_experts).Raw router logtis (post-softmax) that are computed by MoE routers, these terms are used to compute the auxiliary loss for Mixture of Experts models.
-
past_key_values (
tuple(tuple(torch.FloatTensor)), optional, returned whenuse_cache=Trueis passed or whenconfig.use_cache=True) — Tuple oftuple(torch.FloatTensor)of lengthconfig.n_layers, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head))Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
past_key_valuesinput) to speed up sequential decoding. -
hidden_states (
tuple(torch.FloatTensor), optional, returned whenoutput_hidden_states=Trueis passed or whenconfig.output_hidden_states=True) — Tuple oftorch.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 whenoutput_attentions=Trueis passed or whenconfig.output_attentions=True) — Tuple oftorch.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.
The Qwen2MoeForCausalLM 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.
Example:
>>> from transformers import AutoTokenizer, Qwen2MoeForCausalLM
>>> model = Qwen2MoeForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
>>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
>>> 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."Qwen2MoeForSequenceClassification
class transformers.Qwen2MoeForSequenceClassification
< source >( config )
Parameters
- config (Qwen2MoeConfig) — 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 Qwen2MoE Model transformer with a sequence classification head on top (linear layer).
Qwen2MoeForSequenceClassification uses the last token in order to do the classification, as other causal models (e.g. GPT-2) do.
Since it does classification on the last token, it requires to know the position of the last token. If a
pad_token_id is defined in the configuration, it finds the last token that is not a padding token in each row. If
no pad_token_id is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
padding tokens when inputs_embeds are passed instead of input_ids, it does the same (take the last value in
each row of the batch).
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
< source >( input_ids: LongTensor = None attention_mask: Optional = None position_ids: Optional = None past_key_values: Optional = None inputs_embeds: Optional = None labels: Optional = None use_cache: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None return_dict: Optional = None )
Parameters
- input_ids (
torch.LongTensorof shape(batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
- attention_mask (
torch.Tensorof 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.
Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If
past_key_valuesis used, optionally only the lastdecoder_input_idshave to be input (seepast_key_values).If you want to change padding behavior, you should read
modeling_opt._prepare_decoder_attention_maskand modify to your needs. See diagram 1 in the paper for more information on the default strategy.- 1 indicates the head is not masked,
- 0 indicates the head is masked.
- position_ids (
torch.LongTensorof 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]. - past_key_values (
Cacheortuple(tuple(torch.FloatTensor)), 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 thepast_key_valuesreturned by the model at a previous stage of decoding, whenuse_cache=Trueorconfig.use_cache=True.Two formats are allowed:
- a Cache instance;
- Tuple of
tuple(torch.FloatTensor)of lengthconfig.n_layers, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)). This is also known as the legacy cache format.
The model will output the same cache format that is fed as input. If no
past_key_valuesare passed, the legacy cache format will be returned.If
past_key_valuesare used, the user can optionally input only the lastinput_ids(those that don’t have their past key value states given to this model) of shape(batch_size, 1)instead of allinput_idsof shape(batch_size, sequence_length). - inputs_embeds (
torch.FloatTensorof shape(batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passinginput_idsyou can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_idsindices into associated vectors than the model’s internal embedding lookup matrix. - use_cache (
bool, optional) — If set toTrue,past_key_valueskey value states are returned and can be used to speed up decoding (seepast_key_values). - output_attentions (
bool, optional) — Whether or not to return the attentions tensors of all attention layers. Seeattentionsunder returned tensors for more detail. - output_hidden_states (
bool, optional) — Whether or not to return the hidden states of all layers. Seehidden_statesunder returned tensors for more detail. - output_router_logits (
bool, optional) — Whether or not to return the logits of all the routers. They are useful for computing the router loss, and should not be returned during inference. - return_dict (
bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple. - labels (
torch.LongTensorof shape(batch_size,), optional) — Labels for computing the sequence classification/regression loss. Indices should be in[0, ..., config.num_labels - 1]. Ifconfig.num_labels == 1a regression loss is computed (Mean-Square loss), Ifconfig.num_labels > 1a classification loss is computed (Cross-Entropy).
The Qwen2MoeForSequenceClassification 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.