|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
from dataclasses import dataclass |
|
|
from typing import Callable, Optional, Union |
|
|
|
|
|
import torch |
|
|
from torch import nn |
|
|
|
|
|
from ...activations import ACT2FN |
|
|
from ...cache_utils import Cache, DynamicCache |
|
|
from ...generation import GenerationMixin |
|
|
from ...integrations import use_kernel_forward_from_hub |
|
|
from ...masking_utils import create_causal_mask |
|
|
from ...modeling_flash_attention_utils import FlashAttentionKwargs |
|
|
from ...modeling_layers import GradientCheckpointingLayer |
|
|
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, ModelOutput |
|
|
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update |
|
|
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel |
|
|
from ...processing_utils import Unpack |
|
|
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple |
|
|
from ...utils.deprecation import deprecate_kwarg |
|
|
from ...utils.generic import check_model_inputs |
|
|
from ..auto import AutoModel |
|
|
from .configuration_aria import AriaConfig, AriaTextConfig |
|
|
|
|
|
|
|
|
@use_kernel_forward_from_hub("RMSNorm") |
|
|
class AriaTextRMSNorm(nn.Module): |
|
|
def __init__(self, hidden_size, eps=1e-6): |
|
|
""" |
|
|
AriaTextRMSNorm is equivalent to T5LayerNorm |
|
|
""" |
|
|
super().__init__() |
|
|
self.weight = nn.Parameter(torch.ones(hidden_size)) |
|
|
self.variance_epsilon = eps |
|
|
|
|
|
def forward(self, hidden_states): |
|
|
input_dtype = hidden_states.dtype |
|
|
hidden_states = hidden_states.to(torch.float32) |
|
|
variance = hidden_states.pow(2).mean(-1, keepdim=True) |
|
|
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) |
|
|
return self.weight * hidden_states.to(input_dtype) |
|
|
|
|
|
def extra_repr(self): |
|
|
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" |
|
|
|
|
|
|
|
|
class AriaProjectorMLP(nn.Module): |
|
|
""" |
|
|
Feed-Forward Network module for the Aria Projector. |
|
|
|
|
|
Args: |
|
|
in_features (`int`): |
|
|
Input embedding dimension. |
|
|
hidden_features (`int`): |
|
|
Hidden dimension of the feed-forward network. |
|
|
output_dim (`int`): |
|
|
Output dimension. |
|
|
""" |
|
|
|
|
|
def __init__(self, in_features, hidden_features, output_dim): |
|
|
super().__init__() |
|
|
self.linear_in = nn.Linear(in_features, hidden_features, bias=False) |
|
|
self.linear_out = nn.Linear(hidden_features, output_dim, bias=False) |
|
|
self.act = ACT2FN["gelu_new"] |
|
|
|
|
|
def forward(self, hidden_states): |
|
|
hidden_states = self.act(self.linear_in(hidden_states)) |
|
|
hidden_states = self.linear_out(hidden_states) |
|
|
return hidden_states |
|
|
|
|
|
|
|
|
class AriaCrossAttention(nn.Module): |
|
|
""" |
|
|
Aria Cross-Attention module. |
|
|
|
|
|
Args: |
|
|
config (`AriaConfig`): |
|
|
The configuration to use. |
|
|
""" |
|
|
|
|
|
def __init__(self, config: AriaConfig, dropout_rate: float = 0): |
|
|
super().__init__() |
|
|
hidden_size = config.vision_config.hidden_size |
|
|
num_heads = config.vision_config.num_attention_heads |
|
|
self.num_heads = num_heads |
|
|
self.q_proj = nn.Linear(hidden_size, hidden_size, bias=False) |
|
|
self.k_proj = nn.Linear(hidden_size, hidden_size, bias=False) |
|
|
self.v_proj = nn.Linear(hidden_size, hidden_size, bias=False) |
|
|
|
|
|
|
|
|
self.multihead_attn = nn.MultiheadAttention(hidden_size, num_heads, batch_first=True) |
|
|
self.linear = nn.Linear(hidden_size, hidden_size) |
|
|
self.dropout = nn.Dropout(dropout_rate) |
|
|
|
|
|
self.layer_norm = nn.LayerNorm(hidden_size) |
|
|
self.layer_norm_kv = nn.LayerNorm(hidden_size) |
|
|
|
|
|
def forward(self, key_value_states, hidden_states, attn_mask=None): |
|
|
""" |
|
|
Forward pass of the AriaCrossAttention module. |
|
|
|
|
|
Args: |
|
|
key_value_states (`torch.Tensor`): |
|
|
Input tensor for key and value. |
|
|
hidden_states (`torch.Tensor`): |
|
|
Input tensor for query. |
|
|
attn_mask (`torch.Tensor`, *optional*, defaults to None): |
|
|
Attention mask. |
|
|
|
|
|
Returns: |
|
|
torch.Tensor: |
|
|
Output tensor after cross-attention. |
|
|
""" |
|
|
query = self.q_proj(self.layer_norm(hidden_states)) |
|
|
|
|
|
key_value_states = self.layer_norm_kv(key_value_states) |
|
|
key = self.k_proj(key_value_states) |
|
|
value = self.v_proj(key_value_states) |
|
|
|
|
|
attn_output, _ = self.multihead_attn(query, key, value, attn_mask=attn_mask) |
|
|
|
|
|
attn_output = self.dropout(self.linear(attn_output)) |
|
|
|
|
|
return attn_output |
|
|
|
|
|
|
|
|
class AriaProjector(nn.Module): |
|
|
""" |
|
|
Aria Projector module. |
|
|
|
|
|
This module projects vision features into the language model's embedding space, enabling interaction between vision and language components. |
|
|
|
|
|
Args: |
|
|
config (`AriaConfig`): |
|
|
Configuration object for the model. |
|
|
""" |
|
|
|
|
|
def __init__( |
|
|
self, |
|
|
config: AriaConfig, |
|
|
): |
|
|
super().__init__() |
|
|
|
|
|
self.patch_to_query_dict = config.projector_patch_to_query_dict |
|
|
self.in_features = config.vision_config.hidden_size |
|
|
self.num_heads = config.vision_config.num_attention_heads |
|
|
self.kv_dim = config.vision_config.hidden_size |
|
|
self.hidden_features = config.text_config.hidden_size |
|
|
self.output_dim = config.text_config.hidden_size |
|
|
|
|
|
self.query = nn.Parameter(torch.zeros(config.max_value_projector_patch_to_query_dict, self.in_features)) |
|
|
|
|
|
self.cross_attn = AriaCrossAttention(config) |
|
|
|
|
|
self.layer_norm = nn.LayerNorm(self.in_features) |
|
|
self.feed_forward = AriaProjectorMLP(self.in_features, self.hidden_features, self.output_dim) |
|
|
|
|
|
def forward(self, key_value_states: torch.Tensor, attn_mask: Optional[torch.Tensor] = None): |
|
|
""" |
|
|
Forward pass of the Projector module. |
|
|
|
|
|
Args: |
|
|
key_value_states (`torch.Tensor`): |
|
|
Input tensor of shape (batch_size, num_patches, kv_dim). |
|
|
attn_mask (`torch.Tensor`, *optional*, default is None): |
|
|
Attention mask. |
|
|
|
|
|
Returns: |
|
|
`torch.Tensor`: Output tensor of shape (batch_size, query_number, output_dim). |
|
|
""" |
|
|
batch_size, num_patches = key_value_states.shape[0], key_value_states.shape[1] |
|
|
|
|
|
if num_patches not in self.patch_to_query_dict: |
|
|
raise KeyError( |
|
|
f"Number of patches {num_patches} not found in patch_to_query_dict amongst possible values {self.patch_to_query_dict.keys()}." |
|
|
) |
|
|
query_num = self.patch_to_query_dict[num_patches] |
|
|
|
|
|
queries = self.query[:query_num].unsqueeze(0).repeat(batch_size, 1, 1) |
|
|
|
|
|
if attn_mask is not None: |
|
|
attn_mask = attn_mask.repeat_interleave(self.num_heads, 0) |
|
|
attn_mask = attn_mask.unsqueeze(1).expand(-1, queries.size(1), -1) |
|
|
|
|
|
attention_out = self.cross_attn(key_value_states, queries, attn_mask=attn_mask) |
|
|
|
|
|
out = self.feed_forward(self.layer_norm(attention_out)) |
|
|
|
|
|
return out |
|
|
|
|
|
|
|
|
class AriaSharedExpertsMLP(nn.Module): |
|
|
""" |
|
|
Shared Expert MLP for shared experts. |
|
|
|
|
|
Unlike routed experts, shared experts process all tokens without routing. |
|
|
This class reconfigures the intermediate size in comparison to the LlamaMLP. |
|
|
|
|
|
Args: |
|
|
config (`AriaTextConfig`): Configuration object for the Aria language model. |
|
|
""" |
|
|
|
|
|
def __init__(self, config: AriaTextConfig): |
|
|
super().__init__() |
|
|
self.config = config |
|
|
self.hidden_size = config.hidden_size |
|
|
self.intermediate_size = config.intermediate_size * config.moe_num_shared_experts |
|
|
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) |
|
|
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) |
|
|
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias) |
|
|
self.act_fn = ACT2FN[config.hidden_act] |
|
|
|
|
|
def forward(self, x): |
|
|
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) |
|
|
return down_proj |
|
|
|
|
|
|
|
|
def sequential_experts_gemm(token_states, expert_weights, tokens_per_expert): |
|
|
""" |
|
|
Compute the matrix multiplication (GEMM) for each expert sequentially. This approach is computationally inefficient, especially when dealing with a large number of experts. |
|
|
|
|
|
Args: |
|
|
token_states (torch.Tensor): Input tensor of shape (num_tokens, in_features). |
|
|
expert_weights (torch.Tensor): Weight tensor of shape (num_experts, in_features, out_features). |
|
|
tokens_per_expert (torch.Tensor): Number of tokens assigned to each expert. |
|
|
|
|
|
Returns: |
|
|
torch.Tensor: Output tensor of shape (num_tokens, out_features). |
|
|
""" |
|
|
num_tokens = token_states.shape[0] |
|
|
out_features = expert_weights.shape[-1] |
|
|
output = torch.zeros(num_tokens, out_features, dtype=token_states.dtype, device=token_states.device) |
|
|
|
|
|
cumsum_num_tokens = torch.cumsum(tokens_per_expert, dim=0) |
|
|
|
|
|
zero_tensor = torch.zeros(1, dtype=torch.long, device=cumsum_num_tokens.device) |
|
|
cumsum_num_tokens = torch.cat((zero_tensor, cumsum_num_tokens)) |
|
|
|
|
|
for expert_num in range(expert_weights.shape[0]): |
|
|
start = cumsum_num_tokens[expert_num] |
|
|
end = cumsum_num_tokens[expert_num + 1] |
|
|
tokens = token_states[start:end] |
|
|
|
|
|
out = torch.matmul(tokens, expert_weights[expert_num]) |
|
|
output[start:end] = out |
|
|
return output |
|
|
|
|
|
|
|
|
class AriaGroupedExpertsGemm(nn.Module): |
|
|
""" |
|
|
Grouped GEMM (General Matrix Multiplication) module for efficient expert computation. |
|
|
This module utilizes the grouped_gemm library (https://github.com/fanshiqing/grouped_gemm) |
|
|
for optimized performance. If the grouped_gemm library is not installed, it gracefully |
|
|
falls back to a sequential GEMM implementation, which may be slower but ensures |
|
|
functionality. |
|
|
|
|
|
Args: |
|
|
in_features (`int`): |
|
|
Number of input features. |
|
|
out_features (`int`): |
|
|
Number of output features. |
|
|
groups (`int`): |
|
|
Number of expert groups. |
|
|
""" |
|
|
|
|
|
def __init__(self, in_features, out_features, groups): |
|
|
super().__init__() |
|
|
self.in_features = in_features |
|
|
self.out_features = out_features |
|
|
self.groups = groups |
|
|
self.weight = nn.Parameter(torch.empty(groups, in_features, out_features)) |
|
|
|
|
|
def forward(self, input, tokens_per_expert): |
|
|
""" |
|
|
Perform grouped matrix multiplication. |
|
|
|
|
|
Args: |
|
|
input (`torch.Tensor`): |
|
|
Input tensor of shape (num_tokens, in_features). |
|
|
tokens_per_expert (`torch.Tensor`): |
|
|
Number of tokens assigned to each expert. |
|
|
|
|
|
Returns: |
|
|
torch.Tensor: Output tensor of shape (num_tokens, out_features). |
|
|
""" |
|
|
return sequential_experts_gemm( |
|
|
input, |
|
|
self.weight, |
|
|
tokens_per_expert.cpu(), |
|
|
) |
|
|
|
|
|
|
|
|
class AriaGroupedExpertsMLP(nn.Module): |
|
|
""" |
|
|
Grouped MLP module for Mixture of Experts. |
|
|
|
|
|
Args: |
|
|
config (`AriaTextConfig`): |
|
|
Configuration object for the model. |
|
|
""" |
|
|
|
|
|
def __init__(self, config: AriaTextConfig) -> None: |
|
|
super().__init__() |
|
|
self.config = config |
|
|
self.fc1 = AriaGroupedExpertsGemm(config.hidden_size, config.intermediate_size * 2, config.moe_num_experts) |
|
|
self.fc2 = AriaGroupedExpertsGemm(config.intermediate_size, config.hidden_size, config.moe_num_experts) |
|
|
|
|
|
def forward(self, permuted_tokens, tokens_per_expert): |
|
|
""" |
|
|
Forward pass of the Grouped MLP. |
|
|
|
|
|
Args: |
|
|
permuted_tokens (torch.Tensor): Permuted input tokens. |
|
|
tokens_per_expert (torch.Tensor): Number of tokens assigned to each expert. |
|
|
|
|
|
Returns: |
|
|
torch.Tensor: Output tensor after passing through the MLP. |
|
|
""" |
|
|
fc1_output = self.fc1(permuted_tokens, tokens_per_expert) |
|
|
projection, gate = torch.chunk(fc1_output, 2, dim=-1) |
|
|
fc1_output = nn.functional.silu(projection) * gate |
|
|
fc2_output = self.fc2(fc1_output, tokens_per_expert) |
|
|
return fc2_output |
|
|
|
|
|
|
|
|
|
|
|
class AriaTextMoELayer(nn.Module): |
|
|
""" |
|
|
Aria Text Mixture of Experts (MoE) Layer. |
|
|
|
|
|
This layer applies a gating mechanism to route input tokens to different experts. |
|
|
|
|
|
Args: |
|
|
config (`AriaTextConfig`): |
|
|
Configuration object for the text component of the model. |
|
|
""" |
|
|
|
|
|
def __init__(self, config: AriaTextConfig): |
|
|
super().__init__() |
|
|
|
|
|
self.router = nn.Linear(config.hidden_size, config.moe_num_experts, bias=False) |
|
|
self.experts = AriaGroupedExpertsMLP(config) |
|
|
self.shared_experts = AriaSharedExpertsMLP(config) |
|
|
self.config = config |
|
|
|
|
|
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
|
|
""" |
|
|
Forward pass of the MoE Layer. |
|
|
|
|
|
Args: |
|
|
hidden_states (`torch.Tensor`): |
|
|
Input tensor of shape (batch_size, sequence_length, hidden_size). |
|
|
|
|
|
Returns: |
|
|
torch.Tensor: Output tensor after passing through the MoE layer. |
|
|
|
|
|
Process: |
|
|
1. Route tokens to experts using the router. |
|
|
2. Permute tokens based on routing decisions. |
|
|
3. Process tokens through experts. |
|
|
4. Unpermute and combine expert outputs. |
|
|
5. Add shared expert output to the final result. |
|
|
""" |
|
|
original_shape = hidden_states.shape |
|
|
hidden_states = hidden_states.view(-1, hidden_states.size(-1)) |
|
|
|
|
|
|
|
|
logits = self.router(hidden_states) |
|
|
top_logits, top_indices = torch.topk(logits, k=self.config.moe_topk, dim=1) |
|
|
scores = nn.functional.softmax(top_logits, dim=-1) |
|
|
|
|
|
original_dtype = top_indices.dtype |
|
|
|
|
|
tokens_per_expert = torch.histc( |
|
|
top_indices.flatten().to(torch.float32), |
|
|
bins=self.config.moe_num_experts, |
|
|
min=0, |
|
|
max=self.config.moe_num_experts - 1, |
|
|
).to(original_dtype) |
|
|
indices = top_indices |
|
|
|
|
|
|
|
|
flatten_indices = indices.view(-1) |
|
|
sorted_indices = torch.argsort(flatten_indices) |
|
|
permuted_tokens = hidden_states.index_select(0, sorted_indices // self.config.moe_topk) |
|
|
|
|
|
|
|
|
expert_output = self.experts(permuted_tokens, tokens_per_expert) |
|
|
|
|
|
|
|
|
unpermuted_tokens = torch.zeros( |
|
|
(scores.shape[0] * self.config.moe_topk, expert_output.size(1)), |
|
|
dtype=expert_output.dtype, |
|
|
device=expert_output.device, |
|
|
) |
|
|
unpermuted_tokens.index_copy_(0, sorted_indices, expert_output) |
|
|
unpermuted_tokens = unpermuted_tokens.view(-1, self.config.moe_topk, expert_output.size(1)) |
|
|
|
|
|
output = (unpermuted_tokens * scores.unsqueeze(-1)).sum(dim=1).view(original_shape) |
|
|
|
|
|
|
|
|
shared_expert_output = self.shared_experts(hidden_states.view(original_shape)) |
|
|
return output + shared_expert_output |
|
|
|
|
|
|
|
|
def rotate_half(x): |
|
|
"""Rotates half the hidden dims of the input.""" |
|
|
x1 = x[..., : x.shape[-1] // 2] |
|
|
x2 = x[..., x.shape[-1] // 2 :] |
|
|
return torch.cat((-x2, x1), dim=-1) |
|
|
|
|
|
|
|
|
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): |
|
|
"""Applies Rotary Position Embedding to the query and key tensors. |
|
|
|
|
|
Args: |
|
|
q (`torch.Tensor`): The query tensor. |
|
|
k (`torch.Tensor`): The key tensor. |
|
|
cos (`torch.Tensor`): The cosine part of the rotary embedding. |
|
|
sin (`torch.Tensor`): The sine part of the rotary embedding. |
|
|
position_ids (`torch.Tensor`, *optional*): |
|
|
Deprecated and unused. |
|
|
unsqueeze_dim (`int`, *optional*, defaults to 1): |
|
|
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and |
|
|
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note |
|
|
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and |
|
|
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes |
|
|
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have |
|
|
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. |
|
|
Returns: |
|
|
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. |
|
|
""" |
|
|
cos = cos.unsqueeze(unsqueeze_dim) |
|
|
sin = sin.unsqueeze(unsqueeze_dim) |
|
|
q_embed = (q * cos) + (rotate_half(q) * sin) |
|
|
k_embed = (k * cos) + (rotate_half(k) * sin) |
|
|
return q_embed, k_embed |
|
|
|
|
|
|
|
|
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: |
|
|
""" |
|
|
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, |
|
|
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) |
|
|
""" |
|
|
batch, num_key_value_heads, slen, head_dim = hidden_states.shape |
|
|
if n_rep == 1: |
|
|
return hidden_states |
|
|
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) |
|
|
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) |
|
|
|
|
|
|
|
|
def eager_attention_forward( |
|
|
module: nn.Module, |
|
|
query: torch.Tensor, |
|
|
key: torch.Tensor, |
|
|
value: torch.Tensor, |
|
|
attention_mask: Optional[torch.Tensor], |
|
|
scaling: float, |
|
|
dropout: float = 0.0, |
|
|
**kwargs: Unpack[TransformersKwargs], |
|
|
): |
|
|
key_states = repeat_kv(key, module.num_key_value_groups) |
|
|
value_states = repeat_kv(value, module.num_key_value_groups) |
|
|
|
|
|
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling |
|
|
if attention_mask is not None: |
|
|
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] |
|
|
attn_weights = attn_weights + causal_mask |
|
|
|
|
|
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) |
|
|
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) |
|
|
attn_output = torch.matmul(attn_weights, value_states) |
|
|
attn_output = attn_output.transpose(1, 2).contiguous() |
|
|
|
|
|
return attn_output, attn_weights |
|
|
|
|
|
|
|
|
class AriaTextAttention(nn.Module): |
|
|
"""Multi-headed attention from 'Attention Is All You Need' paper""" |
|
|
|
|
|
def __init__(self, config: AriaTextConfig, layer_idx: int): |
|
|
super().__init__() |
|
|
self.config = config |
|
|
self.layer_idx = layer_idx |
|
|
self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads) |
|
|
self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads |
|
|
self.scaling = self.head_dim**-0.5 |
|
|
self.attention_dropout = config.attention_dropout |
|
|
self.is_causal = True |
|
|
|
|
|
self.q_proj = nn.Linear( |
|
|
config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias |
|
|
) |
|
|
self.k_proj = nn.Linear( |
|
|
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias |
|
|
) |
|
|
self.v_proj = nn.Linear( |
|
|
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias |
|
|
) |
|
|
self.o_proj = nn.Linear( |
|
|
config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias |
|
|
) |
|
|
|
|
|
@deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58") |
|
|
def forward( |
|
|
self, |
|
|
hidden_states: torch.Tensor, |
|
|
position_embeddings: tuple[torch.Tensor, torch.Tensor], |
|
|
attention_mask: Optional[torch.Tensor], |
|
|
past_key_values: Optional[Cache] = None, |
|
|
cache_position: Optional[torch.LongTensor] = None, |
|
|
**kwargs: Unpack[TransformersKwargs], |
|
|
) -> tuple[torch.Tensor, torch.Tensor]: |
|
|
input_shape = hidden_states.shape[:-1] |
|
|
hidden_shape = (*input_shape, -1, self.head_dim) |
|
|
|
|
|
query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2) |
|
|
key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2) |
|
|
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2) |
|
|
|
|
|
cos, sin = position_embeddings |
|
|
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) |
|
|
|
|
|
if past_key_values is not None: |
|
|
|
|
|
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} |
|
|
key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs) |
|
|
|
|
|
attention_interface: Callable = eager_attention_forward |
|
|
if self.config._attn_implementation != "eager": |
|
|
attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] |
|
|
|
|
|
attn_output, attn_weights = attention_interface( |
|
|
self, |
|
|
query_states, |
|
|
key_states, |
|
|
value_states, |
|
|
attention_mask, |
|
|
dropout=0.0 if not self.training else self.attention_dropout, |
|
|
scaling=self.scaling, |
|
|
**kwargs, |
|
|
) |
|
|
|
|
|
attn_output = attn_output.reshape(*input_shape, -1).contiguous() |
|
|
attn_output = self.o_proj(attn_output) |
|
|
return attn_output, attn_weights |
|
|
|
|
|
|
|
|
class AriaTextDecoderLayer(GradientCheckpointingLayer): |
|
|
""" |
|
|
Aria Text Decoder Layer. |
|
|
|
|
|
This class defines a single decoder layer in the language model, incorporating self-attention and Mixture of Experts (MoE) feed-forward network. |
|
|
|
|
|
Args: |
|
|
config (`AriaTextConfig`): |
|
|
Configuration object for the text component of the model. |
|
|
layer_idx (`int`): |
|
|
Index of the layer. |
|
|
""" |
|
|
|
|
|
def __init__(self, config: AriaTextConfig, layer_idx: int): |
|
|
super().__init__() |
|
|
self.hidden_size = config.hidden_size |
|
|
|
|
|
self.self_attn = AriaTextAttention(config=config, layer_idx=layer_idx) |
|
|
self.mlp = AriaTextMoELayer(config) |
|
|
self.input_layernorm = AriaTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
|
|
self.post_attention_layernorm = AriaTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
|
|
|
|
|
@deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58") |
|
|
def forward( |
|
|
self, |
|
|
hidden_states: torch.Tensor, |
|
|
attention_mask: Optional[torch.Tensor] = None, |
|
|
position_ids: Optional[torch.LongTensor] = None, |
|
|
past_key_values: Optional[Cache] = None, |
|
|
use_cache: Optional[bool] = False, |
|
|
cache_position: Optional[torch.LongTensor] = None, |
|
|
position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None, |
|
|
**kwargs: Unpack[TransformersKwargs], |
|
|
) -> torch.Tensor: |
|
|
residual = hidden_states |
|
|
hidden_states = self.input_layernorm(hidden_states) |
|
|
|
|
|
hidden_states, _ = self.self_attn( |
|
|
hidden_states=hidden_states, |
|
|
attention_mask=attention_mask, |
|
|
position_ids=position_ids, |
|
|
past_key_values=past_key_values, |
|
|
use_cache=use_cache, |
|
|
cache_position=cache_position, |
|
|
position_embeddings=position_embeddings, |
|
|
**kwargs, |
|
|
) |
|
|
hidden_states = residual + hidden_states |
|
|
|
|
|
|
|
|
residual = hidden_states |
|
|
hidden_states = self.post_attention_layernorm(hidden_states) |
|
|
hidden_states = self.mlp(hidden_states) |
|
|
hidden_states = residual + hidden_states |
|
|
return hidden_states |
|
|
|
|
|
|
|
|
@auto_docstring |
|
|
class AriaTextPreTrainedModel(PreTrainedModel): |
|
|
config: AriaTextConfig |
|
|
base_model_prefix = "model" |
|
|
_no_split_modules = ["AriaTextDecoderLayer", "AriaGroupedExpertsGemm"] |
|
|
supports_gradient_checkpointing = True |
|
|
_skip_keys_device_placement = "past_key_values" |
|
|
_supports_flash_attn = True |
|
|
_supports_sdpa = True |
|
|
|
|
|
_supports_attention_backend = True |
|
|
_can_record_outputs = { |
|
|
"hidden_states": AriaTextDecoderLayer, |
|
|
"attentions": AriaTextAttention, |
|
|
} |
|
|
|
|
|
def _init_weights(self, module): |
|
|
super()._init_weights(module) |
|
|
if isinstance(module, AriaGroupedExpertsGemm): |
|
|
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) |
|
|
|
|
|
|
|
|
@auto_docstring |
|
|
class AriaPreTrainedModel(PreTrainedModel): |
|
|
config: AriaConfig |
|
|
base_model_prefix = "" |
|
|
supports_gradient_checkpointing = True |
|
|
_no_split_modules = ["AriaDecoderLayer"] |
|
|
_skip_keys_device_placement = ["past_key_values"] |
|
|
_supports_flash_attn = True |
|
|
_supports_sdpa = True |
|
|
_supports_flex_attn = True |
|
|
_can_compile_fullgraph = False |
|
|
_supports_attention_backend = True |
|
|
_can_record_outputs = { |
|
|
"hidden_states": AriaTextDecoderLayer, |
|
|
"attentions": AriaTextAttention, |
|
|
} |
|
|
|
|
|
def _init_weights(self, module): |
|
|
super()._init_weights(module) |
|
|
if isinstance(module, AriaProjector): |
|
|
nn.init.trunc_normal_(module.query, std=self.config.initializer_range) |
|
|
|
|
|
|
|
|
class AriaTextRotaryEmbedding(nn.Module): |
|
|
inv_freq: torch.Tensor |
|
|
|
|
|
def __init__(self, config: AriaTextConfig, device=None): |
|
|
super().__init__() |
|
|
|
|
|
if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict): |
|
|
self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) |
|
|
else: |
|
|
self.rope_type = "default" |
|
|
self.max_seq_len_cached = config.max_position_embeddings |
|
|
self.original_max_seq_len = config.max_position_embeddings |
|
|
|
|
|
self.config = config |
|
|
self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] |
|
|
|
|
|
inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device) |
|
|
self.register_buffer("inv_freq", inv_freq, persistent=False) |
|
|
self.original_inv_freq = self.inv_freq |
|
|
|
|
|
@torch.no_grad() |
|
|
@dynamic_rope_update |
|
|
def forward(self, x, position_ids): |
|
|
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) |
|
|
position_ids_expanded = position_ids[:, None, :].float() |
|
|
|
|
|
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu" |
|
|
with torch.autocast(device_type=device_type, enabled=False): |
|
|
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) |
|
|
emb = torch.cat((freqs, freqs), dim=-1) |
|
|
cos = emb.cos() * self.attention_scaling |
|
|
sin = emb.sin() * self.attention_scaling |
|
|
|
|
|
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) |
|
|
|
|
|
|
|
|
@auto_docstring |
|
|
class AriaTextModel(AriaTextPreTrainedModel): |
|
|
def __init__(self, config: AriaTextConfig): |
|
|
super().__init__(config) |
|
|
self.padding_idx = config.pad_token_id |
|
|
self.vocab_size = config.vocab_size |
|
|
|
|
|
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) |
|
|
self.layers = nn.ModuleList( |
|
|
[AriaTextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] |
|
|
) |
|
|
self.norm = AriaTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
|
|
self.rotary_emb = AriaTextRotaryEmbedding(config=config) |
|
|
self.gradient_checkpointing = False |
|
|
|
|
|
|
|
|
self.post_init() |
|
|
|
|
|
@check_model_inputs() |
|
|
@auto_docstring |
|
|
def forward( |
|
|
self, |
|
|
input_ids: Optional[torch.LongTensor] = None, |
|
|
attention_mask: Optional[torch.Tensor] = None, |
|
|
position_ids: Optional[torch.LongTensor] = None, |
|
|
past_key_values: Optional[Cache] = None, |
|
|
inputs_embeds: Optional[torch.FloatTensor] = None, |
|
|
cache_position: Optional[torch.LongTensor] = None, |
|
|
use_cache: Optional[bool] = None, |
|
|
**kwargs: Unpack[TransformersKwargs], |
|
|
) -> BaseModelOutputWithPast: |
|
|
if (input_ids is None) ^ (inputs_embeds is not None): |
|
|
raise ValueError("You must specify exactly one of input_ids or inputs_embeds") |
|
|
|
|
|
if inputs_embeds is None: |
|
|
inputs_embeds: torch.Tensor = self.embed_tokens(input_ids) |
|
|
|
|
|
if use_cache and past_key_values is None: |
|
|
past_key_values = DynamicCache(config=self.config) |
|
|
|
|
|
if cache_position is None: |
|
|
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 |
|
|
cache_position: torch.Tensor = torch.arange( |
|
|
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device |
|
|
) |
|
|
|
|
|
if position_ids is None: |
|
|
position_ids = cache_position.unsqueeze(0) |
|
|
|
|
|
causal_mask = create_causal_mask( |
|
|
config=self.config, |
|
|
input_embeds=inputs_embeds, |
|
|
attention_mask=attention_mask, |
|
|
cache_position=cache_position, |
|
|
past_key_values=past_key_values, |
|
|
position_ids=position_ids, |
|
|
) |
|
|
|
|
|
hidden_states = inputs_embeds |
|
|
position_embeddings = self.rotary_emb(hidden_states, position_ids) |
|
|
|
|
|
for decoder_layer in self.layers[: self.config.num_hidden_layers]: |
|
|
hidden_states = decoder_layer( |
|
|
hidden_states, |
|
|
attention_mask=causal_mask, |
|
|
position_ids=position_ids, |
|
|
past_key_values=past_key_values, |
|
|
cache_position=cache_position, |
|
|
position_embeddings=position_embeddings, |
|
|
**kwargs, |
|
|
) |
|
|
|
|
|
hidden_states = self.norm(hidden_states) |
|
|
return BaseModelOutputWithPast( |
|
|
last_hidden_state=hidden_states, |
|
|
past_key_values=past_key_values, |
|
|
) |
|
|
|
|
|
|
|
|
@auto_docstring |
|
|
class AriaTextForCausalLM(AriaTextPreTrainedModel, GenerationMixin): |
|
|
_tied_weights_keys = ["lm_head.weight"] |
|
|
_tp_plan = {"lm_head": "colwise_rep"} |
|
|
_pp_plan = {"lm_head": (["hidden_states"], ["logits"])} |
|
|
|
|
|
def __init__(self, config: AriaTextConfig): |
|
|
super().__init__(config) |
|
|
self.model = AriaTextModel(config) |
|
|
self.vocab_size = config.vocab_size |
|
|
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) |
|
|
|
|
|
|
|
|
self.post_init() |
|
|
|
|
|
@auto_docstring |
|
|
def forward( |
|
|
self, |
|
|
input_ids: Optional[torch.LongTensor] = None, |
|
|
attention_mask: Optional[torch.Tensor] = None, |
|
|
position_ids: Optional[torch.LongTensor] = None, |
|
|
past_key_values: Optional[Cache] = None, |
|
|
inputs_embeds: Optional[torch.FloatTensor] = None, |
|
|
labels: Optional[torch.LongTensor] = None, |
|
|
use_cache: Optional[bool] = None, |
|
|
cache_position: Optional[torch.LongTensor] = None, |
|
|
logits_to_keep: Union[int, torch.Tensor] = 0, |
|
|
**kwargs: Unpack[TransformersKwargs], |
|
|
) -> CausalLMOutputWithPast: |
|
|
r""" |
|
|
Example: |
|
|
|
|
|
```python |
|
|
>>> from transformers import AutoTokenizer, AriaTextForCausalLM |
|
|
|
|
|
>>> model = AriaTextForCausalLM.from_pretrained("meta-aria_text/AriaText-2-7b-hf") |
|
|
>>> tokenizer = AutoTokenizer.from_pretrained("meta-aria_text/AriaText-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." |
|
|
```""" |
|
|
outputs: BaseModelOutputWithPast = self.model( |
|
|
input_ids=input_ids, |
|
|
attention_mask=attention_mask, |
|
|
position_ids=position_ids, |
|
|
past_key_values=past_key_values, |
|
|
inputs_embeds=inputs_embeds, |
|
|
use_cache=use_cache, |
|
|
cache_position=cache_position, |
|
|
**kwargs, |
|
|
) |
|
|
|
|
|
hidden_states = outputs.last_hidden_state |
|
|
|
|
|
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep |
|
|
logits = self.lm_head(hidden_states[:, slice_indices, :]) |
|
|
|
|
|
loss = None |
|
|
if labels is not None: |
|
|
loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs) |
|
|
|
|
|
return CausalLMOutputWithPast( |
|
|
loss=loss, |
|
|
logits=logits, |
|
|
past_key_values=outputs.past_key_values, |
|
|
hidden_states=outputs.hidden_states, |
|
|
attentions=outputs.attentions, |
|
|
) |
|
|
|
|
|
|
|
|
@dataclass |
|
|
@auto_docstring( |
|
|
custom_intro=""" |
|
|
Base class for Aria causal language model (or autoregressive) outputs. |
|
|
""" |
|
|
) |
|
|
class AriaCausalLMOutputWithPast(ModelOutput): |
|
|
r""" |
|
|
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_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). |
|
|
|
|
|
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. |
|
|
image_hidden_states (`torch.FloatTensor`, *optional*): |
|
|
A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`. |
|
|
image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state. |
|
|
""" |
|
|
|
|
|
loss: Optional[torch.FloatTensor] = None |
|
|
logits: Optional[torch.FloatTensor] = None |
|
|
past_key_values: Optional[Cache] = None |
|
|
hidden_states: Optional[tuple[torch.FloatTensor]] = None |
|
|
attentions: Optional[tuple[torch.FloatTensor]] = None |
|
|
image_hidden_states: Optional[torch.FloatTensor] = None |
|
|
|
|
|
|
|
|
@dataclass |
|
|
@auto_docstring( |
|
|
custom_intro=""" |
|
|
Base class for Aria outputs, with hidden states and attentions. |
|
|
""" |
|
|
) |
|
|
class AriaModelOutputWithPast(BaseModelOutputWithPast): |
|
|
r""" |
|
|
past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): |
|
|
It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). |
|
|
|
|
|
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. |
|
|
image_hidden_states (`torch.FloatTensor`, *optional*): |
|
|
A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`. |
|
|
image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state. |
|
|
""" |
|
|
|
|
|
image_hidden_states: Optional[torch.FloatTensor] = None |
|
|
|
|
|
|
|
|
@auto_docstring( |
|
|
custom_intro=""" |
|
|
The Aria model which consists of a vision backbone and a language model, without a language modeling head. |
|
|
""" |
|
|
) |
|
|
class AriaModel(AriaPreTrainedModel): |
|
|
_checkpoint_conversion_mapping = {"language_model.model": "language_model"} |
|
|
|
|
|
def __init__(self, config: AriaConfig): |
|
|
super().__init__(config) |
|
|
self.vision_tower = AutoModel.from_config(config.vision_config) |
|
|
self.multi_modal_projector = AriaProjector(config) |
|
|
self.language_model = AutoModel.from_config(config.text_config) |
|
|
self.post_init() |
|
|
|
|
|
def get_input_embeddings(self): |
|
|
return self.language_model.get_input_embeddings() |
|
|
|
|
|
def set_input_embeddings(self, value): |
|
|
self.language_model.set_input_embeddings(value) |
|
|
|
|
|
def set_decoder(self, decoder): |
|
|
self.language_model = decoder |
|
|
|
|
|
def get_decoder(self): |
|
|
return self.language_model |
|
|
|
|
|
def get_image_features( |
|
|
self, |
|
|
pixel_values: torch.FloatTensor, |
|
|
pixel_mask: Optional[torch.FloatTensor] = None, |
|
|
vision_feature_layer: int = -1, |
|
|
): |
|
|
""" |
|
|
Obtains image last hidden states from the vision tower and apply multimodal projection. |
|
|
|
|
|
Args: |
|
|
pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`): |
|
|
The tensors corresponding to the input images. |
|
|
pixel_mask (`torch.FloatTensor]`, *optional*): |
|
|
The tensors corresponding to the input image mask. |
|
|
vision_feature_layer (`Union[int, list[int]]`, *optional*): |
|
|
The index of the layer to select the vision feature. If multiple indices are provided, |
|
|
the vision feature of the corresponding indices will be concatenated to form the |
|
|
vision features. |
|
|
Returns: |
|
|
image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`). |
|
|
""" |
|
|
vision_feature_layer = ( |
|
|
vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer |
|
|
) |
|
|
patch_attention_mask = self._create_patch_attention_mask(pixel_mask) |
|
|
image_outputs = self.vision_tower( |
|
|
pixel_values, patch_attention_mask=patch_attention_mask, output_hidden_states=True |
|
|
) |
|
|
image_attn_mask = None |
|
|
if patch_attention_mask is not None: |
|
|
flattened_mask = patch_attention_mask.flatten(1) |
|
|
image_attn_mask = torch.logical_not(flattened_mask) |
|
|
|
|
|
selected_image_feature = image_outputs.hidden_states[vision_feature_layer] |
|
|
image_features = self.multi_modal_projector(selected_image_feature, attn_mask=image_attn_mask) |
|
|
return image_features |
|
|
|
|
|
def get_placeholder_mask( |
|
|
self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor |
|
|
): |
|
|
""" |
|
|
Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is |
|
|
equal to the length of multimodal features. If the lengths are different, an error is raised. |
|
|
""" |
|
|
if input_ids is None: |
|
|
special_image_mask = inputs_embeds == self.get_input_embeddings()( |
|
|
torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device) |
|
|
) |
|
|
special_image_mask = special_image_mask.all(-1) |
|
|
else: |
|
|
special_image_mask = input_ids == self.config.image_token_id |
|
|
|
|
|
n_image_tokens = special_image_mask.sum() |
|
|
special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device) |
|
|
n_image_features = image_features.shape[0] * image_features.shape[1] |
|
|
if inputs_embeds[special_image_mask].numel() != image_features.numel(): |
|
|
raise ValueError( |
|
|
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}" |
|
|
) |
|
|
return special_image_mask |
|
|
|
|
|
@can_return_tuple |
|
|
@auto_docstring |
|
|
def forward( |
|
|
self, |
|
|
input_ids: Optional[torch.LongTensor] = None, |
|
|
pixel_values: Optional[torch.FloatTensor] = None, |
|
|
pixel_mask: Optional[torch.LongTensor] = None, |
|
|
attention_mask: Optional[torch.Tensor] = None, |
|
|
position_ids: Optional[torch.LongTensor] = None, |
|
|
past_key_values: Optional[Cache] = None, |
|
|
inputs_embeds: Optional[torch.FloatTensor] = None, |
|
|
use_cache: Optional[bool] = None, |
|
|
cache_position: Optional[torch.LongTensor] = None, |
|
|
**kwargs: Unpack[FlashAttentionKwargs], |
|
|
) -> Union[tuple, AriaModelOutputWithPast]: |
|
|
if inputs_embeds is None: |
|
|
inputs_embeds = self.get_input_embeddings()(input_ids) |
|
|
|
|
|
|
|
|
if pixel_values is not None and inputs_embeds.shape[1] != 1: |
|
|
image_features = self.get_image_features( |
|
|
pixel_values=pixel_values, |
|
|
pixel_mask=pixel_mask, |
|
|
vision_feature_layer=self.config.vision_feature_layer, |
|
|
) |
|
|
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype) |
|
|
special_image_mask = self.get_placeholder_mask( |
|
|
input_ids, inputs_embeds=inputs_embeds, image_features=image_features |
|
|
) |
|
|
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features) |
|
|
|
|
|
outputs = self.language_model( |
|
|
attention_mask=attention_mask, |
|
|
position_ids=position_ids, |
|
|
past_key_values=past_key_values, |
|
|
inputs_embeds=inputs_embeds, |
|
|
use_cache=use_cache, |
|
|
cache_position=cache_position, |
|
|
**kwargs, |
|
|
) |
|
|
|
|
|
return AriaModelOutputWithPast( |
|
|
last_hidden_state=outputs.last_hidden_state, |
|
|
past_key_values=outputs.past_key_values if use_cache else None, |
|
|
hidden_states=outputs.hidden_states, |
|
|
attentions=outputs.attentions, |
|
|
image_hidden_states=image_features if pixel_values is not None else None, |
|
|
) |
|
|
|
|
|
def _create_patch_attention_mask(self, pixel_mask): |
|
|
if pixel_mask is None: |
|
|
return None |
|
|
|
|
|
patches_subgrid = pixel_mask.unfold( |
|
|
dimension=1, |
|
|
size=self.vision_tower.config.patch_size, |
|
|
step=self.vision_tower.config.patch_size, |
|
|
) |
|
|
patches_subgrid = patches_subgrid.unfold( |
|
|
dimension=2, |
|
|
size=self.vision_tower.config.patch_size, |
|
|
step=self.vision_tower.config.patch_size, |
|
|
) |
|
|
return (patches_subgrid.sum(dim=(-1, -2)) > 0).bool() |
|
|
|
|
|
|
|
|
@auto_docstring( |
|
|
custom_intro=""" |
|
|
Aria model for conditional generation tasks. |
|
|
|
|
|
This model combines a vision tower, a multi-modal projector, and a language model |
|
|
to perform tasks that involve both image and text inputs. |
|
|
""" |
|
|
) |
|
|
class AriaForConditionalGeneration(AriaPreTrainedModel, GenerationMixin): |
|
|
_checkpoint_conversion_mapping = { |
|
|
"^language_model.model": "model.language_model", |
|
|
"^vision_tower": "model.vision_tower", |
|
|
"^multi_modal_projector": "model.multi_modal_projector", |
|
|
"^language_model.lm_head": "lm_head", |
|
|
} |
|
|
_tied_weights_keys = ["lm_head.weight"] |
|
|
|
|
|
def __init__(self, config: AriaConfig): |
|
|
super().__init__(config) |
|
|
self.model = AriaModel(config) |
|
|
self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False) |
|
|
self.post_init() |
|
|
|
|
|
def get_input_embeddings(self): |
|
|
return self.model.get_input_embeddings() |
|
|
|
|
|
def set_input_embeddings(self, value): |
|
|
self.model.set_input_embeddings(value) |
|
|
|
|
|
def get_output_embeddings(self) -> nn.Module: |
|
|
return self.lm_head |
|
|
|
|
|
def set_decoder(self, decoder): |
|
|
self.model.set_decoder(decoder) |
|
|
|
|
|
def get_decoder(self): |
|
|
return self.model.get_decoder() |
|
|
|
|
|
def get_image_features( |
|
|
self, |
|
|
pixel_values: torch.FloatTensor, |
|
|
pixel_mask: Optional[torch.FloatTensor] = None, |
|
|
vision_feature_layer: int = -1, |
|
|
): |
|
|
return self.model.get_image_features( |
|
|
pixel_values=pixel_values, |
|
|
pixel_mask=pixel_mask, |
|
|
vision_feature_layer=vision_feature_layer, |
|
|
) |
|
|
|
|
|
|
|
|
@property |
|
|
def language_model(self): |
|
|
return self.model.language_model |
|
|
|
|
|
@property |
|
|
def vision_tower(self): |
|
|
return self.model.vision_tower |
|
|
|
|
|
@property |
|
|
def multi_modal_projector(self): |
|
|
return self.model.multi_modal_projector |
|
|
|
|
|
@can_return_tuple |
|
|
@auto_docstring |
|
|
def forward( |
|
|
self, |
|
|
input_ids: Optional[torch.LongTensor] = None, |
|
|
pixel_values: Optional[torch.FloatTensor] = None, |
|
|
pixel_mask: Optional[torch.LongTensor] = None, |
|
|
attention_mask: Optional[torch.Tensor] = None, |
|
|
position_ids: Optional[torch.LongTensor] = None, |
|
|
past_key_values: Optional[Cache] = None, |
|
|
inputs_embeds: Optional[torch.FloatTensor] = None, |
|
|
labels: Optional[torch.LongTensor] = None, |
|
|
use_cache: Optional[bool] = None, |
|
|
logits_to_keep: Union[int, torch.Tensor] = 0, |
|
|
cache_position: Optional[torch.LongTensor] = None, |
|
|
**kwargs: Unpack[TransformersKwargs], |
|
|
) -> Union[tuple, AriaCausalLMOutputWithPast]: |
|
|
r""" |
|
|
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 `model.image_token_id` (where `model` is your instance of `AriaForConditionalGeneration`). |
|
|
Tokens with indices set to `model.image_token_id` are ignored (masked), the loss is only |
|
|
computed for the tokens with labels in `[0, ..., config.vocab_size]`. |
|
|
|
|
|
Example: |
|
|
|
|
|
```python |
|
|
>>> import requests |
|
|
>>> import torch |
|
|
>>> from PIL import Image |
|
|
>>> from io import BytesIO |
|
|
|
|
|
>>> from transformers import AutoProcessor, AutoModel |
|
|
>>> from transformers.image_utils import load_image |
|
|
|
|
|
>>> # Note that passing the image urls (instead of the actual pil images) to the processor is also possible |
|
|
>>> image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg") |
|
|
>>> image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg") |
|
|
>>> image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg") |
|
|
|
|
|
>>> processor = AutoProcessor.from_pretrained("Rhymes-AI/Aria") |
|
|
>>> model = AutoModel.from_pretrained("Rhymes-AI/Aria", dtype=torch.bfloat16, device_map="auto") |
|
|
|
|
|
>>> # Create inputs |
|
|
>>> messages = [ |
|
|
... { |
|
|
... "role": "user", |
|
|
... "content": [ |
|
|
... {"type": "image"}, |
|
|
... {"type": "text", "text": "In this image, we can see the city of New York, and more specifically the Statue of Liberty."}, |
|
|
... {"type": "image"}, |
|
|
... {"type": "text", "text": "What can we see in this image?"}, |
|
|
... ] |
|
|
... }, |
|
|
... { |
|
|
... "role": "user", |
|
|
... "content": [ |
|
|
... {"type": "image"}, |
|
|
... {"type": "text", "text": "In which city is that bridge located?"}, |
|
|
... ] |
|
|
... } |
|
|
... ] |
|
|
|
|
|
>>> prompts = [processor.apply_chat_template([message], add_generation_prompt=True) for message in messages] |
|
|
>>> images = [[image1, image2], [image3]] |
|
|
>>> inputs = processor(text=prompts, images=images, padding=True, return_tensors="pt").to(model.device) |
|
|
|
|
|
>>> # Generate |
|
|
>>> generated_ids = model.generate(**inputs, max_new_tokens=256) |
|
|
>>> generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True) |
|
|
|
|
|
>>> print(generated_texts[0]) |
|
|
Assistant: There are buildings, trees, lights, and water visible in this image. |
|
|
|
|
|
>>> print(generated_texts[1]) |
|
|
Assistant: The bridge is in San Francisco. |
|
|
```""" |
|
|
outputs = self.model( |
|
|
input_ids=input_ids, |
|
|
pixel_values=pixel_values, |
|
|
pixel_mask=pixel_mask, |
|
|
attention_mask=attention_mask, |
|
|
position_ids=position_ids, |
|
|
past_key_values=past_key_values, |
|
|
inputs_embeds=inputs_embeds, |
|
|
use_cache=use_cache, |
|
|
cache_position=cache_position, |
|
|
**kwargs, |
|
|
) |
|
|
|
|
|
hidden_states = outputs[0] |
|
|
|
|
|
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep |
|
|
logits = self.lm_head(hidden_states[:, slice_indices, :]) |
|
|
|
|
|
loss = None |
|
|
if labels is not None: |
|
|
loss = self.loss_function( |
|
|
logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs |
|
|
) |
|
|
|
|
|
return AriaCausalLMOutputWithPast( |
|
|
loss=loss, |
|
|
logits=logits, |
|
|
past_key_values=outputs.past_key_values, |
|
|
hidden_states=outputs.hidden_states, |
|
|
attentions=outputs.attentions, |
|
|
) |
|
|
|
|
|
def prepare_inputs_for_generation( |
|
|
self, |
|
|
input_ids, |
|
|
past_key_values=None, |
|
|
inputs_embeds=None, |
|
|
pixel_values=None, |
|
|
pixel_mask=None, |
|
|
attention_mask=None, |
|
|
cache_position=None, |
|
|
logits_to_keep=None, |
|
|
**kwargs, |
|
|
): |
|
|
model_inputs = super().prepare_inputs_for_generation( |
|
|
input_ids, |
|
|
past_key_values=past_key_values, |
|
|
inputs_embeds=inputs_embeds, |
|
|
attention_mask=attention_mask, |
|
|
cache_position=cache_position, |
|
|
logits_to_keep=logits_to_keep, |
|
|
**kwargs, |
|
|
) |
|
|
|
|
|
if cache_position[0] == 0: |
|
|
|
|
|
|
|
|
model_inputs["pixel_values"] = pixel_values |
|
|
model_inputs["pixel_mask"] = pixel_mask |
|
|
|
|
|
return model_inputs |
|
|
|
|
|
|
|
|
__all__ = [ |
|
|
"AriaForConditionalGeneration", |
|
|
"AriaPreTrainedModel", |
|
|
"AriaTextPreTrainedModel", |
|
|
"AriaTextModel", |
|
|
"AriaModel", |
|
|
"AriaTextForCausalLM", |
|
|
] |
|
|
|
