Replace Qwen3 configs/modeling with SDAR version
Browse files- config.json +5 -5
- configuration_qwen3.py → configuration_sdar.py +15 -15
- modeling_qwen3_origin.py +0 -1065
- modeling_qwen3.py → modeling_sdar.py +42 -348
config.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
"architectures": [
|
| 3 |
-
"
|
| 4 |
],
|
| 5 |
"auto_map": {
|
| 6 |
-
"AutoConfig": "
|
| 7 |
-
"AutoModel": "
|
| 8 |
-
"AutoModelForCausalLM": "
|
| 9 |
},
|
| 10 |
"attention_bias": false,
|
| 11 |
"attention_dropout": 0.0,
|
|
@@ -19,7 +19,7 @@
|
|
| 19 |
"intermediate_size": 6144,
|
| 20 |
"max_position_embeddings": 32768,
|
| 21 |
"max_window_layers": 28,
|
| 22 |
-
"model_type": "
|
| 23 |
"num_attention_heads": 16,
|
| 24 |
"num_hidden_layers": 28,
|
| 25 |
"num_key_value_heads": 8,
|
|
|
|
| 1 |
{
|
| 2 |
"architectures": [
|
| 3 |
+
"SDARForCausalLM"
|
| 4 |
],
|
| 5 |
"auto_map": {
|
| 6 |
+
"AutoConfig": "configuration_sdar.SDARConfig",
|
| 7 |
+
"AutoModel": "modeling_sdar.SDARModel",
|
| 8 |
+
"AutoModelForCausalLM": "modeling_sdar.SDARForCausalLM"
|
| 9 |
},
|
| 10 |
"attention_bias": false,
|
| 11 |
"attention_dropout": 0.0,
|
|
|
|
| 19 |
"intermediate_size": 6144,
|
| 20 |
"max_position_embeddings": 32768,
|
| 21 |
"max_window_layers": 28,
|
| 22 |
+
"model_type": "sdar",
|
| 23 |
"num_attention_heads": 16,
|
| 24 |
"num_hidden_layers": 28,
|
| 25 |
"num_key_value_heads": 8,
|
configuration_qwen3.py → configuration_sdar.py
RENAMED
|
@@ -12,7 +12,7 @@
|
|
| 12 |
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
# See the License for the specific language governing permissions and
|
| 14 |
# limitations under the License.
|
| 15 |
-
"""
|
| 16 |
|
| 17 |
from transformers.configuration_utils import PretrainedConfig
|
| 18 |
from transformers.modeling_rope_utils import rope_config_validation
|
|
@@ -22,12 +22,12 @@ from transformers.utils import logging
|
|
| 22 |
logger = logging.get_logger(__name__)
|
| 23 |
|
| 24 |
|
| 25 |
-
class
|
| 26 |
r"""
|
| 27 |
-
This is the configuration class to store the configuration of a [`
|
| 28 |
-
|
| 29 |
with the defaults will yield a similar configuration to that of
|
| 30 |
-
|
| 31 |
|
| 32 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 33 |
documentation from [`PretrainedConfig`] for more information.
|
|
@@ -35,8 +35,8 @@ class Qwen3Config(PretrainedConfig):
|
|
| 35 |
|
| 36 |
Args:
|
| 37 |
vocab_size (`int`, *optional*, defaults to 151936):
|
| 38 |
-
Vocabulary size of the
|
| 39 |
-
`inputs_ids` passed when calling [`
|
| 40 |
hidden_size (`int`, *optional*, defaults to 4096):
|
| 41 |
Dimension of the hidden representations.
|
| 42 |
intermediate_size (`int`, *optional*, defaults to 22016):
|
|
@@ -118,22 +118,22 @@ class Qwen3Config(PretrainedConfig):
|
|
| 118 |
The dropout ratio for the attention probabilities.
|
| 119 |
|
| 120 |
```python
|
| 121 |
-
>>> from transformers import
|
| 122 |
|
| 123 |
-
>>> # Initializing a
|
| 124 |
-
>>> configuration =
|
| 125 |
|
| 126 |
-
>>> # Initializing a model from the
|
| 127 |
-
>>> model =
|
| 128 |
|
| 129 |
>>> # Accessing the model configuration
|
| 130 |
>>> configuration = model.config
|
| 131 |
```"""
|
| 132 |
|
| 133 |
-
model_type = "
|
| 134 |
keys_to_ignore_at_inference = ["past_key_values"]
|
| 135 |
|
| 136 |
-
# Default tensor parallel plan for base model `
|
| 137 |
base_model_tp_plan = {
|
| 138 |
"layers.*.self_attn.q_proj": "colwise",
|
| 139 |
"layers.*.self_attn.k_proj": "colwise",
|
|
@@ -209,4 +209,4 @@ class Qwen3Config(PretrainedConfig):
|
|
| 209 |
)
|
| 210 |
|
| 211 |
|
| 212 |
-
__all__ = ["
|
|
|
|
| 12 |
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
# See the License for the specific language governing permissions and
|
| 14 |
# limitations under the License.
|
| 15 |
+
"""SDAR model configuration"""
|
| 16 |
|
| 17 |
from transformers.configuration_utils import PretrainedConfig
|
| 18 |
from transformers.modeling_rope_utils import rope_config_validation
|
|
|
|
| 22 |
logger = logging.get_logger(__name__)
|
| 23 |
|
| 24 |
|
| 25 |
+
class SDARConfig(PretrainedConfig):
|
| 26 |
r"""
|
| 27 |
+
This is the configuration class to store the configuration of a [`SDARModel`]. It is used to instantiate a
|
| 28 |
+
SDAR model according to the specified arguments, defining the model architecture. Instantiating a configuration
|
| 29 |
with the defaults will yield a similar configuration to that of
|
| 30 |
+
SDAR-1.7B [DiffuOpen/SDAR-1.7B-Chat](https://huggingface.co/DiffuOpen/SDAR-1.7B-Chat/).
|
| 31 |
|
| 32 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 33 |
documentation from [`PretrainedConfig`] for more information.
|
|
|
|
| 35 |
|
| 36 |
Args:
|
| 37 |
vocab_size (`int`, *optional*, defaults to 151936):
|
| 38 |
+
Vocabulary size of the SDAR model. Defines the number of different tokens that can be represented by the
|
| 39 |
+
`inputs_ids` passed when calling [`SDARModel`]
|
| 40 |
hidden_size (`int`, *optional*, defaults to 4096):
|
| 41 |
Dimension of the hidden representations.
|
| 42 |
intermediate_size (`int`, *optional*, defaults to 22016):
|
|
|
|
| 118 |
The dropout ratio for the attention probabilities.
|
| 119 |
|
| 120 |
```python
|
| 121 |
+
>>> from transformers import SDARModel, SDARConfig
|
| 122 |
|
| 123 |
+
>>> # Initializing a SDAR style configuration
|
| 124 |
+
>>> configuration = SDARConfig()
|
| 125 |
|
| 126 |
+
>>> # Initializing a model from the SDAR-8B style configuration
|
| 127 |
+
>>> model = SDARModel(configuration)
|
| 128 |
|
| 129 |
>>> # Accessing the model configuration
|
| 130 |
>>> configuration = model.config
|
| 131 |
```"""
|
| 132 |
|
| 133 |
+
model_type = "sdar"
|
| 134 |
keys_to_ignore_at_inference = ["past_key_values"]
|
| 135 |
|
| 136 |
+
# Default tensor parallel plan for base model `SDAR`
|
| 137 |
base_model_tp_plan = {
|
| 138 |
"layers.*.self_attn.q_proj": "colwise",
|
| 139 |
"layers.*.self_attn.k_proj": "colwise",
|
|
|
|
| 209 |
)
|
| 210 |
|
| 211 |
|
| 212 |
+
__all__ = ["SDARConfig"]
|
modeling_qwen3_origin.py
DELETED
|
@@ -1,1065 +0,0 @@
|
|
| 1 |
-
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
| 2 |
-
# This file was automatically generated from src/transformers/models/qwen3/modular_qwen3.py.
|
| 3 |
-
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
| 4 |
-
# the file from the modular. If any change should be done, please apply the change to the
|
| 5 |
-
# modular_qwen3.py file directly. One of our CI enforces this.
|
| 6 |
-
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
| 7 |
-
# coding=utf-8
|
| 8 |
-
# Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
|
| 9 |
-
#
|
| 10 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 11 |
-
# you may not use this file except in compliance with the License.
|
| 12 |
-
# You may obtain a copy of the License at
|
| 13 |
-
#
|
| 14 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
| 15 |
-
#
|
| 16 |
-
# Unless required by applicable law or agreed to in writing, software
|
| 17 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 18 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 19 |
-
# See the License for the specific language governing permissions and
|
| 20 |
-
# limitations under the License.
|
| 21 |
-
|
| 22 |
-
from typing import Callable, Optional, Tuple, Union
|
| 23 |
-
|
| 24 |
-
import torch
|
| 25 |
-
from torch import nn
|
| 26 |
-
from einops import rearrange
|
| 27 |
-
|
| 28 |
-
from transformers.activations import ACT2FN
|
| 29 |
-
from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
|
| 30 |
-
from transformers.generation import GenerationMixin
|
| 31 |
-
from transformers.integrations import use_kernel_forward_from_hub
|
| 32 |
-
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
|
| 33 |
-
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
|
| 34 |
-
from transformers.modeling_layers import GradientCheckpointingLayer
|
| 35 |
-
from transformers.modeling_outputs import (
|
| 36 |
-
BaseModelOutputWithPast,
|
| 37 |
-
CausalLMOutputWithPast,
|
| 38 |
-
QuestionAnsweringModelOutput,
|
| 39 |
-
SequenceClassifierOutputWithPast,
|
| 40 |
-
TokenClassifierOutput,
|
| 41 |
-
)
|
| 42 |
-
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
|
| 43 |
-
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
| 44 |
-
from transformers.processing_utils import Unpack
|
| 45 |
-
from transformers.utils import LossKwargs, auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging
|
| 46 |
-
from .configuration_qwen3 import Qwen3Config
|
| 47 |
-
|
| 48 |
-
from torch.nn import CrossEntropyLoss
|
| 49 |
-
from fla.modules.activations import swiglu_linear
|
| 50 |
-
from fla.modules import FusedLinearDiffusionCrossEntropyLoss
|
| 51 |
-
from flash_attn.ops.triton.layer_norm import rms_norm_fn as flash_rms_norm
|
| 52 |
-
|
| 53 |
-
if is_torch_flex_attn_available():
|
| 54 |
-
from torch.nn.attention.flex_attention import BlockMask, flex_attention
|
| 55 |
-
|
| 56 |
-
from transformers.integrations.flex_attention import make_flex_block_causal_mask
|
| 57 |
-
|
| 58 |
-
# flex attn needs torch.compile to accelerate
|
| 59 |
-
@torch.compile(fullgraph=True, mode="max-autotune-no-cudagraphs")
|
| 60 |
-
def fused_flex_attention(query, key, value, attention_mask, **kwargs):
|
| 61 |
-
return flex_attention(query, key, value, block_mask=attention_mask, **kwargs)
|
| 62 |
-
|
| 63 |
-
logger = logging.get_logger(__name__)
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
@use_kernel_forward_from_hub("RMSNorm")
|
| 67 |
-
class Qwen3RMSNorm(nn.Module):
|
| 68 |
-
def __init__(self, hidden_size, eps=1e-6):
|
| 69 |
-
"""
|
| 70 |
-
Qwen3RMSNorm is equivalent to T5LayerNorm
|
| 71 |
-
"""
|
| 72 |
-
super().__init__()
|
| 73 |
-
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 74 |
-
self.variance_epsilon = eps
|
| 75 |
-
|
| 76 |
-
def forward(self, hidden_states):
|
| 77 |
-
input_dtype = hidden_states.dtype
|
| 78 |
-
# hidden_states = hidden_states.to(torch.float32)
|
| 79 |
-
# variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
| 80 |
-
# hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
| 81 |
-
return flash_rms_norm(
|
| 82 |
-
x=hidden_states, weight=self.weight, bias=None, eps=self.variance_epsilon).to(input_dtype)
|
| 83 |
-
|
| 84 |
-
def extra_repr(self):
|
| 85 |
-
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
class Qwen3MLP(nn.Module):
|
| 89 |
-
def __init__(self, config):
|
| 90 |
-
super().__init__()
|
| 91 |
-
self.config = config
|
| 92 |
-
self.hidden_size = config.hidden_size
|
| 93 |
-
self.intermediate_size = config.intermediate_size
|
| 94 |
-
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 95 |
-
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 96 |
-
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
| 97 |
-
self.act_fn = ACT2FN[config.hidden_act]
|
| 98 |
-
|
| 99 |
-
def forward(self, x):
|
| 100 |
-
# down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
| 101 |
-
down_proj = swiglu_linear(self.gate_proj(x), self.up_proj(x),
|
| 102 |
-
self.down_proj.weight, self.down_proj.bias)
|
| 103 |
-
return down_proj
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
def rotate_half(x):
|
| 107 |
-
"""Rotates half the hidden dims of the input."""
|
| 108 |
-
x1 = x[..., : x.shape[-1] // 2]
|
| 109 |
-
x2 = x[..., x.shape[-1] // 2 :]
|
| 110 |
-
return torch.cat((-x2, x1), dim=-1)
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
| 114 |
-
"""Applies Rotary Position Embedding to the query and key tensors.
|
| 115 |
-
|
| 116 |
-
Args:
|
| 117 |
-
q (`torch.Tensor`): The query tensor.
|
| 118 |
-
k (`torch.Tensor`): The key tensor.
|
| 119 |
-
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
| 120 |
-
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
| 121 |
-
position_ids (`torch.Tensor`, *optional*):
|
| 122 |
-
Deprecated and unused.
|
| 123 |
-
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
| 124 |
-
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
| 125 |
-
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
| 126 |
-
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
| 127 |
-
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
| 128 |
-
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
| 129 |
-
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
| 130 |
-
Returns:
|
| 131 |
-
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
| 132 |
-
"""
|
| 133 |
-
cos = cos.unsqueeze(unsqueeze_dim)
|
| 134 |
-
sin = sin.unsqueeze(unsqueeze_dim)
|
| 135 |
-
q_embed = (q * cos) + (rotate_half(q) * sin)
|
| 136 |
-
k_embed = (k * cos) + (rotate_half(k) * sin)
|
| 137 |
-
return q_embed, k_embed
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
| 141 |
-
"""
|
| 142 |
-
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
| 143 |
-
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
| 144 |
-
"""
|
| 145 |
-
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
| 146 |
-
if n_rep == 1:
|
| 147 |
-
return hidden_states
|
| 148 |
-
hidden_states = hidden_states[:, :, None, :, :].expand(
|
| 149 |
-
batch, num_key_value_heads, n_rep, slen, head_dim)
|
| 150 |
-
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
def eager_attention_forward(
|
| 154 |
-
module: nn.Module,
|
| 155 |
-
query: torch.Tensor,
|
| 156 |
-
key: torch.Tensor,
|
| 157 |
-
value: torch.Tensor,
|
| 158 |
-
attention_mask: Optional[torch.Tensor],
|
| 159 |
-
scaling: float,
|
| 160 |
-
dropout: float = 0.0,
|
| 161 |
-
**kwargs,
|
| 162 |
-
):
|
| 163 |
-
key_states = repeat_kv(key, module.num_key_value_groups)
|
| 164 |
-
value_states = repeat_kv(value, module.num_key_value_groups)
|
| 165 |
-
|
| 166 |
-
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
| 167 |
-
if attention_mask is not None:
|
| 168 |
-
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
| 169 |
-
attn_weights = attn_weights + causal_mask
|
| 170 |
-
|
| 171 |
-
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
| 172 |
-
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
| 173 |
-
attn_output = torch.matmul(attn_weights, value_states)
|
| 174 |
-
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 175 |
-
|
| 176 |
-
return attn_output, attn_weights
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
class Qwen3Attention(nn.Module):
|
| 180 |
-
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
| 181 |
-
|
| 182 |
-
def __init__(self, config: Qwen3Config, layer_idx: int):
|
| 183 |
-
super().__init__()
|
| 184 |
-
self.config = config
|
| 185 |
-
self.layer_idx = layer_idx
|
| 186 |
-
self.num_attention_heads = config.num_attention_heads
|
| 187 |
-
self.num_key_value_heads = config.num_key_value_heads
|
| 188 |
-
self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
|
| 189 |
-
self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
|
| 190 |
-
self.scaling = self.head_dim**-0.5
|
| 191 |
-
self.attention_dropout = config.attention_dropout
|
| 192 |
-
self.is_causal = False
|
| 193 |
-
|
| 194 |
-
self.q_proj = nn.Linear(
|
| 195 |
-
config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
|
| 196 |
-
)
|
| 197 |
-
self.k_proj = nn.Linear(
|
| 198 |
-
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
| 199 |
-
)
|
| 200 |
-
self.v_proj = nn.Linear(
|
| 201 |
-
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
| 202 |
-
)
|
| 203 |
-
self.o_proj = nn.Linear(
|
| 204 |
-
config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
|
| 205 |
-
)
|
| 206 |
-
self.q_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps) # unlike olmo, only on the head dim!
|
| 207 |
-
self.k_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps) # thus post q_norm does not need reshape
|
| 208 |
-
self.sliding_window = config.sliding_window
|
| 209 |
-
if not (
|
| 210 |
-
self.config.use_sliding_window
|
| 211 |
-
and getattr(self.config, "sliding_window", None) is not None
|
| 212 |
-
and self.layer_idx >= self.config.max_window_layers
|
| 213 |
-
):
|
| 214 |
-
self.sliding_window = None
|
| 215 |
-
|
| 216 |
-
def forward(
|
| 217 |
-
self,
|
| 218 |
-
hidden_states: torch.Tensor,
|
| 219 |
-
position_embeddings: Tuple[torch.Tensor, torch.Tensor],
|
| 220 |
-
attention_mask: Optional[torch.Tensor],
|
| 221 |
-
past_key_value: Optional[Cache] = None,
|
| 222 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 223 |
-
**kwargs: Unpack[FlashAttentionKwargs],
|
| 224 |
-
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 225 |
-
input_shape = hidden_states.shape[:-1]
|
| 226 |
-
bsz, q_len = input_shape
|
| 227 |
-
hidden_shape = (*input_shape, -1, self.head_dim)
|
| 228 |
-
|
| 229 |
-
query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
|
| 230 |
-
key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
|
| 231 |
-
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
| 232 |
-
|
| 233 |
-
cos, sin = position_embeddings
|
| 234 |
-
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
| 235 |
-
|
| 236 |
-
if past_key_value is not None:
|
| 237 |
-
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
| 238 |
-
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
| 239 |
-
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 240 |
-
|
| 241 |
-
attention_interface: Callable = eager_attention_forward
|
| 242 |
-
if self.config._attn_implementation != "eager":
|
| 243 |
-
if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
|
| 244 |
-
logger.warning_once(
|
| 245 |
-
"`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
|
| 246 |
-
'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
| 247 |
-
)
|
| 248 |
-
else:
|
| 249 |
-
attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
|
| 250 |
-
|
| 251 |
-
if self.config._attn_implementation == 'flex_attention':
|
| 252 |
-
# Although there exists `flex_attention_forward` in `AttentionInterface`,
|
| 253 |
-
# we still use our customized `fused_flex_attention` for debugging.
|
| 254 |
-
pad_length = kwargs.get("pad_length", 0)
|
| 255 |
-
# Used for SFT (packing + varlen), seq_len changes at each step
|
| 256 |
-
# seq_len must be divisible by BLOCK_SIZE in flex attn
|
| 257 |
-
pad_q = torch.zeros(
|
| 258 |
-
bsz, self.num_attention_heads, pad_length, self.head_dim, device=query_states.device, dtype=query_states.dtype)
|
| 259 |
-
pad_kv = torch.zeros(
|
| 260 |
-
bsz, self.num_key_value_heads, pad_length, self.head_dim, device=query_states.device, dtype=query_states.dtype)
|
| 261 |
-
attn_output, attn_weights = fused_flex_attention(
|
| 262 |
-
query=torch.cat([query_states, pad_q], dim=2),
|
| 263 |
-
key=torch.cat([key_states, pad_kv], dim=2),
|
| 264 |
-
value=torch.cat([value_states, pad_kv], dim=2),
|
| 265 |
-
attention_mask=attention_mask,
|
| 266 |
-
enable_gqa=True,
|
| 267 |
-
scale=self.scaling,
|
| 268 |
-
return_lse=True
|
| 269 |
-
)
|
| 270 |
-
|
| 271 |
-
attn_output = attn_output[..., :q_len, :].contiguous()
|
| 272 |
-
attn_weights = attn_weights.to(value_states.dtype)
|
| 273 |
-
attn_output = rearrange(attn_output, 'b h l d -> b l (h d)')
|
| 274 |
-
else:
|
| 275 |
-
attn_output, attn_weights = attention_interface(
|
| 276 |
-
self,
|
| 277 |
-
query_states,
|
| 278 |
-
key_states,
|
| 279 |
-
value_states,
|
| 280 |
-
attention_mask,
|
| 281 |
-
dropout=0.0 if not self.training else self.attention_dropout,
|
| 282 |
-
scaling=self.scaling,
|
| 283 |
-
sliding_window=self.sliding_window, # diff with Llama
|
| 284 |
-
**kwargs,
|
| 285 |
-
) # output: [b, l, h, d]
|
| 286 |
-
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
| 287 |
-
attn_output = self.o_proj(attn_output)
|
| 288 |
-
return attn_output, attn_weights
|
| 289 |
-
|
| 290 |
-
|
| 291 |
-
class Qwen3DecoderLayer(GradientCheckpointingLayer):
|
| 292 |
-
def __init__(self, config: Qwen3Config, layer_idx: int):
|
| 293 |
-
super().__init__()
|
| 294 |
-
self.hidden_size = config.hidden_size
|
| 295 |
-
self.self_attn = Qwen3Attention(config=config, layer_idx=layer_idx)
|
| 296 |
-
self.mlp = Qwen3MLP(config)
|
| 297 |
-
self.input_layernorm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 298 |
-
self.post_attention_layernorm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 299 |
-
if (
|
| 300 |
-
config.sliding_window and config._attn_implementation != "flash_attention_2"
|
| 301 |
-
): # diff with Llama is this warning
|
| 302 |
-
logger.warning_once(
|
| 303 |
-
f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
|
| 304 |
-
"unexpected results may be encountered."
|
| 305 |
-
)
|
| 306 |
-
|
| 307 |
-
def forward(
|
| 308 |
-
self,
|
| 309 |
-
hidden_states: torch.Tensor,
|
| 310 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 311 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 312 |
-
past_key_value: Optional[Cache] = None,
|
| 313 |
-
output_attentions: Optional[bool] = False,
|
| 314 |
-
use_cache: Optional[bool] = False,
|
| 315 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 316 |
-
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC
|
| 317 |
-
**kwargs: Unpack[FlashAttentionKwargs],
|
| 318 |
-
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
| 319 |
-
residual = hidden_states
|
| 320 |
-
hidden_states = self.input_layernorm(hidden_states)
|
| 321 |
-
|
| 322 |
-
# Self Attention
|
| 323 |
-
hidden_states, self_attn_weights = self.self_attn(
|
| 324 |
-
hidden_states=hidden_states,
|
| 325 |
-
attention_mask=attention_mask,
|
| 326 |
-
position_ids=position_ids,
|
| 327 |
-
past_key_value=past_key_value,
|
| 328 |
-
output_attentions=output_attentions,
|
| 329 |
-
use_cache=use_cache,
|
| 330 |
-
cache_position=cache_position,
|
| 331 |
-
position_embeddings=position_embeddings,
|
| 332 |
-
**kwargs,
|
| 333 |
-
)
|
| 334 |
-
hidden_states = residual + hidden_states
|
| 335 |
-
|
| 336 |
-
# Fully Connected
|
| 337 |
-
residual = hidden_states
|
| 338 |
-
hidden_states = self.post_attention_layernorm(hidden_states)
|
| 339 |
-
hidden_states = self.mlp(hidden_states)
|
| 340 |
-
hidden_states = residual + hidden_states
|
| 341 |
-
|
| 342 |
-
outputs = (hidden_states,)
|
| 343 |
-
if output_attentions:
|
| 344 |
-
outputs += (self_attn_weights,)
|
| 345 |
-
|
| 346 |
-
return outputs
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
@auto_docstring
|
| 350 |
-
class Qwen3PreTrainedModel(PreTrainedModel):
|
| 351 |
-
config_class = Qwen3Config
|
| 352 |
-
base_model_prefix = "model"
|
| 353 |
-
supports_gradient_checkpointing = True
|
| 354 |
-
_no_split_modules = ["Qwen3DecoderLayer"]
|
| 355 |
-
_skip_keys_device_placement = ["past_key_values"]
|
| 356 |
-
_supports_flash_attn_2 = True
|
| 357 |
-
_supports_sdpa = True
|
| 358 |
-
_supports_flex_attn = True
|
| 359 |
-
_supports_cache_class = True
|
| 360 |
-
_supports_quantized_cache = True
|
| 361 |
-
_supports_static_cache = True
|
| 362 |
-
_supports_attention_backend = True
|
| 363 |
-
|
| 364 |
-
def _init_weights(self, module):
|
| 365 |
-
std = self.config.initializer_range
|
| 366 |
-
if isinstance(module, nn.Linear):
|
| 367 |
-
module.weight.data.normal_(mean=0.0, std=std)
|
| 368 |
-
if module.bias is not None:
|
| 369 |
-
module.bias.data.zero_()
|
| 370 |
-
elif isinstance(module, nn.Embedding):
|
| 371 |
-
module.weight.data.normal_(mean=0.0, std=std)
|
| 372 |
-
if module.padding_idx is not None:
|
| 373 |
-
module.weight.data[module.padding_idx].zero_()
|
| 374 |
-
elif isinstance(module, Qwen3RMSNorm):
|
| 375 |
-
module.weight.data.fill_(1.0)
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
class Qwen3RotaryEmbedding(nn.Module):
|
| 379 |
-
def __init__(self, config: Qwen3Config, device=None):
|
| 380 |
-
super().__init__()
|
| 381 |
-
# BC: "rope_type" was originally "type"
|
| 382 |
-
if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
|
| 383 |
-
self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
|
| 384 |
-
else:
|
| 385 |
-
self.rope_type = "default"
|
| 386 |
-
self.max_seq_len_cached = config.max_position_embeddings
|
| 387 |
-
self.original_max_seq_len = config.max_position_embeddings
|
| 388 |
-
|
| 389 |
-
self.config = config
|
| 390 |
-
self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
|
| 391 |
-
|
| 392 |
-
inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
|
| 393 |
-
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
| 394 |
-
self.original_inv_freq = self.inv_freq
|
| 395 |
-
|
| 396 |
-
@torch.no_grad()
|
| 397 |
-
@dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope)
|
| 398 |
-
def forward(self, x, position_ids):
|
| 399 |
-
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
|
| 400 |
-
position_ids_expanded = position_ids[:, None, :].float()
|
| 401 |
-
|
| 402 |
-
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
|
| 403 |
-
with torch.autocast(device_type=device_type, enabled=False): # Force float32
|
| 404 |
-
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
| 405 |
-
emb = torch.cat((freqs, freqs), dim=-1)
|
| 406 |
-
cos = emb.cos() * self.attention_scaling
|
| 407 |
-
sin = emb.sin() * self.attention_scaling
|
| 408 |
-
|
| 409 |
-
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
| 410 |
-
|
| 411 |
-
|
| 412 |
-
@auto_docstring
|
| 413 |
-
class Qwen3Model(Qwen3PreTrainedModel):
|
| 414 |
-
def __init__(self, config: Qwen3Config):
|
| 415 |
-
super().__init__(config)
|
| 416 |
-
self.padding_idx = config.pad_token_id
|
| 417 |
-
self.vocab_size = config.vocab_size
|
| 418 |
-
|
| 419 |
-
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
| 420 |
-
self.layers = nn.ModuleList(
|
| 421 |
-
[Qwen3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
| 422 |
-
)
|
| 423 |
-
self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 424 |
-
self.rotary_emb = Qwen3RotaryEmbedding(config=config)
|
| 425 |
-
self.gradient_checkpointing = False
|
| 426 |
-
|
| 427 |
-
# Initialize weights and apply final processing
|
| 428 |
-
self.post_init()
|
| 429 |
-
|
| 430 |
-
def get_input_embeddings(self):
|
| 431 |
-
return self.embed_tokens
|
| 432 |
-
|
| 433 |
-
def set_input_embeddings(self, value):
|
| 434 |
-
self.embed_tokens = value
|
| 435 |
-
|
| 436 |
-
@can_return_tuple
|
| 437 |
-
@auto_docstring
|
| 438 |
-
def forward(
|
| 439 |
-
self,
|
| 440 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 441 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 442 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 443 |
-
past_key_values: Optional[Cache] = None,
|
| 444 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 445 |
-
use_cache: Optional[bool] = None,
|
| 446 |
-
output_attentions: Optional[bool] = None,
|
| 447 |
-
output_hidden_states: Optional[bool] = None,
|
| 448 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 449 |
-
**flash_attn_kwargs: Unpack[FlashAttentionKwargs],
|
| 450 |
-
) -> BaseModelOutputWithPast:
|
| 451 |
-
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 452 |
-
output_hidden_states = (
|
| 453 |
-
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 454 |
-
)
|
| 455 |
-
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
| 456 |
-
|
| 457 |
-
if (input_ids is None) ^ (inputs_embeds is not None):
|
| 458 |
-
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
| 459 |
-
|
| 460 |
-
if self.gradient_checkpointing and self.training and use_cache:
|
| 461 |
-
logger.warning_once(
|
| 462 |
-
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
|
| 463 |
-
)
|
| 464 |
-
use_cache = False
|
| 465 |
-
|
| 466 |
-
# TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
|
| 467 |
-
if not isinstance(past_key_values, (type(None), Cache)):
|
| 468 |
-
raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.")
|
| 469 |
-
|
| 470 |
-
if inputs_embeds is None:
|
| 471 |
-
inputs_embeds = self.embed_tokens(input_ids)
|
| 472 |
-
|
| 473 |
-
if use_cache and past_key_values is None:
|
| 474 |
-
past_key_values = DynamicCache()
|
| 475 |
-
|
| 476 |
-
if cache_position is None:
|
| 477 |
-
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
| 478 |
-
cache_position = torch.arange(
|
| 479 |
-
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
|
| 480 |
-
)
|
| 481 |
-
|
| 482 |
-
if position_ids is None:
|
| 483 |
-
position_ids = cache_position.unsqueeze(0)
|
| 484 |
-
|
| 485 |
-
causal_mask = self._update_causal_mask(
|
| 486 |
-
attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
|
| 487 |
-
)
|
| 488 |
-
|
| 489 |
-
hidden_states = inputs_embeds
|
| 490 |
-
|
| 491 |
-
# create position embeddings to be shared across the decoder layers
|
| 492 |
-
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
| 493 |
-
|
| 494 |
-
# decoder layers
|
| 495 |
-
all_hidden_states = () if output_hidden_states else None
|
| 496 |
-
all_self_attns = () if output_attentions else None
|
| 497 |
-
|
| 498 |
-
for decoder_layer in self.layers[: self.config.num_hidden_layers]:
|
| 499 |
-
if output_hidden_states:
|
| 500 |
-
all_hidden_states += (hidden_states,)
|
| 501 |
-
|
| 502 |
-
layer_outputs = decoder_layer(
|
| 503 |
-
hidden_states,
|
| 504 |
-
attention_mask=causal_mask,
|
| 505 |
-
position_ids=position_ids,
|
| 506 |
-
past_key_value=past_key_values,
|
| 507 |
-
output_attentions=output_attentions,
|
| 508 |
-
use_cache=use_cache,
|
| 509 |
-
cache_position=cache_position,
|
| 510 |
-
position_embeddings=position_embeddings,
|
| 511 |
-
**flash_attn_kwargs,
|
| 512 |
-
)
|
| 513 |
-
|
| 514 |
-
hidden_states = layer_outputs[0]
|
| 515 |
-
|
| 516 |
-
if output_attentions:
|
| 517 |
-
all_self_attns += (layer_outputs[1],)
|
| 518 |
-
|
| 519 |
-
hidden_states = self.norm(hidden_states)
|
| 520 |
-
|
| 521 |
-
# add hidden states from the last decoder layer
|
| 522 |
-
if output_hidden_states:
|
| 523 |
-
all_hidden_states += (hidden_states,)
|
| 524 |
-
|
| 525 |
-
return BaseModelOutputWithPast(
|
| 526 |
-
last_hidden_state=hidden_states,
|
| 527 |
-
past_key_values=past_key_values if use_cache else None,
|
| 528 |
-
hidden_states=all_hidden_states,
|
| 529 |
-
attentions=all_self_attns,
|
| 530 |
-
)
|
| 531 |
-
|
| 532 |
-
def _update_causal_mask(
|
| 533 |
-
self,
|
| 534 |
-
attention_mask: Union[torch.Tensor, "BlockMask"],
|
| 535 |
-
input_tensor: torch.Tensor,
|
| 536 |
-
cache_position: torch.Tensor,
|
| 537 |
-
past_key_values: Cache,
|
| 538 |
-
output_attentions: bool = False,
|
| 539 |
-
):
|
| 540 |
-
if self.config._attn_implementation == "flash_attention_2":
|
| 541 |
-
if attention_mask is not None and past_key_values is not None:
|
| 542 |
-
is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
|
| 543 |
-
if is_padding_right:
|
| 544 |
-
raise ValueError(
|
| 545 |
-
"You are attempting to perform batched generation with padding_side='right'"
|
| 546 |
-
" this may lead to unexpected behaviour for Flash Attention version of Qwen3. Make sure to "
|
| 547 |
-
" call `tokenizer.padding_side = 'left'` before tokenizing the input. "
|
| 548 |
-
)
|
| 549 |
-
if attention_mask is not None and 0.0 in attention_mask:
|
| 550 |
-
return attention_mask
|
| 551 |
-
return None
|
| 552 |
-
if self.config._attn_implementation == "flex_attention":
|
| 553 |
-
# Use flex block mask directly
|
| 554 |
-
assert isinstance(attention_mask, BlockMask)
|
| 555 |
-
return attention_mask
|
| 556 |
-
|
| 557 |
-
# For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
|
| 558 |
-
# order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
|
| 559 |
-
# to infer the attention mask.
|
| 560 |
-
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
| 561 |
-
using_static_cache = isinstance(past_key_values, StaticCache)
|
| 562 |
-
using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)
|
| 563 |
-
|
| 564 |
-
# When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
|
| 565 |
-
if (
|
| 566 |
-
self.config._attn_implementation == "sdpa"
|
| 567 |
-
and not (using_static_cache or using_sliding_window_cache)
|
| 568 |
-
and not output_attentions
|
| 569 |
-
):
|
| 570 |
-
if AttentionMaskConverter._ignore_causal_mask_sdpa(
|
| 571 |
-
attention_mask,
|
| 572 |
-
inputs_embeds=input_tensor,
|
| 573 |
-
past_key_values_length=past_seen_tokens,
|
| 574 |
-
sliding_window=self.config.sliding_window,
|
| 575 |
-
is_training=self.training,
|
| 576 |
-
):
|
| 577 |
-
return None
|
| 578 |
-
|
| 579 |
-
dtype = input_tensor.dtype
|
| 580 |
-
min_dtype = torch.finfo(dtype).min
|
| 581 |
-
sequence_length = input_tensor.shape[1]
|
| 582 |
-
# SlidingWindowCache or StaticCache
|
| 583 |
-
if using_sliding_window_cache or using_static_cache:
|
| 584 |
-
target_length = past_key_values.get_max_cache_shape()
|
| 585 |
-
# DynamicCache or no cache
|
| 586 |
-
else:
|
| 587 |
-
target_length = (
|
| 588 |
-
attention_mask.shape[-1]
|
| 589 |
-
if isinstance(attention_mask, torch.Tensor)
|
| 590 |
-
else past_seen_tokens + sequence_length + 1
|
| 591 |
-
)
|
| 592 |
-
|
| 593 |
-
# In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
|
| 594 |
-
causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
|
| 595 |
-
attention_mask,
|
| 596 |
-
sequence_length=sequence_length,
|
| 597 |
-
target_length=target_length,
|
| 598 |
-
dtype=dtype,
|
| 599 |
-
cache_position=cache_position,
|
| 600 |
-
batch_size=input_tensor.shape[0],
|
| 601 |
-
config=self.config,
|
| 602 |
-
past_key_values=past_key_values,
|
| 603 |
-
)
|
| 604 |
-
|
| 605 |
-
if (
|
| 606 |
-
self.config._attn_implementation == "sdpa"
|
| 607 |
-
and attention_mask is not None
|
| 608 |
-
and attention_mask.device.type in ["cuda", "xpu", "npu"]
|
| 609 |
-
and not output_attentions
|
| 610 |
-
):
|
| 611 |
-
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
|
| 612 |
-
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
|
| 613 |
-
# Details: https://github.com/pytorch/pytorch/issues/110213
|
| 614 |
-
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
|
| 615 |
-
|
| 616 |
-
return causal_mask
|
| 617 |
-
|
| 618 |
-
@staticmethod
|
| 619 |
-
def _prepare_4d_causal_attention_mask_with_cache_position(
|
| 620 |
-
attention_mask: torch.Tensor,
|
| 621 |
-
sequence_length: int,
|
| 622 |
-
target_length: int,
|
| 623 |
-
dtype: torch.dtype,
|
| 624 |
-
cache_position: torch.Tensor,
|
| 625 |
-
batch_size: int,
|
| 626 |
-
config: Qwen3Config,
|
| 627 |
-
past_key_values: Cache,
|
| 628 |
-
):
|
| 629 |
-
"""
|
| 630 |
-
Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
|
| 631 |
-
`(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
|
| 632 |
-
|
| 633 |
-
Args:
|
| 634 |
-
attention_mask (`torch.Tensor`):
|
| 635 |
-
A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
|
| 636 |
-
sequence_length (`int`):
|
| 637 |
-
The sequence length being processed.
|
| 638 |
-
target_length (`int`):
|
| 639 |
-
The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
|
| 640 |
-
dtype (`torch.dtype`):
|
| 641 |
-
The dtype to use for the 4D attention mask.
|
| 642 |
-
cache_position (`torch.Tensor`):
|
| 643 |
-
Indices depicting the position of the input sequence tokens in the sequence.
|
| 644 |
-
batch_size (`torch.Tensor`):
|
| 645 |
-
Batch size.
|
| 646 |
-
config (`Qwen3Config`):
|
| 647 |
-
The model's configuration class
|
| 648 |
-
past_key_values (`Cache`):
|
| 649 |
-
The cache class that is being used currently to generate
|
| 650 |
-
"""
|
| 651 |
-
if attention_mask is not None and attention_mask.dim() == 4:
|
| 652 |
-
# In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
|
| 653 |
-
causal_mask = attention_mask
|
| 654 |
-
else:
|
| 655 |
-
min_dtype = torch.finfo(dtype).min
|
| 656 |
-
causal_mask = torch.full(
|
| 657 |
-
(sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
|
| 658 |
-
)
|
| 659 |
-
diagonal_attend_mask = torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
|
| 660 |
-
-1, 1
|
| 661 |
-
)
|
| 662 |
-
text_config = config.get_text_config()
|
| 663 |
-
if getattr(text_config, "use_sliding_window", True) and text_config.sliding_window is not None:
|
| 664 |
-
# if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
|
| 665 |
-
# the check is needed to verify is current checkpoint was trained with sliding window or not
|
| 666 |
-
if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
|
| 667 |
-
sliding_attend_mask = torch.arange(target_length, device=cache_position.device) <= (
|
| 668 |
-
cache_position.reshape(-1, 1) - text_config.sliding_window
|
| 669 |
-
)
|
| 670 |
-
diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
|
| 671 |
-
causal_mask *= diagonal_attend_mask
|
| 672 |
-
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
|
| 673 |
-
if attention_mask is not None:
|
| 674 |
-
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
| 675 |
-
if attention_mask.shape[-1] > target_length:
|
| 676 |
-
attention_mask = attention_mask[:, :target_length]
|
| 677 |
-
mask_length = attention_mask.shape[-1]
|
| 678 |
-
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
|
| 679 |
-
causal_mask.device
|
| 680 |
-
)
|
| 681 |
-
padding_mask = padding_mask == 0
|
| 682 |
-
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
|
| 683 |
-
padding_mask, min_dtype
|
| 684 |
-
)
|
| 685 |
-
return causal_mask
|
| 686 |
-
|
| 687 |
-
|
| 688 |
-
class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
|
| 689 |
-
|
| 690 |
-
|
| 691 |
-
@auto_docstring
|
| 692 |
-
class Qwen3ForCausalLM(Qwen3PreTrainedModel, GenerationMixin):
|
| 693 |
-
_tied_weights_keys = ["lm_head.weight"]
|
| 694 |
-
_tp_plan = {"lm_head": "colwise_rep"}
|
| 695 |
-
_pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
|
| 696 |
-
|
| 697 |
-
def __init__(self, config):
|
| 698 |
-
super().__init__(config)
|
| 699 |
-
self.model = Qwen3Model(config)
|
| 700 |
-
self.vocab_size = config.vocab_size
|
| 701 |
-
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 702 |
-
|
| 703 |
-
# Initialize weights and apply final processing
|
| 704 |
-
self.post_init()
|
| 705 |
-
|
| 706 |
-
def get_input_embeddings(self):
|
| 707 |
-
return self.model.embed_tokens
|
| 708 |
-
|
| 709 |
-
def set_input_embeddings(self, value):
|
| 710 |
-
self.model.embed_tokens = value
|
| 711 |
-
|
| 712 |
-
def get_output_embeddings(self):
|
| 713 |
-
return self.lm_head
|
| 714 |
-
|
| 715 |
-
def set_output_embeddings(self, new_embeddings):
|
| 716 |
-
self.lm_head = new_embeddings
|
| 717 |
-
|
| 718 |
-
def set_decoder(self, decoder):
|
| 719 |
-
self.model = decoder
|
| 720 |
-
|
| 721 |
-
def get_decoder(self):
|
| 722 |
-
return self.model
|
| 723 |
-
|
| 724 |
-
@can_return_tuple
|
| 725 |
-
@auto_docstring
|
| 726 |
-
def forward(
|
| 727 |
-
self,
|
| 728 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 729 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 730 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 731 |
-
past_key_values: Optional[Cache] = None,
|
| 732 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 733 |
-
labels: Optional[torch.LongTensor] = None,
|
| 734 |
-
use_cache: Optional[bool] = None,
|
| 735 |
-
output_attentions: Optional[bool] = None,
|
| 736 |
-
output_hidden_states: Optional[bool] = None,
|
| 737 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 738 |
-
logits_to_keep: Union[int, torch.Tensor] = 0,
|
| 739 |
-
**kwargs: Unpack[KwargsForCausalLM],
|
| 740 |
-
) -> CausalLMOutputWithPast:
|
| 741 |
-
r"""
|
| 742 |
-
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 743 |
-
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
| 744 |
-
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
| 745 |
-
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
| 746 |
-
|
| 747 |
-
Example:
|
| 748 |
-
|
| 749 |
-
```python
|
| 750 |
-
>>> from transformers import AutoTokenizer, Qwen3ForCausalLM
|
| 751 |
-
|
| 752 |
-
>>> model = Qwen3ForCausalLM.from_pretrained("Qwen/Qwen3-8B")
|
| 753 |
-
>>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
|
| 754 |
-
|
| 755 |
-
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
| 756 |
-
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
| 757 |
-
|
| 758 |
-
>>> # Generate
|
| 759 |
-
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
| 760 |
-
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 761 |
-
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
|
| 762 |
-
```"""
|
| 763 |
-
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 764 |
-
output_hidden_states = (
|
| 765 |
-
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 766 |
-
)
|
| 767 |
-
|
| 768 |
-
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
| 769 |
-
outputs: BaseModelOutputWithPast = self.model(
|
| 770 |
-
input_ids=input_ids,
|
| 771 |
-
attention_mask=attention_mask,
|
| 772 |
-
position_ids=position_ids,
|
| 773 |
-
past_key_values=past_key_values,
|
| 774 |
-
inputs_embeds=inputs_embeds,
|
| 775 |
-
use_cache=use_cache,
|
| 776 |
-
output_attentions=output_attentions,
|
| 777 |
-
output_hidden_states=output_hidden_states,
|
| 778 |
-
cache_position=cache_position,
|
| 779 |
-
**kwargs,
|
| 780 |
-
)
|
| 781 |
-
|
| 782 |
-
hidden_states = outputs.last_hidden_state
|
| 783 |
-
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
|
| 784 |
-
logits_to_keep = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
|
| 785 |
-
hidden_states = hidden_states[:, logits_to_keep, :].contiguous()
|
| 786 |
-
fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
|
| 787 |
-
if fuse_linear_and_cross_entropy:
|
| 788 |
-
logits = None
|
| 789 |
-
else:
|
| 790 |
-
logits = self.lm_head(hidden_states)
|
| 791 |
-
|
| 792 |
-
loss = None
|
| 793 |
-
if labels is not None:
|
| 794 |
-
if fuse_linear_and_cross_entropy:
|
| 795 |
-
loss_fct = FusedLinearDiffusionCrossEntropyLoss(
|
| 796 |
-
reduction='sum')
|
| 797 |
-
else:
|
| 798 |
-
loss_fct = CrossEntropyLoss() # nn.CE
|
| 799 |
-
|
| 800 |
-
# you don't have to shift labels
|
| 801 |
-
# labels = labels.to(hidden_states.device)
|
| 802 |
-
# labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], loss_fct.ignore_index)), 1)
|
| 803 |
-
if fuse_linear_and_cross_entropy:
|
| 804 |
-
loss = loss_fct( # it will return (sum_loss, unreduced_loss)
|
| 805 |
-
x=hidden_states, # conduct `view(-1, V)` inside the function
|
| 806 |
-
target=labels,
|
| 807 |
-
weight=self.lm_head.weight,
|
| 808 |
-
bias=self.lm_head.bias,
|
| 809 |
-
p_mask=kwargs['p_mask'],
|
| 810 |
-
)
|
| 811 |
-
else:
|
| 812 |
-
loss = loss_fct(
|
| 813 |
-
logits.view(-1, self.config.vocab_size), labels.view(-1))
|
| 814 |
-
|
| 815 |
-
return CausalLMOutputWithPast(
|
| 816 |
-
loss=loss,
|
| 817 |
-
logits=logits,
|
| 818 |
-
past_key_values=outputs.past_key_values,
|
| 819 |
-
hidden_states=outputs.hidden_states,
|
| 820 |
-
attentions=outputs.attentions,
|
| 821 |
-
)
|
| 822 |
-
|
| 823 |
-
|
| 824 |
-
@auto_docstring(
|
| 825 |
-
custom_intro="""
|
| 826 |
-
The Qwen3 Model transformer with a sequence classification head on top (linear layer).
|
| 827 |
-
|
| 828 |
-
[`Qwen3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
| 829 |
-
(e.g. GPT-2) do.
|
| 830 |
-
|
| 831 |
-
Since it does classification on the last token, it requires to know the position of the last token. If a
|
| 832 |
-
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
| 833 |
-
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
| 834 |
-
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
|
| 835 |
-
each row of the batch).
|
| 836 |
-
"""
|
| 837 |
-
)
|
| 838 |
-
class Qwen3ForSequenceClassification(Qwen3PreTrainedModel):
|
| 839 |
-
def __init__(self, config):
|
| 840 |
-
super().__init__(config)
|
| 841 |
-
self.num_labels = config.num_labels
|
| 842 |
-
self.model = Qwen3Model(config)
|
| 843 |
-
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
|
| 844 |
-
|
| 845 |
-
# Initialize weights and apply final processing
|
| 846 |
-
self.post_init()
|
| 847 |
-
|
| 848 |
-
def get_input_embeddings(self):
|
| 849 |
-
return self.model.embed_tokens
|
| 850 |
-
|
| 851 |
-
def set_input_embeddings(self, value):
|
| 852 |
-
self.model.embed_tokens = value
|
| 853 |
-
|
| 854 |
-
@can_return_tuple
|
| 855 |
-
@auto_docstring
|
| 856 |
-
def forward(
|
| 857 |
-
self,
|
| 858 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 859 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 860 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 861 |
-
past_key_values: Optional[Cache] = None,
|
| 862 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 863 |
-
labels: Optional[torch.LongTensor] = None,
|
| 864 |
-
use_cache: Optional[bool] = None,
|
| 865 |
-
output_attentions: Optional[bool] = None,
|
| 866 |
-
output_hidden_states: Optional[bool] = None,
|
| 867 |
-
) -> SequenceClassifierOutputWithPast:
|
| 868 |
-
r"""
|
| 869 |
-
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 870 |
-
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
| 871 |
-
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
| 872 |
-
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
| 873 |
-
"""
|
| 874 |
-
|
| 875 |
-
transformer_outputs: BaseModelOutputWithPast = self.model(
|
| 876 |
-
input_ids,
|
| 877 |
-
attention_mask=attention_mask,
|
| 878 |
-
position_ids=position_ids,
|
| 879 |
-
past_key_values=past_key_values,
|
| 880 |
-
inputs_embeds=inputs_embeds,
|
| 881 |
-
use_cache=use_cache,
|
| 882 |
-
output_attentions=output_attentions,
|
| 883 |
-
output_hidden_states=output_hidden_states,
|
| 884 |
-
)
|
| 885 |
-
hidden_states = transformer_outputs.last_hidden_state
|
| 886 |
-
logits = self.score(hidden_states)
|
| 887 |
-
|
| 888 |
-
if input_ids is not None:
|
| 889 |
-
batch_size = input_ids.shape[0]
|
| 890 |
-
else:
|
| 891 |
-
batch_size = inputs_embeds.shape[0]
|
| 892 |
-
|
| 893 |
-
if self.config.pad_token_id is None and batch_size != 1:
|
| 894 |
-
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
|
| 895 |
-
if self.config.pad_token_id is None:
|
| 896 |
-
last_non_pad_token = -1
|
| 897 |
-
elif input_ids is not None:
|
| 898 |
-
# To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
|
| 899 |
-
non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32)
|
| 900 |
-
token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32)
|
| 901 |
-
last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
|
| 902 |
-
else:
|
| 903 |
-
last_non_pad_token = -1
|
| 904 |
-
logger.warning_once(
|
| 905 |
-
f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
|
| 906 |
-
"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
|
| 907 |
-
)
|
| 908 |
-
|
| 909 |
-
pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token]
|
| 910 |
-
|
| 911 |
-
loss = None
|
| 912 |
-
if labels is not None:
|
| 913 |
-
loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config)
|
| 914 |
-
|
| 915 |
-
return SequenceClassifierOutputWithPast(
|
| 916 |
-
loss=loss,
|
| 917 |
-
logits=pooled_logits,
|
| 918 |
-
past_key_values=transformer_outputs.past_key_values,
|
| 919 |
-
hidden_states=transformer_outputs.hidden_states,
|
| 920 |
-
attentions=transformer_outputs.attentions,
|
| 921 |
-
)
|
| 922 |
-
|
| 923 |
-
|
| 924 |
-
@auto_docstring
|
| 925 |
-
class Qwen3ForTokenClassification(Qwen3PreTrainedModel):
|
| 926 |
-
def __init__(self, config):
|
| 927 |
-
super().__init__(config)
|
| 928 |
-
self.num_labels = config.num_labels
|
| 929 |
-
self.model = Qwen3Model(config)
|
| 930 |
-
if getattr(config, "classifier_dropout", None) is not None:
|
| 931 |
-
classifier_dropout = config.classifier_dropout
|
| 932 |
-
elif getattr(config, "hidden_dropout", None) is not None:
|
| 933 |
-
classifier_dropout = config.hidden_dropout
|
| 934 |
-
else:
|
| 935 |
-
classifier_dropout = 0.1
|
| 936 |
-
self.dropout = nn.Dropout(classifier_dropout)
|
| 937 |
-
self.score = nn.Linear(config.hidden_size, config.num_labels)
|
| 938 |
-
|
| 939 |
-
# Initialize weights and apply final processing
|
| 940 |
-
self.post_init()
|
| 941 |
-
|
| 942 |
-
def get_input_embeddings(self):
|
| 943 |
-
return self.model.embed_tokens
|
| 944 |
-
|
| 945 |
-
def set_input_embeddings(self, value):
|
| 946 |
-
self.model.embed_tokens = value
|
| 947 |
-
|
| 948 |
-
@can_return_tuple
|
| 949 |
-
@auto_docstring
|
| 950 |
-
def forward(
|
| 951 |
-
self,
|
| 952 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 953 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 954 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 955 |
-
past_key_values: Optional[Cache] = None,
|
| 956 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 957 |
-
labels: Optional[torch.LongTensor] = None,
|
| 958 |
-
use_cache: Optional[bool] = None,
|
| 959 |
-
output_attentions: Optional[bool] = None,
|
| 960 |
-
output_hidden_states: Optional[bool] = None,
|
| 961 |
-
) -> TokenClassifierOutput:
|
| 962 |
-
r"""
|
| 963 |
-
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 964 |
-
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
| 965 |
-
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
| 966 |
-
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
| 967 |
-
"""
|
| 968 |
-
|
| 969 |
-
outputs: BaseModelOutputWithPast = self.model(
|
| 970 |
-
input_ids,
|
| 971 |
-
attention_mask=attention_mask,
|
| 972 |
-
position_ids=position_ids,
|
| 973 |
-
past_key_values=past_key_values,
|
| 974 |
-
inputs_embeds=inputs_embeds,
|
| 975 |
-
use_cache=use_cache,
|
| 976 |
-
output_attentions=output_attentions,
|
| 977 |
-
output_hidden_states=output_hidden_states,
|
| 978 |
-
)
|
| 979 |
-
sequence_output = outputs.last_hidden_state
|
| 980 |
-
sequence_output = self.dropout(sequence_output)
|
| 981 |
-
logits = self.score(sequence_output)
|
| 982 |
-
|
| 983 |
-
loss = None
|
| 984 |
-
if labels is not None:
|
| 985 |
-
loss = self.loss_function(logits, labels, self.config)
|
| 986 |
-
|
| 987 |
-
return TokenClassifierOutput(
|
| 988 |
-
loss=loss,
|
| 989 |
-
logits=logits,
|
| 990 |
-
hidden_states=outputs.hidden_states,
|
| 991 |
-
attentions=outputs.attentions,
|
| 992 |
-
)
|
| 993 |
-
|
| 994 |
-
|
| 995 |
-
@auto_docstring
|
| 996 |
-
class Qwen3ForQuestionAnswering(Qwen3PreTrainedModel):
|
| 997 |
-
base_model_prefix = "transformer"
|
| 998 |
-
|
| 999 |
-
def __init__(self, config):
|
| 1000 |
-
super().__init__(config)
|
| 1001 |
-
self.transformer = Qwen3Model(config)
|
| 1002 |
-
self.qa_outputs = nn.Linear(config.hidden_size, 2)
|
| 1003 |
-
|
| 1004 |
-
# Initialize weights and apply final processing
|
| 1005 |
-
self.post_init()
|
| 1006 |
-
|
| 1007 |
-
def get_input_embeddings(self):
|
| 1008 |
-
return self.transformer.embed_tokens
|
| 1009 |
-
|
| 1010 |
-
def set_input_embeddings(self, value):
|
| 1011 |
-
self.transformer.embed_tokens = value
|
| 1012 |
-
|
| 1013 |
-
@can_return_tuple
|
| 1014 |
-
@auto_docstring
|
| 1015 |
-
def forward(
|
| 1016 |
-
self,
|
| 1017 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 1018 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 1019 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 1020 |
-
past_key_values: Optional[Cache] = None,
|
| 1021 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1022 |
-
start_positions: Optional[torch.LongTensor] = None,
|
| 1023 |
-
end_positions: Optional[torch.LongTensor] = None,
|
| 1024 |
-
output_attentions: Optional[bool] = None,
|
| 1025 |
-
output_hidden_states: Optional[bool] = None,
|
| 1026 |
-
**kwargs,
|
| 1027 |
-
) -> QuestionAnsweringModelOutput:
|
| 1028 |
-
outputs: BaseModelOutputWithPast = self.transformer(
|
| 1029 |
-
input_ids,
|
| 1030 |
-
attention_mask=attention_mask,
|
| 1031 |
-
position_ids=position_ids,
|
| 1032 |
-
past_key_values=past_key_values,
|
| 1033 |
-
inputs_embeds=inputs_embeds,
|
| 1034 |
-
output_attentions=output_attentions,
|
| 1035 |
-
output_hidden_states=output_hidden_states,
|
| 1036 |
-
)
|
| 1037 |
-
|
| 1038 |
-
sequence_output = outputs.last_hidden_state
|
| 1039 |
-
|
| 1040 |
-
logits = self.qa_outputs(sequence_output)
|
| 1041 |
-
start_logits, end_logits = logits.split(1, dim=-1)
|
| 1042 |
-
start_logits = start_logits.squeeze(-1).contiguous()
|
| 1043 |
-
end_logits = end_logits.squeeze(-1).contiguous()
|
| 1044 |
-
|
| 1045 |
-
loss = None
|
| 1046 |
-
if start_positions is not None and end_positions is not None:
|
| 1047 |
-
loss = self.loss_function(start_logits, end_logits, start_positions, end_positions, **kwargs)
|
| 1048 |
-
|
| 1049 |
-
return QuestionAnsweringModelOutput(
|
| 1050 |
-
loss=loss,
|
| 1051 |
-
start_logits=start_logits,
|
| 1052 |
-
end_logits=end_logits,
|
| 1053 |
-
hidden_states=outputs.hidden_states,
|
| 1054 |
-
attentions=outputs.attentions,
|
| 1055 |
-
)
|
| 1056 |
-
|
| 1057 |
-
|
| 1058 |
-
__all__ = [
|
| 1059 |
-
"Qwen3ForCausalLM",
|
| 1060 |
-
"Qwen3ForQuestionAnswering",
|
| 1061 |
-
"Qwen3Model",
|
| 1062 |
-
"Qwen3PreTrainedModel",
|
| 1063 |
-
"Qwen3ForSequenceClassification",
|
| 1064 |
-
"Qwen3ForTokenClassification",
|
| 1065 |
-
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
modeling_qwen3.py → modeling_sdar.py
RENAMED
|
@@ -1,3 +1,5 @@
|
|
|
|
|
|
|
|
| 1 |
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
| 2 |
# This file was automatically generated from src/transformers/models/qwen3/modular_qwen3.py.
|
| 3 |
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
|
@@ -42,7 +44,7 @@ from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_u
|
|
| 42 |
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
| 43 |
from transformers.processing_utils import Unpack
|
| 44 |
from transformers.utils import LossKwargs, auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging
|
| 45 |
-
from .
|
| 46 |
|
| 47 |
from fla.modules.activations import swiglu_linear
|
| 48 |
from fla.modules import (
|
|
@@ -81,10 +83,10 @@ def fused_flex_attention(query, key, value, attention_mask=None, **kwargs):
|
|
| 81 |
|
| 82 |
|
| 83 |
@use_kernel_forward_from_hub("RMSNorm")
|
| 84 |
-
class
|
| 85 |
def __init__(self, hidden_size, eps=1e-6):
|
| 86 |
"""
|
| 87 |
-
|
| 88 |
"""
|
| 89 |
super().__init__()
|
| 90 |
self.weight = nn.Parameter(torch.ones(hidden_size))
|
|
@@ -107,7 +109,7 @@ class Qwen3RMSNorm(nn.Module):
|
|
| 107 |
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
| 108 |
|
| 109 |
|
| 110 |
-
class
|
| 111 |
def __init__(self, config):
|
| 112 |
super().__init__()
|
| 113 |
self.config = config
|
|
@@ -205,10 +207,10 @@ def eager_attention_forward(
|
|
| 205 |
return attn_output, attn_weights
|
| 206 |
|
| 207 |
|
| 208 |
-
class
|
| 209 |
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
| 210 |
|
| 211 |
-
def __init__(self, config:
|
| 212 |
super().__init__()
|
| 213 |
self.config = config
|
| 214 |
self.layer_idx = layer_idx
|
|
@@ -236,9 +238,9 @@ class Qwen3Attention(nn.Module):
|
|
| 236 |
config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
|
| 237 |
)
|
| 238 |
# unlike olmo, only on the head dim!
|
| 239 |
-
self.q_norm =
|
| 240 |
# thus post q_norm does not need reshape
|
| 241 |
-
self.k_norm =
|
| 242 |
self.sliding_window = config.sliding_window
|
| 243 |
if not (
|
| 244 |
self.config.use_sliding_window
|
|
@@ -275,7 +277,8 @@ class Qwen3Attention(nn.Module):
|
|
| 275 |
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
| 276 |
key_states, value_states = past_key_value.update(
|
| 277 |
key_states, value_states, self.layer_idx)
|
| 278 |
-
elif past_key_value is not None and not kwargs.get("store_kv", False) and len(past_key_value) > self.layer_idx
|
|
|
|
| 279 |
past_key_states, past_value_states = past_key_value[self.layer_idx]
|
| 280 |
key_states = torch.cat(
|
| 281 |
[past_key_states, key_states], dim=-2
|
|
@@ -283,75 +286,9 @@ class Qwen3Attention(nn.Module):
|
|
| 283 |
value_states = torch.cat(
|
| 284 |
[past_value_states, value_states], dim=-2
|
| 285 |
)
|
| 286 |
-
|
| 287 |
-
# # sin and cos are specific to RoPE models; cache_position needed for the static cache
|
| 288 |
-
# cache_kwargs = {"sin": sin, "cos": cos,
|
| 289 |
-
# "cache_position": cache_position}
|
| 290 |
-
# key_states, value_states = past_key_value.update(
|
| 291 |
-
# key_states, value_states, self.layer_idx, cache_kwargs)
|
| 292 |
-
|
| 293 |
-
# attention_interface: Callable = eager_attention_forward
|
| 294 |
-
# if self.config._attn_implementation != "eager":
|
| 295 |
-
# if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
|
| 296 |
-
# logger.warning_once(
|
| 297 |
-
# "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
|
| 298 |
-
# 'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
| 299 |
-
# )
|
| 300 |
-
# else:
|
| 301 |
-
# attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
|
| 302 |
-
|
| 303 |
-
# if self.config._attn_implementation == 'flex_attention':
|
| 304 |
-
# # Although `AttentionInterface` has `flex_attention_forward` implementation,
|
| 305 |
-
# # we still use our customized `fused_flex_attention`
|
| 306 |
-
# pad_length = kwargs.get("pad_length", None)
|
| 307 |
-
# if pad_length is not None:
|
| 308 |
-
# # Used for SFT (packing + varlen), seq_len changes at each step
|
| 309 |
-
# # seq_len must be divisible by BLOCK_SIZE in flex attn
|
| 310 |
-
# pad_q = torch.zeros(
|
| 311 |
-
# bsz, self.num_attention_heads, pad_length, self.head_dim, device=query_states.device, dtype=query_states.dtype)
|
| 312 |
-
# pad_kv = torch.zeros(
|
| 313 |
-
# bsz, self.num_key_value_heads, pad_length, self.head_dim, device=query_states.device, dtype=query_states.dtype)
|
| 314 |
-
# attn_output, attn_weights = fused_flex_attention(
|
| 315 |
-
# query=torch.cat([query_states, pad_q], dim=2),
|
| 316 |
-
# key=torch.cat([key_states, pad_kv], dim=2),
|
| 317 |
-
# value=torch.cat([value_states, pad_kv], dim=2),
|
| 318 |
-
# attention_mask=attention_mask,
|
| 319 |
-
# enable_gqa=True,
|
| 320 |
-
# scale=self.scaling,
|
| 321 |
-
# return_lse=True
|
| 322 |
-
# )
|
| 323 |
-
# attn_output = attn_output[..., :q_len,
|
| 324 |
-
# :].transpose(1, 2).contiguous()
|
| 325 |
-
# attn_weights = attn_weights.to(value_states.dtype)
|
| 326 |
-
# else:
|
| 327 |
-
# attn_output, attn_weights = fused_flex_attention(
|
| 328 |
-
# query=query_states,
|
| 329 |
-
# key=key_states,
|
| 330 |
-
# value=value_states,
|
| 331 |
-
# attention_mask=attention_mask,
|
| 332 |
-
# enable_gqa=True,
|
| 333 |
-
# scale=self.scaling,
|
| 334 |
-
# return_lse=True
|
| 335 |
-
# )
|
| 336 |
-
# attn_output = attn_output.transpose(1, 2).contiguous()
|
| 337 |
-
# attn_weights = attn_weights.to(value_states.dtype)
|
| 338 |
-
# else:
|
| 339 |
-
# attn_output, attn_weights = attention_interface(
|
| 340 |
-
# self,
|
| 341 |
-
# query_states,
|
| 342 |
-
# key_states,
|
| 343 |
-
# value_states,
|
| 344 |
-
# attention_mask,
|
| 345 |
-
# dropout=0.0 if not self.training else self.attention_dropout,
|
| 346 |
-
# scaling=self.scaling,
|
| 347 |
-
# sliding_window=self.sliding_window, # diff with Llama
|
| 348 |
-
# **kwargs,
|
| 349 |
-
# )
|
| 350 |
-
# q: (b, h, l, d); k,v: (b, h', l, d); attn_output: (b, l, h, d);
|
| 351 |
-
# key_states = repeat_kv(key_states, 2)
|
| 352 |
-
# value_states = repeat_kv(value_states, 2)
|
| 353 |
attention_mask = attention_mask.bool() if attention_mask is not None else None
|
| 354 |
-
if torch.all(attention_mask):
|
| 355 |
query_states = query_states.transpose(1, 2)
|
| 356 |
key_states = key_states.transpose(1, 2)
|
| 357 |
value_states = value_states.transpose(1, 2)
|
|
@@ -362,7 +299,7 @@ class Qwen3Attention(nn.Module):
|
|
| 362 |
causal=False,
|
| 363 |
softmax_scale=self.scaling)
|
| 364 |
|
| 365 |
-
else:
|
| 366 |
attn_output = F.scaled_dot_product_attention(
|
| 367 |
query=query_states,
|
| 368 |
key=key_states,
|
|
@@ -378,15 +315,15 @@ class Qwen3Attention(nn.Module):
|
|
| 378 |
return attn_output, None #, attn_weights
|
| 379 |
|
| 380 |
|
| 381 |
-
class
|
| 382 |
-
def __init__(self, config:
|
| 383 |
super().__init__()
|
| 384 |
self.hidden_size = config.hidden_size
|
| 385 |
-
self.self_attn =
|
| 386 |
-
self.mlp =
|
| 387 |
-
self.input_layernorm =
|
| 388 |
config.hidden_size, eps=config.rms_norm_eps)
|
| 389 |
-
self.post_attention_layernorm =
|
| 390 |
config.hidden_size, eps=config.rms_norm_eps)
|
| 391 |
if (
|
| 392 |
config.sliding_window and config._attn_implementation != "flash_attention_2"
|
|
@@ -443,11 +380,11 @@ class Qwen3DecoderLayer(GradientCheckpointingLayer):
|
|
| 443 |
|
| 444 |
|
| 445 |
@auto_docstring
|
| 446 |
-
class
|
| 447 |
-
config_class =
|
| 448 |
base_model_prefix = "model"
|
| 449 |
supports_gradient_checkpointing = True
|
| 450 |
-
_no_split_modules = ["
|
| 451 |
_skip_keys_device_placement = ["past_key_values"]
|
| 452 |
_supports_flash_attn_2 = True
|
| 453 |
_supports_sdpa = True
|
|
@@ -467,12 +404,12 @@ class Qwen3PreTrainedModel(PreTrainedModel):
|
|
| 467 |
module.weight.data.normal_(mean=0.0, std=std)
|
| 468 |
if module.padding_idx is not None:
|
| 469 |
module.weight.data[module.padding_idx].zero_()
|
| 470 |
-
elif isinstance(module,
|
| 471 |
module.weight.data.fill_(1.0)
|
| 472 |
|
| 473 |
|
| 474 |
-
class
|
| 475 |
-
def __init__(self, config:
|
| 476 |
super().__init__()
|
| 477 |
# BC: "rope_type" was originally "type"
|
| 478 |
if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
|
|
@@ -512,8 +449,8 @@ class Qwen3RotaryEmbedding(nn.Module):
|
|
| 512 |
|
| 513 |
|
| 514 |
@auto_docstring
|
| 515 |
-
class
|
| 516 |
-
def __init__(self, config:
|
| 517 |
super().__init__(config)
|
| 518 |
self.padding_idx = config.pad_token_id
|
| 519 |
self.vocab_size = config.vocab_size
|
|
@@ -521,11 +458,11 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
| 521 |
self.embed_tokens = nn.Embedding(
|
| 522 |
config.vocab_size, config.hidden_size, self.padding_idx)
|
| 523 |
self.layers = nn.ModuleList(
|
| 524 |
-
[
|
| 525 |
for layer_idx in range(config.num_hidden_layers)]
|
| 526 |
)
|
| 527 |
-
self.norm =
|
| 528 |
-
self.rotary_emb =
|
| 529 |
self.gradient_checkpointing = False
|
| 530 |
|
| 531 |
# Initialize weights and apply final processing
|
|
@@ -745,7 +682,7 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
| 745 |
dtype: torch.dtype,
|
| 746 |
cache_position: torch.Tensor,
|
| 747 |
batch_size: int,
|
| 748 |
-
config:
|
| 749 |
past_key_values: Cache,
|
| 750 |
):
|
| 751 |
"""
|
|
@@ -765,7 +702,7 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
| 765 |
Indices depicting the position of the input sequence tokens in the sequence.
|
| 766 |
batch_size (`torch.Tensor`):
|
| 767 |
Batch size.
|
| 768 |
-
config (`
|
| 769 |
The model's configuration class
|
| 770 |
past_key_values (`Cache`):
|
| 771 |
The cache class that is being used currently to generate
|
|
@@ -814,14 +751,14 @@ class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs):
|
|
| 814 |
|
| 815 |
|
| 816 |
@auto_docstring
|
| 817 |
-
class
|
| 818 |
_tied_weights_keys = ["lm_head.weight"]
|
| 819 |
_tp_plan = {"lm_head": "colwise_rep"}
|
| 820 |
_pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
|
| 821 |
|
| 822 |
def __init__(self, config):
|
| 823 |
super().__init__(config)
|
| 824 |
-
self.model =
|
| 825 |
self.vocab_size = config.vocab_size
|
| 826 |
self.lm_head = nn.Linear(
|
| 827 |
config.hidden_size, config.vocab_size, bias=False)
|
|
@@ -873,10 +810,10 @@ class Qwen3ForCausalLM(Qwen3PreTrainedModel, GenerationMixin):
|
|
| 873 |
Example:
|
| 874 |
|
| 875 |
```python
|
| 876 |
-
>>> from transformers import AutoTokenizer,
|
| 877 |
|
| 878 |
-
>>> model =
|
| 879 |
-
>>> tokenizer = AutoTokenizer.from_pretrained("
|
| 880 |
|
| 881 |
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
| 882 |
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
|
@@ -958,251 +895,8 @@ class Qwen3ForCausalLM(Qwen3PreTrainedModel, GenerationMixin):
|
|
| 958 |
)
|
| 959 |
|
| 960 |
|
| 961 |
-
@auto_docstring(
|
| 962 |
-
custom_intro="""
|
| 963 |
-
The Qwen3 Model transformer with a sequence classification head on top (linear layer).
|
| 964 |
-
|
| 965 |
-
[`Qwen3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
| 966 |
-
(e.g. GPT-2) do.
|
| 967 |
-
|
| 968 |
-
Since it does classification on the last token, it requires to know the position of the last token. If a
|
| 969 |
-
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
| 970 |
-
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
| 971 |
-
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
|
| 972 |
-
each row of the batch).
|
| 973 |
-
"""
|
| 974 |
-
)
|
| 975 |
-
class Qwen3ForSequenceClassification(Qwen3PreTrainedModel):
|
| 976 |
-
def __init__(self, config):
|
| 977 |
-
super().__init__(config)
|
| 978 |
-
self.num_labels = config.num_labels
|
| 979 |
-
self.model = Qwen3Model(config)
|
| 980 |
-
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
|
| 981 |
-
|
| 982 |
-
# Initialize weights and apply final processing
|
| 983 |
-
self.post_init()
|
| 984 |
-
|
| 985 |
-
def get_input_embeddings(self):
|
| 986 |
-
return self.model.embed_tokens
|
| 987 |
-
|
| 988 |
-
def set_input_embeddings(self, value):
|
| 989 |
-
self.model.embed_tokens = value
|
| 990 |
-
|
| 991 |
-
@can_return_tuple
|
| 992 |
-
@auto_docstring
|
| 993 |
-
def forward(
|
| 994 |
-
self,
|
| 995 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 996 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 997 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 998 |
-
past_key_values: Optional[Cache] = None,
|
| 999 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1000 |
-
labels: Optional[torch.LongTensor] = None,
|
| 1001 |
-
use_cache: Optional[bool] = None,
|
| 1002 |
-
output_attentions: Optional[bool] = None,
|
| 1003 |
-
output_hidden_states: Optional[bool] = None,
|
| 1004 |
-
) -> SequenceClassifierOutputWithPast:
|
| 1005 |
-
r"""
|
| 1006 |
-
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 1007 |
-
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
| 1008 |
-
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
| 1009 |
-
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
| 1010 |
-
"""
|
| 1011 |
-
|
| 1012 |
-
transformer_outputs: BaseModelOutputWithPast = self.model(
|
| 1013 |
-
input_ids,
|
| 1014 |
-
attention_mask=attention_mask,
|
| 1015 |
-
position_ids=position_ids,
|
| 1016 |
-
past_key_values=past_key_values,
|
| 1017 |
-
inputs_embeds=inputs_embeds,
|
| 1018 |
-
use_cache=use_cache,
|
| 1019 |
-
output_attentions=output_attentions,
|
| 1020 |
-
output_hidden_states=output_hidden_states,
|
| 1021 |
-
)
|
| 1022 |
-
hidden_states = transformer_outputs.last_hidden_state
|
| 1023 |
-
logits = self.score(hidden_states)
|
| 1024 |
-
|
| 1025 |
-
if input_ids is not None:
|
| 1026 |
-
batch_size = input_ids.shape[0]
|
| 1027 |
-
else:
|
| 1028 |
-
batch_size = inputs_embeds.shape[0]
|
| 1029 |
-
|
| 1030 |
-
if self.config.pad_token_id is None and batch_size != 1:
|
| 1031 |
-
raise ValueError(
|
| 1032 |
-
"Cannot handle batch sizes > 1 if no padding token is defined.")
|
| 1033 |
-
if self.config.pad_token_id is None:
|
| 1034 |
-
last_non_pad_token = -1
|
| 1035 |
-
elif input_ids is not None:
|
| 1036 |
-
# To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
|
| 1037 |
-
non_pad_mask = (input_ids != self.config.pad_token_id).to(
|
| 1038 |
-
logits.device, torch.int32)
|
| 1039 |
-
token_indices = torch.arange(
|
| 1040 |
-
input_ids.shape[-1], device=logits.device, dtype=torch.int32)
|
| 1041 |
-
last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
|
| 1042 |
-
else:
|
| 1043 |
-
last_non_pad_token = -1
|
| 1044 |
-
logger.warning_once(
|
| 1045 |
-
f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
|
| 1046 |
-
"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
|
| 1047 |
-
)
|
| 1048 |
-
|
| 1049 |
-
pooled_logits = logits[torch.arange(
|
| 1050 |
-
batch_size, device=logits.device), last_non_pad_token]
|
| 1051 |
-
|
| 1052 |
-
loss = None
|
| 1053 |
-
if labels is not None:
|
| 1054 |
-
loss = self.loss_function(
|
| 1055 |
-
logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config)
|
| 1056 |
-
|
| 1057 |
-
return SequenceClassifierOutputWithPast(
|
| 1058 |
-
loss=loss,
|
| 1059 |
-
logits=pooled_logits,
|
| 1060 |
-
past_key_values=transformer_outputs.past_key_values,
|
| 1061 |
-
hidden_states=transformer_outputs.hidden_states,
|
| 1062 |
-
attentions=transformer_outputs.attentions,
|
| 1063 |
-
)
|
| 1064 |
-
|
| 1065 |
-
|
| 1066 |
-
@auto_docstring
|
| 1067 |
-
class Qwen3ForTokenClassification(Qwen3PreTrainedModel):
|
| 1068 |
-
def __init__(self, config):
|
| 1069 |
-
super().__init__(config)
|
| 1070 |
-
self.num_labels = config.num_labels
|
| 1071 |
-
self.model = Qwen3Model(config)
|
| 1072 |
-
if getattr(config, "classifier_dropout", None) is not None:
|
| 1073 |
-
classifier_dropout = config.classifier_dropout
|
| 1074 |
-
elif getattr(config, "hidden_dropout", None) is not None:
|
| 1075 |
-
classifier_dropout = config.hidden_dropout
|
| 1076 |
-
else:
|
| 1077 |
-
classifier_dropout = 0.1
|
| 1078 |
-
self.dropout = nn.Dropout(classifier_dropout)
|
| 1079 |
-
self.score = nn.Linear(config.hidden_size, config.num_labels)
|
| 1080 |
-
|
| 1081 |
-
# Initialize weights and apply final processing
|
| 1082 |
-
self.post_init()
|
| 1083 |
-
|
| 1084 |
-
def get_input_embeddings(self):
|
| 1085 |
-
return self.model.embed_tokens
|
| 1086 |
-
|
| 1087 |
-
def set_input_embeddings(self, value):
|
| 1088 |
-
self.model.embed_tokens = value
|
| 1089 |
-
|
| 1090 |
-
@can_return_tuple
|
| 1091 |
-
@auto_docstring
|
| 1092 |
-
def forward(
|
| 1093 |
-
self,
|
| 1094 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 1095 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 1096 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 1097 |
-
past_key_values: Optional[Cache] = None,
|
| 1098 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1099 |
-
labels: Optional[torch.LongTensor] = None,
|
| 1100 |
-
use_cache: Optional[bool] = None,
|
| 1101 |
-
output_attentions: Optional[bool] = None,
|
| 1102 |
-
output_hidden_states: Optional[bool] = None,
|
| 1103 |
-
) -> TokenClassifierOutput:
|
| 1104 |
-
r"""
|
| 1105 |
-
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 1106 |
-
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
| 1107 |
-
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
| 1108 |
-
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
| 1109 |
-
"""
|
| 1110 |
-
|
| 1111 |
-
outputs: BaseModelOutputWithPast = self.model(
|
| 1112 |
-
input_ids,
|
| 1113 |
-
attention_mask=attention_mask,
|
| 1114 |
-
position_ids=position_ids,
|
| 1115 |
-
past_key_values=past_key_values,
|
| 1116 |
-
inputs_embeds=inputs_embeds,
|
| 1117 |
-
use_cache=use_cache,
|
| 1118 |
-
output_attentions=output_attentions,
|
| 1119 |
-
output_hidden_states=output_hidden_states,
|
| 1120 |
-
)
|
| 1121 |
-
sequence_output = outputs.last_hidden_state
|
| 1122 |
-
sequence_output = self.dropout(sequence_output)
|
| 1123 |
-
logits = self.score(sequence_output)
|
| 1124 |
-
|
| 1125 |
-
loss = None
|
| 1126 |
-
if labels is not None:
|
| 1127 |
-
loss = self.loss_function(logits, labels, self.config)
|
| 1128 |
-
|
| 1129 |
-
return TokenClassifierOutput(
|
| 1130 |
-
loss=loss,
|
| 1131 |
-
logits=logits,
|
| 1132 |
-
hidden_states=outputs.hidden_states,
|
| 1133 |
-
attentions=outputs.attentions,
|
| 1134 |
-
)
|
| 1135 |
-
|
| 1136 |
-
|
| 1137 |
-
@auto_docstring
|
| 1138 |
-
class Qwen3ForQuestionAnswering(Qwen3PreTrainedModel):
|
| 1139 |
-
base_model_prefix = "transformer"
|
| 1140 |
-
|
| 1141 |
-
def __init__(self, config):
|
| 1142 |
-
super().__init__(config)
|
| 1143 |
-
self.transformer = Qwen3Model(config)
|
| 1144 |
-
self.qa_outputs = nn.Linear(config.hidden_size, 2)
|
| 1145 |
-
|
| 1146 |
-
# Initialize weights and apply final processing
|
| 1147 |
-
self.post_init()
|
| 1148 |
-
|
| 1149 |
-
def get_input_embeddings(self):
|
| 1150 |
-
return self.transformer.embed_tokens
|
| 1151 |
-
|
| 1152 |
-
def set_input_embeddings(self, value):
|
| 1153 |
-
self.transformer.embed_tokens = value
|
| 1154 |
-
|
| 1155 |
-
@can_return_tuple
|
| 1156 |
-
@auto_docstring
|
| 1157 |
-
def forward(
|
| 1158 |
-
self,
|
| 1159 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 1160 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 1161 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 1162 |
-
past_key_values: Optional[Cache] = None,
|
| 1163 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1164 |
-
start_positions: Optional[torch.LongTensor] = None,
|
| 1165 |
-
end_positions: Optional[torch.LongTensor] = None,
|
| 1166 |
-
output_attentions: Optional[bool] = None,
|
| 1167 |
-
output_hidden_states: Optional[bool] = None,
|
| 1168 |
-
**kwargs,
|
| 1169 |
-
) -> QuestionAnsweringModelOutput:
|
| 1170 |
-
outputs: BaseModelOutputWithPast = self.transformer(
|
| 1171 |
-
input_ids,
|
| 1172 |
-
attention_mask=attention_mask,
|
| 1173 |
-
position_ids=position_ids,
|
| 1174 |
-
past_key_values=past_key_values,
|
| 1175 |
-
inputs_embeds=inputs_embeds,
|
| 1176 |
-
output_attentions=output_attentions,
|
| 1177 |
-
output_hidden_states=output_hidden_states,
|
| 1178 |
-
)
|
| 1179 |
-
|
| 1180 |
-
sequence_output = outputs.last_hidden_state
|
| 1181 |
-
|
| 1182 |
-
logits = self.qa_outputs(sequence_output)
|
| 1183 |
-
start_logits, end_logits = logits.split(1, dim=-1)
|
| 1184 |
-
start_logits = start_logits.squeeze(-1).contiguous()
|
| 1185 |
-
end_logits = end_logits.squeeze(-1).contiguous()
|
| 1186 |
-
|
| 1187 |
-
loss = None
|
| 1188 |
-
if start_positions is not None and end_positions is not None:
|
| 1189 |
-
loss = self.loss_function(
|
| 1190 |
-
start_logits, end_logits, start_positions, end_positions, **kwargs)
|
| 1191 |
-
|
| 1192 |
-
return QuestionAnsweringModelOutput(
|
| 1193 |
-
loss=loss,
|
| 1194 |
-
start_logits=start_logits,
|
| 1195 |
-
end_logits=end_logits,
|
| 1196 |
-
hidden_states=outputs.hidden_states,
|
| 1197 |
-
attentions=outputs.attentions,
|
| 1198 |
-
)
|
| 1199 |
-
|
| 1200 |
-
|
| 1201 |
__all__ = [
|
| 1202 |
-
"
|
| 1203 |
-
"
|
| 1204 |
-
"
|
| 1205 |
-
"Qwen3PreTrainedModel",
|
| 1206 |
-
"Qwen3ForSequenceClassification",
|
| 1207 |
-
"Qwen3ForTokenClassification",
|
| 1208 |
]
|
|
|
|
| 1 |
+
# This file is modified based on https://github.com/huggingface/transformers/blob/v4.52.4/src/transformers/models/qwen3/modeling_qwen3.py.
|
| 2 |
+
#
|
| 3 |
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
| 4 |
# This file was automatically generated from src/transformers/models/qwen3/modular_qwen3.py.
|
| 5 |
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
|
|
|
| 44 |
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
| 45 |
from transformers.processing_utils import Unpack
|
| 46 |
from transformers.utils import LossKwargs, auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging
|
| 47 |
+
from .configuration_sdar import SDARConfig
|
| 48 |
|
| 49 |
from fla.modules.activations import swiglu_linear
|
| 50 |
from fla.modules import (
|
|
|
|
| 83 |
|
| 84 |
|
| 85 |
@use_kernel_forward_from_hub("RMSNorm")
|
| 86 |
+
class SDARRMSNorm(nn.Module):
|
| 87 |
def __init__(self, hidden_size, eps=1e-6):
|
| 88 |
"""
|
| 89 |
+
SDARRMSNorm is equivalent to T5LayerNorm
|
| 90 |
"""
|
| 91 |
super().__init__()
|
| 92 |
self.weight = nn.Parameter(torch.ones(hidden_size))
|
|
|
|
| 109 |
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
| 110 |
|
| 111 |
|
| 112 |
+
class SDARMLP(nn.Module):
|
| 113 |
def __init__(self, config):
|
| 114 |
super().__init__()
|
| 115 |
self.config = config
|
|
|
|
| 207 |
return attn_output, attn_weights
|
| 208 |
|
| 209 |
|
| 210 |
+
class SDARAttention(nn.Module):
|
| 211 |
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
| 212 |
|
| 213 |
+
def __init__(self, config: SDARConfig, layer_idx: int):
|
| 214 |
super().__init__()
|
| 215 |
self.config = config
|
| 216 |
self.layer_idx = layer_idx
|
|
|
|
| 238 |
config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
|
| 239 |
)
|
| 240 |
# unlike olmo, only on the head dim!
|
| 241 |
+
self.q_norm = SDARRMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
| 242 |
# thus post q_norm does not need reshape
|
| 243 |
+
self.k_norm = SDARRMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
| 244 |
self.sliding_window = config.sliding_window
|
| 245 |
if not (
|
| 246 |
self.config.use_sliding_window
|
|
|
|
| 277 |
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
| 278 |
key_states, value_states = past_key_value.update(
|
| 279 |
key_states, value_states, self.layer_idx)
|
| 280 |
+
elif past_key_value is not None and not kwargs.get("store_kv", False) and len(past_key_value) > self.layer_idx:
|
| 281 |
+
# only retrive, do not store kv
|
| 282 |
past_key_states, past_value_states = past_key_value[self.layer_idx]
|
| 283 |
key_states = torch.cat(
|
| 284 |
[past_key_states, key_states], dim=-2
|
|
|
|
| 286 |
value_states = torch.cat(
|
| 287 |
[past_value_states, value_states], dim=-2
|
| 288 |
)
|
| 289 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 290 |
attention_mask = attention_mask.bool() if attention_mask is not None else None
|
| 291 |
+
if torch.all(attention_mask): # decoding
|
| 292 |
query_states = query_states.transpose(1, 2)
|
| 293 |
key_states = key_states.transpose(1, 2)
|
| 294 |
value_states = value_states.transpose(1, 2)
|
|
|
|
| 299 |
causal=False,
|
| 300 |
softmax_scale=self.scaling)
|
| 301 |
|
| 302 |
+
else: # prefilling
|
| 303 |
attn_output = F.scaled_dot_product_attention(
|
| 304 |
query=query_states,
|
| 305 |
key=key_states,
|
|
|
|
| 315 |
return attn_output, None #, attn_weights
|
| 316 |
|
| 317 |
|
| 318 |
+
class SDARDecoderLayer(GradientCheckpointingLayer):
|
| 319 |
+
def __init__(self, config: SDARConfig, layer_idx: int):
|
| 320 |
super().__init__()
|
| 321 |
self.hidden_size = config.hidden_size
|
| 322 |
+
self.self_attn = SDARAttention(config=config, layer_idx=layer_idx)
|
| 323 |
+
self.mlp = SDARMLP(config)
|
| 324 |
+
self.input_layernorm = SDARRMSNorm(
|
| 325 |
config.hidden_size, eps=config.rms_norm_eps)
|
| 326 |
+
self.post_attention_layernorm = SDARRMSNorm(
|
| 327 |
config.hidden_size, eps=config.rms_norm_eps)
|
| 328 |
if (
|
| 329 |
config.sliding_window and config._attn_implementation != "flash_attention_2"
|
|
|
|
| 380 |
|
| 381 |
|
| 382 |
@auto_docstring
|
| 383 |
+
class SDARPreTrainedModel(PreTrainedModel):
|
| 384 |
+
config_class = SDARConfig
|
| 385 |
base_model_prefix = "model"
|
| 386 |
supports_gradient_checkpointing = True
|
| 387 |
+
_no_split_modules = ["SDARDecoderLayer"]
|
| 388 |
_skip_keys_device_placement = ["past_key_values"]
|
| 389 |
_supports_flash_attn_2 = True
|
| 390 |
_supports_sdpa = True
|
|
|
|
| 404 |
module.weight.data.normal_(mean=0.0, std=std)
|
| 405 |
if module.padding_idx is not None:
|
| 406 |
module.weight.data[module.padding_idx].zero_()
|
| 407 |
+
elif isinstance(module, SDARRMSNorm):
|
| 408 |
module.weight.data.fill_(1.0)
|
| 409 |
|
| 410 |
|
| 411 |
+
class SDARRotaryEmbedding(nn.Module):
|
| 412 |
+
def __init__(self, config: SDARConfig, device=None):
|
| 413 |
super().__init__()
|
| 414 |
# BC: "rope_type" was originally "type"
|
| 415 |
if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
|
|
|
|
| 449 |
|
| 450 |
|
| 451 |
@auto_docstring
|
| 452 |
+
class SDARModel(SDARPreTrainedModel):
|
| 453 |
+
def __init__(self, config: SDARConfig):
|
| 454 |
super().__init__(config)
|
| 455 |
self.padding_idx = config.pad_token_id
|
| 456 |
self.vocab_size = config.vocab_size
|
|
|
|
| 458 |
self.embed_tokens = nn.Embedding(
|
| 459 |
config.vocab_size, config.hidden_size, self.padding_idx)
|
| 460 |
self.layers = nn.ModuleList(
|
| 461 |
+
[SDARDecoderLayer(config, layer_idx)
|
| 462 |
for layer_idx in range(config.num_hidden_layers)]
|
| 463 |
)
|
| 464 |
+
self.norm = SDARRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 465 |
+
self.rotary_emb = SDARRotaryEmbedding(config=config)
|
| 466 |
self.gradient_checkpointing = False
|
| 467 |
|
| 468 |
# Initialize weights and apply final processing
|
|
|
|
| 682 |
dtype: torch.dtype,
|
| 683 |
cache_position: torch.Tensor,
|
| 684 |
batch_size: int,
|
| 685 |
+
config: SDARConfig,
|
| 686 |
past_key_values: Cache,
|
| 687 |
):
|
| 688 |
"""
|
|
|
|
| 702 |
Indices depicting the position of the input sequence tokens in the sequence.
|
| 703 |
batch_size (`torch.Tensor`):
|
| 704 |
Batch size.
|
| 705 |
+
config (`SDARConfig`):
|
| 706 |
The model's configuration class
|
| 707 |
past_key_values (`Cache`):
|
| 708 |
The cache class that is being used currently to generate
|
|
|
|
| 751 |
|
| 752 |
|
| 753 |
@auto_docstring
|
| 754 |
+
class SDARForCausalLM(SDARPreTrainedModel, GenerationMixin):
|
| 755 |
_tied_weights_keys = ["lm_head.weight"]
|
| 756 |
_tp_plan = {"lm_head": "colwise_rep"}
|
| 757 |
_pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
|
| 758 |
|
| 759 |
def __init__(self, config):
|
| 760 |
super().__init__(config)
|
| 761 |
+
self.model = SDARModel(config)
|
| 762 |
self.vocab_size = config.vocab_size
|
| 763 |
self.lm_head = nn.Linear(
|
| 764 |
config.hidden_size, config.vocab_size, bias=False)
|
|
|
|
| 810 |
Example:
|
| 811 |
|
| 812 |
```python
|
| 813 |
+
>>> from transformers import AutoTokenizer, SDARForCausalLM
|
| 814 |
|
| 815 |
+
>>> model = SDARForCausalLM.from_pretrained("DiffuOpen/SDAR-1.7B-Chat")
|
| 816 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("DiffuOpen/SDAR-1.7B-Chat")
|
| 817 |
|
| 818 |
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
| 819 |
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
|
|
|
| 895 |
)
|
| 896 |
|
| 897 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 898 |
__all__ = [
|
| 899 |
+
"SDARForCausalLM",
|
| 900 |
+
"SDARModel",
|
| 901 |
+
"SDARPreTrainedModel",
|
|
|
|
|
|
|
|
|
|
| 902 |
]
|