Upload 16 files
Browse files- .gitattributes +1 -0
- InfiMed.py +522 -0
- added_tokens.json +30 -0
- chat_template.jinja +85 -0
- config.json +70 -0
- configuration_qwen3.py +212 -0
- configuration_siglip.py +265 -0
- merges.txt +0 -0
- model-00001-of-00002.safetensors +3 -0
- model-00002-of-00002.safetensors +3 -0
- model.safetensors.index.json +858 -0
- modeling_siglip.py +1585 -0
- preprocessor_config.json +24 -0
- special_tokens_map.json +34 -0
- tokenizer.json +3 -0
- tokenizer_config.json +253 -0
- vocab.json +0 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
InfiMed.py
ADDED
|
@@ -0,0 +1,522 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from pathlib import Path
|
| 2 |
+
from typing import Any, Callable, List, Optional, Tuple, Union
|
| 3 |
+
|
| 4 |
+
import torch
|
| 5 |
+
import torch.nn as nn
|
| 6 |
+
import os
|
| 7 |
+
from accelerate import PartialState
|
| 8 |
+
import PIL
|
| 9 |
+
from transformers import PreTrainedModel, PretrainedConfig, GenerationConfig, AutoTokenizer, LlamaTokenizerFast
|
| 10 |
+
from transformers.utils import WEIGHTS_NAME, SAFE_WEIGHTS_NAME
|
| 11 |
+
from transformers import Qwen3ForCausalLM, SiglipImageProcessor
|
| 12 |
+
from safetensors.torch import load_file
|
| 13 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
| 14 |
+
from modeling_siglip import SiglipVisionModel
|
| 15 |
+
from configuration_siglip import SiglipVisionConfig
|
| 16 |
+
from configuration_qwen3 import Qwen3Config
|
| 17 |
+
from abc import ABC, abstractmethod
|
| 18 |
+
from einops import rearrange
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
IGNORE_INDEX = -100
|
| 22 |
+
IMAGE_TOKEN_INDEX = -200
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class PromptBuilder(ABC):
|
| 26 |
+
def __init__(self, system_prompt: Optional[str] = None) -> None:
|
| 27 |
+
# Only some models define a system prompt => let subclasses handle this logic!
|
| 28 |
+
self.system_prompt = system_prompt
|
| 29 |
+
|
| 30 |
+
@abstractmethod
|
| 31 |
+
def add_turn(self, role: str, message: str) -> str: ...
|
| 32 |
+
|
| 33 |
+
@abstractmethod
|
| 34 |
+
def get_potential_prompt(self, user_msg: str) -> None: ...
|
| 35 |
+
|
| 36 |
+
@abstractmethod
|
| 37 |
+
def get_prompt(self) -> str: ...
|
| 38 |
+
|
| 39 |
+
class Qwen3PromptBuilder(PromptBuilder):
|
| 40 |
+
def __init__(self, system_prompt: Optional[str] = None) -> None:
|
| 41 |
+
super().__init__(system_prompt)
|
| 42 |
+
self.system_prompt = "<s><|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
|
| 43 |
+
self.bos, self.eos = "<s>", "<|im_end|>"
|
| 44 |
+
|
| 45 |
+
# Get role-specific "wrap" functions
|
| 46 |
+
self.wrap_human = lambda msg: f"<|im_start|>user\n{msg}<|im_end|>assistant\n"
|
| 47 |
+
self.wrap_gpt = lambda msg: f"{msg if msg != '' else ' '}{self.eos}\n"
|
| 48 |
+
|
| 49 |
+
# === `self.prompt` gets built up over multiple turns ===
|
| 50 |
+
self.prompt, self.turn_count = "", 0
|
| 51 |
+
|
| 52 |
+
def add_turn(self, role: str, message: str) -> str:
|
| 53 |
+
# assert (role == "human") if (self.turn_count % 2 == 0) else (role == "gpt")
|
| 54 |
+
message = message.strip() #.replace("<image>", "").strip()
|
| 55 |
+
|
| 56 |
+
# Special Handling for "system" prompt (turn_count == 0)
|
| 57 |
+
if self.turn_count == 0:
|
| 58 |
+
sys_message = self.system_prompt + self.wrap_human(message)
|
| 59 |
+
wrapped_message = sys_message
|
| 60 |
+
elif (self.turn_count % 2) == 0:
|
| 61 |
+
human_message = self.wrap_human(message)
|
| 62 |
+
wrapped_message = human_message
|
| 63 |
+
else:
|
| 64 |
+
gpt_message = self.wrap_gpt(message)
|
| 65 |
+
wrapped_message = gpt_message
|
| 66 |
+
|
| 67 |
+
# Update Prompt
|
| 68 |
+
self.prompt += wrapped_message
|
| 69 |
+
|
| 70 |
+
|
| 71 |
+
# Bump Turn Counter
|
| 72 |
+
self.turn_count += 1
|
| 73 |
+
|
| 74 |
+
# Return "wrapped_message" (effective string added to context)
|
| 75 |
+
return wrapped_message
|
| 76 |
+
|
| 77 |
+
def get_potential_prompt(self, message: str) -> None:
|
| 78 |
+
# Assumes that it's always the user's (human's) turn!
|
| 79 |
+
prompt_copy = str(self.prompt)
|
| 80 |
+
|
| 81 |
+
# Special Handling for "system" prompt (turn_count == 0)
|
| 82 |
+
if self.turn_count == 0:
|
| 83 |
+
sys_message = self.system_prompt + self.wrap_human(message)
|
| 84 |
+
prompt_copy += sys_message
|
| 85 |
+
|
| 86 |
+
else:
|
| 87 |
+
human_message = self.wrap_human(message)
|
| 88 |
+
prompt_copy += human_message
|
| 89 |
+
|
| 90 |
+
# return prompt_copy.removeprefix(self.bos).rstrip()
|
| 91 |
+
return prompt_copy.rstrip()
|
| 92 |
+
|
| 93 |
+
def get_prompt(self) -> str:
|
| 94 |
+
# Remove prefix <bos> (if exists) because it gets auto-inserted by tokenizer!
|
| 95 |
+
# return self.prompt.removeprefix(self.bos).rstrip()
|
| 96 |
+
return self.prompt.rstrip()
|
| 97 |
+
|
| 98 |
+
class InfiMedConfig(PretrainedConfig):
|
| 99 |
+
|
| 100 |
+
def __init__(
|
| 101 |
+
self,
|
| 102 |
+
vision_config=None,
|
| 103 |
+
llm_config=None,
|
| 104 |
+
run_dir: str = None,
|
| 105 |
+
load_precision: str = "bf16",
|
| 106 |
+
max_length: int = 128,
|
| 107 |
+
temperature: float = 1.0,
|
| 108 |
+
**kwargs
|
| 109 |
+
):
|
| 110 |
+
if vision_config is None:
|
| 111 |
+
vision_config = {}
|
| 112 |
+
print(
|
| 113 |
+
'vision_config is None. Initializing the SiglipVisionConfig with default values.')
|
| 114 |
+
|
| 115 |
+
if llm_config is None:
|
| 116 |
+
llm_config = {'architectures': ['Qwen3ForCausalLM']}
|
| 117 |
+
print(
|
| 118 |
+
'llm_config is None. Initializing the Qwen3Config config with default values')
|
| 119 |
+
|
| 120 |
+
self.vision_config = SiglipVisionConfig(**vision_config)
|
| 121 |
+
if llm_config['architectures'][0] == 'Qwen3ForCausalLM':
|
| 122 |
+
self.llm_config = Qwen3Config(**llm_config)
|
| 123 |
+
else:
|
| 124 |
+
raise ValueError('Unsupported architecture: {}'.format(
|
| 125 |
+
llm_config['architectures'][0]))
|
| 126 |
+
self.run_dir = run_dir
|
| 127 |
+
self.load_precision = load_precision
|
| 128 |
+
self.max_length = max_length
|
| 129 |
+
self.temperature = temperature
|
| 130 |
+
super().__init__(**kwargs)
|
| 131 |
+
|
| 132 |
+
class AvgPoolProjector(nn.Module):
|
| 133 |
+
def __init__(
|
| 134 |
+
self,
|
| 135 |
+
layer_num: int = 2,
|
| 136 |
+
query_num: int = 144,
|
| 137 |
+
mm_hidden_size: int = 1024,
|
| 138 |
+
llm_hidden_size: int = 4096,
|
| 139 |
+
):
|
| 140 |
+
super().__init__()
|
| 141 |
+
self.layer_num = layer_num
|
| 142 |
+
self.query_num = query_num
|
| 143 |
+
self.mm_hidden_size = mm_hidden_size
|
| 144 |
+
self.llm_hidden_size = llm_hidden_size
|
| 145 |
+
self.build_net()
|
| 146 |
+
|
| 147 |
+
def build_net(self):
|
| 148 |
+
hw = int(self.query_num ** 0.5)
|
| 149 |
+
sampler = nn.AdaptiveAvgPool2d((hw, hw))
|
| 150 |
+
self.sampler = sampler
|
| 151 |
+
modules = [nn.Linear(self.mm_hidden_size, self.llm_hidden_size)]
|
| 152 |
+
for _ in range(1, self.layer_num):
|
| 153 |
+
modules.append(nn.GELU())
|
| 154 |
+
modules.append(nn.Linear(self.llm_hidden_size, self.llm_hidden_size))
|
| 155 |
+
self.mlp_projector = nn.Sequential(*modules)
|
| 156 |
+
print(f"patch size {hw} average pooling layer initialized")
|
| 157 |
+
|
| 158 |
+
def forward(self, visual_feat: torch.Tensor) -> torch.Tensor:
|
| 159 |
+
batch_size, seq_len, h_dim = visual_feat.shape
|
| 160 |
+
hw = int(seq_len ** 0.5)
|
| 161 |
+
shaped_visual_feat = rearrange(visual_feat, "b (h w) d -> b d h w", h=hw, w=hw)
|
| 162 |
+
pooled_visual_feat = self.sampler(shaped_visual_feat)
|
| 163 |
+
reshaped_visual_feat = rearrange(pooled_visual_feat, "b d h w -> b (h w) d")
|
| 164 |
+
output_feat = self.mlp_projector(reshaped_visual_feat)
|
| 165 |
+
return output_feat
|
| 166 |
+
|
| 167 |
+
class InfiMed(PreTrainedModel):
|
| 168 |
+
config_class = InfiMedConfig
|
| 169 |
+
|
| 170 |
+
def __init__(self, config: InfiMedConfig, vision_model=None, language_model=None):
|
| 171 |
+
super().__init__(config)
|
| 172 |
+
self.run_dir = Path(config.run_dir) if config.run_dir else None
|
| 173 |
+
self.model_dtype = {"fp32": torch.float32, "fp16": torch.float16, "bf16": torch.bfloat16}[config.load_precision]
|
| 174 |
+
self.distributed_state = PartialState()
|
| 175 |
+
self.max_new_tokens = config.max_length
|
| 176 |
+
self.temperature = config.temperature
|
| 177 |
+
self.top_p = config.top_p
|
| 178 |
+
self.repetition_penalty = config.repetition_penalty
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
if vision_model is not None:
|
| 182 |
+
self.vision_model = vision_model
|
| 183 |
+
else:
|
| 184 |
+
self.vision_model = SiglipVisionModel.from_pretrained(config.vision_config._name_or_path, hidden_act = "gelu")
|
| 185 |
+
|
| 186 |
+
if language_model is not None:
|
| 187 |
+
self.language_model = language_model
|
| 188 |
+
self.config.llm_config = language_model.config
|
| 189 |
+
else:
|
| 190 |
+
if config.llm_config.architectures[0] == 'Qwen3ForCausalLM':
|
| 191 |
+
self.language_model = Qwen3ForCausalLM.from_pretrained(config.llm_config._name_or_path, pad_token_id = 151670, bos_token_id = 128245, eos_token_id = 151645, tie_word_embeddings = False)
|
| 192 |
+
else:
|
| 193 |
+
raise NotImplementedError(
|
| 194 |
+
f'{config.llm_config.architectures[0]} is not implemented.')
|
| 195 |
+
|
| 196 |
+
self.tokenizer = AutoTokenizer.from_pretrained(config.llm_config._name_or_path, use_fast=True)
|
| 197 |
+
self.tokenizer.add_special_tokens({"additional_special_tokens": ["<|endofchunk|>", "<s>", "<|pad|>"]})
|
| 198 |
+
self.tokenizer.pad_token = "<|pad|>"
|
| 199 |
+
self.tokenizer.bos_token = "<s>"
|
| 200 |
+
|
| 201 |
+
self.offset = 1 if self.tokenizer.encode("\n")[0] == self.tokenizer.bos_token_id else 0
|
| 202 |
+
|
| 203 |
+
if "finetune" in config.run_dir:
|
| 204 |
+
self.arch_specifier = "full-align+729-avgpool"
|
| 205 |
+
else:
|
| 206 |
+
self.arch_specifier = "no-align+avgpool"
|
| 207 |
+
|
| 208 |
+
if self.arch_specifier.split("+")[-1].split("-")[0] != "avgpool":
|
| 209 |
+
query_dim = int(self.arch_specifier.split("+")[-1].split("-")[0])
|
| 210 |
+
else:
|
| 211 |
+
query_dim = 144
|
| 212 |
+
self.projector = AvgPoolProjector(query_num=query_dim, mm_hidden_size=config.vision_config.hidden_size, llm_hidden_size=config.llm_config.hidden_size)
|
| 213 |
+
|
| 214 |
+
self.vision_backbone_requires_grad = False
|
| 215 |
+
|
| 216 |
+
self.img_context_token_id = 151655
|
| 217 |
+
|
| 218 |
+
self.image_processor = SiglipImageProcessor.from_pretrained(
|
| 219 |
+
config.vision_config._name_or_path,
|
| 220 |
+
size={"height": 384, "width": 384},
|
| 221 |
+
resample=PIL.Image.Resampling.BICUBIC,
|
| 222 |
+
crop_size={"height": 384, "width": 384},
|
| 223 |
+
do_center_crop=True,
|
| 224 |
+
do_normalize=True,
|
| 225 |
+
image_mean=[0.5, 0.5, 0.5],
|
| 226 |
+
image_std=[0.5, 0.5, 0.5],
|
| 227 |
+
do_convert_rgb=True
|
| 228 |
+
)
|
| 229 |
+
|
| 230 |
+
|
| 231 |
+
@classmethod
|
| 232 |
+
# load model from .pt file
|
| 233 |
+
def from_pretrained_ckpt(cls, pretrained_model_name_or_path, *args, **kwargs):
|
| 234 |
+
config = InfiMedConfig.from_pretrained(pretrained_model_name_or_path, *args, **kwargs)
|
| 235 |
+
model = cls(config)
|
| 236 |
+
ckpt_base_path = os.path.join(os.path.dirname(pretrained_model_name_or_path), "checkpoints")
|
| 237 |
+
if (Path(ckpt_base_path) / SAFE_WEIGHTS_NAME).exists():
|
| 238 |
+
state_dict = load_file(Path(ckpt_base_path) / SAFE_WEIGHTS_NAME)
|
| 239 |
+
elif (Path(ckpt_base_path) / WEIGHTS_NAME).exists():
|
| 240 |
+
state_dict = torch.load(Path(ckpt_base_path) / WEIGHTS_NAME, map_location="cpu")["model"]
|
| 241 |
+
elif (Path(ckpt_base_path) / "latest-checkpoint.pt").exists():
|
| 242 |
+
state_dict = torch.load(Path(ckpt_base_path) / "latest-checkpoint.pt", map_location="cpu")["model"]
|
| 243 |
+
else:
|
| 244 |
+
raise FileNotFoundError("No model weights found in the directory.")
|
| 245 |
+
if "vision_backbone" in state_dict:
|
| 246 |
+
model.vision_model.load_state_dict(state_dict["vision_backbone"])
|
| 247 |
+
|
| 248 |
+
new_state_dict = {}
|
| 249 |
+
for key, value in state_dict["llm_backbone"].items():
|
| 250 |
+
new_key = key.replace("llm.", "")
|
| 251 |
+
new_state_dict[new_key] = value
|
| 252 |
+
model.language_model.load_state_dict(new_state_dict)
|
| 253 |
+
model.projector.load_state_dict(state_dict["projector"])
|
| 254 |
+
|
| 255 |
+
model.to("cuda", dtype=torch.bfloat16)
|
| 256 |
+
|
| 257 |
+
model.requires_grad_(False)
|
| 258 |
+
model.eval()
|
| 259 |
+
return model
|
| 260 |
+
|
| 261 |
+
def save_checkpoint(self, save_path):
|
| 262 |
+
os.makedirs(save_path, exist_ok=True)
|
| 263 |
+
self.save_pretrained(save_path)
|
| 264 |
+
self.tokenizer.save_pretrained(save_path)
|
| 265 |
+
self.image_processor.save_pretrained(save_path)
|
| 266 |
+
|
| 267 |
+
|
| 268 |
+
def process_messages(self,messages):
|
| 269 |
+
prompt_builder = Qwen3PromptBuilder()
|
| 270 |
+
if "image" in messages:
|
| 271 |
+
processed_prompt = "<image>" + "\n" + messages['prompt'].replace("<image>", '')
|
| 272 |
+
elif "images" in messages:
|
| 273 |
+
processed_prompt = ""
|
| 274 |
+
for i, image in enumerate(messages['images']):
|
| 275 |
+
processed_prompt += f"<image_{i+1}>: "
|
| 276 |
+
processed_prompt += "\n" + messages['prompt'].replace("<image>", '')
|
| 277 |
+
|
| 278 |
+
msg = prompt_builder.add_turn("user", processed_prompt)
|
| 279 |
+
msg = msg.strip()
|
| 280 |
+
|
| 281 |
+
if isinstance(self.tokenizer, LlamaTokenizerFast):
|
| 282 |
+
msg = msg.rstrip()
|
| 283 |
+
else:
|
| 284 |
+
pass
|
| 285 |
+
|
| 286 |
+
turn_input_ids, _ = tokenizer_image_token(msg, self.tokenizer)
|
| 287 |
+
|
| 288 |
+
result = []
|
| 289 |
+
for x in turn_input_ids:
|
| 290 |
+
if x == -200:
|
| 291 |
+
result.extend([self.img_context_token_id] * 729)
|
| 292 |
+
else:
|
| 293 |
+
result.append(x)
|
| 294 |
+
|
| 295 |
+
turn_input_ids = result
|
| 296 |
+
|
| 297 |
+
input_ids = torch.tensor(turn_input_ids)
|
| 298 |
+
|
| 299 |
+
input_ids = input_ids[: self.tokenizer.model_max_length]
|
| 300 |
+
|
| 301 |
+
input_ids = input_ids.unsqueeze(0)
|
| 302 |
+
|
| 303 |
+
if "image" in messages:
|
| 304 |
+
pixel_values = self.image_processor(images=messages["image"], return_tensors="pt")["pixel_values"]
|
| 305 |
+
else:
|
| 306 |
+
pixel_values = None
|
| 307 |
+
|
| 308 |
+
input_ids = input_ids.to("cuda")
|
| 309 |
+
pixel_values = pixel_values.to("cuda") if pixel_values is not None else None
|
| 310 |
+
|
| 311 |
+
return input_ids, pixel_values
|
| 312 |
+
|
| 313 |
+
def forward(
|
| 314 |
+
self,
|
| 315 |
+
pixel_values: torch.FloatTensor,
|
| 316 |
+
input_ids: torch.LongTensor = None,
|
| 317 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 318 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 319 |
+
image_flags: Optional[torch.LongTensor] = None,
|
| 320 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 321 |
+
labels: Optional[torch.LongTensor] = None,
|
| 322 |
+
use_cache: Optional[bool] = None,
|
| 323 |
+
output_attentions: Optional[bool] = None,
|
| 324 |
+
output_hidden_states: Optional[bool] = None,
|
| 325 |
+
return_dict: Optional[bool] = None,
|
| 326 |
+
) -> CausalLMOutputWithPast:
|
| 327 |
+
|
| 328 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 329 |
+
|
| 330 |
+
vit_embeds = self.extract_feature(pixel_values)
|
| 331 |
+
|
| 332 |
+
input_embeds = self.language_model.get_input_embeddings()(input_ids)
|
| 333 |
+
|
| 334 |
+
vit_batch_size = pixel_values.shape[0]
|
| 335 |
+
|
| 336 |
+
B, N, C = input_embeds.shape
|
| 337 |
+
input_embeds = input_embeds.reshape(B * N, C)
|
| 338 |
+
|
| 339 |
+
|
| 340 |
+
input_ids = input_ids.reshape(B * N)
|
| 341 |
+
selected = (input_ids == self.img_context_token_id)
|
| 342 |
+
try:
|
| 343 |
+
input_embeds[selected] = input_embeds[selected] * \
|
| 344 |
+
0.0 + vit_embeds.reshape(-1, C)
|
| 345 |
+
except Exception as e:
|
| 346 |
+
vit_embeds = vit_embeds.reshape(-1, C)
|
| 347 |
+
print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, '
|
| 348 |
+
f'vit_embeds.shape={vit_embeds.shape}')
|
| 349 |
+
n_token = selected.sum()
|
| 350 |
+
input_embeds[selected] = input_embeds[selected] * \
|
| 351 |
+
0.0 + vit_embeds[:n_token]
|
| 352 |
+
|
| 353 |
+
input_embeds = input_embeds.reshape(B, N, C)
|
| 354 |
+
|
| 355 |
+
if attention_mask is None:
|
| 356 |
+
batch_size = input_embeds.shape[0]
|
| 357 |
+
max_len = input_embeds.shape[1]
|
| 358 |
+
attention_mask = torch.zeros((batch_size, max_len), device=input_embeds.device).bool()
|
| 359 |
+
for index in range(batch_size):
|
| 360 |
+
if getattr(self.tokenizer, 'tokenizer_padding_side', 'right') == 'left':
|
| 361 |
+
attention_mask[index, -max_len:] = True
|
| 362 |
+
else:
|
| 363 |
+
attention_mask[index, :max_len] = True
|
| 364 |
+
|
| 365 |
+
|
| 366 |
+
outputs = self.language_model(
|
| 367 |
+
inputs_embeds=input_embeds,
|
| 368 |
+
attention_mask=attention_mask,
|
| 369 |
+
position_ids=position_ids,
|
| 370 |
+
past_key_values=past_key_values,
|
| 371 |
+
use_cache=use_cache,
|
| 372 |
+
output_attentions=output_attentions,
|
| 373 |
+
output_hidden_states=output_hidden_states,
|
| 374 |
+
return_dict=return_dict,
|
| 375 |
+
)
|
| 376 |
+
logits = outputs.logits
|
| 377 |
+
|
| 378 |
+
loss = None
|
| 379 |
+
if labels is not None:
|
| 380 |
+
# Shift so that tokens < n predict n
|
| 381 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
| 382 |
+
shift_labels = labels[..., 1:].contiguous()
|
| 383 |
+
# Flatten the tokens
|
| 384 |
+
loss_fct = CrossEntropyLoss()
|
| 385 |
+
shift_logits = shift_logits.view(-1,
|
| 386 |
+
self.language_model.config.vocab_size)
|
| 387 |
+
shift_labels = shift_labels.view(-1)
|
| 388 |
+
# Enable model parallelism
|
| 389 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
| 390 |
+
loss = loss_fct(shift_logits, shift_labels)
|
| 391 |
+
|
| 392 |
+
if not return_dict:
|
| 393 |
+
output = (logits,) + outputs[1:]
|
| 394 |
+
return (loss,) + output if loss is not None else output
|
| 395 |
+
|
| 396 |
+
return CausalLMOutputWithPast(
|
| 397 |
+
loss=loss,
|
| 398 |
+
logits=logits,
|
| 399 |
+
past_key_values=outputs.past_key_values,
|
| 400 |
+
hidden_states=outputs.hidden_states,
|
| 401 |
+
attentions=outputs.attentions,
|
| 402 |
+
)
|
| 403 |
+
|
| 404 |
+
|
| 405 |
+
|
| 406 |
+
def extract_feature(self, pixel_values):
|
| 407 |
+
vit_embeds = self.vision_model(
|
| 408 |
+
pixel_values=pixel_values,
|
| 409 |
+
output_hidden_states=True,
|
| 410 |
+
return_dict=True).hidden_states[-2]
|
| 411 |
+
|
| 412 |
+
h = w = int(vit_embeds.shape[1] ** 0.5)
|
| 413 |
+
vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
|
| 414 |
+
vit_embeds = vit_embeds.reshape(
|
| 415 |
+
vit_embeds.shape[0], -1, vit_embeds.shape[-1])
|
| 416 |
+
vit_embeds = self.projector(vit_embeds)
|
| 417 |
+
return vit_embeds
|
| 418 |
+
|
| 419 |
+
|
| 420 |
+
@torch.no_grad()
|
| 421 |
+
def generate(
|
| 422 |
+
self,
|
| 423 |
+
pixel_values: Optional[torch.FloatTensor] = None,
|
| 424 |
+
input_ids: Optional[torch.FloatTensor] = None,
|
| 425 |
+
attention_mask: Optional[torch.LongTensor] = None,
|
| 426 |
+
visual_features: Optional[torch.FloatTensor] = None,
|
| 427 |
+
generation_config: Optional[GenerationConfig] = None,
|
| 428 |
+
output_hidden_states: Optional[bool] = None,
|
| 429 |
+
return_dict: Optional[bool] = None,
|
| 430 |
+
**generate_kwargs,
|
| 431 |
+
) -> torch.LongTensor:
|
| 432 |
+
|
| 433 |
+
assert self.img_context_token_id is not None
|
| 434 |
+
if pixel_values is not None:
|
| 435 |
+
if visual_features is not None:
|
| 436 |
+
vit_embeds = visual_features
|
| 437 |
+
else:
|
| 438 |
+
vit_embeds = self.extract_feature(pixel_values)
|
| 439 |
+
input_embeds = self.language_model.get_input_embeddings()(input_ids)
|
| 440 |
+
B, N, C = input_embeds.shape
|
| 441 |
+
input_embeds = input_embeds.reshape(B * N, C)
|
| 442 |
+
|
| 443 |
+
input_ids = input_ids.reshape(B * N)
|
| 444 |
+
selected = (input_ids == self.img_context_token_id)
|
| 445 |
+
assert selected.sum() != 0
|
| 446 |
+
input_embeds[selected] = vit_embeds.reshape(
|
| 447 |
+
-1, C).to(input_embeds.device)
|
| 448 |
+
|
| 449 |
+
input_embeds = input_embeds.reshape(B, N, C)
|
| 450 |
+
else:
|
| 451 |
+
input_embeds = self.language_model.get_input_embeddings()(input_ids)
|
| 452 |
+
|
| 453 |
+
|
| 454 |
+
if attention_mask is None:
|
| 455 |
+
batch_size = input_embeds.shape[0]
|
| 456 |
+
max_len = input_embeds.shape[1]
|
| 457 |
+
attention_mask = torch.zeros((batch_size, max_len), device=input_embeds.device).bool()
|
| 458 |
+
for index in range(batch_size):
|
| 459 |
+
if getattr(self.tokenizer, 'tokenizer_padding_side', 'right') == 'left':
|
| 460 |
+
attention_mask[index, -max_len:] = True
|
| 461 |
+
else:
|
| 462 |
+
attention_mask[index, :max_len] = True
|
| 463 |
+
|
| 464 |
+
outputs = self.language_model.generate(
|
| 465 |
+
# input_ids=origin_input_ids,
|
| 466 |
+
inputs_embeds=input_embeds,
|
| 467 |
+
attention_mask=attention_mask,
|
| 468 |
+
generation_config=generation_config,
|
| 469 |
+
output_hidden_states=output_hidden_states,
|
| 470 |
+
# return_dict=return_dict,
|
| 471 |
+
use_cache=True,
|
| 472 |
+
**generate_kwargs,
|
| 473 |
+
)
|
| 474 |
+
|
| 475 |
+
return outputs
|
| 476 |
+
|
| 477 |
+
|
| 478 |
+
|
| 479 |
+
@torch.no_grad()
|
| 480 |
+
def generate_output(self,messages):
|
| 481 |
+
input_ids, pixel_values = self.process_messages(messages)
|
| 482 |
+
do_sample = False if self.temperature == 0 else True
|
| 483 |
+
generated_ids = self.generate(pixel_values=pixel_values, input_ids=input_ids, temperature=self.temperature,top_p=self.top_p,repetition_penalty=self.repetition_penalty,max_new_tokens=self.max_new_tokens,do_sample = do_sample)
|
| 484 |
+
generated_ids_trimmed = generated_ids
|
| 485 |
+
output_text = self.tokenizer.batch_decode(
|
| 486 |
+
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
| 487 |
+
)
|
| 488 |
+
return output_text[0]
|
| 489 |
+
|
| 490 |
+
def generate_outputs(self,messages_list):
|
| 491 |
+
res = []
|
| 492 |
+
for messages in messages_list:
|
| 493 |
+
result = self.generate_output(messages)
|
| 494 |
+
res.append(result)
|
| 495 |
+
return res
|
| 496 |
+
|
| 497 |
+
|
| 498 |
+
def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
|
| 499 |
+
prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]
|
| 500 |
+
|
| 501 |
+
def insert_separator(X, sep):
|
| 502 |
+
return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]
|
| 503 |
+
|
| 504 |
+
input_ids = []
|
| 505 |
+
labels = []
|
| 506 |
+
offset = 0
|
| 507 |
+
if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
|
| 508 |
+
offset = 1
|
| 509 |
+
input_ids.append(prompt_chunks[0][0])
|
| 510 |
+
labels.append(prompt_chunks[0][0])
|
| 511 |
+
|
| 512 |
+
for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
|
| 513 |
+
input_ids.extend(x[offset:])
|
| 514 |
+
|
| 515 |
+
for x in insert_separator(prompt_chunks, [IGNORE_INDEX] * (offset + 1)):
|
| 516 |
+
labels.extend(x[offset:])
|
| 517 |
+
|
| 518 |
+
if return_tensors is not None:
|
| 519 |
+
if return_tensors == 'pt':
|
| 520 |
+
return torch.tensor(input_ids, dtype=torch.long), torch.tensor(labels, dtype=torch.long)
|
| 521 |
+
raise ValueError(f'Unsupported tensor type: {return_tensors}')
|
| 522 |
+
return input_ids, labels
|
added_tokens.json
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"</think>": 151668,
|
| 3 |
+
"</tool_call>": 151658,
|
| 4 |
+
"</tool_response>": 151666,
|
| 5 |
+
"<think>": 151667,
|
| 6 |
+
"<tool_call>": 151657,
|
| 7 |
+
"<tool_response>": 151665,
|
| 8 |
+
"<|box_end|>": 151649,
|
| 9 |
+
"<|box_start|>": 151648,
|
| 10 |
+
"<|endofchunk|>": 151669,
|
| 11 |
+
"<|endoftext|>": 151643,
|
| 12 |
+
"<|file_sep|>": 151664,
|
| 13 |
+
"<|fim_middle|>": 151660,
|
| 14 |
+
"<|fim_pad|>": 151662,
|
| 15 |
+
"<|fim_prefix|>": 151659,
|
| 16 |
+
"<|fim_suffix|>": 151661,
|
| 17 |
+
"<|im_end|>": 151645,
|
| 18 |
+
"<|im_start|>": 151644,
|
| 19 |
+
"<|image_pad|>": 151655,
|
| 20 |
+
"<|object_ref_end|>": 151647,
|
| 21 |
+
"<|object_ref_start|>": 151646,
|
| 22 |
+
"<|pad|>": 151670,
|
| 23 |
+
"<|quad_end|>": 151651,
|
| 24 |
+
"<|quad_start|>": 151650,
|
| 25 |
+
"<|repo_name|>": 151663,
|
| 26 |
+
"<|video_pad|>": 151656,
|
| 27 |
+
"<|vision_end|>": 151653,
|
| 28 |
+
"<|vision_pad|>": 151654,
|
| 29 |
+
"<|vision_start|>": 151652
|
| 30 |
+
}
|
chat_template.jinja
ADDED
|
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{%- if tools %}
|
| 2 |
+
{{- '<|im_start|>system\n' }}
|
| 3 |
+
{%- if messages[0].role == 'system' %}
|
| 4 |
+
{{- messages[0].content + '\n\n' }}
|
| 5 |
+
{%- endif %}
|
| 6 |
+
{{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
|
| 7 |
+
{%- for tool in tools %}
|
| 8 |
+
{{- "\n" }}
|
| 9 |
+
{{- tool | tojson }}
|
| 10 |
+
{%- endfor %}
|
| 11 |
+
{{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
|
| 12 |
+
{%- else %}
|
| 13 |
+
{%- if messages[0].role == 'system' %}
|
| 14 |
+
{{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
|
| 15 |
+
{%- endif %}
|
| 16 |
+
{%- endif %}
|
| 17 |
+
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
|
| 18 |
+
{%- for message in messages[::-1] %}
|
| 19 |
+
{%- set index = (messages|length - 1) - loop.index0 %}
|
| 20 |
+
{%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
|
| 21 |
+
{%- set ns.multi_step_tool = false %}
|
| 22 |
+
{%- set ns.last_query_index = index %}
|
| 23 |
+
{%- endif %}
|
| 24 |
+
{%- endfor %}
|
| 25 |
+
{%- for message in messages %}
|
| 26 |
+
{%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
|
| 27 |
+
{{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
|
| 28 |
+
{%- elif message.role == "assistant" %}
|
| 29 |
+
{%- set content = message.content %}
|
| 30 |
+
{%- set reasoning_content = '' %}
|
| 31 |
+
{%- if message.reasoning_content is defined and message.reasoning_content is not none %}
|
| 32 |
+
{%- set reasoning_content = message.reasoning_content %}
|
| 33 |
+
{%- else %}
|
| 34 |
+
{%- if '</think>' in message.content %}
|
| 35 |
+
{%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
|
| 36 |
+
{%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
|
| 37 |
+
{%- endif %}
|
| 38 |
+
{%- endif %}
|
| 39 |
+
{%- if loop.index0 > ns.last_query_index %}
|
| 40 |
+
{%- if loop.last or (not loop.last and reasoning_content) %}
|
| 41 |
+
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
|
| 42 |
+
{%- else %}
|
| 43 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 44 |
+
{%- endif %}
|
| 45 |
+
{%- else %}
|
| 46 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 47 |
+
{%- endif %}
|
| 48 |
+
{%- if message.tool_calls %}
|
| 49 |
+
{%- for tool_call in message.tool_calls %}
|
| 50 |
+
{%- if (loop.first and content) or (not loop.first) %}
|
| 51 |
+
{{- '\n' }}
|
| 52 |
+
{%- endif %}
|
| 53 |
+
{%- if tool_call.function %}
|
| 54 |
+
{%- set tool_call = tool_call.function %}
|
| 55 |
+
{%- endif %}
|
| 56 |
+
{{- '<tool_call>\n{"name": "' }}
|
| 57 |
+
{{- tool_call.name }}
|
| 58 |
+
{{- '", "arguments": ' }}
|
| 59 |
+
{%- if tool_call.arguments is string %}
|
| 60 |
+
{{- tool_call.arguments }}
|
| 61 |
+
{%- else %}
|
| 62 |
+
{{- tool_call.arguments | tojson }}
|
| 63 |
+
{%- endif %}
|
| 64 |
+
{{- '}\n</tool_call>' }}
|
| 65 |
+
{%- endfor %}
|
| 66 |
+
{%- endif %}
|
| 67 |
+
{{- '<|im_end|>\n' }}
|
| 68 |
+
{%- elif message.role == "tool" %}
|
| 69 |
+
{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
|
| 70 |
+
{{- '<|im_start|>user' }}
|
| 71 |
+
{%- endif %}
|
| 72 |
+
{{- '\n<tool_response>\n' }}
|
| 73 |
+
{{- message.content }}
|
| 74 |
+
{{- '\n</tool_response>' }}
|
| 75 |
+
{%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
|
| 76 |
+
{{- '<|im_end|>\n' }}
|
| 77 |
+
{%- endif %}
|
| 78 |
+
{%- endif %}
|
| 79 |
+
{%- endfor %}
|
| 80 |
+
{%- if add_generation_prompt %}
|
| 81 |
+
{{- '<|im_start|>assistant\n' }}
|
| 82 |
+
{%- if enable_thinking is defined and enable_thinking is false %}
|
| 83 |
+
{{- '<think>\n\n</think>\n\n' }}
|
| 84 |
+
{%- endif %}
|
| 85 |
+
{%- endif %}
|
config.json
ADDED
|
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"InfiMed"
|
| 4 |
+
],
|
| 5 |
+
"llm_config": {
|
| 6 |
+
"_name_or_path": "/lustre/projects/polyullm/models/Qwen3/Qwen3-4B",
|
| 7 |
+
"architectures": [
|
| 8 |
+
"Qwen3ForCausalLM"
|
| 9 |
+
],
|
| 10 |
+
"attention_bias": false,
|
| 11 |
+
"attention_dropout": 0.0,
|
| 12 |
+
"bos_token_id": 151643,
|
| 13 |
+
"eos_token_id": 151645,
|
| 14 |
+
"head_dim": 128,
|
| 15 |
+
"hidden_act": "silu",
|
| 16 |
+
"hidden_size": 2560,
|
| 17 |
+
"initializer_range": 0.02,
|
| 18 |
+
"intermediate_size": 9728,
|
| 19 |
+
"max_position_embeddings": 40960,
|
| 20 |
+
"max_window_layers": 36,
|
| 21 |
+
"model_type": "qwen3",
|
| 22 |
+
"num_attention_heads": 32,
|
| 23 |
+
"num_hidden_layers": 36,
|
| 24 |
+
"num_key_value_heads": 8,
|
| 25 |
+
"rms_norm_eps": 1e-06,
|
| 26 |
+
"rope_scaling": null,
|
| 27 |
+
"rope_theta": 1000000,
|
| 28 |
+
"sliding_window": null,
|
| 29 |
+
"tie_word_embeddings": true,
|
| 30 |
+
"torch_dtype": "bfloat16",
|
| 31 |
+
"use_cache": true,
|
| 32 |
+
"use_sliding_window": false,
|
| 33 |
+
"vocab_size": 151936
|
| 34 |
+
},
|
| 35 |
+
"load_precision": "bf16",
|
| 36 |
+
"max_length": 32,
|
| 37 |
+
"repetition_penalty": 1.2,
|
| 38 |
+
"run_dir": "qwen3-4b-instruct-continue-training+stage-finetune+x7",
|
| 39 |
+
"seed": 7,
|
| 40 |
+
"stage": "finetune",
|
| 41 |
+
"temperature": 0.0,
|
| 42 |
+
"top_p": 0.0001,
|
| 43 |
+
"torch_dtype": "bfloat16",
|
| 44 |
+
"trackers": [
|
| 45 |
+
"jsonl",
|
| 46 |
+
"wandb"
|
| 47 |
+
],
|
| 48 |
+
"transformers_version": "4.52.4",
|
| 49 |
+
"vision_config": {
|
| 50 |
+
"_name_or_path": "/home/projects/polyullm/guanghao/train_code/siglip-so400m-patch14-384",
|
| 51 |
+
"architectures": [
|
| 52 |
+
"SiglipModel"
|
| 53 |
+
],
|
| 54 |
+
"attention_dropout": 0.0,
|
| 55 |
+
"hidden_act": "gelu_pytorch_tanh",
|
| 56 |
+
"hidden_size": 1152,
|
| 57 |
+
"image_size": 384,
|
| 58 |
+
"initializer_factor": 1.0,
|
| 59 |
+
"intermediate_size": 4304,
|
| 60 |
+
"layer_norm_eps": 1e-06,
|
| 61 |
+
"model_type": "siglip_vision_model",
|
| 62 |
+
"num_attention_heads": 16,
|
| 63 |
+
"num_channels": 3,
|
| 64 |
+
"num_hidden_layers": 27,
|
| 65 |
+
"patch_size": 14,
|
| 66 |
+
"torch_dtype": "bfloat16"
|
| 67 |
+
},
|
| 68 |
+
"wandb_entity": null,
|
| 69 |
+
"wandb_project": "mmpretrain"
|
| 70 |
+
}
|
configuration_qwen3.py
ADDED
|
@@ -0,0 +1,212 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 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 |
+
"""Qwen3 model configuration"""
|
| 16 |
+
|
| 17 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 18 |
+
from transformers.modeling_rope_utils import rope_config_validation
|
| 19 |
+
from transformers.utils import logging
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
logger = logging.get_logger(__name__)
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class Qwen3Config(PretrainedConfig):
|
| 26 |
+
r"""
|
| 27 |
+
This is the configuration class to store the configuration of a [`Qwen3Model`]. It is used to instantiate a
|
| 28 |
+
Qwen3 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 |
+
Qwen3-8B [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B).
|
| 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.
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
Args:
|
| 37 |
+
vocab_size (`int`, *optional*, defaults to 151936):
|
| 38 |
+
Vocabulary size of the Qwen3 model. Defines the number of different tokens that can be represented by the
|
| 39 |
+
`inputs_ids` passed when calling [`Qwen3Model`]
|
| 40 |
+
hidden_size (`int`, *optional*, defaults to 4096):
|
| 41 |
+
Dimension of the hidden representations.
|
| 42 |
+
intermediate_size (`int`, *optional*, defaults to 22016):
|
| 43 |
+
Dimension of the MLP representations.
|
| 44 |
+
num_hidden_layers (`int`, *optional*, defaults to 32):
|
| 45 |
+
Number of hidden layers in the Transformer encoder.
|
| 46 |
+
num_attention_heads (`int`, *optional*, defaults to 32):
|
| 47 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
| 48 |
+
num_key_value_heads (`int`, *optional*, defaults to 32):
|
| 49 |
+
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
| 50 |
+
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
| 51 |
+
`num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
|
| 52 |
+
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
|
| 53 |
+
by meanpooling all the original heads within that group. For more details checkout [this
|
| 54 |
+
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
|
| 55 |
+
head_dim (`int`, *optional*, defaults to 128):
|
| 56 |
+
The attention head dimension.
|
| 57 |
+
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
| 58 |
+
The non-linear activation function (function or string) in the decoder.
|
| 59 |
+
max_position_embeddings (`int`, *optional*, defaults to 32768):
|
| 60 |
+
The maximum sequence length that this model might ever be used with.
|
| 61 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
| 62 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
| 63 |
+
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
|
| 64 |
+
The epsilon used by the rms normalization layers.
|
| 65 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
| 66 |
+
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
| 67 |
+
relevant if `config.is_decoder=True`.
|
| 68 |
+
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
| 69 |
+
Whether the model's input and output word embeddings should be tied.
|
| 70 |
+
rope_theta (`float`, *optional*, defaults to 10000.0):
|
| 71 |
+
The base period of the RoPE embeddings.
|
| 72 |
+
rope_scaling (`Dict`, *optional*):
|
| 73 |
+
Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
|
| 74 |
+
and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
|
| 75 |
+
accordingly.
|
| 76 |
+
Expected contents:
|
| 77 |
+
`rope_type` (`str`):
|
| 78 |
+
The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
|
| 79 |
+
'llama3'], with 'default' being the original RoPE implementation.
|
| 80 |
+
`factor` (`float`, *optional*):
|
| 81 |
+
Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
|
| 82 |
+
most scaling types, a `factor` of x will enable the model to handle sequences of length x *
|
| 83 |
+
original maximum pre-trained length.
|
| 84 |
+
`original_max_position_embeddings` (`int`, *optional*):
|
| 85 |
+
Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
|
| 86 |
+
pretraining.
|
| 87 |
+
`attention_factor` (`float`, *optional*):
|
| 88 |
+
Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
|
| 89 |
+
computation. If unspecified, it defaults to value recommended by the implementation, using the
|
| 90 |
+
`factor` field to infer the suggested value.
|
| 91 |
+
`beta_fast` (`float`, *optional*):
|
| 92 |
+
Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
|
| 93 |
+
ramp function. If unspecified, it defaults to 32.
|
| 94 |
+
`beta_slow` (`float`, *optional*):
|
| 95 |
+
Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
|
| 96 |
+
ramp function. If unspecified, it defaults to 1.
|
| 97 |
+
`short_factor` (`List[float]`, *optional*):
|
| 98 |
+
Only used with 'longrope'. The scaling factor to be applied to short contexts (<
|
| 99 |
+
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
|
| 100 |
+
size divided by the number of attention heads divided by 2
|
| 101 |
+
`long_factor` (`List[float]`, *optional*):
|
| 102 |
+
Only used with 'longrope'. The scaling factor to be applied to long contexts (<
|
| 103 |
+
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
|
| 104 |
+
size divided by the number of attention heads divided by 2
|
| 105 |
+
`low_freq_factor` (`float`, *optional*):
|
| 106 |
+
Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
|
| 107 |
+
`high_freq_factor` (`float`, *optional*):
|
| 108 |
+
Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
|
| 109 |
+
attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
|
| 110 |
+
Whether to use a bias in the query, key, value and output projection layers during self-attention.
|
| 111 |
+
use_sliding_window (`bool`, *optional*, defaults to `False`):
|
| 112 |
+
Whether to use sliding window attention.
|
| 113 |
+
sliding_window (`int`, *optional*, defaults to 4096):
|
| 114 |
+
Sliding window attention (SWA) window size. If not specified, will default to `4096`.
|
| 115 |
+
max_window_layers (`int`, *optional*, defaults to 28):
|
| 116 |
+
The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
|
| 117 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 118 |
+
The dropout ratio for the attention probabilities.
|
| 119 |
+
|
| 120 |
+
```python
|
| 121 |
+
>>> from transformers import Qwen3Model, Qwen3Config
|
| 122 |
+
|
| 123 |
+
>>> # Initializing a Qwen3 style configuration
|
| 124 |
+
>>> configuration = Qwen3Config()
|
| 125 |
+
|
| 126 |
+
>>> # Initializing a model from the Qwen3-8B style configuration
|
| 127 |
+
>>> model = Qwen3Model(configuration)
|
| 128 |
+
|
| 129 |
+
>>> # Accessing the model configuration
|
| 130 |
+
>>> configuration = model.config
|
| 131 |
+
```"""
|
| 132 |
+
|
| 133 |
+
model_type = "qwen3"
|
| 134 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 135 |
+
|
| 136 |
+
# Default tensor parallel plan for base model `Qwen3`
|
| 137 |
+
base_model_tp_plan = {
|
| 138 |
+
"layers.*.self_attn.q_proj": "colwise",
|
| 139 |
+
"layers.*.self_attn.k_proj": "colwise",
|
| 140 |
+
"layers.*.self_attn.v_proj": "colwise",
|
| 141 |
+
"layers.*.self_attn.o_proj": "rowwise",
|
| 142 |
+
"layers.*.mlp.gate_proj": "colwise",
|
| 143 |
+
"layers.*.mlp.up_proj": "colwise",
|
| 144 |
+
"layers.*.mlp.down_proj": "rowwise",
|
| 145 |
+
}
|
| 146 |
+
base_model_pp_plan = {
|
| 147 |
+
"embed_tokens": (["input_ids"], ["inputs_embeds"]),
|
| 148 |
+
"layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
|
| 149 |
+
"norm": (["hidden_states"], ["hidden_states"]),
|
| 150 |
+
}
|
| 151 |
+
|
| 152 |
+
def __init__(
|
| 153 |
+
self,
|
| 154 |
+
vocab_size=151936,
|
| 155 |
+
hidden_size=4096,
|
| 156 |
+
intermediate_size=22016,
|
| 157 |
+
num_hidden_layers=32,
|
| 158 |
+
num_attention_heads=32,
|
| 159 |
+
num_key_value_heads=32,
|
| 160 |
+
head_dim=128,
|
| 161 |
+
hidden_act="silu",
|
| 162 |
+
max_position_embeddings=32768,
|
| 163 |
+
initializer_range=0.02,
|
| 164 |
+
rms_norm_eps=1e-6,
|
| 165 |
+
use_cache=True,
|
| 166 |
+
tie_word_embeddings=False,
|
| 167 |
+
rope_theta=10000.0,
|
| 168 |
+
rope_scaling=None,
|
| 169 |
+
attention_bias=False,
|
| 170 |
+
use_sliding_window=False,
|
| 171 |
+
sliding_window=4096,
|
| 172 |
+
max_window_layers=28,
|
| 173 |
+
attention_dropout=0.0,
|
| 174 |
+
**kwargs,
|
| 175 |
+
):
|
| 176 |
+
self.vocab_size = vocab_size
|
| 177 |
+
self.max_position_embeddings = max_position_embeddings
|
| 178 |
+
self.hidden_size = hidden_size
|
| 179 |
+
self.intermediate_size = intermediate_size
|
| 180 |
+
self.num_hidden_layers = num_hidden_layers
|
| 181 |
+
self.num_attention_heads = num_attention_heads
|
| 182 |
+
self.use_sliding_window = use_sliding_window
|
| 183 |
+
self.sliding_window = sliding_window # we check `use_sliding_window` in the modeling code
|
| 184 |
+
self.max_window_layers = max_window_layers
|
| 185 |
+
|
| 186 |
+
# for backward compatibility
|
| 187 |
+
if num_key_value_heads is None:
|
| 188 |
+
num_key_value_heads = num_attention_heads
|
| 189 |
+
|
| 190 |
+
self.num_key_value_heads = num_key_value_heads
|
| 191 |
+
self.head_dim = head_dim
|
| 192 |
+
self.hidden_act = hidden_act
|
| 193 |
+
self.initializer_range = initializer_range
|
| 194 |
+
self.rms_norm_eps = rms_norm_eps
|
| 195 |
+
self.use_cache = use_cache
|
| 196 |
+
self.rope_theta = rope_theta
|
| 197 |
+
self.rope_scaling = rope_scaling
|
| 198 |
+
self.attention_bias = attention_bias
|
| 199 |
+
self.attention_dropout = attention_dropout
|
| 200 |
+
# Validate the correctness of rotary position embeddings parameters
|
| 201 |
+
# BC: if there is a 'type' field, move it to 'rope_type'.
|
| 202 |
+
if self.rope_scaling is not None and "type" in self.rope_scaling:
|
| 203 |
+
self.rope_scaling["rope_type"] = self.rope_scaling["type"]
|
| 204 |
+
rope_config_validation(self)
|
| 205 |
+
|
| 206 |
+
super().__init__(
|
| 207 |
+
tie_word_embeddings=tie_word_embeddings,
|
| 208 |
+
**kwargs,
|
| 209 |
+
)
|
| 210 |
+
|
| 211 |
+
|
| 212 |
+
__all__ = ["Qwen3Config"]
|
configuration_siglip.py
ADDED
|
@@ -0,0 +1,265 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 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 |
+
"""Siglip model configuration"""
|
| 16 |
+
|
| 17 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 18 |
+
from transformers.utils import logging
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
logger = logging.get_logger(__name__)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
class SiglipTextConfig(PretrainedConfig):
|
| 25 |
+
r"""
|
| 26 |
+
This is the configuration class to store the configuration of a [`SiglipTextModel`]. It is used to instantiate a
|
| 27 |
+
Siglip text encoder according to the specified arguments, defining the model architecture. Instantiating a
|
| 28 |
+
configuration with the defaults will yield a similar configuration to that of the text encoder of the Siglip
|
| 29 |
+
[google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
|
| 30 |
+
|
| 31 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 32 |
+
documentation from [`PretrainedConfig`] for more information.
|
| 33 |
+
|
| 34 |
+
Args:
|
| 35 |
+
vocab_size (`int`, *optional*, defaults to 32000):
|
| 36 |
+
Vocabulary size of the Siglip text model. Defines the number of different tokens that can be represented by
|
| 37 |
+
the `inputs_ids` passed when calling [`SiglipModel`].
|
| 38 |
+
hidden_size (`int`, *optional*, defaults to 768):
|
| 39 |
+
Dimensionality of the encoder layers and the pooler layer.
|
| 40 |
+
intermediate_size (`int`, *optional*, defaults to 3072):
|
| 41 |
+
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
|
| 42 |
+
num_hidden_layers (`int`, *optional*, defaults to 12):
|
| 43 |
+
Number of hidden layers in the Transformer encoder.
|
| 44 |
+
num_attention_heads (`int`, *optional*, defaults to 12):
|
| 45 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
| 46 |
+
max_position_embeddings (`int`, *optional*, defaults to 64):
|
| 47 |
+
The maximum sequence length that this model might ever be used with. Typically set this to something large
|
| 48 |
+
just in case (e.g., 512 or 1024 or 2048).
|
| 49 |
+
hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
|
| 50 |
+
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
|
| 51 |
+
`"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
|
| 52 |
+
layer_norm_eps (`float`, *optional*, defaults to 1e-06):
|
| 53 |
+
The epsilon used by the layer normalization layers.
|
| 54 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 55 |
+
The dropout ratio for the attention probabilities.
|
| 56 |
+
pad_token_id (`int`, *optional*, defaults to 1):
|
| 57 |
+
The id of the padding token in the vocabulary.
|
| 58 |
+
bos_token_id (`int`, *optional*, defaults to 49406):
|
| 59 |
+
The id of the beginning-of-sequence token in the vocabulary.
|
| 60 |
+
eos_token_id (`int`, *optional*, defaults to 49407):
|
| 61 |
+
The id of the end-of-sequence token in the vocabulary.
|
| 62 |
+
|
| 63 |
+
Example:
|
| 64 |
+
|
| 65 |
+
```python
|
| 66 |
+
>>> from transformers import SiglipTextConfig, SiglipTextModel
|
| 67 |
+
|
| 68 |
+
>>> # Initializing a SiglipTextConfig with google/siglip-base-patch16-224 style configuration
|
| 69 |
+
>>> configuration = SiglipTextConfig()
|
| 70 |
+
|
| 71 |
+
>>> # Initializing a SiglipTextModel (with random weights) from the google/siglip-base-patch16-224 style configuration
|
| 72 |
+
>>> model = SiglipTextModel(configuration)
|
| 73 |
+
|
| 74 |
+
>>> # Accessing the model configuration
|
| 75 |
+
>>> configuration = model.config
|
| 76 |
+
```"""
|
| 77 |
+
|
| 78 |
+
model_type = "siglip_text_model"
|
| 79 |
+
base_config_key = "text_config"
|
| 80 |
+
|
| 81 |
+
def __init__(
|
| 82 |
+
self,
|
| 83 |
+
vocab_size=32000,
|
| 84 |
+
hidden_size=768,
|
| 85 |
+
intermediate_size=3072,
|
| 86 |
+
num_hidden_layers=12,
|
| 87 |
+
num_attention_heads=12,
|
| 88 |
+
max_position_embeddings=64,
|
| 89 |
+
hidden_act="gelu_pytorch_tanh",
|
| 90 |
+
layer_norm_eps=1e-6,
|
| 91 |
+
attention_dropout=0.0,
|
| 92 |
+
# This differs from `CLIPTokenizer`'s default and from openai/siglip
|
| 93 |
+
# See https://github.com/huggingface/transformers/pull/24773#issuecomment-1632287538
|
| 94 |
+
pad_token_id=1,
|
| 95 |
+
bos_token_id=49406,
|
| 96 |
+
eos_token_id=49407,
|
| 97 |
+
**kwargs,
|
| 98 |
+
):
|
| 99 |
+
super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
|
| 100 |
+
|
| 101 |
+
self.vocab_size = vocab_size
|
| 102 |
+
self.hidden_size = hidden_size
|
| 103 |
+
self.intermediate_size = intermediate_size
|
| 104 |
+
self.num_hidden_layers = num_hidden_layers
|
| 105 |
+
self.num_attention_heads = num_attention_heads
|
| 106 |
+
self.max_position_embeddings = max_position_embeddings
|
| 107 |
+
self.layer_norm_eps = layer_norm_eps
|
| 108 |
+
self.hidden_act = hidden_act
|
| 109 |
+
self.attention_dropout = attention_dropout
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
class SiglipVisionConfig(PretrainedConfig):
|
| 113 |
+
r"""
|
| 114 |
+
This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
|
| 115 |
+
Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a
|
| 116 |
+
configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip
|
| 117 |
+
[google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
|
| 118 |
+
|
| 119 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 120 |
+
documentation from [`PretrainedConfig`] for more information.
|
| 121 |
+
|
| 122 |
+
Args:
|
| 123 |
+
hidden_size (`int`, *optional*, defaults to 768):
|
| 124 |
+
Dimensionality of the encoder layers and the pooler layer.
|
| 125 |
+
intermediate_size (`int`, *optional*, defaults to 3072):
|
| 126 |
+
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
|
| 127 |
+
num_hidden_layers (`int`, *optional*, defaults to 12):
|
| 128 |
+
Number of hidden layers in the Transformer encoder.
|
| 129 |
+
num_attention_heads (`int`, *optional*, defaults to 12):
|
| 130 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
| 131 |
+
num_channels (`int`, *optional*, defaults to 3):
|
| 132 |
+
Number of channels in the input images.
|
| 133 |
+
image_size (`int`, *optional*, defaults to 224):
|
| 134 |
+
The size (resolution) of each image.
|
| 135 |
+
patch_size (`int`, *optional*, defaults to 16):
|
| 136 |
+
The size (resolution) of each patch.
|
| 137 |
+
hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
|
| 138 |
+
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
|
| 139 |
+
`"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
|
| 140 |
+
layer_norm_eps (`float`, *optional*, defaults to 1e-06):
|
| 141 |
+
The epsilon used by the layer normalization layers.
|
| 142 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 143 |
+
The dropout ratio for the attention probabilities.
|
| 144 |
+
|
| 145 |
+
Example:
|
| 146 |
+
|
| 147 |
+
```python
|
| 148 |
+
>>> from transformers import SiglipVisionConfig, SiglipVisionModel
|
| 149 |
+
|
| 150 |
+
>>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
|
| 151 |
+
>>> configuration = SiglipVisionConfig()
|
| 152 |
+
|
| 153 |
+
>>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
|
| 154 |
+
>>> model = SiglipVisionModel(configuration)
|
| 155 |
+
|
| 156 |
+
>>> # Accessing the model configuration
|
| 157 |
+
>>> configuration = model.config
|
| 158 |
+
```"""
|
| 159 |
+
|
| 160 |
+
model_type = "siglip_vision_model"
|
| 161 |
+
base_config_key = "vision_config"
|
| 162 |
+
|
| 163 |
+
def __init__(
|
| 164 |
+
self,
|
| 165 |
+
hidden_size=768,
|
| 166 |
+
intermediate_size=3072,
|
| 167 |
+
num_hidden_layers=12,
|
| 168 |
+
num_attention_heads=12,
|
| 169 |
+
num_channels=3,
|
| 170 |
+
image_size=224,
|
| 171 |
+
patch_size=16,
|
| 172 |
+
hidden_act="gelu_pytorch_tanh",
|
| 173 |
+
layer_norm_eps=1e-6,
|
| 174 |
+
attention_dropout=0.0,
|
| 175 |
+
**kwargs,
|
| 176 |
+
):
|
| 177 |
+
super().__init__(**kwargs)
|
| 178 |
+
|
| 179 |
+
self.hidden_size = hidden_size
|
| 180 |
+
self.intermediate_size = intermediate_size
|
| 181 |
+
self.num_hidden_layers = num_hidden_layers
|
| 182 |
+
self.num_attention_heads = num_attention_heads
|
| 183 |
+
self.num_channels = num_channels
|
| 184 |
+
self.patch_size = patch_size
|
| 185 |
+
self.image_size = image_size
|
| 186 |
+
self.attention_dropout = attention_dropout
|
| 187 |
+
self.layer_norm_eps = layer_norm_eps
|
| 188 |
+
self.hidden_act = hidden_act
|
| 189 |
+
|
| 190 |
+
|
| 191 |
+
class SiglipConfig(PretrainedConfig):
|
| 192 |
+
r"""
|
| 193 |
+
[`SiglipConfig`] is the configuration class to store the configuration of a [`SiglipModel`]. It is used to
|
| 194 |
+
instantiate a Siglip model according to the specified arguments, defining the text model and vision model configs.
|
| 195 |
+
Instantiating a configuration with the defaults will yield a similar configuration to that of the Siglip
|
| 196 |
+
[google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
|
| 197 |
+
|
| 198 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 199 |
+
documentation from [`PretrainedConfig`] for more information.
|
| 200 |
+
|
| 201 |
+
Args:
|
| 202 |
+
text_config (`dict`, *optional*):
|
| 203 |
+
Dictionary of configuration options used to initialize [`SiglipTextConfig`].
|
| 204 |
+
vision_config (`dict`, *optional*):
|
| 205 |
+
Dictionary of configuration options used to initialize [`SiglipVisionConfig`].
|
| 206 |
+
kwargs (*optional*):
|
| 207 |
+
Dictionary of keyword arguments.
|
| 208 |
+
|
| 209 |
+
Example:
|
| 210 |
+
|
| 211 |
+
```python
|
| 212 |
+
>>> from transformers import SiglipConfig, SiglipModel
|
| 213 |
+
|
| 214 |
+
>>> # Initializing a SiglipConfig with google/siglip-base-patch16-224 style configuration
|
| 215 |
+
>>> configuration = SiglipConfig()
|
| 216 |
+
|
| 217 |
+
>>> # Initializing a SiglipModel (with random weights) from the google/siglip-base-patch16-224 style configuration
|
| 218 |
+
>>> model = SiglipModel(configuration)
|
| 219 |
+
|
| 220 |
+
>>> # Accessing the model configuration
|
| 221 |
+
>>> configuration = model.config
|
| 222 |
+
|
| 223 |
+
>>> # We can also initialize a SiglipConfig from a SiglipTextConfig and a SiglipVisionConfig
|
| 224 |
+
>>> from transformers import SiglipTextConfig, SiglipVisionConfig
|
| 225 |
+
|
| 226 |
+
>>> # Initializing a SiglipText and SiglipVision configuration
|
| 227 |
+
>>> config_text = SiglipTextConfig()
|
| 228 |
+
>>> config_vision = SiglipVisionConfig()
|
| 229 |
+
|
| 230 |
+
>>> config = SiglipConfig.from_text_vision_configs(config_text, config_vision)
|
| 231 |
+
```"""
|
| 232 |
+
|
| 233 |
+
model_type = "siglip"
|
| 234 |
+
sub_configs = {"text_config": SiglipTextConfig, "vision_config": SiglipVisionConfig}
|
| 235 |
+
|
| 236 |
+
def __init__(self, text_config=None, vision_config=None, **kwargs):
|
| 237 |
+
super().__init__(**kwargs)
|
| 238 |
+
|
| 239 |
+
if text_config is None:
|
| 240 |
+
text_config = {}
|
| 241 |
+
logger.info("`text_config` is `None`. Initializing the `SiglipTextConfig` with default values.")
|
| 242 |
+
|
| 243 |
+
if vision_config is None:
|
| 244 |
+
vision_config = {}
|
| 245 |
+
logger.info("`vision_config` is `None`. initializing the `SiglipVisionConfig` with default values.")
|
| 246 |
+
|
| 247 |
+
self.text_config = SiglipTextConfig(**text_config)
|
| 248 |
+
self.vision_config = SiglipVisionConfig(**vision_config)
|
| 249 |
+
|
| 250 |
+
self.initializer_factor = 1.0
|
| 251 |
+
|
| 252 |
+
@classmethod
|
| 253 |
+
def from_text_vision_configs(cls, text_config: SiglipTextConfig, vision_config: SiglipVisionConfig, **kwargs):
|
| 254 |
+
r"""
|
| 255 |
+
Instantiate a [`SiglipConfig`] (or a derived class) from siglip text model configuration and siglip vision
|
| 256 |
+
model configuration.
|
| 257 |
+
|
| 258 |
+
Returns:
|
| 259 |
+
[`SiglipConfig`]: An instance of a configuration object
|
| 260 |
+
"""
|
| 261 |
+
|
| 262 |
+
return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
|
| 263 |
+
|
| 264 |
+
|
| 265 |
+
__all__ = ["SiglipConfig", "SiglipTextConfig", "SiglipVisionConfig"]
|
merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
model-00001-of-00002.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1031cef38e5bc62149fd5d1b68bd67d00d7d723a82180c97fd17090c70e63449
|
| 3 |
+
size 4966471968
|
model-00002-of-00002.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d3b8d072383149ce00cc8183badf0f5abf42ee7fa2e4a795dcfed5208e18b438
|
| 3 |
+
size 4731957576
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,858 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 9698315392
|
| 4 |
+
},
|
| 5 |
+
"weight_map": {
|
| 6 |
+
"language_model.lm_head.weight": "model-00002-of-00002.safetensors",
|
| 7 |
+
"language_model.model.embed_tokens.weight": "model-00001-of-00002.safetensors",
|
| 8 |
+
"language_model.model.layers.0.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 9 |
+
"language_model.model.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 10 |
+
"language_model.model.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 11 |
+
"language_model.model.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 12 |
+
"language_model.model.layers.0.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 13 |
+
"language_model.model.layers.0.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 14 |
+
"language_model.model.layers.0.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 15 |
+
"language_model.model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 16 |
+
"language_model.model.layers.0.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 17 |
+
"language_model.model.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 18 |
+
"language_model.model.layers.0.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 19 |
+
"language_model.model.layers.1.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 20 |
+
"language_model.model.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 21 |
+
"language_model.model.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 22 |
+
"language_model.model.layers.1.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 23 |
+
"language_model.model.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 24 |
+
"language_model.model.layers.1.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 25 |
+
"language_model.model.layers.1.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 26 |
+
"language_model.model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 27 |
+
"language_model.model.layers.1.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 28 |
+
"language_model.model.layers.1.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 29 |
+
"language_model.model.layers.1.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 30 |
+
"language_model.model.layers.10.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 31 |
+
"language_model.model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 32 |
+
"language_model.model.layers.10.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 33 |
+
"language_model.model.layers.10.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 34 |
+
"language_model.model.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 35 |
+
"language_model.model.layers.10.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 36 |
+
"language_model.model.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 37 |
+
"language_model.model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 38 |
+
"language_model.model.layers.10.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 39 |
+
"language_model.model.layers.10.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 40 |
+
"language_model.model.layers.10.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 41 |
+
"language_model.model.layers.11.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 42 |
+
"language_model.model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 43 |
+
"language_model.model.layers.11.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 44 |
+
"language_model.model.layers.11.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 45 |
+
"language_model.model.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 46 |
+
"language_model.model.layers.11.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 47 |
+
"language_model.model.layers.11.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 48 |
+
"language_model.model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 49 |
+
"language_model.model.layers.11.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 50 |
+
"language_model.model.layers.11.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 51 |
+
"language_model.model.layers.11.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 52 |
+
"language_model.model.layers.12.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 53 |
+
"language_model.model.layers.12.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 54 |
+
"language_model.model.layers.12.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 55 |
+
"language_model.model.layers.12.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 56 |
+
"language_model.model.layers.12.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 57 |
+
"language_model.model.layers.12.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 58 |
+
"language_model.model.layers.12.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 59 |
+
"language_model.model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 60 |
+
"language_model.model.layers.12.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 61 |
+
"language_model.model.layers.12.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 62 |
+
"language_model.model.layers.12.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 63 |
+
"language_model.model.layers.13.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 64 |
+
"language_model.model.layers.13.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 65 |
+
"language_model.model.layers.13.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 66 |
+
"language_model.model.layers.13.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 67 |
+
"language_model.model.layers.13.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 68 |
+
"language_model.model.layers.13.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 69 |
+
"language_model.model.layers.13.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 70 |
+
"language_model.model.layers.13.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 71 |
+
"language_model.model.layers.13.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 72 |
+
"language_model.model.layers.13.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 73 |
+
"language_model.model.layers.13.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 74 |
+
"language_model.model.layers.14.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 75 |
+
"language_model.model.layers.14.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 76 |
+
"language_model.model.layers.14.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 77 |
+
"language_model.model.layers.14.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 78 |
+
"language_model.model.layers.14.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 79 |
+
"language_model.model.layers.14.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 80 |
+
"language_model.model.layers.14.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 81 |
+
"language_model.model.layers.14.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 82 |
+
"language_model.model.layers.14.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 83 |
+
"language_model.model.layers.14.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 84 |
+
"language_model.model.layers.14.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 85 |
+
"language_model.model.layers.15.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 86 |
+
"language_model.model.layers.15.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 87 |
+
"language_model.model.layers.15.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 88 |
+
"language_model.model.layers.15.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 89 |
+
"language_model.model.layers.15.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 90 |
+
"language_model.model.layers.15.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 91 |
+
"language_model.model.layers.15.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 92 |
+
"language_model.model.layers.15.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 93 |
+
"language_model.model.layers.15.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 94 |
+
"language_model.model.layers.15.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 95 |
+
"language_model.model.layers.15.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 96 |
+
"language_model.model.layers.16.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 97 |
+
"language_model.model.layers.16.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 98 |
+
"language_model.model.layers.16.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 99 |
+
"language_model.model.layers.16.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 100 |
+
"language_model.model.layers.16.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 101 |
+
"language_model.model.layers.16.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 102 |
+
"language_model.model.layers.16.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 103 |
+
"language_model.model.layers.16.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 104 |
+
"language_model.model.layers.16.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 105 |
+
"language_model.model.layers.16.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 106 |
+
"language_model.model.layers.16.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 107 |
+
"language_model.model.layers.17.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 108 |
+
"language_model.model.layers.17.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 109 |
+
"language_model.model.layers.17.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 110 |
+
"language_model.model.layers.17.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 111 |
+
"language_model.model.layers.17.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 112 |
+
"language_model.model.layers.17.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 113 |
+
"language_model.model.layers.17.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 114 |
+
"language_model.model.layers.17.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 115 |
+
"language_model.model.layers.17.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 116 |
+
"language_model.model.layers.17.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 117 |
+
"language_model.model.layers.17.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 118 |
+
"language_model.model.layers.18.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 119 |
+
"language_model.model.layers.18.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 120 |
+
"language_model.model.layers.18.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 121 |
+
"language_model.model.layers.18.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 122 |
+
"language_model.model.layers.18.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 123 |
+
"language_model.model.layers.18.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 124 |
+
"language_model.model.layers.18.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 125 |
+
"language_model.model.layers.18.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 126 |
+
"language_model.model.layers.18.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 127 |
+
"language_model.model.layers.18.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 128 |
+
"language_model.model.layers.18.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 129 |
+
"language_model.model.layers.19.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 130 |
+
"language_model.model.layers.19.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 131 |
+
"language_model.model.layers.19.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 132 |
+
"language_model.model.layers.19.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 133 |
+
"language_model.model.layers.19.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 134 |
+
"language_model.model.layers.19.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 135 |
+
"language_model.model.layers.19.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 136 |
+
"language_model.model.layers.19.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 137 |
+
"language_model.model.layers.19.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 138 |
+
"language_model.model.layers.19.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 139 |
+
"language_model.model.layers.19.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 140 |
+
"language_model.model.layers.2.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 141 |
+
"language_model.model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 142 |
+
"language_model.model.layers.2.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 143 |
+
"language_model.model.layers.2.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 144 |
+
"language_model.model.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 145 |
+
"language_model.model.layers.2.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 146 |
+
"language_model.model.layers.2.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 147 |
+
"language_model.model.layers.2.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 148 |
+
"language_model.model.layers.2.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 149 |
+
"language_model.model.layers.2.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 150 |
+
"language_model.model.layers.2.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 151 |
+
"language_model.model.layers.20.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 152 |
+
"language_model.model.layers.20.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 153 |
+
"language_model.model.layers.20.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 154 |
+
"language_model.model.layers.20.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 155 |
+
"language_model.model.layers.20.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 156 |
+
"language_model.model.layers.20.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 157 |
+
"language_model.model.layers.20.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 158 |
+
"language_model.model.layers.20.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 159 |
+
"language_model.model.layers.20.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 160 |
+
"language_model.model.layers.20.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 161 |
+
"language_model.model.layers.20.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 162 |
+
"language_model.model.layers.21.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 163 |
+
"language_model.model.layers.21.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 164 |
+
"language_model.model.layers.21.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 165 |
+
"language_model.model.layers.21.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 166 |
+
"language_model.model.layers.21.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 167 |
+
"language_model.model.layers.21.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 168 |
+
"language_model.model.layers.21.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 169 |
+
"language_model.model.layers.21.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 170 |
+
"language_model.model.layers.21.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 171 |
+
"language_model.model.layers.21.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 172 |
+
"language_model.model.layers.21.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 173 |
+
"language_model.model.layers.22.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 174 |
+
"language_model.model.layers.22.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 175 |
+
"language_model.model.layers.22.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 176 |
+
"language_model.model.layers.22.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 177 |
+
"language_model.model.layers.22.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 178 |
+
"language_model.model.layers.22.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 179 |
+
"language_model.model.layers.22.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 180 |
+
"language_model.model.layers.22.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 181 |
+
"language_model.model.layers.22.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 182 |
+
"language_model.model.layers.22.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 183 |
+
"language_model.model.layers.22.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 184 |
+
"language_model.model.layers.23.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 185 |
+
"language_model.model.layers.23.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 186 |
+
"language_model.model.layers.23.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 187 |
+
"language_model.model.layers.23.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 188 |
+
"language_model.model.layers.23.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 189 |
+
"language_model.model.layers.23.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 190 |
+
"language_model.model.layers.23.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 191 |
+
"language_model.model.layers.23.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 192 |
+
"language_model.model.layers.23.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 193 |
+
"language_model.model.layers.23.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 194 |
+
"language_model.model.layers.23.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 195 |
+
"language_model.model.layers.24.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 196 |
+
"language_model.model.layers.24.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 197 |
+
"language_model.model.layers.24.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 198 |
+
"language_model.model.layers.24.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 199 |
+
"language_model.model.layers.24.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 200 |
+
"language_model.model.layers.24.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 201 |
+
"language_model.model.layers.24.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 202 |
+
"language_model.model.layers.24.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 203 |
+
"language_model.model.layers.24.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 204 |
+
"language_model.model.layers.24.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 205 |
+
"language_model.model.layers.24.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 206 |
+
"language_model.model.layers.25.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 207 |
+
"language_model.model.layers.25.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 208 |
+
"language_model.model.layers.25.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 209 |
+
"language_model.model.layers.25.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 210 |
+
"language_model.model.layers.25.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 211 |
+
"language_model.model.layers.25.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 212 |
+
"language_model.model.layers.25.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 213 |
+
"language_model.model.layers.25.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 214 |
+
"language_model.model.layers.25.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 215 |
+
"language_model.model.layers.25.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 216 |
+
"language_model.model.layers.25.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 217 |
+
"language_model.model.layers.26.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 218 |
+
"language_model.model.layers.26.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 219 |
+
"language_model.model.layers.26.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 220 |
+
"language_model.model.layers.26.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 221 |
+
"language_model.model.layers.26.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 222 |
+
"language_model.model.layers.26.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 223 |
+
"language_model.model.layers.26.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 224 |
+
"language_model.model.layers.26.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 225 |
+
"language_model.model.layers.26.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 226 |
+
"language_model.model.layers.26.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 227 |
+
"language_model.model.layers.26.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 228 |
+
"language_model.model.layers.27.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 229 |
+
"language_model.model.layers.27.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 230 |
+
"language_model.model.layers.27.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 231 |
+
"language_model.model.layers.27.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 232 |
+
"language_model.model.layers.27.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 233 |
+
"language_model.model.layers.27.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 234 |
+
"language_model.model.layers.27.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 235 |
+
"language_model.model.layers.27.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 236 |
+
"language_model.model.layers.27.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 237 |
+
"language_model.model.layers.27.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 238 |
+
"language_model.model.layers.27.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 239 |
+
"language_model.model.layers.28.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 240 |
+
"language_model.model.layers.28.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 241 |
+
"language_model.model.layers.28.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 242 |
+
"language_model.model.layers.28.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 243 |
+
"language_model.model.layers.28.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 244 |
+
"language_model.model.layers.28.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 245 |
+
"language_model.model.layers.28.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 246 |
+
"language_model.model.layers.28.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 247 |
+
"language_model.model.layers.28.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 248 |
+
"language_model.model.layers.28.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 249 |
+
"language_model.model.layers.28.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 250 |
+
"language_model.model.layers.29.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 251 |
+
"language_model.model.layers.29.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 252 |
+
"language_model.model.layers.29.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 253 |
+
"language_model.model.layers.29.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 254 |
+
"language_model.model.layers.29.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 255 |
+
"language_model.model.layers.29.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 256 |
+
"language_model.model.layers.29.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 257 |
+
"language_model.model.layers.29.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 258 |
+
"language_model.model.layers.29.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 259 |
+
"language_model.model.layers.29.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 260 |
+
"language_model.model.layers.29.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 261 |
+
"language_model.model.layers.3.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 262 |
+
"language_model.model.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 263 |
+
"language_model.model.layers.3.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 264 |
+
"language_model.model.layers.3.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 265 |
+
"language_model.model.layers.3.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 266 |
+
"language_model.model.layers.3.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 267 |
+
"language_model.model.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 268 |
+
"language_model.model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 269 |
+
"language_model.model.layers.3.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 270 |
+
"language_model.model.layers.3.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 271 |
+
"language_model.model.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 272 |
+
"language_model.model.layers.30.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 273 |
+
"language_model.model.layers.30.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 274 |
+
"language_model.model.layers.30.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 275 |
+
"language_model.model.layers.30.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 276 |
+
"language_model.model.layers.30.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 277 |
+
"language_model.model.layers.30.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 278 |
+
"language_model.model.layers.30.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 279 |
+
"language_model.model.layers.30.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 280 |
+
"language_model.model.layers.30.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 281 |
+
"language_model.model.layers.30.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 282 |
+
"language_model.model.layers.30.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 283 |
+
"language_model.model.layers.31.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 284 |
+
"language_model.model.layers.31.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 285 |
+
"language_model.model.layers.31.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 286 |
+
"language_model.model.layers.31.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 287 |
+
"language_model.model.layers.31.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 288 |
+
"language_model.model.layers.31.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 289 |
+
"language_model.model.layers.31.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 290 |
+
"language_model.model.layers.31.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 291 |
+
"language_model.model.layers.31.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 292 |
+
"language_model.model.layers.31.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 293 |
+
"language_model.model.layers.31.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 294 |
+
"language_model.model.layers.32.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 295 |
+
"language_model.model.layers.32.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 296 |
+
"language_model.model.layers.32.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 297 |
+
"language_model.model.layers.32.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 298 |
+
"language_model.model.layers.32.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 299 |
+
"language_model.model.layers.32.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 300 |
+
"language_model.model.layers.32.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 301 |
+
"language_model.model.layers.32.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 302 |
+
"language_model.model.layers.32.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 303 |
+
"language_model.model.layers.32.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 304 |
+
"language_model.model.layers.32.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 305 |
+
"language_model.model.layers.33.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 306 |
+
"language_model.model.layers.33.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 307 |
+
"language_model.model.layers.33.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 308 |
+
"language_model.model.layers.33.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 309 |
+
"language_model.model.layers.33.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 310 |
+
"language_model.model.layers.33.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 311 |
+
"language_model.model.layers.33.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 312 |
+
"language_model.model.layers.33.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 313 |
+
"language_model.model.layers.33.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 314 |
+
"language_model.model.layers.33.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 315 |
+
"language_model.model.layers.33.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 316 |
+
"language_model.model.layers.34.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 317 |
+
"language_model.model.layers.34.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 318 |
+
"language_model.model.layers.34.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 319 |
+
"language_model.model.layers.34.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 320 |
+
"language_model.model.layers.34.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 321 |
+
"language_model.model.layers.34.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 322 |
+
"language_model.model.layers.34.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 323 |
+
"language_model.model.layers.34.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 324 |
+
"language_model.model.layers.34.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 325 |
+
"language_model.model.layers.34.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 326 |
+
"language_model.model.layers.34.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 327 |
+
"language_model.model.layers.35.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 328 |
+
"language_model.model.layers.35.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 329 |
+
"language_model.model.layers.35.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 330 |
+
"language_model.model.layers.35.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 331 |
+
"language_model.model.layers.35.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 332 |
+
"language_model.model.layers.35.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 333 |
+
"language_model.model.layers.35.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 334 |
+
"language_model.model.layers.35.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 335 |
+
"language_model.model.layers.35.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 336 |
+
"language_model.model.layers.35.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 337 |
+
"language_model.model.layers.35.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 338 |
+
"language_model.model.layers.4.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 339 |
+
"language_model.model.layers.4.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 340 |
+
"language_model.model.layers.4.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 341 |
+
"language_model.model.layers.4.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 342 |
+
"language_model.model.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 343 |
+
"language_model.model.layers.4.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 344 |
+
"language_model.model.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 345 |
+
"language_model.model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 346 |
+
"language_model.model.layers.4.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 347 |
+
"language_model.model.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 348 |
+
"language_model.model.layers.4.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 349 |
+
"language_model.model.layers.5.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 350 |
+
"language_model.model.layers.5.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 351 |
+
"language_model.model.layers.5.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 352 |
+
"language_model.model.layers.5.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 353 |
+
"language_model.model.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 354 |
+
"language_model.model.layers.5.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 355 |
+
"language_model.model.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 356 |
+
"language_model.model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 357 |
+
"language_model.model.layers.5.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 358 |
+
"language_model.model.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 359 |
+
"language_model.model.layers.5.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 360 |
+
"language_model.model.layers.6.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 361 |
+
"language_model.model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 362 |
+
"language_model.model.layers.6.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 363 |
+
"language_model.model.layers.6.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 364 |
+
"language_model.model.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 365 |
+
"language_model.model.layers.6.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 366 |
+
"language_model.model.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 367 |
+
"language_model.model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 368 |
+
"language_model.model.layers.6.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 369 |
+
"language_model.model.layers.6.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 370 |
+
"language_model.model.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 371 |
+
"language_model.model.layers.7.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 372 |
+
"language_model.model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 373 |
+
"language_model.model.layers.7.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 374 |
+
"language_model.model.layers.7.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 375 |
+
"language_model.model.layers.7.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 376 |
+
"language_model.model.layers.7.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 377 |
+
"language_model.model.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 378 |
+
"language_model.model.layers.7.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 379 |
+
"language_model.model.layers.7.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 380 |
+
"language_model.model.layers.7.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 381 |
+
"language_model.model.layers.7.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 382 |
+
"language_model.model.layers.8.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 383 |
+
"language_model.model.layers.8.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 384 |
+
"language_model.model.layers.8.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 385 |
+
"language_model.model.layers.8.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 386 |
+
"language_model.model.layers.8.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 387 |
+
"language_model.model.layers.8.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 388 |
+
"language_model.model.layers.8.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 389 |
+
"language_model.model.layers.8.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 390 |
+
"language_model.model.layers.8.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 391 |
+
"language_model.model.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 392 |
+
"language_model.model.layers.8.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 393 |
+
"language_model.model.layers.9.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 394 |
+
"language_model.model.layers.9.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 395 |
+
"language_model.model.layers.9.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 396 |
+
"language_model.model.layers.9.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 397 |
+
"language_model.model.layers.9.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 398 |
+
"language_model.model.layers.9.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 399 |
+
"language_model.model.layers.9.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 400 |
+
"language_model.model.layers.9.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 401 |
+
"language_model.model.layers.9.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 402 |
+
"language_model.model.layers.9.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 403 |
+
"language_model.model.layers.9.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 404 |
+
"language_model.model.norm.weight": "model-00002-of-00002.safetensors",
|
| 405 |
+
"projector.mlp_projector.0.bias": "model-00002-of-00002.safetensors",
|
| 406 |
+
"projector.mlp_projector.0.weight": "model-00002-of-00002.safetensors",
|
| 407 |
+
"projector.mlp_projector.2.bias": "model-00002-of-00002.safetensors",
|
| 408 |
+
"projector.mlp_projector.2.weight": "model-00002-of-00002.safetensors",
|
| 409 |
+
"vision_model.vision_model.embeddings.patch_embedding.bias": "model-00001-of-00002.safetensors",
|
| 410 |
+
"vision_model.vision_model.embeddings.patch_embedding.weight": "model-00001-of-00002.safetensors",
|
| 411 |
+
"vision_model.vision_model.embeddings.position_embedding.weight": "model-00001-of-00002.safetensors",
|
| 412 |
+
"vision_model.vision_model.encoder.layers.0.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 413 |
+
"vision_model.vision_model.encoder.layers.0.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 414 |
+
"vision_model.vision_model.encoder.layers.0.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 415 |
+
"vision_model.vision_model.encoder.layers.0.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 416 |
+
"vision_model.vision_model.encoder.layers.0.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 417 |
+
"vision_model.vision_model.encoder.layers.0.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 418 |
+
"vision_model.vision_model.encoder.layers.0.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 419 |
+
"vision_model.vision_model.encoder.layers.0.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 420 |
+
"vision_model.vision_model.encoder.layers.0.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 421 |
+
"vision_model.vision_model.encoder.layers.0.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 422 |
+
"vision_model.vision_model.encoder.layers.0.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 423 |
+
"vision_model.vision_model.encoder.layers.0.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 424 |
+
"vision_model.vision_model.encoder.layers.0.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 425 |
+
"vision_model.vision_model.encoder.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 426 |
+
"vision_model.vision_model.encoder.layers.0.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 427 |
+
"vision_model.vision_model.encoder.layers.0.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 428 |
+
"vision_model.vision_model.encoder.layers.1.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 429 |
+
"vision_model.vision_model.encoder.layers.1.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 430 |
+
"vision_model.vision_model.encoder.layers.1.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 431 |
+
"vision_model.vision_model.encoder.layers.1.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 432 |
+
"vision_model.vision_model.encoder.layers.1.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 433 |
+
"vision_model.vision_model.encoder.layers.1.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 434 |
+
"vision_model.vision_model.encoder.layers.1.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 435 |
+
"vision_model.vision_model.encoder.layers.1.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 436 |
+
"vision_model.vision_model.encoder.layers.1.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 437 |
+
"vision_model.vision_model.encoder.layers.1.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 438 |
+
"vision_model.vision_model.encoder.layers.1.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 439 |
+
"vision_model.vision_model.encoder.layers.1.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 440 |
+
"vision_model.vision_model.encoder.layers.1.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 441 |
+
"vision_model.vision_model.encoder.layers.1.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 442 |
+
"vision_model.vision_model.encoder.layers.1.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 443 |
+
"vision_model.vision_model.encoder.layers.1.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 444 |
+
"vision_model.vision_model.encoder.layers.10.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 445 |
+
"vision_model.vision_model.encoder.layers.10.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 446 |
+
"vision_model.vision_model.encoder.layers.10.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 447 |
+
"vision_model.vision_model.encoder.layers.10.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 448 |
+
"vision_model.vision_model.encoder.layers.10.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 449 |
+
"vision_model.vision_model.encoder.layers.10.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 450 |
+
"vision_model.vision_model.encoder.layers.10.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 451 |
+
"vision_model.vision_model.encoder.layers.10.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 452 |
+
"vision_model.vision_model.encoder.layers.10.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 453 |
+
"vision_model.vision_model.encoder.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 454 |
+
"vision_model.vision_model.encoder.layers.10.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 455 |
+
"vision_model.vision_model.encoder.layers.10.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 456 |
+
"vision_model.vision_model.encoder.layers.10.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 457 |
+
"vision_model.vision_model.encoder.layers.10.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 458 |
+
"vision_model.vision_model.encoder.layers.10.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 459 |
+
"vision_model.vision_model.encoder.layers.10.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 460 |
+
"vision_model.vision_model.encoder.layers.11.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 461 |
+
"vision_model.vision_model.encoder.layers.11.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 462 |
+
"vision_model.vision_model.encoder.layers.11.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 463 |
+
"vision_model.vision_model.encoder.layers.11.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 464 |
+
"vision_model.vision_model.encoder.layers.11.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 465 |
+
"vision_model.vision_model.encoder.layers.11.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 466 |
+
"vision_model.vision_model.encoder.layers.11.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 467 |
+
"vision_model.vision_model.encoder.layers.11.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 468 |
+
"vision_model.vision_model.encoder.layers.11.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 469 |
+
"vision_model.vision_model.encoder.layers.11.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 470 |
+
"vision_model.vision_model.encoder.layers.11.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 471 |
+
"vision_model.vision_model.encoder.layers.11.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 472 |
+
"vision_model.vision_model.encoder.layers.11.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 473 |
+
"vision_model.vision_model.encoder.layers.11.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 474 |
+
"vision_model.vision_model.encoder.layers.11.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 475 |
+
"vision_model.vision_model.encoder.layers.11.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 476 |
+
"vision_model.vision_model.encoder.layers.12.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 477 |
+
"vision_model.vision_model.encoder.layers.12.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 478 |
+
"vision_model.vision_model.encoder.layers.12.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 479 |
+
"vision_model.vision_model.encoder.layers.12.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 480 |
+
"vision_model.vision_model.encoder.layers.12.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 481 |
+
"vision_model.vision_model.encoder.layers.12.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 482 |
+
"vision_model.vision_model.encoder.layers.12.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 483 |
+
"vision_model.vision_model.encoder.layers.12.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 484 |
+
"vision_model.vision_model.encoder.layers.12.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 485 |
+
"vision_model.vision_model.encoder.layers.12.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 486 |
+
"vision_model.vision_model.encoder.layers.12.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 487 |
+
"vision_model.vision_model.encoder.layers.12.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 488 |
+
"vision_model.vision_model.encoder.layers.12.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 489 |
+
"vision_model.vision_model.encoder.layers.12.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 490 |
+
"vision_model.vision_model.encoder.layers.12.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 491 |
+
"vision_model.vision_model.encoder.layers.12.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 492 |
+
"vision_model.vision_model.encoder.layers.13.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 493 |
+
"vision_model.vision_model.encoder.layers.13.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 494 |
+
"vision_model.vision_model.encoder.layers.13.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 495 |
+
"vision_model.vision_model.encoder.layers.13.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 496 |
+
"vision_model.vision_model.encoder.layers.13.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 497 |
+
"vision_model.vision_model.encoder.layers.13.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 498 |
+
"vision_model.vision_model.encoder.layers.13.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 499 |
+
"vision_model.vision_model.encoder.layers.13.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 500 |
+
"vision_model.vision_model.encoder.layers.13.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 501 |
+
"vision_model.vision_model.encoder.layers.13.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 502 |
+
"vision_model.vision_model.encoder.layers.13.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 503 |
+
"vision_model.vision_model.encoder.layers.13.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 504 |
+
"vision_model.vision_model.encoder.layers.13.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 505 |
+
"vision_model.vision_model.encoder.layers.13.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 506 |
+
"vision_model.vision_model.encoder.layers.13.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 507 |
+
"vision_model.vision_model.encoder.layers.13.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 508 |
+
"vision_model.vision_model.encoder.layers.14.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 509 |
+
"vision_model.vision_model.encoder.layers.14.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 510 |
+
"vision_model.vision_model.encoder.layers.14.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 511 |
+
"vision_model.vision_model.encoder.layers.14.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 512 |
+
"vision_model.vision_model.encoder.layers.14.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 513 |
+
"vision_model.vision_model.encoder.layers.14.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 514 |
+
"vision_model.vision_model.encoder.layers.14.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 515 |
+
"vision_model.vision_model.encoder.layers.14.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 516 |
+
"vision_model.vision_model.encoder.layers.14.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 517 |
+
"vision_model.vision_model.encoder.layers.14.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 518 |
+
"vision_model.vision_model.encoder.layers.14.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 519 |
+
"vision_model.vision_model.encoder.layers.14.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 520 |
+
"vision_model.vision_model.encoder.layers.14.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 521 |
+
"vision_model.vision_model.encoder.layers.14.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 522 |
+
"vision_model.vision_model.encoder.layers.14.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 523 |
+
"vision_model.vision_model.encoder.layers.14.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 524 |
+
"vision_model.vision_model.encoder.layers.15.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 525 |
+
"vision_model.vision_model.encoder.layers.15.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 526 |
+
"vision_model.vision_model.encoder.layers.15.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 527 |
+
"vision_model.vision_model.encoder.layers.15.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 528 |
+
"vision_model.vision_model.encoder.layers.15.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 529 |
+
"vision_model.vision_model.encoder.layers.15.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 530 |
+
"vision_model.vision_model.encoder.layers.15.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 531 |
+
"vision_model.vision_model.encoder.layers.15.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 532 |
+
"vision_model.vision_model.encoder.layers.15.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 533 |
+
"vision_model.vision_model.encoder.layers.15.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 534 |
+
"vision_model.vision_model.encoder.layers.15.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 535 |
+
"vision_model.vision_model.encoder.layers.15.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 536 |
+
"vision_model.vision_model.encoder.layers.15.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 537 |
+
"vision_model.vision_model.encoder.layers.15.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 538 |
+
"vision_model.vision_model.encoder.layers.15.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 539 |
+
"vision_model.vision_model.encoder.layers.15.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 540 |
+
"vision_model.vision_model.encoder.layers.16.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 541 |
+
"vision_model.vision_model.encoder.layers.16.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 542 |
+
"vision_model.vision_model.encoder.layers.16.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 543 |
+
"vision_model.vision_model.encoder.layers.16.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 544 |
+
"vision_model.vision_model.encoder.layers.16.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 545 |
+
"vision_model.vision_model.encoder.layers.16.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 546 |
+
"vision_model.vision_model.encoder.layers.16.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 547 |
+
"vision_model.vision_model.encoder.layers.16.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 548 |
+
"vision_model.vision_model.encoder.layers.16.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 549 |
+
"vision_model.vision_model.encoder.layers.16.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 550 |
+
"vision_model.vision_model.encoder.layers.16.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 551 |
+
"vision_model.vision_model.encoder.layers.16.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 552 |
+
"vision_model.vision_model.encoder.layers.16.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 553 |
+
"vision_model.vision_model.encoder.layers.16.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 554 |
+
"vision_model.vision_model.encoder.layers.16.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 555 |
+
"vision_model.vision_model.encoder.layers.16.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 556 |
+
"vision_model.vision_model.encoder.layers.17.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 557 |
+
"vision_model.vision_model.encoder.layers.17.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 558 |
+
"vision_model.vision_model.encoder.layers.17.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 559 |
+
"vision_model.vision_model.encoder.layers.17.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 560 |
+
"vision_model.vision_model.encoder.layers.17.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 561 |
+
"vision_model.vision_model.encoder.layers.17.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 562 |
+
"vision_model.vision_model.encoder.layers.17.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 563 |
+
"vision_model.vision_model.encoder.layers.17.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 564 |
+
"vision_model.vision_model.encoder.layers.17.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 565 |
+
"vision_model.vision_model.encoder.layers.17.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 566 |
+
"vision_model.vision_model.encoder.layers.17.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 567 |
+
"vision_model.vision_model.encoder.layers.17.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 568 |
+
"vision_model.vision_model.encoder.layers.17.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 569 |
+
"vision_model.vision_model.encoder.layers.17.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 570 |
+
"vision_model.vision_model.encoder.layers.17.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 571 |
+
"vision_model.vision_model.encoder.layers.17.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 572 |
+
"vision_model.vision_model.encoder.layers.18.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 573 |
+
"vision_model.vision_model.encoder.layers.18.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 574 |
+
"vision_model.vision_model.encoder.layers.18.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 575 |
+
"vision_model.vision_model.encoder.layers.18.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 576 |
+
"vision_model.vision_model.encoder.layers.18.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 577 |
+
"vision_model.vision_model.encoder.layers.18.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 578 |
+
"vision_model.vision_model.encoder.layers.18.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 579 |
+
"vision_model.vision_model.encoder.layers.18.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 580 |
+
"vision_model.vision_model.encoder.layers.18.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 581 |
+
"vision_model.vision_model.encoder.layers.18.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 582 |
+
"vision_model.vision_model.encoder.layers.18.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 583 |
+
"vision_model.vision_model.encoder.layers.18.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 584 |
+
"vision_model.vision_model.encoder.layers.18.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 585 |
+
"vision_model.vision_model.encoder.layers.18.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 586 |
+
"vision_model.vision_model.encoder.layers.18.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 587 |
+
"vision_model.vision_model.encoder.layers.18.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 588 |
+
"vision_model.vision_model.encoder.layers.19.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 589 |
+
"vision_model.vision_model.encoder.layers.19.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 590 |
+
"vision_model.vision_model.encoder.layers.19.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 591 |
+
"vision_model.vision_model.encoder.layers.19.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 592 |
+
"vision_model.vision_model.encoder.layers.19.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 593 |
+
"vision_model.vision_model.encoder.layers.19.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 594 |
+
"vision_model.vision_model.encoder.layers.19.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 595 |
+
"vision_model.vision_model.encoder.layers.19.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 596 |
+
"vision_model.vision_model.encoder.layers.19.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 597 |
+
"vision_model.vision_model.encoder.layers.19.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 598 |
+
"vision_model.vision_model.encoder.layers.19.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 599 |
+
"vision_model.vision_model.encoder.layers.19.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 600 |
+
"vision_model.vision_model.encoder.layers.19.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 601 |
+
"vision_model.vision_model.encoder.layers.19.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 602 |
+
"vision_model.vision_model.encoder.layers.19.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 603 |
+
"vision_model.vision_model.encoder.layers.19.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 604 |
+
"vision_model.vision_model.encoder.layers.2.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 605 |
+
"vision_model.vision_model.encoder.layers.2.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 606 |
+
"vision_model.vision_model.encoder.layers.2.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 607 |
+
"vision_model.vision_model.encoder.layers.2.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 608 |
+
"vision_model.vision_model.encoder.layers.2.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 609 |
+
"vision_model.vision_model.encoder.layers.2.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 610 |
+
"vision_model.vision_model.encoder.layers.2.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 611 |
+
"vision_model.vision_model.encoder.layers.2.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 612 |
+
"vision_model.vision_model.encoder.layers.2.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 613 |
+
"vision_model.vision_model.encoder.layers.2.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 614 |
+
"vision_model.vision_model.encoder.layers.2.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 615 |
+
"vision_model.vision_model.encoder.layers.2.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 616 |
+
"vision_model.vision_model.encoder.layers.2.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 617 |
+
"vision_model.vision_model.encoder.layers.2.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 618 |
+
"vision_model.vision_model.encoder.layers.2.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 619 |
+
"vision_model.vision_model.encoder.layers.2.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 620 |
+
"vision_model.vision_model.encoder.layers.20.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 621 |
+
"vision_model.vision_model.encoder.layers.20.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 622 |
+
"vision_model.vision_model.encoder.layers.20.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 623 |
+
"vision_model.vision_model.encoder.layers.20.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 624 |
+
"vision_model.vision_model.encoder.layers.20.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 625 |
+
"vision_model.vision_model.encoder.layers.20.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 626 |
+
"vision_model.vision_model.encoder.layers.20.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 627 |
+
"vision_model.vision_model.encoder.layers.20.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 628 |
+
"vision_model.vision_model.encoder.layers.20.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 629 |
+
"vision_model.vision_model.encoder.layers.20.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 630 |
+
"vision_model.vision_model.encoder.layers.20.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 631 |
+
"vision_model.vision_model.encoder.layers.20.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 632 |
+
"vision_model.vision_model.encoder.layers.20.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 633 |
+
"vision_model.vision_model.encoder.layers.20.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 634 |
+
"vision_model.vision_model.encoder.layers.20.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 635 |
+
"vision_model.vision_model.encoder.layers.20.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 636 |
+
"vision_model.vision_model.encoder.layers.21.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 637 |
+
"vision_model.vision_model.encoder.layers.21.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 638 |
+
"vision_model.vision_model.encoder.layers.21.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 639 |
+
"vision_model.vision_model.encoder.layers.21.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 640 |
+
"vision_model.vision_model.encoder.layers.21.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 641 |
+
"vision_model.vision_model.encoder.layers.21.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 642 |
+
"vision_model.vision_model.encoder.layers.21.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 643 |
+
"vision_model.vision_model.encoder.layers.21.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 644 |
+
"vision_model.vision_model.encoder.layers.21.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 645 |
+
"vision_model.vision_model.encoder.layers.21.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 646 |
+
"vision_model.vision_model.encoder.layers.21.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 647 |
+
"vision_model.vision_model.encoder.layers.21.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 648 |
+
"vision_model.vision_model.encoder.layers.21.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 649 |
+
"vision_model.vision_model.encoder.layers.21.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 650 |
+
"vision_model.vision_model.encoder.layers.21.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 651 |
+
"vision_model.vision_model.encoder.layers.21.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 652 |
+
"vision_model.vision_model.encoder.layers.22.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 653 |
+
"vision_model.vision_model.encoder.layers.22.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 654 |
+
"vision_model.vision_model.encoder.layers.22.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 655 |
+
"vision_model.vision_model.encoder.layers.22.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 656 |
+
"vision_model.vision_model.encoder.layers.22.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 657 |
+
"vision_model.vision_model.encoder.layers.22.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 658 |
+
"vision_model.vision_model.encoder.layers.22.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 659 |
+
"vision_model.vision_model.encoder.layers.22.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 660 |
+
"vision_model.vision_model.encoder.layers.22.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 661 |
+
"vision_model.vision_model.encoder.layers.22.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 662 |
+
"vision_model.vision_model.encoder.layers.22.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 663 |
+
"vision_model.vision_model.encoder.layers.22.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 664 |
+
"vision_model.vision_model.encoder.layers.22.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 665 |
+
"vision_model.vision_model.encoder.layers.22.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 666 |
+
"vision_model.vision_model.encoder.layers.22.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 667 |
+
"vision_model.vision_model.encoder.layers.22.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 668 |
+
"vision_model.vision_model.encoder.layers.23.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 669 |
+
"vision_model.vision_model.encoder.layers.23.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 670 |
+
"vision_model.vision_model.encoder.layers.23.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 671 |
+
"vision_model.vision_model.encoder.layers.23.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 672 |
+
"vision_model.vision_model.encoder.layers.23.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 673 |
+
"vision_model.vision_model.encoder.layers.23.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 674 |
+
"vision_model.vision_model.encoder.layers.23.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 675 |
+
"vision_model.vision_model.encoder.layers.23.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 676 |
+
"vision_model.vision_model.encoder.layers.23.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 677 |
+
"vision_model.vision_model.encoder.layers.23.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 678 |
+
"vision_model.vision_model.encoder.layers.23.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 679 |
+
"vision_model.vision_model.encoder.layers.23.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 680 |
+
"vision_model.vision_model.encoder.layers.23.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 681 |
+
"vision_model.vision_model.encoder.layers.23.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 682 |
+
"vision_model.vision_model.encoder.layers.23.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 683 |
+
"vision_model.vision_model.encoder.layers.23.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 684 |
+
"vision_model.vision_model.encoder.layers.24.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 685 |
+
"vision_model.vision_model.encoder.layers.24.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 686 |
+
"vision_model.vision_model.encoder.layers.24.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 687 |
+
"vision_model.vision_model.encoder.layers.24.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 688 |
+
"vision_model.vision_model.encoder.layers.24.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 689 |
+
"vision_model.vision_model.encoder.layers.24.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 690 |
+
"vision_model.vision_model.encoder.layers.24.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 691 |
+
"vision_model.vision_model.encoder.layers.24.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 692 |
+
"vision_model.vision_model.encoder.layers.24.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 693 |
+
"vision_model.vision_model.encoder.layers.24.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 694 |
+
"vision_model.vision_model.encoder.layers.24.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 695 |
+
"vision_model.vision_model.encoder.layers.24.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 696 |
+
"vision_model.vision_model.encoder.layers.24.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 697 |
+
"vision_model.vision_model.encoder.layers.24.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 698 |
+
"vision_model.vision_model.encoder.layers.24.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 699 |
+
"vision_model.vision_model.encoder.layers.24.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 700 |
+
"vision_model.vision_model.encoder.layers.25.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 701 |
+
"vision_model.vision_model.encoder.layers.25.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 702 |
+
"vision_model.vision_model.encoder.layers.25.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 703 |
+
"vision_model.vision_model.encoder.layers.25.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 704 |
+
"vision_model.vision_model.encoder.layers.25.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 705 |
+
"vision_model.vision_model.encoder.layers.25.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 706 |
+
"vision_model.vision_model.encoder.layers.25.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 707 |
+
"vision_model.vision_model.encoder.layers.25.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 708 |
+
"vision_model.vision_model.encoder.layers.25.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 709 |
+
"vision_model.vision_model.encoder.layers.25.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 710 |
+
"vision_model.vision_model.encoder.layers.25.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 711 |
+
"vision_model.vision_model.encoder.layers.25.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 712 |
+
"vision_model.vision_model.encoder.layers.25.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 713 |
+
"vision_model.vision_model.encoder.layers.25.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 714 |
+
"vision_model.vision_model.encoder.layers.25.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 715 |
+
"vision_model.vision_model.encoder.layers.25.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 716 |
+
"vision_model.vision_model.encoder.layers.26.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 717 |
+
"vision_model.vision_model.encoder.layers.26.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 718 |
+
"vision_model.vision_model.encoder.layers.26.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 719 |
+
"vision_model.vision_model.encoder.layers.26.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 720 |
+
"vision_model.vision_model.encoder.layers.26.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 721 |
+
"vision_model.vision_model.encoder.layers.26.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 722 |
+
"vision_model.vision_model.encoder.layers.26.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 723 |
+
"vision_model.vision_model.encoder.layers.26.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 724 |
+
"vision_model.vision_model.encoder.layers.26.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 725 |
+
"vision_model.vision_model.encoder.layers.26.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 726 |
+
"vision_model.vision_model.encoder.layers.26.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 727 |
+
"vision_model.vision_model.encoder.layers.26.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 728 |
+
"vision_model.vision_model.encoder.layers.26.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 729 |
+
"vision_model.vision_model.encoder.layers.26.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 730 |
+
"vision_model.vision_model.encoder.layers.26.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 731 |
+
"vision_model.vision_model.encoder.layers.26.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 732 |
+
"vision_model.vision_model.encoder.layers.3.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 733 |
+
"vision_model.vision_model.encoder.layers.3.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 734 |
+
"vision_model.vision_model.encoder.layers.3.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 735 |
+
"vision_model.vision_model.encoder.layers.3.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 736 |
+
"vision_model.vision_model.encoder.layers.3.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 737 |
+
"vision_model.vision_model.encoder.layers.3.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 738 |
+
"vision_model.vision_model.encoder.layers.3.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 739 |
+
"vision_model.vision_model.encoder.layers.3.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 740 |
+
"vision_model.vision_model.encoder.layers.3.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 741 |
+
"vision_model.vision_model.encoder.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 742 |
+
"vision_model.vision_model.encoder.layers.3.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 743 |
+
"vision_model.vision_model.encoder.layers.3.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 744 |
+
"vision_model.vision_model.encoder.layers.3.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 745 |
+
"vision_model.vision_model.encoder.layers.3.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 746 |
+
"vision_model.vision_model.encoder.layers.3.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 747 |
+
"vision_model.vision_model.encoder.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 748 |
+
"vision_model.vision_model.encoder.layers.4.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 749 |
+
"vision_model.vision_model.encoder.layers.4.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 750 |
+
"vision_model.vision_model.encoder.layers.4.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 751 |
+
"vision_model.vision_model.encoder.layers.4.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 752 |
+
"vision_model.vision_model.encoder.layers.4.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 753 |
+
"vision_model.vision_model.encoder.layers.4.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 754 |
+
"vision_model.vision_model.encoder.layers.4.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 755 |
+
"vision_model.vision_model.encoder.layers.4.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 756 |
+
"vision_model.vision_model.encoder.layers.4.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 757 |
+
"vision_model.vision_model.encoder.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 758 |
+
"vision_model.vision_model.encoder.layers.4.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 759 |
+
"vision_model.vision_model.encoder.layers.4.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 760 |
+
"vision_model.vision_model.encoder.layers.4.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 761 |
+
"vision_model.vision_model.encoder.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 762 |
+
"vision_model.vision_model.encoder.layers.4.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 763 |
+
"vision_model.vision_model.encoder.layers.4.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 764 |
+
"vision_model.vision_model.encoder.layers.5.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 765 |
+
"vision_model.vision_model.encoder.layers.5.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 766 |
+
"vision_model.vision_model.encoder.layers.5.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 767 |
+
"vision_model.vision_model.encoder.layers.5.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 768 |
+
"vision_model.vision_model.encoder.layers.5.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 769 |
+
"vision_model.vision_model.encoder.layers.5.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 770 |
+
"vision_model.vision_model.encoder.layers.5.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 771 |
+
"vision_model.vision_model.encoder.layers.5.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 772 |
+
"vision_model.vision_model.encoder.layers.5.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 773 |
+
"vision_model.vision_model.encoder.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 774 |
+
"vision_model.vision_model.encoder.layers.5.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 775 |
+
"vision_model.vision_model.encoder.layers.5.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 776 |
+
"vision_model.vision_model.encoder.layers.5.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 777 |
+
"vision_model.vision_model.encoder.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 778 |
+
"vision_model.vision_model.encoder.layers.5.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 779 |
+
"vision_model.vision_model.encoder.layers.5.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 780 |
+
"vision_model.vision_model.encoder.layers.6.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 781 |
+
"vision_model.vision_model.encoder.layers.6.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 782 |
+
"vision_model.vision_model.encoder.layers.6.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 783 |
+
"vision_model.vision_model.encoder.layers.6.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 784 |
+
"vision_model.vision_model.encoder.layers.6.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 785 |
+
"vision_model.vision_model.encoder.layers.6.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 786 |
+
"vision_model.vision_model.encoder.layers.6.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 787 |
+
"vision_model.vision_model.encoder.layers.6.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 788 |
+
"vision_model.vision_model.encoder.layers.6.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 789 |
+
"vision_model.vision_model.encoder.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 790 |
+
"vision_model.vision_model.encoder.layers.6.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 791 |
+
"vision_model.vision_model.encoder.layers.6.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 792 |
+
"vision_model.vision_model.encoder.layers.6.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 793 |
+
"vision_model.vision_model.encoder.layers.6.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 794 |
+
"vision_model.vision_model.encoder.layers.6.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 795 |
+
"vision_model.vision_model.encoder.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 796 |
+
"vision_model.vision_model.encoder.layers.7.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 797 |
+
"vision_model.vision_model.encoder.layers.7.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 798 |
+
"vision_model.vision_model.encoder.layers.7.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 799 |
+
"vision_model.vision_model.encoder.layers.7.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 800 |
+
"vision_model.vision_model.encoder.layers.7.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 801 |
+
"vision_model.vision_model.encoder.layers.7.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 802 |
+
"vision_model.vision_model.encoder.layers.7.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 803 |
+
"vision_model.vision_model.encoder.layers.7.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 804 |
+
"vision_model.vision_model.encoder.layers.7.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 805 |
+
"vision_model.vision_model.encoder.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 806 |
+
"vision_model.vision_model.encoder.layers.7.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 807 |
+
"vision_model.vision_model.encoder.layers.7.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 808 |
+
"vision_model.vision_model.encoder.layers.7.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 809 |
+
"vision_model.vision_model.encoder.layers.7.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 810 |
+
"vision_model.vision_model.encoder.layers.7.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 811 |
+
"vision_model.vision_model.encoder.layers.7.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 812 |
+
"vision_model.vision_model.encoder.layers.8.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 813 |
+
"vision_model.vision_model.encoder.layers.8.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 814 |
+
"vision_model.vision_model.encoder.layers.8.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 815 |
+
"vision_model.vision_model.encoder.layers.8.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 816 |
+
"vision_model.vision_model.encoder.layers.8.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 817 |
+
"vision_model.vision_model.encoder.layers.8.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 818 |
+
"vision_model.vision_model.encoder.layers.8.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 819 |
+
"vision_model.vision_model.encoder.layers.8.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 820 |
+
"vision_model.vision_model.encoder.layers.8.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 821 |
+
"vision_model.vision_model.encoder.layers.8.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 822 |
+
"vision_model.vision_model.encoder.layers.8.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 823 |
+
"vision_model.vision_model.encoder.layers.8.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 824 |
+
"vision_model.vision_model.encoder.layers.8.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 825 |
+
"vision_model.vision_model.encoder.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 826 |
+
"vision_model.vision_model.encoder.layers.8.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 827 |
+
"vision_model.vision_model.encoder.layers.8.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 828 |
+
"vision_model.vision_model.encoder.layers.9.layer_norm1.bias": "model-00001-of-00002.safetensors",
|
| 829 |
+
"vision_model.vision_model.encoder.layers.9.layer_norm1.weight": "model-00001-of-00002.safetensors",
|
| 830 |
+
"vision_model.vision_model.encoder.layers.9.layer_norm2.bias": "model-00001-of-00002.safetensors",
|
| 831 |
+
"vision_model.vision_model.encoder.layers.9.layer_norm2.weight": "model-00001-of-00002.safetensors",
|
| 832 |
+
"vision_model.vision_model.encoder.layers.9.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 833 |
+
"vision_model.vision_model.encoder.layers.9.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 834 |
+
"vision_model.vision_model.encoder.layers.9.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 835 |
+
"vision_model.vision_model.encoder.layers.9.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 836 |
+
"vision_model.vision_model.encoder.layers.9.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 837 |
+
"vision_model.vision_model.encoder.layers.9.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 838 |
+
"vision_model.vision_model.encoder.layers.9.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 839 |
+
"vision_model.vision_model.encoder.layers.9.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 840 |
+
"vision_model.vision_model.encoder.layers.9.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 841 |
+
"vision_model.vision_model.encoder.layers.9.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 842 |
+
"vision_model.vision_model.encoder.layers.9.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 843 |
+
"vision_model.vision_model.encoder.layers.9.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 844 |
+
"vision_model.vision_model.head.attention.in_proj_bias": "model-00001-of-00002.safetensors",
|
| 845 |
+
"vision_model.vision_model.head.attention.in_proj_weight": "model-00001-of-00002.safetensors",
|
| 846 |
+
"vision_model.vision_model.head.attention.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 847 |
+
"vision_model.vision_model.head.attention.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 848 |
+
"vision_model.vision_model.head.layernorm.bias": "model-00001-of-00002.safetensors",
|
| 849 |
+
"vision_model.vision_model.head.layernorm.weight": "model-00001-of-00002.safetensors",
|
| 850 |
+
"vision_model.vision_model.head.mlp.fc1.bias": "model-00001-of-00002.safetensors",
|
| 851 |
+
"vision_model.vision_model.head.mlp.fc1.weight": "model-00001-of-00002.safetensors",
|
| 852 |
+
"vision_model.vision_model.head.mlp.fc2.bias": "model-00001-of-00002.safetensors",
|
| 853 |
+
"vision_model.vision_model.head.mlp.fc2.weight": "model-00001-of-00002.safetensors",
|
| 854 |
+
"vision_model.vision_model.head.probe": "model-00001-of-00002.safetensors",
|
| 855 |
+
"vision_model.vision_model.post_layernorm.bias": "model-00001-of-00002.safetensors",
|
| 856 |
+
"vision_model.vision_model.post_layernorm.weight": "model-00001-of-00002.safetensors"
|
| 857 |
+
}
|
| 858 |
+
}
|
modeling_siglip.py
ADDED
|
@@ -0,0 +1,1585 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2024 Google AI and The HuggingFace Team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 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 |
+
"""PyTorch Siglip model."""
|
| 16 |
+
|
| 17 |
+
import math
|
| 18 |
+
import warnings
|
| 19 |
+
from dataclasses import dataclass
|
| 20 |
+
from typing import Any, Optional, Tuple, Union
|
| 21 |
+
|
| 22 |
+
import numpy as np
|
| 23 |
+
import torch
|
| 24 |
+
import torch.utils.checkpoint
|
| 25 |
+
from torch import nn
|
| 26 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
| 27 |
+
from torch.nn.init import _calculate_fan_in_and_fan_out
|
| 28 |
+
|
| 29 |
+
from transformers.activations import ACT2FN
|
| 30 |
+
from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask
|
| 31 |
+
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
|
| 32 |
+
from transformers.modeling_utils import PreTrainedModel
|
| 33 |
+
from transformers.utils import (
|
| 34 |
+
ModelOutput,
|
| 35 |
+
add_start_docstrings,
|
| 36 |
+
add_start_docstrings_to_model_forward,
|
| 37 |
+
is_flash_attn_2_available,
|
| 38 |
+
is_flash_attn_greater_or_equal_2_10,
|
| 39 |
+
logging,
|
| 40 |
+
replace_return_docstrings,
|
| 41 |
+
torch_int,
|
| 42 |
+
)
|
| 43 |
+
from configuration_siglip import SiglipConfig, SiglipTextConfig, SiglipVisionConfig
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
if is_flash_attn_2_available():
|
| 47 |
+
from transformers.modeling_flash_attention_utils import _flash_attention_forward
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
logger = logging.get_logger(__name__)
|
| 51 |
+
|
| 52 |
+
# General docstring
|
| 53 |
+
_CONFIG_FOR_DOC = "SiglipConfig"
|
| 54 |
+
_CHECKPOINT_FOR_DOC = "google/siglip-base-patch16-224"
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
def _trunc_normal_(tensor, mean, std, a, b):
|
| 58 |
+
# Cut & paste from PyTorch official master until it's in a few official releases - RW
|
| 59 |
+
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
|
| 60 |
+
def norm_cdf(x):
|
| 61 |
+
# Computes standard normal cumulative distribution function
|
| 62 |
+
return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
|
| 63 |
+
|
| 64 |
+
if (mean < a - 2 * std) or (mean > b + 2 * std):
|
| 65 |
+
warnings.warn(
|
| 66 |
+
"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
|
| 67 |
+
"The distribution of values may be incorrect.",
|
| 68 |
+
stacklevel=2,
|
| 69 |
+
)
|
| 70 |
+
|
| 71 |
+
# Values are generated by using a truncated uniform distribution and
|
| 72 |
+
# then using the inverse CDF for the normal distribution.
|
| 73 |
+
# Get upper and lower cdf values
|
| 74 |
+
l = norm_cdf((a - mean) / std)
|
| 75 |
+
u = norm_cdf((b - mean) / std)
|
| 76 |
+
|
| 77 |
+
# Uniformly fill tensor with values from [l, u], then translate to
|
| 78 |
+
# [2l-1, 2u-1].
|
| 79 |
+
tensor.uniform_(2 * l - 1, 2 * u - 1)
|
| 80 |
+
|
| 81 |
+
# Use inverse cdf transform for normal distribution to get truncated
|
| 82 |
+
# standard normal
|
| 83 |
+
tensor.erfinv_()
|
| 84 |
+
|
| 85 |
+
# Transform to proper mean, std
|
| 86 |
+
tensor.mul_(std * math.sqrt(2.0))
|
| 87 |
+
tensor.add_(mean)
|
| 88 |
+
|
| 89 |
+
# Clamp to ensure it's in the proper range
|
| 90 |
+
tensor.clamp_(min=a, max=b)
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
def trunc_normal_tf_(
|
| 94 |
+
tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
|
| 95 |
+
) -> torch.Tensor:
|
| 96 |
+
"""Fills the input Tensor with values drawn from a truncated
|
| 97 |
+
normal distribution. The values are effectively drawn from the
|
| 98 |
+
normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
|
| 99 |
+
with values outside :math:`[a, b]` redrawn until they are within
|
| 100 |
+
the bounds. The method used for generating the random values works
|
| 101 |
+
best when :math:`a \\leq \text{mean} \\leq b`.
|
| 102 |
+
|
| 103 |
+
NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
|
| 104 |
+
bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
|
| 105 |
+
and the result is subsequently scaled and shifted by the mean and std args.
|
| 106 |
+
|
| 107 |
+
Args:
|
| 108 |
+
tensor: an n-dimensional `torch.Tensor`
|
| 109 |
+
mean: the mean of the normal distribution
|
| 110 |
+
std: the standard deviation of the normal distribution
|
| 111 |
+
a: the minimum cutoff value
|
| 112 |
+
b: the maximum cutoff value
|
| 113 |
+
"""
|
| 114 |
+
with torch.no_grad():
|
| 115 |
+
_trunc_normal_(tensor, 0, 1.0, a, b)
|
| 116 |
+
tensor.mul_(std).add_(mean)
|
| 117 |
+
|
| 118 |
+
|
| 119 |
+
def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
|
| 120 |
+
fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
|
| 121 |
+
if mode == "fan_in":
|
| 122 |
+
denom = fan_in
|
| 123 |
+
elif mode == "fan_out":
|
| 124 |
+
denom = fan_out
|
| 125 |
+
elif mode == "fan_avg":
|
| 126 |
+
denom = (fan_in + fan_out) / 2
|
| 127 |
+
|
| 128 |
+
variance = scale / denom
|
| 129 |
+
|
| 130 |
+
if distribution == "truncated_normal":
|
| 131 |
+
# constant is stddev of standard normal truncated to (-2, 2)
|
| 132 |
+
trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
|
| 133 |
+
elif distribution == "normal":
|
| 134 |
+
with torch.no_grad():
|
| 135 |
+
tensor.normal_(std=math.sqrt(variance))
|
| 136 |
+
elif distribution == "uniform":
|
| 137 |
+
bound = math.sqrt(3 * variance)
|
| 138 |
+
with torch.no_grad():
|
| 139 |
+
tensor.uniform_(-bound, bound)
|
| 140 |
+
else:
|
| 141 |
+
raise ValueError(f"invalid distribution {distribution}")
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
def lecun_normal_(tensor):
|
| 145 |
+
variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
|
| 146 |
+
|
| 147 |
+
|
| 148 |
+
def default_flax_embed_init(tensor):
|
| 149 |
+
variance_scaling_(tensor, mode="fan_in", distribution="normal")
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
@dataclass
|
| 153 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->Siglip
|
| 154 |
+
class SiglipVisionModelOutput(ModelOutput):
|
| 155 |
+
"""
|
| 156 |
+
Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
|
| 157 |
+
|
| 158 |
+
Args:
|
| 159 |
+
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
|
| 160 |
+
The image embeddings obtained by applying the projection layer to the pooler_output.
|
| 161 |
+
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
| 162 |
+
Sequence of hidden-states at the output of the last layer of the model.
|
| 163 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
| 164 |
+
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
| 165 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
| 166 |
+
|
| 167 |
+
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
| 168 |
+
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
| 169 |
+
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
| 170 |
+
sequence_length)`.
|
| 171 |
+
|
| 172 |
+
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
|
| 173 |
+
heads.
|
| 174 |
+
"""
|
| 175 |
+
|
| 176 |
+
image_embeds: Optional[torch.FloatTensor] = None
|
| 177 |
+
last_hidden_state: torch.FloatTensor = None
|
| 178 |
+
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
|
| 179 |
+
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
|
| 180 |
+
|
| 181 |
+
|
| 182 |
+
@dataclass
|
| 183 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPTextModelOutput with CLIP->Siglip
|
| 184 |
+
class SiglipTextModelOutput(ModelOutput):
|
| 185 |
+
"""
|
| 186 |
+
Base class for text model's outputs that also contains a pooling of the last hidden states.
|
| 187 |
+
|
| 188 |
+
Args:
|
| 189 |
+
text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
|
| 190 |
+
The text embeddings obtained by applying the projection layer to the pooler_output.
|
| 191 |
+
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
| 192 |
+
Sequence of hidden-states at the output of the last layer of the model.
|
| 193 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
| 194 |
+
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
| 195 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
| 196 |
+
|
| 197 |
+
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
| 198 |
+
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
| 199 |
+
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
| 200 |
+
sequence_length)`.
|
| 201 |
+
|
| 202 |
+
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
|
| 203 |
+
heads.
|
| 204 |
+
"""
|
| 205 |
+
|
| 206 |
+
text_embeds: Optional[torch.FloatTensor] = None
|
| 207 |
+
last_hidden_state: torch.FloatTensor = None
|
| 208 |
+
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
|
| 209 |
+
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
|
| 210 |
+
|
| 211 |
+
|
| 212 |
+
@dataclass
|
| 213 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPOutput with CLIP->Siglip
|
| 214 |
+
class SiglipOutput(ModelOutput):
|
| 215 |
+
"""
|
| 216 |
+
Args:
|
| 217 |
+
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
|
| 218 |
+
Contrastive loss for image-text similarity.
|
| 219 |
+
logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
|
| 220 |
+
The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
|
| 221 |
+
similarity scores.
|
| 222 |
+
logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
|
| 223 |
+
The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
|
| 224 |
+
similarity scores.
|
| 225 |
+
text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
|
| 226 |
+
The text embeddings obtained by applying the projection layer to the pooled output of [`SiglipTextModel`].
|
| 227 |
+
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
|
| 228 |
+
The image embeddings obtained by applying the projection layer to the pooled output of [`SiglipVisionModel`].
|
| 229 |
+
text_model_output (`BaseModelOutputWithPooling`):
|
| 230 |
+
The output of the [`SiglipTextModel`].
|
| 231 |
+
vision_model_output (`BaseModelOutputWithPooling`):
|
| 232 |
+
The output of the [`SiglipVisionModel`].
|
| 233 |
+
"""
|
| 234 |
+
|
| 235 |
+
loss: Optional[torch.FloatTensor] = None
|
| 236 |
+
logits_per_image: torch.FloatTensor = None
|
| 237 |
+
logits_per_text: torch.FloatTensor = None
|
| 238 |
+
text_embeds: torch.FloatTensor = None
|
| 239 |
+
image_embeds: torch.FloatTensor = None
|
| 240 |
+
text_model_output: BaseModelOutputWithPooling = None
|
| 241 |
+
vision_model_output: BaseModelOutputWithPooling = None
|
| 242 |
+
|
| 243 |
+
def to_tuple(self) -> Tuple[Any]:
|
| 244 |
+
return tuple(
|
| 245 |
+
self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
|
| 246 |
+
for k in self.keys()
|
| 247 |
+
)
|
| 248 |
+
|
| 249 |
+
|
| 250 |
+
class SiglipVisionEmbeddings(nn.Module):
|
| 251 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 252 |
+
super().__init__()
|
| 253 |
+
self.config = config
|
| 254 |
+
self.embed_dim = config.hidden_size
|
| 255 |
+
self.image_size = config.image_size
|
| 256 |
+
self.patch_size = config.patch_size
|
| 257 |
+
|
| 258 |
+
self.patch_embedding = nn.Conv2d(
|
| 259 |
+
in_channels=config.num_channels,
|
| 260 |
+
out_channels=self.embed_dim,
|
| 261 |
+
kernel_size=self.patch_size,
|
| 262 |
+
stride=self.patch_size,
|
| 263 |
+
padding="valid",
|
| 264 |
+
)
|
| 265 |
+
|
| 266 |
+
self.num_patches = (self.image_size // self.patch_size) ** 2
|
| 267 |
+
self.num_positions = self.num_patches
|
| 268 |
+
self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
|
| 269 |
+
self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
|
| 270 |
+
|
| 271 |
+
def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
|
| 272 |
+
"""
|
| 273 |
+
This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
|
| 274 |
+
images. This method is also adapted to support torch.jit tracing and no class embeddings.
|
| 275 |
+
|
| 276 |
+
Adapted from:
|
| 277 |
+
- https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
|
| 278 |
+
- https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
|
| 279 |
+
"""
|
| 280 |
+
|
| 281 |
+
num_patches = embeddings.shape[1]
|
| 282 |
+
num_positions = self.position_embedding.weight.shape[0]
|
| 283 |
+
|
| 284 |
+
# always interpolate when tracing to ensure the exported model works for dynamic input shapes
|
| 285 |
+
if not torch.jit.is_tracing() and num_patches == num_positions and height == width:
|
| 286 |
+
return self.position_embedding(self.position_ids)
|
| 287 |
+
|
| 288 |
+
patch_pos_embed = self.position_embedding.weight.unsqueeze(0)
|
| 289 |
+
|
| 290 |
+
dim = embeddings.shape[-1]
|
| 291 |
+
|
| 292 |
+
new_height = height // self.patch_size
|
| 293 |
+
new_width = width // self.patch_size
|
| 294 |
+
|
| 295 |
+
sqrt_num_positions = torch_int(num_positions**0.5)
|
| 296 |
+
patch_pos_embed = patch_pos_embed.reshape(1, sqrt_num_positions, sqrt_num_positions, dim)
|
| 297 |
+
patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
|
| 298 |
+
|
| 299 |
+
patch_pos_embed = nn.functional.interpolate(
|
| 300 |
+
patch_pos_embed,
|
| 301 |
+
size=(new_height, new_width),
|
| 302 |
+
mode="bicubic",
|
| 303 |
+
align_corners=False,
|
| 304 |
+
)
|
| 305 |
+
|
| 306 |
+
patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
|
| 307 |
+
return patch_pos_embed
|
| 308 |
+
|
| 309 |
+
def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor:
|
| 310 |
+
_, _, height, width = pixel_values.shape
|
| 311 |
+
target_dtype = self.patch_embedding.weight.dtype
|
| 312 |
+
# import pdb;pdb.set_trace()
|
| 313 |
+
patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype)) # shape = [*, width, grid, grid]
|
| 314 |
+
embeddings = patch_embeds.flatten(2).transpose(1, 2)
|
| 315 |
+
|
| 316 |
+
if interpolate_pos_encoding:
|
| 317 |
+
embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
|
| 318 |
+
else:
|
| 319 |
+
embeddings = embeddings + self.position_embedding(self.position_ids)
|
| 320 |
+
return embeddings
|
| 321 |
+
|
| 322 |
+
|
| 323 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPTextEmbeddings with CLIP->Siglip
|
| 324 |
+
class SiglipTextEmbeddings(nn.Module):
|
| 325 |
+
def __init__(self, config: SiglipTextConfig):
|
| 326 |
+
super().__init__()
|
| 327 |
+
embed_dim = config.hidden_size
|
| 328 |
+
|
| 329 |
+
self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
|
| 330 |
+
self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
|
| 331 |
+
|
| 332 |
+
# position_ids (1, len position emb) is contiguous in memory and exported when serialized
|
| 333 |
+
self.register_buffer(
|
| 334 |
+
"position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
|
| 335 |
+
)
|
| 336 |
+
|
| 337 |
+
def forward(
|
| 338 |
+
self,
|
| 339 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 340 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 341 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 342 |
+
) -> torch.Tensor:
|
| 343 |
+
seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
|
| 344 |
+
max_position_embedding = self.position_embedding.weight.shape[0]
|
| 345 |
+
|
| 346 |
+
if seq_length > max_position_embedding:
|
| 347 |
+
raise ValueError(
|
| 348 |
+
f"Sequence length must be less than max_position_embeddings (got `sequence length`: "
|
| 349 |
+
f"{seq_length} and max_position_embeddings: {max_position_embedding}"
|
| 350 |
+
)
|
| 351 |
+
|
| 352 |
+
if position_ids is None:
|
| 353 |
+
position_ids = self.position_ids[:, :seq_length]
|
| 354 |
+
|
| 355 |
+
if inputs_embeds is None:
|
| 356 |
+
inputs_embeds = self.token_embedding(input_ids)
|
| 357 |
+
|
| 358 |
+
position_embeddings = self.position_embedding(position_ids)
|
| 359 |
+
embeddings = inputs_embeds + position_embeddings
|
| 360 |
+
|
| 361 |
+
return embeddings
|
| 362 |
+
|
| 363 |
+
|
| 364 |
+
class SiglipAttention(nn.Module):
|
| 365 |
+
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
| 366 |
+
|
| 367 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
|
| 368 |
+
def __init__(self, config):
|
| 369 |
+
super().__init__()
|
| 370 |
+
self.config = config
|
| 371 |
+
self.embed_dim = config.hidden_size
|
| 372 |
+
self.num_heads = config.num_attention_heads
|
| 373 |
+
self.head_dim = self.embed_dim // self.num_heads
|
| 374 |
+
if self.head_dim * self.num_heads != self.embed_dim:
|
| 375 |
+
raise ValueError(
|
| 376 |
+
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
|
| 377 |
+
f" {self.num_heads})."
|
| 378 |
+
)
|
| 379 |
+
self.scale = self.head_dim**-0.5
|
| 380 |
+
self.dropout = config.attention_dropout
|
| 381 |
+
|
| 382 |
+
self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
| 383 |
+
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
| 384 |
+
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
| 385 |
+
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
| 386 |
+
|
| 387 |
+
def forward(
|
| 388 |
+
self,
|
| 389 |
+
hidden_states: torch.Tensor,
|
| 390 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 391 |
+
output_attentions: Optional[bool] = False,
|
| 392 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
|
| 393 |
+
"""Input shape: Batch x Time x Channel"""
|
| 394 |
+
|
| 395 |
+
batch_size, q_len, _ = hidden_states.size()
|
| 396 |
+
|
| 397 |
+
query_states = self.q_proj(hidden_states)
|
| 398 |
+
key_states = self.k_proj(hidden_states)
|
| 399 |
+
value_states = self.v_proj(hidden_states)
|
| 400 |
+
|
| 401 |
+
query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 402 |
+
key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 403 |
+
value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 404 |
+
|
| 405 |
+
k_v_seq_len = key_states.shape[-2]
|
| 406 |
+
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale
|
| 407 |
+
|
| 408 |
+
if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
|
| 409 |
+
raise ValueError(
|
| 410 |
+
f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
|
| 411 |
+
f" {attn_weights.size()}"
|
| 412 |
+
)
|
| 413 |
+
|
| 414 |
+
if attention_mask is not None:
|
| 415 |
+
if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
|
| 416 |
+
raise ValueError(
|
| 417 |
+
f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
|
| 418 |
+
)
|
| 419 |
+
attn_weights = attn_weights + attention_mask
|
| 420 |
+
|
| 421 |
+
# upcast attention to fp32
|
| 422 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
| 423 |
+
attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
|
| 424 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
| 425 |
+
|
| 426 |
+
if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
|
| 427 |
+
raise ValueError(
|
| 428 |
+
f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
|
| 429 |
+
f" {attn_output.size()}"
|
| 430 |
+
)
|
| 431 |
+
|
| 432 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 433 |
+
attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
|
| 434 |
+
|
| 435 |
+
attn_output = self.out_proj(attn_output)
|
| 436 |
+
|
| 437 |
+
return attn_output, attn_weights
|
| 438 |
+
|
| 439 |
+
|
| 440 |
+
class SiglipFlashAttention2(SiglipAttention):
|
| 441 |
+
"""
|
| 442 |
+
SiglipAttention flash attention module. This module inherits from `SiglipAttention` as the weights of the module stays
|
| 443 |
+
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
|
| 444 |
+
flash attention and deal with padding tokens in case the input contains any of them.
|
| 445 |
+
"""
|
| 446 |
+
|
| 447 |
+
is_causal = False
|
| 448 |
+
|
| 449 |
+
def __init__(self, *args, **kwargs):
|
| 450 |
+
super().__init__(*args, **kwargs)
|
| 451 |
+
|
| 452 |
+
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
|
| 453 |
+
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
|
| 454 |
+
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
|
| 455 |
+
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
| 456 |
+
|
| 457 |
+
# Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
|
| 458 |
+
def forward(
|
| 459 |
+
self,
|
| 460 |
+
hidden_states: torch.Tensor,
|
| 461 |
+
attention_mask: Optional[torch.LongTensor] = None,
|
| 462 |
+
output_attentions: bool = False,
|
| 463 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 464 |
+
output_attentions = False
|
| 465 |
+
|
| 466 |
+
batch_size, q_len, _ = hidden_states.size()
|
| 467 |
+
|
| 468 |
+
query_states = self.q_proj(hidden_states)
|
| 469 |
+
key_states = self.k_proj(hidden_states)
|
| 470 |
+
value_states = self.v_proj(hidden_states)
|
| 471 |
+
|
| 472 |
+
# Flash attention requires the input to have the shape
|
| 473 |
+
# batch_size x seq_length x head_dim x hidden_dim
|
| 474 |
+
# therefore we just need to keep the original shape
|
| 475 |
+
query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 476 |
+
key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 477 |
+
value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 478 |
+
|
| 479 |
+
# TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
|
| 480 |
+
# to be able to avoid many of these transpose/reshape/view.
|
| 481 |
+
query_states = query_states.transpose(1, 2)
|
| 482 |
+
key_states = key_states.transpose(1, 2)
|
| 483 |
+
value_states = value_states.transpose(1, 2)
|
| 484 |
+
|
| 485 |
+
dropout_rate = self.dropout if self.training else 0.0
|
| 486 |
+
|
| 487 |
+
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
| 488 |
+
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
| 489 |
+
# cast them back in the correct dtype just to be sure everything works as expected.
|
| 490 |
+
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
|
| 491 |
+
# in fp32.
|
| 492 |
+
|
| 493 |
+
input_dtype = query_states.dtype
|
| 494 |
+
if input_dtype == torch.float32:
|
| 495 |
+
if torch.is_autocast_enabled():
|
| 496 |
+
target_dtype = torch.get_autocast_gpu_dtype()
|
| 497 |
+
# Handle the case where the model is quantized
|
| 498 |
+
elif hasattr(self.config, "_pre_quantization_dtype"):
|
| 499 |
+
target_dtype = self.config._pre_quantization_dtype
|
| 500 |
+
else:
|
| 501 |
+
target_dtype = self.q_proj.weight.dtype
|
| 502 |
+
|
| 503 |
+
logger.warning_once(
|
| 504 |
+
f"The input hidden states seems to be silently casted in float32, this might be related to"
|
| 505 |
+
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
| 506 |
+
f" {target_dtype}."
|
| 507 |
+
)
|
| 508 |
+
|
| 509 |
+
query_states = query_states.to(target_dtype)
|
| 510 |
+
key_states = key_states.to(target_dtype)
|
| 511 |
+
value_states = value_states.to(target_dtype)
|
| 512 |
+
|
| 513 |
+
attn_output = _flash_attention_forward(
|
| 514 |
+
query_states,
|
| 515 |
+
key_states,
|
| 516 |
+
value_states,
|
| 517 |
+
attention_mask,
|
| 518 |
+
q_len,
|
| 519 |
+
dropout=dropout_rate,
|
| 520 |
+
is_causal=self.is_causal,
|
| 521 |
+
use_top_left_mask=self._flash_attn_uses_top_left_mask,
|
| 522 |
+
)
|
| 523 |
+
|
| 524 |
+
attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim).contiguous()
|
| 525 |
+
attn_output = self.out_proj(attn_output)
|
| 526 |
+
|
| 527 |
+
if not output_attentions:
|
| 528 |
+
attn_weights = None
|
| 529 |
+
|
| 530 |
+
return attn_output, attn_weights
|
| 531 |
+
|
| 532 |
+
|
| 533 |
+
class SiglipSdpaAttention(SiglipAttention):
|
| 534 |
+
"""
|
| 535 |
+
Siglip attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
| 536 |
+
`SiglipAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
|
| 537 |
+
SDPA API.
|
| 538 |
+
"""
|
| 539 |
+
|
| 540 |
+
is_causal = False
|
| 541 |
+
|
| 542 |
+
# Adapted from SiglipAttention.forward and transformers.models.llama.modeling_llama.LlamaSdpaAttention.forward
|
| 543 |
+
def forward(
|
| 544 |
+
self,
|
| 545 |
+
hidden_states: torch.Tensor,
|
| 546 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 547 |
+
output_attentions: Optional[bool] = False,
|
| 548 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
|
| 549 |
+
if output_attentions:
|
| 550 |
+
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
| 551 |
+
logger.warning_once(
|
| 552 |
+
"SiglipModel is using SiglipSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
|
| 553 |
+
'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
| 554 |
+
)
|
| 555 |
+
return super().forward(
|
| 556 |
+
hidden_states=hidden_states,
|
| 557 |
+
attention_mask=attention_mask,
|
| 558 |
+
output_attentions=output_attentions,
|
| 559 |
+
)
|
| 560 |
+
|
| 561 |
+
batch_size, q_len, _ = hidden_states.size()
|
| 562 |
+
|
| 563 |
+
query_states = self.q_proj(hidden_states)
|
| 564 |
+
key_states = self.k_proj(hidden_states)
|
| 565 |
+
value_states = self.v_proj(hidden_states)
|
| 566 |
+
|
| 567 |
+
query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 568 |
+
key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 569 |
+
value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 570 |
+
|
| 571 |
+
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
| 572 |
+
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
| 573 |
+
if query_states.device.type == "cuda" and attention_mask is not None:
|
| 574 |
+
query_states = query_states.contiguous()
|
| 575 |
+
key_states = key_states.contiguous()
|
| 576 |
+
value_states = value_states.contiguous()
|
| 577 |
+
|
| 578 |
+
# We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
|
| 579 |
+
# in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
|
| 580 |
+
is_causal = True if self.is_causal and q_len > 1 else False
|
| 581 |
+
|
| 582 |
+
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
| 583 |
+
query_states,
|
| 584 |
+
key_states,
|
| 585 |
+
value_states,
|
| 586 |
+
attn_mask=attention_mask,
|
| 587 |
+
dropout_p=self.dropout if self.training else 0.0,
|
| 588 |
+
is_causal=is_causal,
|
| 589 |
+
)
|
| 590 |
+
|
| 591 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 592 |
+
attn_output = attn_output.view(batch_size, q_len, self.embed_dim)
|
| 593 |
+
|
| 594 |
+
attn_output = self.out_proj(attn_output)
|
| 595 |
+
|
| 596 |
+
return attn_output, None
|
| 597 |
+
|
| 598 |
+
|
| 599 |
+
SIGLIP_ATTENTION_CLASSES = {
|
| 600 |
+
"eager": SiglipAttention,
|
| 601 |
+
"flash_attention_2": SiglipFlashAttention2,
|
| 602 |
+
"sdpa": SiglipSdpaAttention,
|
| 603 |
+
}
|
| 604 |
+
|
| 605 |
+
|
| 606 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
|
| 607 |
+
class SiglipMLP(nn.Module):
|
| 608 |
+
def __init__(self, config):
|
| 609 |
+
super().__init__()
|
| 610 |
+
self.config = config
|
| 611 |
+
self.activation_fn = ACT2FN[config.hidden_act]
|
| 612 |
+
self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
|
| 613 |
+
self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
|
| 614 |
+
|
| 615 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 616 |
+
hidden_states = self.fc1(hidden_states)
|
| 617 |
+
hidden_states = self.activation_fn(hidden_states)
|
| 618 |
+
hidden_states = self.fc2(hidden_states)
|
| 619 |
+
return hidden_states
|
| 620 |
+
|
| 621 |
+
|
| 622 |
+
class SiglipEncoderLayer(nn.Module):
|
| 623 |
+
def __init__(self, config: SiglipConfig):
|
| 624 |
+
super().__init__()
|
| 625 |
+
self.embed_dim = config.hidden_size
|
| 626 |
+
self.self_attn = SIGLIP_ATTENTION_CLASSES[config._attn_implementation](config=config)
|
| 627 |
+
self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
|
| 628 |
+
self.mlp = SiglipMLP(config)
|
| 629 |
+
self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
|
| 630 |
+
|
| 631 |
+
# Ignore copy
|
| 632 |
+
def forward(
|
| 633 |
+
self,
|
| 634 |
+
hidden_states: torch.Tensor,
|
| 635 |
+
attention_mask: torch.Tensor,
|
| 636 |
+
output_attentions: Optional[bool] = False,
|
| 637 |
+
) -> Tuple[torch.FloatTensor]:
|
| 638 |
+
"""
|
| 639 |
+
Args:
|
| 640 |
+
hidden_states (`torch.FloatTensor`):
|
| 641 |
+
Input to the layer of shape `(batch, seq_len, embed_dim)`.
|
| 642 |
+
attention_mask (`torch.FloatTensor`):
|
| 643 |
+
Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
|
| 644 |
+
output_attentions (`bool`, *optional*, defaults to `False`):
|
| 645 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 646 |
+
returned tensors for more detail.
|
| 647 |
+
"""
|
| 648 |
+
residual = hidden_states
|
| 649 |
+
|
| 650 |
+
hidden_states = self.layer_norm1(hidden_states)
|
| 651 |
+
hidden_states, attn_weights = self.self_attn(
|
| 652 |
+
hidden_states=hidden_states,
|
| 653 |
+
attention_mask=attention_mask,
|
| 654 |
+
output_attentions=output_attentions,
|
| 655 |
+
)
|
| 656 |
+
hidden_states = residual + hidden_states
|
| 657 |
+
|
| 658 |
+
residual = hidden_states
|
| 659 |
+
hidden_states = self.layer_norm2(hidden_states)
|
| 660 |
+
hidden_states = self.mlp(hidden_states)
|
| 661 |
+
hidden_states = residual + hidden_states
|
| 662 |
+
|
| 663 |
+
outputs = (hidden_states,)
|
| 664 |
+
|
| 665 |
+
if output_attentions:
|
| 666 |
+
outputs += (attn_weights,)
|
| 667 |
+
|
| 668 |
+
return outputs
|
| 669 |
+
|
| 670 |
+
|
| 671 |
+
class SiglipPreTrainedModel(PreTrainedModel):
|
| 672 |
+
"""
|
| 673 |
+
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
| 674 |
+
models.
|
| 675 |
+
"""
|
| 676 |
+
|
| 677 |
+
config_class = SiglipConfig
|
| 678 |
+
base_model_prefix = "siglip"
|
| 679 |
+
supports_gradient_checkpointing = True
|
| 680 |
+
|
| 681 |
+
_no_split_modules = [
|
| 682 |
+
"SiglipTextEmbeddings",
|
| 683 |
+
"SiglipEncoderLayer",
|
| 684 |
+
"SiglipVisionEmbeddings",
|
| 685 |
+
"SiglipEncoderLayer",
|
| 686 |
+
"SiglipMultiheadAttentionPoolingHead",
|
| 687 |
+
]
|
| 688 |
+
_supports_flash_attn_2 = True
|
| 689 |
+
_supports_sdpa = True
|
| 690 |
+
|
| 691 |
+
def _init_weights(self, module):
|
| 692 |
+
"""Initialize the weights"""
|
| 693 |
+
if isinstance(module, SiglipVisionEmbeddings):
|
| 694 |
+
width = (
|
| 695 |
+
self.config.vision_config.hidden_size
|
| 696 |
+
if isinstance(self.config, SiglipConfig)
|
| 697 |
+
else self.config.hidden_size
|
| 698 |
+
)
|
| 699 |
+
nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
|
| 700 |
+
elif isinstance(module, nn.Embedding):
|
| 701 |
+
default_flax_embed_init(module.weight)
|
| 702 |
+
elif isinstance(module, SiglipAttention):
|
| 703 |
+
nn.init.xavier_uniform_(module.q_proj.weight)
|
| 704 |
+
nn.init.xavier_uniform_(module.k_proj.weight)
|
| 705 |
+
nn.init.xavier_uniform_(module.v_proj.weight)
|
| 706 |
+
nn.init.xavier_uniform_(module.out_proj.weight)
|
| 707 |
+
nn.init.zeros_(module.q_proj.bias)
|
| 708 |
+
nn.init.zeros_(module.k_proj.bias)
|
| 709 |
+
nn.init.zeros_(module.v_proj.bias)
|
| 710 |
+
nn.init.zeros_(module.out_proj.bias)
|
| 711 |
+
elif isinstance(module, SiglipMLP):
|
| 712 |
+
nn.init.xavier_uniform_(module.fc1.weight)
|
| 713 |
+
nn.init.xavier_uniform_(module.fc2.weight)
|
| 714 |
+
nn.init.normal_(module.fc1.bias, std=1e-6)
|
| 715 |
+
nn.init.normal_(module.fc2.bias, std=1e-6)
|
| 716 |
+
elif isinstance(module, SiglipMultiheadAttentionPoolingHead):
|
| 717 |
+
nn.init.xavier_uniform_(module.probe.data)
|
| 718 |
+
nn.init.xavier_uniform_(module.attention.in_proj_weight.data)
|
| 719 |
+
nn.init.zeros_(module.attention.in_proj_bias.data)
|
| 720 |
+
elif isinstance(module, SiglipModel):
|
| 721 |
+
logit_scale_init = torch.log(torch.tensor(1.0))
|
| 722 |
+
module.logit_scale.data.fill_(logit_scale_init)
|
| 723 |
+
module.logit_bias.data.zero_()
|
| 724 |
+
elif isinstance(module, SiglipForImageClassification):
|
| 725 |
+
nn.init.normal_(
|
| 726 |
+
module.classifier.weight,
|
| 727 |
+
std=self.config.vision_config.hidden_size**-0.5 * self.config.initializer_factor,
|
| 728 |
+
)
|
| 729 |
+
elif isinstance(module, (nn.Linear, nn.Conv2d)):
|
| 730 |
+
lecun_normal_(module.weight)
|
| 731 |
+
if module.bias is not None:
|
| 732 |
+
nn.init.zeros_(module.bias)
|
| 733 |
+
elif isinstance(module, nn.LayerNorm):
|
| 734 |
+
module.bias.data.zero_()
|
| 735 |
+
module.weight.data.fill_(1.0)
|
| 736 |
+
|
| 737 |
+
|
| 738 |
+
SIGLIP_START_DOCSTRING = r"""
|
| 739 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
| 740 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
| 741 |
+
etc.)
|
| 742 |
+
|
| 743 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
| 744 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
| 745 |
+
and behavior.
|
| 746 |
+
|
| 747 |
+
Parameters:
|
| 748 |
+
config ([`SiglipConfig`]): Model configuration class with all the parameters of the model.
|
| 749 |
+
Initializing with a config file does not load the weights associated with the model, only the
|
| 750 |
+
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
| 751 |
+
"""
|
| 752 |
+
|
| 753 |
+
SIGLIP_TEXT_INPUTS_DOCSTRING = r"""
|
| 754 |
+
Args:
|
| 755 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
| 756 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
| 757 |
+
it.
|
| 758 |
+
|
| 759 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 760 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 761 |
+
|
| 762 |
+
[What are input IDs?](../glossary#input-ids)
|
| 763 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 764 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
| 765 |
+
|
| 766 |
+
- 1 for tokens that are **not masked**,
|
| 767 |
+
- 0 for tokens that are **masked**.
|
| 768 |
+
|
| 769 |
+
[What are attention masks?](../glossary#attention-mask)
|
| 770 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 771 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
| 772 |
+
config.max_position_embeddings - 1]`.
|
| 773 |
+
|
| 774 |
+
[What are position IDs?](../glossary#position-ids)
|
| 775 |
+
output_attentions (`bool`, *optional*):
|
| 776 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
| 777 |
+
tensors for more detail.
|
| 778 |
+
output_hidden_states (`bool`, *optional*):
|
| 779 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
| 780 |
+
more detail.
|
| 781 |
+
return_dict (`bool`, *optional*):
|
| 782 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 783 |
+
"""
|
| 784 |
+
|
| 785 |
+
SIGLIP_VISION_INPUTS_DOCSTRING = r"""
|
| 786 |
+
Args:
|
| 787 |
+
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
|
| 788 |
+
Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
|
| 789 |
+
[`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
|
| 790 |
+
output_attentions (`bool`, *optional*):
|
| 791 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
| 792 |
+
tensors for more detail.
|
| 793 |
+
output_hidden_states (`bool`, *optional*):
|
| 794 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
| 795 |
+
more detail.
|
| 796 |
+
interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
|
| 797 |
+
Whether to interpolate the pre-trained position encodings.
|
| 798 |
+
return_dict (`bool`, *optional*):
|
| 799 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 800 |
+
"""
|
| 801 |
+
|
| 802 |
+
SIGLIP_INPUTS_DOCSTRING = r"""
|
| 803 |
+
Args:
|
| 804 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
| 805 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
| 806 |
+
it.
|
| 807 |
+
|
| 808 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 809 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 810 |
+
|
| 811 |
+
[What are input IDs?](../glossary#input-ids)
|
| 812 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 813 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
| 814 |
+
|
| 815 |
+
- 1 for tokens that are **not masked**,
|
| 816 |
+
- 0 for tokens that are **masked**.
|
| 817 |
+
|
| 818 |
+
[What are attention masks?](../glossary#attention-mask)
|
| 819 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 820 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
| 821 |
+
config.max_position_embeddings - 1]`.
|
| 822 |
+
|
| 823 |
+
[What are position IDs?](../glossary#position-ids)
|
| 824 |
+
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
|
| 825 |
+
Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
|
| 826 |
+
[`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
|
| 827 |
+
return_loss (`bool`, *optional*):
|
| 828 |
+
Whether or not to return the contrastive loss.
|
| 829 |
+
output_attentions (`bool`, *optional*):
|
| 830 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
| 831 |
+
tensors for more detail.
|
| 832 |
+
output_hidden_states (`bool`, *optional*):
|
| 833 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
| 834 |
+
more detail.
|
| 835 |
+
interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
|
| 836 |
+
Whether to interpolate the pre-trained position encodings.
|
| 837 |
+
return_dict (`bool`, *optional*):
|
| 838 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 839 |
+
"""
|
| 840 |
+
|
| 841 |
+
|
| 842 |
+
# Copied from transformers.models.altclip.modeling_altclip.AltCLIPEncoder with AltCLIP->Siglip
|
| 843 |
+
class SiglipEncoder(nn.Module):
|
| 844 |
+
"""
|
| 845 |
+
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
|
| 846 |
+
[`SiglipEncoderLayer`].
|
| 847 |
+
|
| 848 |
+
Args:
|
| 849 |
+
config: SiglipConfig
|
| 850 |
+
"""
|
| 851 |
+
|
| 852 |
+
def __init__(self, config: SiglipConfig):
|
| 853 |
+
super().__init__()
|
| 854 |
+
self.config = config
|
| 855 |
+
self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
|
| 856 |
+
self.gradient_checkpointing = False
|
| 857 |
+
|
| 858 |
+
# Ignore copy
|
| 859 |
+
def forward(
|
| 860 |
+
self,
|
| 861 |
+
inputs_embeds,
|
| 862 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 863 |
+
output_attentions: Optional[bool] = None,
|
| 864 |
+
output_hidden_states: Optional[bool] = None,
|
| 865 |
+
return_dict: Optional[bool] = None,
|
| 866 |
+
) -> Union[Tuple, BaseModelOutput]:
|
| 867 |
+
r"""
|
| 868 |
+
Args:
|
| 869 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
| 870 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
|
| 871 |
+
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
|
| 872 |
+
than the model's internal embedding lookup matrix.
|
| 873 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 874 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
| 875 |
+
|
| 876 |
+
- 1 for tokens that are **not masked**,
|
| 877 |
+
- 0 for tokens that are **masked**.
|
| 878 |
+
|
| 879 |
+
[What are attention masks?](../glossary#attention-mask)
|
| 880 |
+
output_attentions (`bool`, *optional*):
|
| 881 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 882 |
+
returned tensors for more detail.
|
| 883 |
+
output_hidden_states (`bool`, *optional*):
|
| 884 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
|
| 885 |
+
for more detail.
|
| 886 |
+
return_dict (`bool`, *optional*):
|
| 887 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 888 |
+
"""
|
| 889 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 890 |
+
output_hidden_states = (
|
| 891 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 892 |
+
)
|
| 893 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 894 |
+
|
| 895 |
+
encoder_states = () if output_hidden_states else None
|
| 896 |
+
all_attentions = () if output_attentions else None
|
| 897 |
+
|
| 898 |
+
hidden_states = inputs_embeds
|
| 899 |
+
for encoder_layer in self.layers:
|
| 900 |
+
if output_hidden_states:
|
| 901 |
+
encoder_states = encoder_states + (hidden_states,)
|
| 902 |
+
if self.gradient_checkpointing and self.training:
|
| 903 |
+
layer_outputs = self._gradient_checkpointing_func(
|
| 904 |
+
encoder_layer.__call__,
|
| 905 |
+
hidden_states,
|
| 906 |
+
attention_mask,
|
| 907 |
+
output_attentions,
|
| 908 |
+
)
|
| 909 |
+
else:
|
| 910 |
+
layer_outputs = encoder_layer(
|
| 911 |
+
hidden_states,
|
| 912 |
+
attention_mask,
|
| 913 |
+
output_attentions=output_attentions,
|
| 914 |
+
)
|
| 915 |
+
|
| 916 |
+
hidden_states = layer_outputs[0]
|
| 917 |
+
|
| 918 |
+
if output_attentions:
|
| 919 |
+
all_attentions = all_attentions + (layer_outputs[1],)
|
| 920 |
+
|
| 921 |
+
if output_hidden_states:
|
| 922 |
+
encoder_states = encoder_states + (hidden_states,)
|
| 923 |
+
|
| 924 |
+
if not return_dict:
|
| 925 |
+
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
|
| 926 |
+
return BaseModelOutput(
|
| 927 |
+
last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
|
| 928 |
+
)
|
| 929 |
+
|
| 930 |
+
|
| 931 |
+
class SiglipTextTransformer(nn.Module):
|
| 932 |
+
def __init__(self, config: SiglipTextConfig):
|
| 933 |
+
super().__init__()
|
| 934 |
+
self.config = config
|
| 935 |
+
embed_dim = config.hidden_size
|
| 936 |
+
self.embeddings = SiglipTextEmbeddings(config)
|
| 937 |
+
self.encoder = SiglipEncoder(config)
|
| 938 |
+
self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
|
| 939 |
+
|
| 940 |
+
self.head = nn.Linear(embed_dim, embed_dim)
|
| 941 |
+
self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
|
| 942 |
+
|
| 943 |
+
@add_start_docstrings_to_model_forward(SIGLIP_TEXT_INPUTS_DOCSTRING)
|
| 944 |
+
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=SiglipTextConfig)
|
| 945 |
+
def forward(
|
| 946 |
+
self,
|
| 947 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 948 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 949 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 950 |
+
output_attentions: Optional[bool] = None,
|
| 951 |
+
output_hidden_states: Optional[bool] = None,
|
| 952 |
+
return_dict: Optional[bool] = None,
|
| 953 |
+
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
| 954 |
+
r"""
|
| 955 |
+
Returns:
|
| 956 |
+
|
| 957 |
+
"""
|
| 958 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 959 |
+
output_hidden_states = (
|
| 960 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 961 |
+
)
|
| 962 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 963 |
+
|
| 964 |
+
if input_ids is None:
|
| 965 |
+
raise ValueError("You have to specify input_ids")
|
| 966 |
+
|
| 967 |
+
input_shape = input_ids.size()
|
| 968 |
+
input_ids = input_ids.view(-1, input_shape[-1])
|
| 969 |
+
|
| 970 |
+
hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)
|
| 971 |
+
|
| 972 |
+
# note: SigLIP's text model does not use a causal mask, unlike the original CLIP model.
|
| 973 |
+
# expand attention_mask
|
| 974 |
+
if attention_mask is not None and not self._use_flash_attention_2:
|
| 975 |
+
# [batch_size, seq_len] -> [batch_size, 1, tgt_seq_len, src_seq_len]
|
| 976 |
+
attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
|
| 977 |
+
|
| 978 |
+
encoder_outputs = self.encoder(
|
| 979 |
+
inputs_embeds=hidden_states,
|
| 980 |
+
attention_mask=attention_mask,
|
| 981 |
+
output_attentions=output_attentions,
|
| 982 |
+
output_hidden_states=output_hidden_states,
|
| 983 |
+
return_dict=return_dict,
|
| 984 |
+
)
|
| 985 |
+
|
| 986 |
+
last_hidden_state = encoder_outputs[0]
|
| 987 |
+
last_hidden_state = self.final_layer_norm(last_hidden_state)
|
| 988 |
+
|
| 989 |
+
# Assuming "sticky" EOS tokenization, last token is always EOS.
|
| 990 |
+
pooled_output = last_hidden_state[:, -1, :]
|
| 991 |
+
pooled_output = self.head(pooled_output)
|
| 992 |
+
|
| 993 |
+
if not return_dict:
|
| 994 |
+
return (last_hidden_state, pooled_output) + encoder_outputs[1:]
|
| 995 |
+
|
| 996 |
+
return BaseModelOutputWithPooling(
|
| 997 |
+
last_hidden_state=last_hidden_state,
|
| 998 |
+
pooler_output=pooled_output,
|
| 999 |
+
hidden_states=encoder_outputs.hidden_states,
|
| 1000 |
+
attentions=encoder_outputs.attentions,
|
| 1001 |
+
)
|
| 1002 |
+
|
| 1003 |
+
|
| 1004 |
+
@add_start_docstrings(
|
| 1005 |
+
"""The text model from SigLIP without any head or projection on top.""",
|
| 1006 |
+
SIGLIP_START_DOCSTRING,
|
| 1007 |
+
)
|
| 1008 |
+
class SiglipTextModel(SiglipPreTrainedModel):
|
| 1009 |
+
config_class = SiglipTextConfig
|
| 1010 |
+
|
| 1011 |
+
def __init__(self, config: SiglipTextConfig):
|
| 1012 |
+
super().__init__(config)
|
| 1013 |
+
self.text_model = SiglipTextTransformer(config)
|
| 1014 |
+
# Initialize weights and apply final processing
|
| 1015 |
+
self.post_init()
|
| 1016 |
+
|
| 1017 |
+
def get_input_embeddings(self) -> nn.Module:
|
| 1018 |
+
return self.text_model.embeddings.token_embedding
|
| 1019 |
+
|
| 1020 |
+
def set_input_embeddings(self, value):
|
| 1021 |
+
self.text_model.embeddings.token_embedding = value
|
| 1022 |
+
|
| 1023 |
+
@add_start_docstrings_to_model_forward(SIGLIP_TEXT_INPUTS_DOCSTRING)
|
| 1024 |
+
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=SiglipTextConfig)
|
| 1025 |
+
def forward(
|
| 1026 |
+
self,
|
| 1027 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 1028 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1029 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 1030 |
+
output_attentions: Optional[bool] = None,
|
| 1031 |
+
output_hidden_states: Optional[bool] = None,
|
| 1032 |
+
return_dict: Optional[bool] = None,
|
| 1033 |
+
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
| 1034 |
+
r"""
|
| 1035 |
+
Returns:
|
| 1036 |
+
|
| 1037 |
+
Examples:
|
| 1038 |
+
|
| 1039 |
+
```python
|
| 1040 |
+
>>> from transformers import AutoTokenizer, SiglipTextModel
|
| 1041 |
+
|
| 1042 |
+
>>> model = SiglipTextModel.from_pretrained("google/siglip-base-patch16-224")
|
| 1043 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("google/siglip-base-patch16-224")
|
| 1044 |
+
|
| 1045 |
+
>>> # important: make sure to set padding="max_length" as that's how the model was trained
|
| 1046 |
+
>>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding="max_length", return_tensors="pt")
|
| 1047 |
+
|
| 1048 |
+
>>> outputs = model(**inputs)
|
| 1049 |
+
>>> last_hidden_state = outputs.last_hidden_state
|
| 1050 |
+
>>> pooled_output = outputs.pooler_output # pooled (EOS token) states
|
| 1051 |
+
```"""
|
| 1052 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1053 |
+
|
| 1054 |
+
return self.text_model(
|
| 1055 |
+
input_ids=input_ids,
|
| 1056 |
+
attention_mask=attention_mask,
|
| 1057 |
+
position_ids=position_ids,
|
| 1058 |
+
output_attentions=output_attentions,
|
| 1059 |
+
output_hidden_states=output_hidden_states,
|
| 1060 |
+
return_dict=return_dict,
|
| 1061 |
+
)
|
| 1062 |
+
|
| 1063 |
+
|
| 1064 |
+
class SiglipVisionTransformer(nn.Module):
|
| 1065 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 1066 |
+
super().__init__()
|
| 1067 |
+
self.config = config
|
| 1068 |
+
embed_dim = config.hidden_size
|
| 1069 |
+
|
| 1070 |
+
self.embeddings = SiglipVisionEmbeddings(config)
|
| 1071 |
+
self.encoder = SiglipEncoder(config)
|
| 1072 |
+
self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
|
| 1073 |
+
self.use_head = True if not hasattr(config, "vision_use_head") else config.vision_use_head
|
| 1074 |
+
if self.use_head:
|
| 1075 |
+
self.head = SiglipMultiheadAttentionPoolingHead(config)
|
| 1076 |
+
|
| 1077 |
+
@add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
|
| 1078 |
+
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=SiglipVisionConfig)
|
| 1079 |
+
def forward(
|
| 1080 |
+
self,
|
| 1081 |
+
pixel_values,
|
| 1082 |
+
output_attentions: Optional[bool] = None,
|
| 1083 |
+
output_hidden_states: Optional[bool] = None,
|
| 1084 |
+
return_dict: Optional[bool] = None,
|
| 1085 |
+
interpolate_pos_encoding: Optional[bool] = False,
|
| 1086 |
+
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
| 1087 |
+
r"""
|
| 1088 |
+
Returns:
|
| 1089 |
+
|
| 1090 |
+
"""
|
| 1091 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1092 |
+
output_hidden_states = (
|
| 1093 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1094 |
+
)
|
| 1095 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1096 |
+
|
| 1097 |
+
hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)
|
| 1098 |
+
|
| 1099 |
+
encoder_outputs = self.encoder(
|
| 1100 |
+
inputs_embeds=hidden_states,
|
| 1101 |
+
output_attentions=output_attentions,
|
| 1102 |
+
output_hidden_states=output_hidden_states,
|
| 1103 |
+
return_dict=return_dict,
|
| 1104 |
+
)
|
| 1105 |
+
|
| 1106 |
+
last_hidden_state = encoder_outputs[0]
|
| 1107 |
+
last_hidden_state = self.post_layernorm(last_hidden_state)
|
| 1108 |
+
|
| 1109 |
+
pooler_output = self.head(last_hidden_state) if self.use_head else None
|
| 1110 |
+
if not return_dict:
|
| 1111 |
+
return (last_hidden_state, pooler_output) + encoder_outputs[1:]
|
| 1112 |
+
|
| 1113 |
+
return BaseModelOutputWithPooling(
|
| 1114 |
+
last_hidden_state=last_hidden_state,
|
| 1115 |
+
pooler_output=pooler_output,
|
| 1116 |
+
hidden_states=encoder_outputs.hidden_states,
|
| 1117 |
+
attentions=encoder_outputs.attentions,
|
| 1118 |
+
)
|
| 1119 |
+
|
| 1120 |
+
|
| 1121 |
+
class SiglipMultiheadAttentionPoolingHead(nn.Module):
|
| 1122 |
+
"""Multihead Attention Pooling."""
|
| 1123 |
+
|
| 1124 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 1125 |
+
super().__init__()
|
| 1126 |
+
|
| 1127 |
+
self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
|
| 1128 |
+
self.attention = torch.nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
|
| 1129 |
+
self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
| 1130 |
+
self.mlp = SiglipMLP(config)
|
| 1131 |
+
|
| 1132 |
+
def forward(self, hidden_state):
|
| 1133 |
+
batch_size = hidden_state.shape[0]
|
| 1134 |
+
probe = self.probe.repeat(batch_size, 1, 1)
|
| 1135 |
+
|
| 1136 |
+
hidden_state = self.attention(probe, hidden_state, hidden_state)[0]
|
| 1137 |
+
|
| 1138 |
+
residual = hidden_state
|
| 1139 |
+
hidden_state = self.layernorm(hidden_state)
|
| 1140 |
+
hidden_state = residual + self.mlp(hidden_state)
|
| 1141 |
+
|
| 1142 |
+
return hidden_state[:, 0]
|
| 1143 |
+
|
| 1144 |
+
|
| 1145 |
+
@add_start_docstrings(
|
| 1146 |
+
"""The vision model from SigLIP without any head or projection on top.""",
|
| 1147 |
+
SIGLIP_START_DOCSTRING,
|
| 1148 |
+
)
|
| 1149 |
+
class SiglipVisionModel(SiglipPreTrainedModel):
|
| 1150 |
+
config_class = SiglipVisionConfig
|
| 1151 |
+
main_input_name = "pixel_values"
|
| 1152 |
+
|
| 1153 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 1154 |
+
super().__init__(config)
|
| 1155 |
+
|
| 1156 |
+
self.vision_model = SiglipVisionTransformer(config)
|
| 1157 |
+
|
| 1158 |
+
# Initialize weights and apply final processing
|
| 1159 |
+
self.post_init()
|
| 1160 |
+
|
| 1161 |
+
def get_input_embeddings(self) -> nn.Module:
|
| 1162 |
+
return self.vision_model.embeddings.patch_embedding
|
| 1163 |
+
|
| 1164 |
+
@add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
|
| 1165 |
+
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=SiglipVisionConfig)
|
| 1166 |
+
def forward(
|
| 1167 |
+
self,
|
| 1168 |
+
pixel_values,
|
| 1169 |
+
output_attentions: Optional[bool] = None,
|
| 1170 |
+
output_hidden_states: Optional[bool] = None,
|
| 1171 |
+
return_dict: Optional[bool] = None,
|
| 1172 |
+
interpolate_pos_encoding: bool = False,
|
| 1173 |
+
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
| 1174 |
+
r"""
|
| 1175 |
+
Returns:
|
| 1176 |
+
|
| 1177 |
+
Examples:
|
| 1178 |
+
|
| 1179 |
+
```python
|
| 1180 |
+
>>> from PIL import Image
|
| 1181 |
+
>>> import requests
|
| 1182 |
+
>>> from transformers import AutoProcessor, SiglipVisionModel
|
| 1183 |
+
|
| 1184 |
+
>>> model = SiglipVisionModel.from_pretrained("google/siglip-base-patch16-224")
|
| 1185 |
+
>>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
|
| 1186 |
+
|
| 1187 |
+
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
| 1188 |
+
>>> image = Image.open(requests.get(url, stream=True).raw)
|
| 1189 |
+
|
| 1190 |
+
>>> inputs = processor(images=image, return_tensors="pt")
|
| 1191 |
+
|
| 1192 |
+
>>> outputs = model(**inputs)
|
| 1193 |
+
>>> last_hidden_state = outputs.last_hidden_state
|
| 1194 |
+
>>> pooled_output = outputs.pooler_output # pooled features
|
| 1195 |
+
```"""
|
| 1196 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1197 |
+
|
| 1198 |
+
return self.vision_model(
|
| 1199 |
+
pixel_values=pixel_values,
|
| 1200 |
+
output_attentions=output_attentions,
|
| 1201 |
+
output_hidden_states=output_hidden_states,
|
| 1202 |
+
return_dict=return_dict,
|
| 1203 |
+
interpolate_pos_encoding=interpolate_pos_encoding,
|
| 1204 |
+
)
|
| 1205 |
+
|
| 1206 |
+
|
| 1207 |
+
@add_start_docstrings(SIGLIP_START_DOCSTRING)
|
| 1208 |
+
class SiglipModel(SiglipPreTrainedModel):
|
| 1209 |
+
config_class = SiglipConfig
|
| 1210 |
+
|
| 1211 |
+
def __init__(self, config: SiglipConfig):
|
| 1212 |
+
super().__init__(config)
|
| 1213 |
+
|
| 1214 |
+
if not isinstance(config.text_config, SiglipTextConfig):
|
| 1215 |
+
raise TypeError(
|
| 1216 |
+
"config.text_config is expected to be of type SiglipTextConfig but is of type"
|
| 1217 |
+
f" {type(config.text_config)}."
|
| 1218 |
+
)
|
| 1219 |
+
|
| 1220 |
+
if not isinstance(config.vision_config, SiglipVisionConfig):
|
| 1221 |
+
raise TypeError(
|
| 1222 |
+
"config.vision_config is expected to be of type SiglipVisionConfig but is of type"
|
| 1223 |
+
f" {type(config.vision_config)}."
|
| 1224 |
+
)
|
| 1225 |
+
|
| 1226 |
+
text_config = config.text_config
|
| 1227 |
+
vision_config = config.vision_config
|
| 1228 |
+
|
| 1229 |
+
# First, initialize the text and vision models with proper attention implementation
|
| 1230 |
+
text_model = SiglipTextModel._from_config(text_config)
|
| 1231 |
+
vision_model = SiglipVisionModel._from_config(vision_config)
|
| 1232 |
+
|
| 1233 |
+
# Second, get the text and vision submodules (for backward compatibility)
|
| 1234 |
+
self.text_model = text_model.text_model
|
| 1235 |
+
self.vision_model = vision_model.vision_model
|
| 1236 |
+
|
| 1237 |
+
self.logit_scale = nn.Parameter(torch.randn(1))
|
| 1238 |
+
self.logit_bias = nn.Parameter(torch.randn(1))
|
| 1239 |
+
|
| 1240 |
+
# Initialize weights and apply final processing
|
| 1241 |
+
self.post_init()
|
| 1242 |
+
|
| 1243 |
+
@add_start_docstrings_to_model_forward(SIGLIP_TEXT_INPUTS_DOCSTRING)
|
| 1244 |
+
def get_text_features(
|
| 1245 |
+
self,
|
| 1246 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 1247 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1248 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 1249 |
+
output_attentions: Optional[bool] = None,
|
| 1250 |
+
output_hidden_states: Optional[bool] = None,
|
| 1251 |
+
return_dict: Optional[bool] = None,
|
| 1252 |
+
) -> torch.FloatTensor:
|
| 1253 |
+
r"""
|
| 1254 |
+
Returns:
|
| 1255 |
+
text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
|
| 1256 |
+
applying the projection layer to the pooled output of [`SiglipTextModel`].
|
| 1257 |
+
|
| 1258 |
+
Examples:
|
| 1259 |
+
|
| 1260 |
+
```python
|
| 1261 |
+
>>> from transformers import AutoTokenizer, AutoModel
|
| 1262 |
+
>>> import torch
|
| 1263 |
+
|
| 1264 |
+
>>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
|
| 1265 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("google/siglip-base-patch16-224")
|
| 1266 |
+
|
| 1267 |
+
>>> # important: make sure to set padding="max_length" as that's how the model was trained
|
| 1268 |
+
>>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding="max_length", return_tensors="pt")
|
| 1269 |
+
>>> with torch.no_grad():
|
| 1270 |
+
... text_features = model.get_text_features(**inputs)
|
| 1271 |
+
```"""
|
| 1272 |
+
# Use SigLIP model's config for some fields (if specified) instead of those of vision & text components.
|
| 1273 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1274 |
+
output_hidden_states = (
|
| 1275 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1276 |
+
)
|
| 1277 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1278 |
+
|
| 1279 |
+
text_outputs = self.text_model(
|
| 1280 |
+
input_ids=input_ids,
|
| 1281 |
+
attention_mask=attention_mask,
|
| 1282 |
+
position_ids=position_ids,
|
| 1283 |
+
output_attentions=output_attentions,
|
| 1284 |
+
output_hidden_states=output_hidden_states,
|
| 1285 |
+
return_dict=return_dict,
|
| 1286 |
+
)
|
| 1287 |
+
|
| 1288 |
+
pooled_output = text_outputs[1]
|
| 1289 |
+
|
| 1290 |
+
return pooled_output
|
| 1291 |
+
|
| 1292 |
+
@add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
|
| 1293 |
+
def get_image_features(
|
| 1294 |
+
self,
|
| 1295 |
+
pixel_values: Optional[torch.FloatTensor] = None,
|
| 1296 |
+
output_attentions: Optional[bool] = None,
|
| 1297 |
+
output_hidden_states: Optional[bool] = None,
|
| 1298 |
+
return_dict: Optional[bool] = None,
|
| 1299 |
+
interpolate_pos_encoding: bool = False,
|
| 1300 |
+
) -> torch.FloatTensor:
|
| 1301 |
+
r"""
|
| 1302 |
+
Returns:
|
| 1303 |
+
image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
|
| 1304 |
+
applying the projection layer to the pooled output of [`SiglipVisionModel`].
|
| 1305 |
+
|
| 1306 |
+
Examples:
|
| 1307 |
+
|
| 1308 |
+
```python
|
| 1309 |
+
>>> from PIL import Image
|
| 1310 |
+
>>> import requests
|
| 1311 |
+
>>> from transformers import AutoProcessor, AutoModel
|
| 1312 |
+
>>> import torch
|
| 1313 |
+
|
| 1314 |
+
>>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
|
| 1315 |
+
>>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
|
| 1316 |
+
|
| 1317 |
+
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
| 1318 |
+
>>> image = Image.open(requests.get(url, stream=True).raw)
|
| 1319 |
+
|
| 1320 |
+
>>> inputs = processor(images=image, return_tensors="pt")
|
| 1321 |
+
|
| 1322 |
+
>>> with torch.no_grad():
|
| 1323 |
+
... image_features = model.get_image_features(**inputs)
|
| 1324 |
+
```"""
|
| 1325 |
+
# Use SiglipModel's config for some fields (if specified) instead of those of vision & text components.
|
| 1326 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1327 |
+
output_hidden_states = (
|
| 1328 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1329 |
+
)
|
| 1330 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1331 |
+
|
| 1332 |
+
vision_outputs = self.vision_model(
|
| 1333 |
+
pixel_values=pixel_values,
|
| 1334 |
+
output_attentions=output_attentions,
|
| 1335 |
+
output_hidden_states=output_hidden_states,
|
| 1336 |
+
return_dict=return_dict,
|
| 1337 |
+
interpolate_pos_encoding=interpolate_pos_encoding,
|
| 1338 |
+
)
|
| 1339 |
+
|
| 1340 |
+
pooled_output = vision_outputs[1]
|
| 1341 |
+
|
| 1342 |
+
return pooled_output
|
| 1343 |
+
|
| 1344 |
+
@add_start_docstrings_to_model_forward(SIGLIP_INPUTS_DOCSTRING)
|
| 1345 |
+
@replace_return_docstrings(output_type=SiglipOutput, config_class=SiglipConfig)
|
| 1346 |
+
def forward(
|
| 1347 |
+
self,
|
| 1348 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 1349 |
+
pixel_values: Optional[torch.FloatTensor] = None,
|
| 1350 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1351 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1352 |
+
return_loss: Optional[bool] = None,
|
| 1353 |
+
output_attentions: Optional[bool] = None,
|
| 1354 |
+
output_hidden_states: Optional[bool] = None,
|
| 1355 |
+
return_dict: Optional[bool] = None,
|
| 1356 |
+
interpolate_pos_encoding: bool = False,
|
| 1357 |
+
) -> Union[Tuple, SiglipOutput]:
|
| 1358 |
+
r"""
|
| 1359 |
+
Returns:
|
| 1360 |
+
|
| 1361 |
+
Examples:
|
| 1362 |
+
|
| 1363 |
+
```python
|
| 1364 |
+
>>> from PIL import Image
|
| 1365 |
+
>>> import requests
|
| 1366 |
+
>>> from transformers import AutoProcessor, AutoModel
|
| 1367 |
+
>>> import torch
|
| 1368 |
+
|
| 1369 |
+
>>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
|
| 1370 |
+
>>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
|
| 1371 |
+
|
| 1372 |
+
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
| 1373 |
+
>>> image = Image.open(requests.get(url, stream=True).raw)
|
| 1374 |
+
|
| 1375 |
+
>>> texts = ["a photo of 2 cats", "a photo of 2 dogs"]
|
| 1376 |
+
>>> # important: we pass `padding=max_length` since the model was trained with this
|
| 1377 |
+
>>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt")
|
| 1378 |
+
|
| 1379 |
+
>>> with torch.no_grad():
|
| 1380 |
+
... outputs = model(**inputs)
|
| 1381 |
+
|
| 1382 |
+
>>> logits_per_image = outputs.logits_per_image
|
| 1383 |
+
>>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
|
| 1384 |
+
>>> print(f"{probs[0][0]:.1%} that image 0 is '{texts[0]}'")
|
| 1385 |
+
31.9% that image 0 is 'a photo of 2 cats'
|
| 1386 |
+
```"""
|
| 1387 |
+
# Use SigLIP model's config for some fields (if specified) instead of those of vision & text components.
|
| 1388 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1389 |
+
output_hidden_states = (
|
| 1390 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1391 |
+
)
|
| 1392 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1393 |
+
|
| 1394 |
+
vision_outputs = self.vision_model(
|
| 1395 |
+
pixel_values=pixel_values,
|
| 1396 |
+
output_attentions=output_attentions,
|
| 1397 |
+
output_hidden_states=output_hidden_states,
|
| 1398 |
+
return_dict=return_dict,
|
| 1399 |
+
interpolate_pos_encoding=interpolate_pos_encoding,
|
| 1400 |
+
)
|
| 1401 |
+
|
| 1402 |
+
text_outputs = self.text_model(
|
| 1403 |
+
input_ids=input_ids,
|
| 1404 |
+
attention_mask=attention_mask,
|
| 1405 |
+
position_ids=position_ids,
|
| 1406 |
+
output_attentions=output_attentions,
|
| 1407 |
+
output_hidden_states=output_hidden_states,
|
| 1408 |
+
return_dict=return_dict,
|
| 1409 |
+
)
|
| 1410 |
+
|
| 1411 |
+
image_embeds = vision_outputs[1]
|
| 1412 |
+
text_embeds = text_outputs[1]
|
| 1413 |
+
|
| 1414 |
+
# normalized features
|
| 1415 |
+
image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
|
| 1416 |
+
text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
|
| 1417 |
+
|
| 1418 |
+
# cosine similarity as logits
|
| 1419 |
+
logits_per_text = (
|
| 1420 |
+
torch.matmul(text_embeds, image_embeds.t().to(text_embeds.device)) * self.logit_scale.exp()
|
| 1421 |
+
+ self.logit_bias
|
| 1422 |
+
)
|
| 1423 |
+
logits_per_image = logits_per_text.t()
|
| 1424 |
+
|
| 1425 |
+
loss = None
|
| 1426 |
+
if return_loss:
|
| 1427 |
+
# Adapted from https://github.com/google-research/big_vision/blob/01edb81a4716f93a48be43b3a4af14e29cdb3a7f/big_vision/trainers/proj/image_text/siglip.py#L287
|
| 1428 |
+
eye = torch.eye(logits_per_text.size(0), device=logits_per_text.device)
|
| 1429 |
+
m1_diag1 = -torch.ones_like(logits_per_text) + 2 * eye
|
| 1430 |
+
loglik = torch.nn.functional.logsigmoid(m1_diag1 * logits_per_text)
|
| 1431 |
+
nll = -torch.sum(loglik, dim=-1)
|
| 1432 |
+
loss = nll.mean()
|
| 1433 |
+
|
| 1434 |
+
if not return_dict:
|
| 1435 |
+
output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
|
| 1436 |
+
return ((loss,) + output) if loss is not None else output
|
| 1437 |
+
|
| 1438 |
+
return SiglipOutput(
|
| 1439 |
+
loss=loss,
|
| 1440 |
+
logits_per_image=logits_per_image,
|
| 1441 |
+
logits_per_text=logits_per_text,
|
| 1442 |
+
text_embeds=text_embeds,
|
| 1443 |
+
image_embeds=image_embeds,
|
| 1444 |
+
text_model_output=text_outputs,
|
| 1445 |
+
vision_model_output=vision_outputs,
|
| 1446 |
+
)
|
| 1447 |
+
|
| 1448 |
+
|
| 1449 |
+
@add_start_docstrings(
|
| 1450 |
+
"""
|
| 1451 |
+
SigLIP vision encoder with an image classification head on top (a linear layer on top of the pooled final hidden states of
|
| 1452 |
+
the patch tokens) e.g. for ImageNet.
|
| 1453 |
+
""",
|
| 1454 |
+
SIGLIP_START_DOCSTRING,
|
| 1455 |
+
)
|
| 1456 |
+
class SiglipForImageClassification(SiglipPreTrainedModel):
|
| 1457 |
+
main_input_name = "pixel_values"
|
| 1458 |
+
|
| 1459 |
+
def __init__(self, config: SiglipConfig) -> None:
|
| 1460 |
+
super().__init__(config)
|
| 1461 |
+
|
| 1462 |
+
self.num_labels = config.num_labels
|
| 1463 |
+
|
| 1464 |
+
# Create the vision model with proper attention
|
| 1465 |
+
# and take only vision_model submodule (for backward compatibility)
|
| 1466 |
+
vision_model = SiglipVisionModel._from_config(config.vision_config)
|
| 1467 |
+
self.vision_model = vision_model.vision_model
|
| 1468 |
+
|
| 1469 |
+
# Classifier head
|
| 1470 |
+
self.classifier = (
|
| 1471 |
+
nn.Linear(config.vision_config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity()
|
| 1472 |
+
)
|
| 1473 |
+
|
| 1474 |
+
# Initialize weights and apply final processing
|
| 1475 |
+
self.post_init()
|
| 1476 |
+
|
| 1477 |
+
@add_start_docstrings_to_model_forward(SIGLIP_INPUTS_DOCSTRING)
|
| 1478 |
+
@replace_return_docstrings(output_type=ImageClassifierOutput, config_class=_CONFIG_FOR_DOC)
|
| 1479 |
+
def forward(
|
| 1480 |
+
self,
|
| 1481 |
+
pixel_values: Optional[torch.Tensor] = None,
|
| 1482 |
+
labels: Optional[torch.Tensor] = None,
|
| 1483 |
+
output_attentions: Optional[bool] = None,
|
| 1484 |
+
output_hidden_states: Optional[bool] = None,
|
| 1485 |
+
return_dict: Optional[bool] = None,
|
| 1486 |
+
interpolate_pos_encoding: bool = False,
|
| 1487 |
+
) -> Union[tuple, ImageClassifierOutput]:
|
| 1488 |
+
r"""
|
| 1489 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 1490 |
+
Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
|
| 1491 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
| 1492 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
| 1493 |
+
|
| 1494 |
+
Returns:
|
| 1495 |
+
|
| 1496 |
+
Examples:
|
| 1497 |
+
|
| 1498 |
+
```python
|
| 1499 |
+
>>> from transformers import AutoImageProcessor, SiglipForImageClassification
|
| 1500 |
+
>>> import torch
|
| 1501 |
+
>>> from PIL import Image
|
| 1502 |
+
>>> import requests
|
| 1503 |
+
|
| 1504 |
+
>>> torch.manual_seed(3) # doctest: +IGNORE_RESULT
|
| 1505 |
+
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
| 1506 |
+
>>> image = Image.open(requests.get(url, stream=True).raw)
|
| 1507 |
+
|
| 1508 |
+
>>> # note: we are loading a `SiglipModel` from the hub here,
|
| 1509 |
+
>>> # so the head will be randomly initialized, hence the predictions will be random if seed is not set above.
|
| 1510 |
+
>>> image_processor = AutoImageProcessor.from_pretrained("google/siglip-base-patch16-224")
|
| 1511 |
+
>>> model = SiglipForImageClassification.from_pretrained("google/siglip-base-patch16-224")
|
| 1512 |
+
|
| 1513 |
+
>>> inputs = image_processor(images=image, return_tensors="pt")
|
| 1514 |
+
>>> outputs = model(**inputs)
|
| 1515 |
+
>>> logits = outputs.logits
|
| 1516 |
+
>>> # model predicts one of the two classes
|
| 1517 |
+
>>> predicted_class_idx = logits.argmax(-1).item()
|
| 1518 |
+
>>> print("Predicted class:", model.config.id2label[predicted_class_idx])
|
| 1519 |
+
Predicted class: LABEL_1
|
| 1520 |
+
```"""
|
| 1521 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1522 |
+
output_hidden_states = (
|
| 1523 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1524 |
+
)
|
| 1525 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1526 |
+
|
| 1527 |
+
outputs = self.vision_model(
|
| 1528 |
+
pixel_values,
|
| 1529 |
+
output_attentions=output_attentions,
|
| 1530 |
+
output_hidden_states=output_hidden_states,
|
| 1531 |
+
return_dict=return_dict,
|
| 1532 |
+
interpolate_pos_encoding=interpolate_pos_encoding,
|
| 1533 |
+
)
|
| 1534 |
+
|
| 1535 |
+
sequence_output = outputs[0]
|
| 1536 |
+
|
| 1537 |
+
# average pool the patch tokens
|
| 1538 |
+
sequence_output = torch.mean(sequence_output, dim=1)
|
| 1539 |
+
# apply classifier
|
| 1540 |
+
logits = self.classifier(sequence_output)
|
| 1541 |
+
|
| 1542 |
+
loss = None
|
| 1543 |
+
if labels is not None:
|
| 1544 |
+
# move labels to correct device to enable model parallelism
|
| 1545 |
+
labels = labels.to(logits.device)
|
| 1546 |
+
if self.config.problem_type is None:
|
| 1547 |
+
if self.num_labels == 1:
|
| 1548 |
+
self.config.problem_type = "regression"
|
| 1549 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
| 1550 |
+
self.config.problem_type = "single_label_classification"
|
| 1551 |
+
else:
|
| 1552 |
+
self.config.problem_type = "multi_label_classification"
|
| 1553 |
+
|
| 1554 |
+
if self.config.problem_type == "regression":
|
| 1555 |
+
loss_fct = MSELoss()
|
| 1556 |
+
if self.num_labels == 1:
|
| 1557 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
| 1558 |
+
else:
|
| 1559 |
+
loss = loss_fct(logits, labels)
|
| 1560 |
+
elif self.config.problem_type == "single_label_classification":
|
| 1561 |
+
loss_fct = CrossEntropyLoss()
|
| 1562 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
| 1563 |
+
elif self.config.problem_type == "multi_label_classification":
|
| 1564 |
+
loss_fct = BCEWithLogitsLoss()
|
| 1565 |
+
loss = loss_fct(logits, labels)
|
| 1566 |
+
|
| 1567 |
+
if not return_dict:
|
| 1568 |
+
output = (logits,) + outputs[2:]
|
| 1569 |
+
return ((loss,) + output) if loss is not None else output
|
| 1570 |
+
|
| 1571 |
+
return ImageClassifierOutput(
|
| 1572 |
+
loss=loss,
|
| 1573 |
+
logits=logits,
|
| 1574 |
+
hidden_states=outputs.hidden_states,
|
| 1575 |
+
attentions=outputs.attentions,
|
| 1576 |
+
)
|
| 1577 |
+
|
| 1578 |
+
|
| 1579 |
+
__all__ = [
|
| 1580 |
+
"SiglipModel",
|
| 1581 |
+
"SiglipPreTrainedModel",
|
| 1582 |
+
"SiglipTextModel",
|
| 1583 |
+
"SiglipVisionModel",
|
| 1584 |
+
"SiglipForImageClassification",
|
| 1585 |
+
]
|
preprocessor_config.json
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"do_convert_rgb": true,
|
| 3 |
+
"do_normalize": true,
|
| 4 |
+
"do_rescale": true,
|
| 5 |
+
"do_resize": true,
|
| 6 |
+
"image_mean": [
|
| 7 |
+
0.5,
|
| 8 |
+
0.5,
|
| 9 |
+
0.5
|
| 10 |
+
],
|
| 11 |
+
"image_processor_type": "SiglipImageProcessor",
|
| 12 |
+
"image_std": [
|
| 13 |
+
0.5,
|
| 14 |
+
0.5,
|
| 15 |
+
0.5
|
| 16 |
+
],
|
| 17 |
+
"processor_class": "SiglipProcessor",
|
| 18 |
+
"resample": 3,
|
| 19 |
+
"rescale_factor": 0.00392156862745098,
|
| 20 |
+
"size": {
|
| 21 |
+
"height": 384,
|
| 22 |
+
"width": 384
|
| 23 |
+
}
|
| 24 |
+
}
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"additional_special_tokens": [
|
| 3 |
+
{
|
| 4 |
+
"content": "<|endofchunk|>",
|
| 5 |
+
"lstrip": false,
|
| 6 |
+
"normalized": false,
|
| 7 |
+
"rstrip": false,
|
| 8 |
+
"single_word": false
|
| 9 |
+
},
|
| 10 |
+
{
|
| 11 |
+
"content": "<s>",
|
| 12 |
+
"lstrip": false,
|
| 13 |
+
"normalized": false,
|
| 14 |
+
"rstrip": false,
|
| 15 |
+
"single_word": false
|
| 16 |
+
},
|
| 17 |
+
{
|
| 18 |
+
"content": "<|pad|>",
|
| 19 |
+
"lstrip": false,
|
| 20 |
+
"normalized": false,
|
| 21 |
+
"rstrip": false,
|
| 22 |
+
"single_word": false
|
| 23 |
+
}
|
| 24 |
+
],
|
| 25 |
+
"bos_token": "<s>",
|
| 26 |
+
"eos_token": {
|
| 27 |
+
"content": "<|im_end|>",
|
| 28 |
+
"lstrip": false,
|
| 29 |
+
"normalized": false,
|
| 30 |
+
"rstrip": false,
|
| 31 |
+
"single_word": false
|
| 32 |
+
},
|
| 33 |
+
"pad_token": "<|pad|>"
|
| 34 |
+
}
|
tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:da321834704c685bd2e638824b76f3beb841556558a6df5b8a43bf2ad31f7b94
|
| 3 |
+
size 11423209
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,253 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": false,
|
| 3 |
+
"add_prefix_space": false,
|
| 4 |
+
"added_tokens_decoder": {
|
| 5 |
+
"128245": {
|
| 6 |
+
"content": "<s>",
|
| 7 |
+
"lstrip": false,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false,
|
| 11 |
+
"special": true
|
| 12 |
+
},
|
| 13 |
+
"151643": {
|
| 14 |
+
"content": "<|endoftext|>",
|
| 15 |
+
"lstrip": false,
|
| 16 |
+
"normalized": false,
|
| 17 |
+
"rstrip": false,
|
| 18 |
+
"single_word": false,
|
| 19 |
+
"special": true
|
| 20 |
+
},
|
| 21 |
+
"151644": {
|
| 22 |
+
"content": "<|im_start|>",
|
| 23 |
+
"lstrip": false,
|
| 24 |
+
"normalized": false,
|
| 25 |
+
"rstrip": false,
|
| 26 |
+
"single_word": false,
|
| 27 |
+
"special": true
|
| 28 |
+
},
|
| 29 |
+
"151645": {
|
| 30 |
+
"content": "<|im_end|>",
|
| 31 |
+
"lstrip": false,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false,
|
| 35 |
+
"special": true
|
| 36 |
+
},
|
| 37 |
+
"151646": {
|
| 38 |
+
"content": "<|object_ref_start|>",
|
| 39 |
+
"lstrip": false,
|
| 40 |
+
"normalized": false,
|
| 41 |
+
"rstrip": false,
|
| 42 |
+
"single_word": false,
|
| 43 |
+
"special": true
|
| 44 |
+
},
|
| 45 |
+
"151647": {
|
| 46 |
+
"content": "<|object_ref_end|>",
|
| 47 |
+
"lstrip": false,
|
| 48 |
+
"normalized": false,
|
| 49 |
+
"rstrip": false,
|
| 50 |
+
"single_word": false,
|
| 51 |
+
"special": true
|
| 52 |
+
},
|
| 53 |
+
"151648": {
|
| 54 |
+
"content": "<|box_start|>",
|
| 55 |
+
"lstrip": false,
|
| 56 |
+
"normalized": false,
|
| 57 |
+
"rstrip": false,
|
| 58 |
+
"single_word": false,
|
| 59 |
+
"special": true
|
| 60 |
+
},
|
| 61 |
+
"151649": {
|
| 62 |
+
"content": "<|box_end|>",
|
| 63 |
+
"lstrip": false,
|
| 64 |
+
"normalized": false,
|
| 65 |
+
"rstrip": false,
|
| 66 |
+
"single_word": false,
|
| 67 |
+
"special": true
|
| 68 |
+
},
|
| 69 |
+
"151650": {
|
| 70 |
+
"content": "<|quad_start|>",
|
| 71 |
+
"lstrip": false,
|
| 72 |
+
"normalized": false,
|
| 73 |
+
"rstrip": false,
|
| 74 |
+
"single_word": false,
|
| 75 |
+
"special": true
|
| 76 |
+
},
|
| 77 |
+
"151651": {
|
| 78 |
+
"content": "<|quad_end|>",
|
| 79 |
+
"lstrip": false,
|
| 80 |
+
"normalized": false,
|
| 81 |
+
"rstrip": false,
|
| 82 |
+
"single_word": false,
|
| 83 |
+
"special": true
|
| 84 |
+
},
|
| 85 |
+
"151652": {
|
| 86 |
+
"content": "<|vision_start|>",
|
| 87 |
+
"lstrip": false,
|
| 88 |
+
"normalized": false,
|
| 89 |
+
"rstrip": false,
|
| 90 |
+
"single_word": false,
|
| 91 |
+
"special": true
|
| 92 |
+
},
|
| 93 |
+
"151653": {
|
| 94 |
+
"content": "<|vision_end|>",
|
| 95 |
+
"lstrip": false,
|
| 96 |
+
"normalized": false,
|
| 97 |
+
"rstrip": false,
|
| 98 |
+
"single_word": false,
|
| 99 |
+
"special": true
|
| 100 |
+
},
|
| 101 |
+
"151654": {
|
| 102 |
+
"content": "<|vision_pad|>",
|
| 103 |
+
"lstrip": false,
|
| 104 |
+
"normalized": false,
|
| 105 |
+
"rstrip": false,
|
| 106 |
+
"single_word": false,
|
| 107 |
+
"special": true
|
| 108 |
+
},
|
| 109 |
+
"151655": {
|
| 110 |
+
"content": "<|image_pad|>",
|
| 111 |
+
"lstrip": false,
|
| 112 |
+
"normalized": false,
|
| 113 |
+
"rstrip": false,
|
| 114 |
+
"single_word": false,
|
| 115 |
+
"special": true
|
| 116 |
+
},
|
| 117 |
+
"151656": {
|
| 118 |
+
"content": "<|video_pad|>",
|
| 119 |
+
"lstrip": false,
|
| 120 |
+
"normalized": false,
|
| 121 |
+
"rstrip": false,
|
| 122 |
+
"single_word": false,
|
| 123 |
+
"special": true
|
| 124 |
+
},
|
| 125 |
+
"151657": {
|
| 126 |
+
"content": "<tool_call>",
|
| 127 |
+
"lstrip": false,
|
| 128 |
+
"normalized": false,
|
| 129 |
+
"rstrip": false,
|
| 130 |
+
"single_word": false,
|
| 131 |
+
"special": false
|
| 132 |
+
},
|
| 133 |
+
"151658": {
|
| 134 |
+
"content": "</tool_call>",
|
| 135 |
+
"lstrip": false,
|
| 136 |
+
"normalized": false,
|
| 137 |
+
"rstrip": false,
|
| 138 |
+
"single_word": false,
|
| 139 |
+
"special": false
|
| 140 |
+
},
|
| 141 |
+
"151659": {
|
| 142 |
+
"content": "<|fim_prefix|>",
|
| 143 |
+
"lstrip": false,
|
| 144 |
+
"normalized": false,
|
| 145 |
+
"rstrip": false,
|
| 146 |
+
"single_word": false,
|
| 147 |
+
"special": false
|
| 148 |
+
},
|
| 149 |
+
"151660": {
|
| 150 |
+
"content": "<|fim_middle|>",
|
| 151 |
+
"lstrip": false,
|
| 152 |
+
"normalized": false,
|
| 153 |
+
"rstrip": false,
|
| 154 |
+
"single_word": false,
|
| 155 |
+
"special": false
|
| 156 |
+
},
|
| 157 |
+
"151661": {
|
| 158 |
+
"content": "<|fim_suffix|>",
|
| 159 |
+
"lstrip": false,
|
| 160 |
+
"normalized": false,
|
| 161 |
+
"rstrip": false,
|
| 162 |
+
"single_word": false,
|
| 163 |
+
"special": false
|
| 164 |
+
},
|
| 165 |
+
"151662": {
|
| 166 |
+
"content": "<|fim_pad|>",
|
| 167 |
+
"lstrip": false,
|
| 168 |
+
"normalized": false,
|
| 169 |
+
"rstrip": false,
|
| 170 |
+
"single_word": false,
|
| 171 |
+
"special": false
|
| 172 |
+
},
|
| 173 |
+
"151663": {
|
| 174 |
+
"content": "<|repo_name|>",
|
| 175 |
+
"lstrip": false,
|
| 176 |
+
"normalized": false,
|
| 177 |
+
"rstrip": false,
|
| 178 |
+
"single_word": false,
|
| 179 |
+
"special": false
|
| 180 |
+
},
|
| 181 |
+
"151664": {
|
| 182 |
+
"content": "<|file_sep|>",
|
| 183 |
+
"lstrip": false,
|
| 184 |
+
"normalized": false,
|
| 185 |
+
"rstrip": false,
|
| 186 |
+
"single_word": false,
|
| 187 |
+
"special": false
|
| 188 |
+
},
|
| 189 |
+
"151665": {
|
| 190 |
+
"content": "<tool_response>",
|
| 191 |
+
"lstrip": false,
|
| 192 |
+
"normalized": false,
|
| 193 |
+
"rstrip": false,
|
| 194 |
+
"single_word": false,
|
| 195 |
+
"special": false
|
| 196 |
+
},
|
| 197 |
+
"151666": {
|
| 198 |
+
"content": "</tool_response>",
|
| 199 |
+
"lstrip": false,
|
| 200 |
+
"normalized": false,
|
| 201 |
+
"rstrip": false,
|
| 202 |
+
"single_word": false,
|
| 203 |
+
"special": false
|
| 204 |
+
},
|
| 205 |
+
"151667": {
|
| 206 |
+
"content": "<think>",
|
| 207 |
+
"lstrip": false,
|
| 208 |
+
"normalized": false,
|
| 209 |
+
"rstrip": false,
|
| 210 |
+
"single_word": false,
|
| 211 |
+
"special": false
|
| 212 |
+
},
|
| 213 |
+
"151668": {
|
| 214 |
+
"content": "</think>",
|
| 215 |
+
"lstrip": false,
|
| 216 |
+
"normalized": false,
|
| 217 |
+
"rstrip": false,
|
| 218 |
+
"single_word": false,
|
| 219 |
+
"special": false
|
| 220 |
+
},
|
| 221 |
+
"151669": {
|
| 222 |
+
"content": "<|endofchunk|>",
|
| 223 |
+
"lstrip": false,
|
| 224 |
+
"normalized": false,
|
| 225 |
+
"rstrip": false,
|
| 226 |
+
"single_word": false,
|
| 227 |
+
"special": true
|
| 228 |
+
},
|
| 229 |
+
"151670": {
|
| 230 |
+
"content": "<|pad|>",
|
| 231 |
+
"lstrip": false,
|
| 232 |
+
"normalized": false,
|
| 233 |
+
"rstrip": false,
|
| 234 |
+
"single_word": false,
|
| 235 |
+
"special": true
|
| 236 |
+
}
|
| 237 |
+
},
|
| 238 |
+
"additional_special_tokens": [
|
| 239 |
+
"<|endofchunk|>",
|
| 240 |
+
"<s>",
|
| 241 |
+
"<|pad|>"
|
| 242 |
+
],
|
| 243 |
+
"bos_token": "<s>",
|
| 244 |
+
"clean_up_tokenization_spaces": false,
|
| 245 |
+
"eos_token": "<|im_end|>",
|
| 246 |
+
"errors": "replace",
|
| 247 |
+
"extra_special_tokens": {},
|
| 248 |
+
"model_max_length": 131072,
|
| 249 |
+
"pad_token": "<|pad|>",
|
| 250 |
+
"split_special_tokens": false,
|
| 251 |
+
"tokenizer_class": "Qwen2Tokenizer",
|
| 252 |
+
"unk_token": null
|
| 253 |
+
}
|
vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|