File size: 7,716 Bytes
07d61c1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 |
# coding=utf-8
# Copyright 2025-present, the HuggingFace Inc. Team and AIRAS Inc. Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import torch
import json
from pathlib import Path
from transformers import AutoModelForCausalLM, AutoTokenizer
from ctransformers import AutoModelForCausalLM as GGUFModel
from models.sapnous import SapnousT1Config
def load_safetensors_state_dict(model_path, weight_map):
"""Load state dict from safetensors shards with custom metadata handling."""
import safetensors
from safetensors.torch import load_file
state_dict = {}
metadata = {}
# Load and validate each shard
for param_name, shard_file in weight_map['weight_map'].items():
shard_path = os.path.join(model_path, shard_file)
if not os.path.exists(shard_path):
raise OSError(f"Missing weight shard: {shard_path}")
try:
# Load shard with metadata
shard_dict = load_file(shard_path)
shard_metadata = safetensors.safe_open(shard_path, framework="pt").metadata()
if shard_metadata:
metadata.update(shard_metadata)
# Add tensors to state dict
for key, tensor in shard_dict.items():
if key in state_dict:
raise ValueError(f"Duplicate parameter {key} found in multiple shards")
state_dict[key] = tensor
except Exception as e:
raise OSError(f"Error loading shard {shard_file}: {str(e)}")
# Add metadata to state dict
if metadata:
state_dict['_metadata'] = metadata
return state_dict
return state_dict
def convert_to_gguf(model_path, output_path):
# Load configuration and weight map
config_path = os.path.join(model_path, 'config.json')
weight_map_path = os.path.join(model_path, 'model.safetensors.index.json')
if not os.path.exists(config_path):
raise OSError(f"Missing config file: {config_path}")
if not os.path.exists(weight_map_path):
raise OSError(f"Missing weight map file: {weight_map_path}")
with open(config_path, 'r') as f:
config = json.load(f)
with open(weight_map_path, 'r') as f:
weight_map = json.load(f)
# Validate weight map structure
if 'weight_map' not in weight_map:
raise ValueError("Invalid weight map format: missing 'weight_map' key")
if 'metadata' not in weight_map:
raise ValueError("Invalid weight map format: missing 'metadata' key")
# Load the model and tokenizer with vision-language support
model = AutoModelForCausalLM.from_pretrained(
model_path,
trust_remote_code=True,
device_map=None, # Disable device mapping for conversion
torch_dtype=torch.float16, # Use FP16 for memory efficiency
low_cpu_mem_usage=True, # Enable low CPU memory usage
local_files_only=True, # Use local files only
ignore_mismatched_sizes=True, # Bypass size validation
use_safetensors=True, # Explicitly enable safetensors
use_auth_token=False # Disable auth token
)
tokenizer = AutoTokenizer.from_pretrained(
model_path,
trust_remote_code=True
)
# Get model configuration
config = model.config
if not isinstance(config, SapnousT1Config):
raise ValueError("Model must be a SapnousT1 model")
# Save in intermediate format
model.save_pretrained(output_path, safe_serialization=True)
tokenizer.save_pretrained(output_path)
# Convert to GGUF using custom SapnousT1 architecture settings
gguf_model = GGUFModel.from_pretrained(
output_path,
model_type='sapnous_t1', # Custom architecture type
gpu_layers=0, # CPU only for conversion
config={
'context_length': config.sliding_window,
'attention_type': 'multihead', # Custom attention implementation
'num_attention_heads': config.num_attention_heads,
'num_key_value_heads': config.num_key_value_heads,
'hidden_size': config.hidden_size,
'intermediate_size': config.intermediate_size,
'max_position_embeddings': config.max_position_embeddings,
'vocab_size': config.vocab_size,
'num_hidden_layers': config.num_hidden_layers,
'rms_norm_eps': config.rms_norm_eps,
'rope_theta': config.rope_theta,
# Vision model parameters
'vision_config': {
'hidden_size': config.vision_hidden_size,
'num_hidden_layers': config.vision_layers,
'num_attention_heads': config.vision_heads,
'intermediate_size': config.vision_intermediate_size,
'patch_size': config.patch_size,
'image_size': config.image_size
}
}
)
print(f"Model converted and saved to {output_path}")
return gguf_model
def convert_to_hf(gguf_path, output_path):
"""Convert GGUF model back to Hugging Face format"""
# Load GGUF model configuration
config_path = Path(gguf_path) / "config.json"
with open(config_path, 'r') as f:
gguf_config = json.load(f)
# Create SapnousT1 configuration
config = SapnousT1Config(
vocab_size=gguf_config['vocab_size'],
hidden_size=gguf_config['hidden_size'],
num_hidden_layers=gguf_config['num_hidden_layers'],
num_attention_heads=gguf_config['num_attention_heads'],
num_key_value_heads=gguf_config['num_key_value_heads'],
intermediate_size=gguf_config['intermediate_size'],
max_position_embeddings=gguf_config['max_position_embeddings'],
rms_norm_eps=gguf_config['rms_norm_eps'],
rope_theta=gguf_config['rope_theta'],
# Vision configuration
vision_hidden_size=gguf_config['vision_config']['hidden_size'],
vision_layers=gguf_config['vision_config']['num_hidden_layers'],
vision_heads=gguf_config['vision_config']['num_attention_heads'],
vision_intermediate_size=gguf_config['vision_config']['intermediate_size'],
patch_size=gguf_config['vision_config']['patch_size'],
image_size=gguf_config['vision_config']['image_size']
)
# Load GGUF model
gguf_model = GGUFModel.from_pretrained(gguf_path)
# Convert weights to HF format
model = AutoModelForCausalLM.from_config(config)
model.load_state_dict(gguf_model.state_dict())
# Save converted model
model.save_pretrained(output_path)
print(f"Model converted back to Hugging Face format at {output_path}")
return model
if __name__ == '__main__':
model_path = os.path.dirname(os.path.abspath(__file__))
output_path = os.path.join(model_path, 'gguf_model')
if not os.path.exists(output_path):
os.makedirs(output_path)
convert_to_gguf(model_path, output_path) |