Text Classification
Transformers
ONNX
Safetensors
English
qwen3_5_text
text-generation
safety
guardrail
coding-agent
tool-call
command-review
permission
qwen3.5
Instructions to use openvulcan/vulcan-decision-shell-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use openvulcan/vulcan-decision-shell-v1 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="openvulcan/vulcan-decision-shell-v1")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("openvulcan/vulcan-decision-shell-v1") model = AutoModelForCausalLM.from_pretrained("openvulcan/vulcan-decision-shell-v1", device_map="auto") - Notebooks
- Google Colab
- Kaggle
File size: 5,373 Bytes
6ab931e | 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 | """Load verified v8055 INT16 storage and mixed INT8 variants with explicit bindings.
使用显式绑定加载已核验的v8055 INT16存储与混合INT8变体。
"""
import hashlib
import json
import sys
from pathlib import Path
def sha(path):
"""Return a complete SHA-256 for the exact file supplied.
返回所给确切文件的完整SHA-256。
"""
with Path(path).open('rb') as stream:
return hashlib.file_digest(stream,'sha256').hexdigest()
def map_name(canonical):
"""Translate only the proven Qwen checkpoint prefix and reject other namespaces.
仅转换已证明的Qwen检查点前缀,其他命名空间明确拒绝。
"""
prefix='model.language_model.'
if not canonical.startswith(prefix):
raise ValueError('Unexpected canonical Qwen parameter namespace')
return 'model.'+canonical[len(prefix):]
def load_variant(name,base):
"""Return a strictly loaded CPU model; the caller must run every neural operation on CUDA.
返回严格加载的CPU模型;调用方必须将所有神经计算放在CUDA。
"""
import torch
from safetensors.torch import load_file
from transformers import AutoConfig,Qwen3_5ForCausalLM
base=Path(base)
expected='ce8e905daaa5726b2ef162a08120955c52700d3ed686ca2f429ce8b01a60cb35'
if name=='int16_storage':
folder=base/'variants/int16-storage'
metadata=json.loads((folder/'int16-storage-manifest.json').read_bytes())
path=folder/'weights-int16.safetensors'
if metadata['source_model_sha256']!=expected or sha(path)!='c80ddd299149daca2516659d65d0c7f575a30daceaa096ff82123bd279cf8aca':
raise ValueError('The exact INT16 storage binding changed')
packed=load_file(str(path),device='cpu')
state={}
for canonical,entry in metadata['mapping'].items():
if entry['kind']=='symmetric_rowwise_int16':
value=packed[entry['storage']].float()*packed[entry['scale']]
elif entry['kind']=='unchanged_float32':
value=packed[entry['storage']]
else:
raise ValueError('Unknown INT16 storage mapping kind')
if list(value.shape)!=entry['shape'] or value.dtype!=torch.float32:
raise ValueError('The dequantized tensor contract changed')
state[map_name(canonical)]=value
if len(state)!=320 or sum(value.numel() for value in state.values())!=752393024:
raise ValueError('INT16 dequantization does not reconstruct the full model')
config=AutoConfig.from_pretrained(folder,local_files_only=True)
with torch.device('cpu'):
model=Qwen3_5ForCausalLM(config)
model.tie_weights()
if set(dict(model.named_parameters()))!=set(state) or not model.lm_head.weight is model.get_input_embeddings().weight:
raise ValueError('The confirmed320-parameter namespace or tied head changed')
state['lm_head.weight']=state['model.embed_tokens.weight']
model.load_state_dict(state,strict=True,assign=True)
model.tie_weights()
if any(parameter.is_meta for parameter in model.parameters()) or any(value.is_meta for _,value in model.named_buffers()):
raise ValueError('A model parameter or derived rotary buffer remains uninitialized')
if model.config._attn_implementation!='sdpa':
raise ValueError('The confirmed native SDPA implementation changed')
return model
if name=='int8_mixed_delta':
import torchao
from torchao.dtypes.affine_quantized_tensor import AffineQuantizedTensor
if torchao.__version__!='0.14.1':
raise ValueError('The pinned standalone TorchAO version changed')
path=base/'variants/int8-mixed-delta/weights-int8-torchao-delta.pt'
if sha(base/'model.safetensors')!=expected or sha(path)!='2d31ebd846f58266f134f3536cdcbefd8576e0b5e1b7607fcc49d1e0a41e7332':
raise ValueError('The exact INT8 base/delta binding changed')
model,loading=Qwen3_5ForCausalLM.from_pretrained(base,dtype=torch.float32,attn_implementation='sdpa',local_files_only=True,output_loading_info=True)
if any(loading[key] for key in ('missing_keys','unexpected_keys','mismatched_keys','error_msgs')):
raise ValueError('The INT8 FP32 base did not load exactly')
packed=torch.load(path,weights_only=True,map_location='cpu')
expected_names={module_name+'.weight' for module_name,module in model.named_modules() if isinstance(module,torch.nn.Linear) and ('.mlp.' in module_name or '.self_attn.' in module_name)}
if len(packed)!=96 or {map_name(key) for key in packed}!=expected_names:
raise ValueError('The96 protected-mixed INT8 linear bindings changed')
for canonical,value in packed.items():
if not isinstance(value,AffineQuantizedTensor) or value.tensor_impl.int_data.dtype!=torch.int8:
raise ValueError('A packed weight is not the declared INT8 tensor')
module=model.get_submodule(map_name(canonical)[:-len('.weight')])
if list(module.weight.shape)!=list(value.shape):
raise ValueError('An INT8 linear shape changed')
module.weight=torch.nn.Parameter(value,requires_grad=False)
return model
raise ValueError('Undeclared precision variant: '+name)
|