# https://huggingface.co/Qwen/Qwen3-8B/tree/main
model_name = "/Volumes/models/Qwen3-8B/"
hf_token = "..."
tokenizer = AutoTokenizer.from_pretrained(model_name, token=hf_token)
tokenizer.pad_token_id = 0
# device_map="auto" will distribute the model over multiple GPUs
# this notebook was run on a runpod 3xA100—the cuda:0 device will need to have enough spare memory
# to do inference on for this notebook to work
model = AutoModelForCausalLM.from_pretrained(
model_name, torch_dtype=torch.float16, device_map="mps:0", token=hf_token
)
wrapped_model = model
model = ControlModel(wrapped_model, list(range(1, 35)))
def chat_template_unparse(messages: list[tuple[str, str]]) -> str:
template = []
for role, content in messages:
template.append(
f"<|im_start|>{role}\n\n{content}<|im_end|>"
)
if messages[-1][0] != "assistant":
# prefill assistant prefix
template.append("<|im_start|>assistant\n\n")
return "".join(template)
with open("./data/all_truncated_outputs.json") as f:
output_suffixes = json.load(f)
truncated_output_suffixes = [
tokenizer.convert_tokens_to_string(tokens[:i])
for tokens in (tokenizer.tokenize(s) for s in output_suffixes)
for i in range(1, len(tokens))
]
truncated_output_suffixes_512 = [
tokenizer.convert_tokens_to_string(tokens[:i])
for tokens in (tokenizer.tokenize(s) for s in output_suffixes[:512])
for i in range(1, len(tokens))
]
def make_dataset(
template: str,
positive_personas: list[str],
negative_personas: list[str],
suffix_list: list[str],
) -> list[DatasetEntry]:
dataset = []
for suffix in suffix_list:
for positive_persona, negative_persona in zip(
positive_personas, negative_personas
):
positive_template = template.format(persona=positive_persona)
negative_template = template.format(persona=negative_persona)
dataset.append(
DatasetEntry(
positive=f"{positive_template}{suffix}",
negative=f"{negative_template}{suffix}",
)
)
return dataset
def make_vector(template, pos, neg):
ds = make_dataset(
template,
pos, neg,
truncated_output_suffixes,
)
model.reset()
vec = ControlVector.train(
model, tokenizer, ds, batch_size=32, method="pca_center"
)
return vec
vec = make_vector(
chat_template_unparse([("user", "{persona}")]),
["Pretend you're a happy person making statements about the world"],
["Pretend you're a depressed person making statements about the world"]
)---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
Cell In[12], line 1
----> 1 vec = make_vector(
2 chat_template_unparse([("user", "{persona}")]),
3 ["Pretend you're a happy person making statements about the world"],
4 ["Pretend you're a depressed person making statements about the world"]
5 )
Cell In[10], line 8, in make_vector(template, pos, neg)
2 ds = make_dataset(
3 template,
4 pos, neg,
5 truncated_output_suffixes,
6 )
7 model.reset()
----> 8 vec = ControlVector.train(
9 model, tokenizer, ds, batch_size=32, method="pca_center"
10 )
11 return vec
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/repeng/extract.py:53](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/repeng/extract.py#line=52), in ControlVector.train(cls, model, tokenizer, dataset, **kwargs)
36 """
37 Train a ControlVector for a given model and tokenizer using the provided dataset.
38
(...) 50 ControlVector: The trained vector.
51 """
52 with torch.inference_mode():
---> 53 dirs = read_representations(
54 model,
55 tokenizer,
56 dataset,
57 **kwargs,
58 )
59 return cls(model_type=model.config.model_type, directions=dirs)
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/repeng/extract.py:267](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/repeng/extract.py#line=266), in read_representations(model, tokenizer, inputs, hidden_layers, batch_size, method, transform_hiddens)
264 # the order is [positive, negative, positive, negative, ...]
265 train_strs = [s for ex in inputs for s in (ex.positive, ex.negative)]
--> 267 layer_hiddens = batched_get_hiddens(
268 model, tokenizer, train_strs, hidden_layers, batch_size
269 )
271 if transform_hiddens is not None:
272 layer_hiddens = transform_hiddens(layer_hiddens)
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/repeng/extract.py:350](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/repeng/extract.py#line=349), in batched_get_hiddens(model, tokenizer, inputs, hidden_layers, batch_size)
348 encoded_batch = tokenizer(batch, padding=True, return_tensors="pt")
349 encoded_batch = encoded_batch.to(model.device)
--> 350 out = model(**encoded_batch, output_hidden_states=True)
351 attention_mask = encoded_batch["attention_mask"]
352 for i in range(len(batch)):
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/repeng/control.py:121](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/repeng/control.py#line=120), in ControlModel.__call__(self, *args, **kwargs)
120 def __call__(self, *args, **kwargs):
--> 121 return self.model(*args, **kwargs)
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py:1751](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py#line=1750), in Module._wrapped_call_impl(self, *args, **kwargs)
1749 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1750 else:
-> 1751 return self._call_impl(*args, **kwargs)
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py:1762](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py#line=1761), in Module._call_impl(self, *args, **kwargs)
1757 # If we don't have any hooks, we want to skip the rest of the logic in
1758 # this function, and just call forward.
1759 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1760 or _global_backward_pre_hooks or _global_backward_hooks
1761 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1762 return forward_call(*args, **kwargs)
1764 result = None
1765 called_always_called_hooks = set()
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/transformers/utils/generic.py:943](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/transformers/utils/generic.py#line=942), in can_return_tuple.<locals>.wrapper(self, *args, **kwargs)
940 set_attribute_for_modules(self, "_is_top_level_module", False)
942 try:
--> 943 output = func(self, *args, **kwargs)
944 if is_requested_to_return_tuple or (is_configured_to_return_tuple and is_top_level_module):
945 output = output.to_tuple()
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/transformers/models/qwen3/modeling_qwen3.py:570](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/transformers/models/qwen3/modeling_qwen3.py#line=569), in Qwen3ForCausalLM.forward(self, input_ids, attention_mask, position_ids, past_key_values, inputs_embeds, labels, use_cache, output_attentions, output_hidden_states, cache_position, logits_to_keep, **kwargs)
565 output_hidden_states = (
566 output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
567 )
569 # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
--> 570 outputs: BaseModelOutputWithPast = self.model(
571 input_ids=input_ids,
572 attention_mask=attention_mask,
573 position_ids=position_ids,
574 past_key_values=past_key_values,
575 inputs_embeds=inputs_embeds,
576 use_cache=use_cache,
577 output_attentions=output_attentions,
578 output_hidden_states=output_hidden_states,
579 cache_position=cache_position,
580 **kwargs,
581 )
583 hidden_states = outputs.last_hidden_state
584 # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py:1751](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py#line=1750), in Module._wrapped_call_impl(self, *args, **kwargs)
1749 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1750 else:
-> 1751 return self._call_impl(*args, **kwargs)
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py:1762](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py#line=1761), in Module._call_impl(self, *args, **kwargs)
1757 # If we don't have any hooks, we want to skip the rest of the logic in
1758 # this function, and just call forward.
1759 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1760 or _global_backward_pre_hooks or _global_backward_hooks
1761 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1762 return forward_call(*args, **kwargs)
1764 result = None
1765 called_always_called_hooks = set()
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/transformers/utils/generic.py:943](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/transformers/utils/generic.py#line=942), in can_return_tuple.<locals>.wrapper(self, *args, **kwargs)
940 set_attribute_for_modules(self, "_is_top_level_module", False)
942 try:
--> 943 output = func(self, *args, **kwargs)
944 if is_requested_to_return_tuple or (is_configured_to_return_tuple and is_top_level_module):
945 output = output.to_tuple()
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/transformers/models/qwen3/modeling_qwen3.py:460](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/transformers/models/qwen3/modeling_qwen3.py#line=459), in Qwen3Model.forward(self, input_ids, attention_mask, position_ids, past_key_values, inputs_embeds, use_cache, output_attentions, output_hidden_states, cache_position, **flash_attn_kwargs)
455 if output_hidden_states:
456 all_hidden_states += (hidden_states,)
458 layer_outputs = decoder_layer(
459 hidden_states,
--> 460 attention_mask=causal_mask_mapping[decoder_layer.attention_type],
461 position_ids=position_ids,
462 past_key_value=past_key_values,
463 output_attentions=output_attentions,
464 use_cache=use_cache,
465 cache_position=cache_position,
466 position_embeddings=position_embeddings,
467 **flash_attn_kwargs,
468 )
470 hidden_states = layer_outputs[0]
472 if output_attentions:
File [~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py:1940](http://metta:8888/lab/tree/~/.cache/uv/archive-v0/V2To1YSJ6Tv55EOIzYrge/lib/python3.12/site-packages/torch/nn/modules/module.py#line=1939), in Module.__getattr__(self, name)
1938 if name in modules:
1939 return modules[name]
-> 1940 raise AttributeError(
1941 f"'{type(self).__name__}' object has no attribute '{name}'"
1942 )
AttributeError: 'ControlModule' object has no attribute 'attention_type'
