The nnsight package enables interpreting and manipulating the internals of deep learned models. Read our paper!
Install this package through pip by running:
pip install nnsight
Here is a simple example where we run the nnsight API locally on gpt2 and save the hidden states of the last layer:
from nnsight import LanguageModel
model = LanguageModel('openai-community/gpt2', device_map='auto')
with model.trace('The Eiffel Tower is in the city of') as tracer:
hidden_states = model.transformer.h[-1].output[0].save()
output = model.output.save()Lets go over this piece by piece.
We import the LanguageModel object from the nnsight module and create a gpt2 model using the huggingface repo ID for gpt2, 'openai-community/gpt2'. This accepts arguments to create the model including device_map to specify which device to run on.
from nnsight import LanguageModel
model = LanguageModel('openai-community/gpt2',device_map='auto')Then, we create a tracing context block by calling .trace(...) on the model object. This denotes we want to run the model with our prompt.
with model.trace('The Eiffel Tower is in the city of') as tracer:Now calling .trace(...) does not actually initialize or run the model. Only after the tracing` block is exited, is the actual model loaded and ran. All operations in the block are "proxies" which essentially creates a graph of operations we wish to carry out later.
Within this context, all operations/interventions will be applied to the processing of the given prompt.
hidden_states = model.transformer.h[-1].output[0].save()On this line were saying, access the last layer of the transformer model.transformer.h[-1], access its output .output, index it at 0 .output[0], and save it .save()
A few things, we can see the module tree of the model by printing the model. This allows us to know what attributes to access to get to the module we need.
Running print(model) results in:
GPT2LMHeadModel(
(transformer): GPT2Model(
(wte): Embedding(50257, 768)
(wpe): Embedding(1024, 768)
(drop): Dropout(p=0.1, inplace=False)
(h): ModuleList(
(0-11): 12 x GPT2Block(
(ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
(attn): GPT2Attention(
(c_attn): Conv1D()
(c_proj): Conv1D()
(attn_dropout): Dropout(p=0.1, inplace=False)
(resid_dropout): Dropout(p=0.1, inplace=False)
)
(ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
(mlp): GPT2MLP(
(c_fc): Conv1D()
(c_proj): Conv1D()
(act): NewGELUActivation()
(dropout): Dropout(p=0.1, inplace=False)
)
)
)
(ln_f): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
)
(lm_head): Linear(in_features=768, out_features=50257, bias=False)
)
.output returns a proxy for the output of this module. This essentially means were saying, when we get to the output of this module during inference, grab it and perform any operations we define on it (which also become proxies). There are two operational proxies here, one for getting the 0th index of the output, and one for saving the output. We take the 0th index because the output of gpt2 transformer layers are a tuple where the first index are the actual hidden states (last two indicies are from attention). We can call .shape on any proxies to get what shape the value will eventually be.
Running print(model.transformer.h[-1].output.shape) returns (torch.Size([1, 10, 768]), (torch.Size([1, 12, 10, 64]), torch.Size([1, 12, 10, 64])))
During processing of the intervention computational graph we are building, when the value of a proxy is no longer ever needed, its value is dereferenced and destroyed. However calling .save() on the proxy informs the computation graph to save the value of this proxy and never destroy it, allowing us to access to value after generation.
After exiting the generator context, the model is ran with the specified arguments and intervention graph. output is populated with the actual output and hidden_states will contain the hidden value.
print(output)
print(hidden_states)returns:
tensor([[ 464, 412, 733, 417, 8765, 318, 287, 262, 1748, 286, 6342]],
device='cuda:0')
tensor([[[ 0.0505, -0.1728, -0.1690, ..., -1.0096, 0.1280, -1.0687],
[ 8.7494, 2.9057, 5.3024, ..., -8.0418, 1.2964, -2.8677],
[ 0.2960, 4.6686, -3.6642, ..., 0.2391, -2.6064, 3.2263],
...,
[ 2.1537, 6.8917, 3.8651, ..., 0.0588, -1.9866, 5.9188],
[-0.4460, 7.4285, -9.3065, ..., 2.0528, -2.7946, 0.5556],
[ 6.6286, 1.7258, 4.7969, ..., 7.6714, 3.0682, 2.0481]]],
device='cuda:0')
Most basic operations and torch operations work on proxies and are added to the computation graph.
from nnsight import LanguageModel
import torch
model = LanguageModel('openai-community/gpt2', device_map='cuda')
with model.trace('The Eiffel Tower is in the city of'):
hidden_states_pre = model.transformer.h[-1].output[0].save()
hs_sum = torch.sum(hidden_states_pre).save()
hs_edited = hidden_states_pre + hs_sum
hs_edited = hs_edited.save()
print(hidden_states_pre)
print(hs_sum)
print(hs_edited)In this example we get the sum of the hidden states and add them to the hidden_states themselves (for whatever reason). By saving the various steps, we can see how the values change.
tensor([[[ 0.0505, -0.1728, -0.1690, ..., -1.0096, 0.1280, -1.0687],
[ 8.7494, 2.9057, 5.3024, ..., -8.0418, 1.2964, -2.8677],
[ 0.2960, 4.6686, -3.6642, ..., 0.2391, -2.6064, 3.2263],
...,
[ 2.1537, 6.8917, 3.8651, ..., 0.0588, -1.9866, 5.9188],
[-0.4460, 7.4285, -9.3065, ..., 2.0528, -2.7946, 0.5556],
[ 6.6286, 1.7258, 4.7969, ..., 7.6714, 3.0682, 2.0481]]],
device='cuda:0')
tensor(501.2957, device='cuda:0')
tensor([[[501.3461, 501.1229, 501.1267, ..., 500.2860, 501.4237, 500.2270],
[510.0451, 504.2014, 506.5981, ..., 493.2538, 502.5920, 498.4279],
[501.5916, 505.9643, 497.6315, ..., 501.5348, 498.6892, 504.5219],
...,
[503.4493, 508.1874, 505.1607, ..., 501.3545, 499.3091, 507.2145],
[500.8496, 508.7242, 491.9892, ..., 503.3485, 498.5010, 501.8512],
[507.9242, 503.0215, 506.0926, ..., 508.9671, 504.3639, 503.3438]]],
device='cuda:0')
We often not only want to see whats happening during computation, but intervene and edit the flow of information.
from nnsight import LanguageModel
import torch
model = LanguageModel('openai-community/gpt2', device_map='cuda')
with model.trace('The Eiffel Tower is in the city of') as tracer:
hidden_states_pre = model.transformer.h[-1].mlp.output.clone().save()
noise = (0.001**0.5)*torch.randn(hidden_states_pre.shape)
model.transformer.h[-1].mlp.output = hidden_states_pre + noise
hidden_states_post = model.transformer.h[-1].mlp.output.save()
print(hidden_states_pre)
print(hidden_states_post)In this example, we create a tensor of noise to add to the hidden states. We then add it, use the assigment = operator to update the value of .output with these new noised activations.
We can see the change in the results:
tensor([[[ 0.0505, -0.1728, -0.1690, ..., -1.0096, 0.1280, -1.0687],
[ 8.7494, 2.9057, 5.3024, ..., -8.0418, 1.2964, -2.8677],
[ 0.2960, 4.6686, -3.6642, ..., 0.2391, -2.6064, 3.2263],
...,
[ 2.1537, 6.8917, 3.8651, ..., 0.0588, -1.9866, 5.9188],
[-0.4460, 7.4285, -9.3065, ..., 2.0528, -2.7946, 0.5556],
[ 6.6286, 1.7258, 4.7969, ..., 7.6714, 3.0682, 2.0481]]],
device='cuda:0')
tensor([[[ 0.0674, -0.1741, -0.1771, ..., -0.9811, 0.1972, -1.0645],
[ 8.7080, 2.9067, 5.2924, ..., -8.0253, 1.2729, -2.8419],
[ 0.2611, 4.6911, -3.6434, ..., 0.2295, -2.6007, 3.2635],
...,
[ 2.1859, 6.9242, 3.8666, ..., 0.0556, -2.0282, 5.8863],
[-0.4568, 7.4101, -9.3698, ..., 2.0630, -2.7971, 0.5522],
[ 6.6764, 1.7416, 4.8027, ..., 7.6507, 3.0754, 2.0218]]],
device='cuda:0')
When generating more than one token, use .generate(...) and .next() on the module you want to get the next value of to denote following interventions should be applied to the subsequent generations.
Here we again generate using gpt2, but generate three tokens and save the hidden states of the last layer for each one:
from nnsight import LanguageModel
model = LanguageModel('openai-community/gpt2', device_map='cuda')
with model.generate('The Eiffel Tower is in the city of', max_new_tokens=3) as tracer:
hidden_states1 = model.transformer.h[-1].output[0].save()
invoker.next()
hidden_states2 = model.transformer.h[-1].next().output[0].save()
invoker.next()
hidden_states3 = model.transformer.h[-1].next().output[0].save()Intervention operations work cross prompt! Use two invocations within the same generation block and operations can work between them.
You can do this by not passing a prompt into .trace/.generate, but by calling .invoke(...) on the created tracer object.
In this case, we grab the token embeddings coming from the first prompt, "Madison square garden is located in the city of New" and replace the embeddings of the second prompt with them.
from nnsight import LanguageModel
model = LanguageModel('openai-community/gpt2', device_map='cuda')
with model.generate(max_new_tokens=3) as tracer:
with tracer.invoke("Madison square garden is located in the city of New"):
embeddings = model.transformer.wte.output
with tracer.invoke("_ _ _ _ _ _ _ _ _ _"):
model.transformer.wte.output = embeddings
output = model.generator.output.save()
print(model.tokenizer.decode(output[0]))
print(model.tokenizer.decode(output[1]))This results in:
Madison square garden is located in the city of New York City.
_ _ _ _ _ _ _ _ _ _ York City.
We also could have entered a pre-saved embedding tensor as shown here:
from nnsight import LanguageModel
model = LanguageModel('openai-community/gpt2', device_map='cuda')
with model.generate(max_new_tokens=3) as tracer:
with tracer.invoke("Madison square garden is located in the city of New") as invoker:
embeddings = model.transformer.wte.output.save()
with model.generate(max_new_tokens=3) as tracer:
with tracer.invoke("_ _ _ _ _ _ _ _ _ _") as invoker:
model.transformer.wte.output = embeddings.valueAnother thing we can do is apply modules in the model's module tree at any point during computation, even if it's out of order.
from nnsight import LanguageModel
import torch
model = LanguageModel("openai-community/gpt2", device_map='cuda')
with model.generate('The Eiffel Tower is in the city of') as generator:
hidden_states = model.transformer.h[-1].output[0]
hidden_states = model.lm_head(model.transformer.ln_f(hidden_states)).save()
tokens = torch.softmax(hidden_states, dim=2).argmax(dim=2).save()
print(hidden_states)
print(tokens)
print(model.tokenizer.decode(tokens[0]))Here we get the hidden states of the last layer like usual. We also chain apply model.transformer.ln_f and model.lm_head in order to "decode" the hidden states into vocabularly space.
Applying softmax and then argmax allows us to then transform the vocabulary space hidden states into actually tokens which we can then use the tokenizer to decode.
The output looks like:
tensor([[[ -36.2874, -35.0114, -38.0793, ..., -40.5163, -41.3759,
-34.9193],
[ -68.8886, -70.1562, -71.8408, ..., -80.4195, -78.2552,
-71.1206],
[ -82.2950, -81.6519, -83.9941, ..., -94.4878, -94.5194,
-85.6998],
...,
[-113.8675, -111.8628, -113.6634, ..., -116.7652, -114.8267,
-112.3621],
[ -81.8531, -83.3006, -91.8192, ..., -92.9943, -89.8382,
-85.6898],
[-103.9307, -102.5054, -105.1563, ..., -109.3099, -110.4195,
-103.1395]]], device='cuda:0')
tensor([[ 198, 12, 417, 8765, 318, 257, 262, 3504, 7372, 6342]],
device='cuda:0')
-el Tower is a the middle centre Paris
More examples can be found at nnsight.net
If you use nnsight in your research, please cite using the following
@article{fiottokaufman2024nnsightndifdemocratizingaccess,
title={NNsight and NDIF: Democratizing Access to Foundation Model Internals},
author={Jaden Fiotto-Kaufman and Alexander R Loftus and Eric Todd and Jannik Brinkmann and Caden Juang and Koyena Pal and Can Rager and Aaron Mueller and Samuel Marks and Arnab Sen Sharma and Francesca Lucchetti and Michael Ripa and Adam Belfki and Nikhil Prakash and Sumeet Multani and Carla Brodley and Arjun Guha and Jonathan Bell and Byron Wallace and David Bau},
year={2024},
eprint={2407.14561},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2407.14561},
}