There was a bug in split_text() (in autogpt/processing/text.py). The original algorithm split a webpage into paragraphs, and kept adding paragraphs into a "chunk" list, until it hit some max_length. Then it yielded that chunk, and started the next chunk by simply adding the next paragraph.The issue is, that some paragraphs are huge, without any newline in them, and this "let's add a paragraph as the first element of a chunk" is what caused the error in all of the issues. The other issue was that split_text() ignored BROWSE_CHUNK_MAX_LENGTH The third issue was that BROWSE_CHUNK_MAX_LENGTH was understood as the number of characters, whereas in reality it should really use the number of tokens.

Same problem can be found in tf.keras.layers.ZeroPadding3D. Here is the repo code: padding = 16 data_format = None __input___0_tensor = tf.random.uniform([1, 1, 2, 2, 3], minval=0.8510533546655319,maxval=3.0,dtype=tf.float64) __input___0 = tf.identity(__input___0_tensor) ZeroPadding3D_class = tf.keras.layers.ZeroPadding3D(padding=padding,data_format=data_format) layer = ZeroPadding3D_class inputs = __input___0 with tf.GradientTape() as g: g.watch(inputs) res = layer(inputs) print(res.shape) grad = g.jacobian(res, inputs) print(grad)

"nvcc --compiler-bindir /path/to/clang" sets __clang__ while compiling CUDA code. This causes gpu_device_functions.h to think it is being compiled with Clang and try to use a Clang-specific function.

Findings:

The Likely Cause:

Root Cause Identified

Root Cause Analysis

Successfully fixed the TypeError that occurred when the DataForSEO API returns an unexpected response structure where items could be None.

'm running into same, we need to limit certain chunks I think. Should be able to change chunk size to fix, not sure if that'll fix the total token amount, but we can make the summaries smaller too.

Here is how you can check if there is some bottleneck on your machine: Instead of running ./run.bat (or ./run.sh) you can run: python -m cProfile -o profile.pickle -m autogpt

This is a prompting issue and a limitation of LLMs

I also think is that, I find that plugins_config (at line 178 Auto-GPT-stable/autogpt/plugins/init.py) is always empty, no matter if is configured "correctly", which is not clear how they need to properly be defined considering every plugin developer use a different name for the plugins (ones with dashes, others without, etc..) besides some are not yet implementing the template plugin class. but also in what I read say that this is not yet strictly necessary. They talk about a "as long as they are in the correct (NEW) format" but it is not clear either what this is.

Whenever it has completed auto-tasks (I usually do tasks in blocks of 50 (y -50) to (y -200). You can type in a message to it instead of typing y -xx or n (to exit). It will say it doesn't understand but typically "fixes" itself and sometimes will accept what you've written.

Try using something different than the local memory. I downloaded the code 5 days ago so I don't know if it has been changed but inside the config.py file in the scripts folder on line 75, the memory backend is hardcoded to local. Change local to pinecone and use a pinecone API key if you want.

Both libraries use signals as a identifier which leads to a namespace collision.

OOM when allocating tensor with shape[50982027264] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc [[{{node gradient_tape/UnsortedSegmentSum/pfor/UnsortedSegmentSum}}]]

It may be that the size is too large and the memory is overflowing.

Hi, By default the colab notebook is using tensorflow v2.17 which contains keras3.0 which was causing the error. Could you please try to import keras2.0 with the below commands.