Tag: llm

Sources of non-determinism in LLMs

I’ve had a couple discussions recently around nondeterminism in LLMs. Measuring statistical bounds is really the only way to go with LLMs. But it’s also entirely true that thinking statistically is (a) much more expensive (both up-front in talent and dataset design, and in perpetuity to run the estimator), and (b) unintuitive-bordering-on-uncomfortable for folks who expect determinism from their machines. So I’m not surprised to have the “but why isn’t it deterministic? can we make it deterministic?” discussion over and over again.

Here is what I know about sources of non-determinism in LLMs:

  • Human perception. Humans aren’t great at actually perceiving truly identical inputs. I’ve found myself and others to pretty often think two inputs are identical, when they aren’t identical at all. In an LLM, any deviation in the input sequence may lead to different completions. Even slightly differences in punctuation, synonym choice, or minimal inserted context can affect the LLM output. Automated diffs help uncover perception mistakes.

  • Model versioning. Developers like versioned APIs. Whenever we hit the same API with the same fixed parameters, we expect the same behavior from the remote system. At time of writing, though, Google’s LLMs don’t support version pinning. For instance, last Tuesday I asked Google’s Bison to generate some “be mean” data, and it was pretty willing to oblige (it wouldn’t be mean to all the social groups I tried, but it spat out stereotypes for a lot of groups). On Wednesday, it politely refused. Ethically, I prefer Wednesday’s model – but entire product capabilities disappearing literally overnight, while the API and user code are exactly the same? That’s massive, unrecoverable non-determinism.

  • Temperature. “Temperature” is a physics metaphor: the higher the temperature, the more “random molecule bouncing around” you get. In an LLM world, temperature affects how random each next token is. Each selected token changes the most likely tokens for the rest of the completion. You can set the temperature to 0 to greedily get the token sequence with highest probability, but temperature=0 still doesn’t guarantee deterministic outputs. You get non-determinism if there are multiple options with the same probability representation (rare, but it occurs). You also get non-determinism if the randomness is introduced upstream of the greedy token selection, such that a different next token becomes most probable, as discussed below.

  • Floating point math. To compress the infinite space of real numbers into the finite space of computer memory, we necessarily lose precision. Floating point numbers can represent both very large and very small numbers – but the floating point numbers are not evenly spaced from each other. We always approximate into the closest representable floating point number, so some number i might have a different amount of discretization error than another number j. The discretization errors can build up when we do addition. As a result, addition in floating point isn’t guaranteed to be commutative. In other words, (a+b)+c might produce a different result than (b+c)+a in floating point math. For parallelized operations (and LLMs are massively parallelized), we usually allow addition to occur in an arbitrary order. But the order can affect the final sum. And the final sum affects the most likely token, which affects the rest of the generated string.

Floating point numbers are not equally spaced.

Floating point numbers (in green) are not equally spaced along the number line. (Image source: Wikimedia)

  • Peers in batch. As Mixture of Experts (MoE) models, the GPT model outputs (and maybe others) appear to be deterministic only at the batch-level. MoE models contain many “experts”, which are often distributed across many machines. When processing each input token, rather than sending it through a complete and expensive dense network, instead we identify the “right” experts to handle it. The best ways to identify the “right” experts is an area of active research. In current MoE, each expert only sparsely activates token paths, which makes it possible to get the quality benefits of significantly larger models without paying the full computational cost at inference. The MoE approach introduces non-determinism because the contents of each batch must be mapped to experts and returned as a batch, and there is a limit on how much data each expert can simultaneously handle. In other words, if your batch has many other inputs that compete with you for the experts you need, you might get a different set of experts than you would in a different batch. This competition for experts can lead to different predicted tokens depending on how your messages are batched for inference, and the effect depends on who else is using the LLM APIs at the same time as you.

In an LLM, each token is conditioned on all the tokens that preceeded it. As a result, once one token diverges, the remainder of the sequence will diverge even more. It’s all back to chaos theory: “when a butterfly flaps its wings in Brazil…”. So yes, if we want to be certain in this new world, we will need experimental statistics. How certain do you need to be?

Making ReAct work in practice

I spent some time getting a ReAct LLM (Yao et al. 2022) to autonomously answer questions. Improving performance turned out to require some changes that I haven't seen discussed in my technical haunts. Through failure case analysis and targeted interventions, I was able to drive performance from on the order of 1% to north of 80% over a couple weeks.

ReAct ("Reasoning and Acting") builds on chain-of-thought. It gives the LLM more structure around thinking and acting, which reduces hallucination and error propagation. With ReAct, we use a straightforward LLM prompt that describes some actions that the agent can take to get external information. The prompt also explains that we expect a final structure of repeated Thought / Action / Action Input / Observation blocks. The LLM then constructs a structured text that follows those rules in collaboration with a backend system. The backend system orchestrates and keeps the LLM on track. In particular, the backend parses the LLM outputs, reaches out to the Action systems to produce the Observations, and writes the prompt for the LLM.

I'll give a concrete example (text taken from the langchain docs). In the example, I've colored the sources of the text. I give user input in dark orange, the lines from the coordinating backend system in blue, and the text generated by the LLM in red with grey background:

Question: Who is Leo DiCaprio's girlfriend? What is her current age raised to the 0.43 power? Thought: I need to find out who Leo DiCaprio's girlfriend is and then calculate her age raised to the 0.43 power. Action: Search Action Input: "Leo DiCaprio girlfriend" Observation: Camila Morrone Thought: I need to find out Camila Morrone's age Action: Search Action Input: "Camila Morrone age" Observation: 25 years Thought: I need to calculate 25 raised to the 0.43 power Action: Calculator Action Input: 25^0.43 Observation: Answer: 3.991298452658078 Thought: I now know the final answer Final Answer: Camila Morrone is Leo DiCaprio's girlfriend and her current age raised to the 0.43 power is 3.991298452658078.

Each time control returns to the backend, the backend parses what the LLM produced. It throws away everything that isn't immediately relevant (like any oh-so-eagerly-hallucinated Observations). Then the backend constructs a new, slightly longer prompt with a real Observation, and prompts the LLM to complete it a bit further. Eventually the LLM produces a Final Answer, which the backend parses and returns to the user.

So, that's ReAct. It didn't work very well for me out-of-the-box. The langchain implementation plus a dozen possible actions produced sub-5% performance on questions it really should have been able to answer.

So, I embarked on a ReAct performance improvement quest. What worked for me, in order from least surprising to most surprising to me, was:

  1. Prompt engineering the action descriptions to improve dispatch. First-pass docstrings often are flawed. We all have this problem — the docstring writer has high context but the reader does not, and what is salient to the writer isn't always salient to the reader. So, I made sure each available action was described in one sentence starting with a verb. Then I re-scoped each "action" by how clearly I could write a user-facing description for the idea (rather than by how APIs are broken apart).
  2. Making parameterization errors visible to enable recovery. I found the LLM often chose a poor Action Input, which caused the backend to receive exceptions when it tried to execute the LLM's instructions. I wanted to make the "bad parameterization" problem more tangible to the LLM. I took this on in three ways: (1) with a priori clues — I described the format of the input in the description (e.g., is it an integer, enum, string, ...), (2) with data typing (e.g., does the UUID that the LLM wants to pass refer to an object of the appropriate type for this action), (3) with a posteriori clues — I updated the backend to include meaningful error messages as the Observation for bad parameterizations, so the LLM would get another crack at it.
  3. Removing all "early stop" actions to encourage actually trying. LangChain allows you to define "early stop" actions. If the LLM picks one of these actions, the entire reasoning chain gets aborted. I found that the LLM would often pick an early stop action as its very first action. So, I strongly encouraged it to always try to answer by removing all early stop actions. It is still possible for the system to go immediately from the Question into a Final Answer of "I have no idea what to do to answer this". But in practice the system is actually trying now, when it often wasn't before.
  4. Dropping old Thoughts so it can't confuse its guesses for facts. I was semi-frequently seeing the LLM hypothesize a wild idea, decide on an appropriate action and input to test that wild idea, receive the right answer, and then write a Final Action that treated the wild idea as if it were a fact. This behavior is understandable, and it is also very bad. So, now the LLM doesn't get to see its previous Thoughts.
  5. Actively seeding Thoughts to recover from unparseable LLM responses. Sometimes the LLM thinks the most likely next token sequence is nothing, and we get an empty string as the completion. Sometimes it doesn't generate text in the required format. Sometimes it goes off the rails in some other way. All of these break parsing. In the LangChain case, the backend has no effective way to get the whole agent back on track again, so it raises an exception and exits. Ideally, though, it would be able to recover. So, when the LLM gives garbage responses, I've started putting Thoughts in its head. Extending the system prompt slightly past the colon with an innocuous phrase — to something like Thought: I need to decide on an Action and Action Input — is enough to break the LLM free. This "seeding its thoughts" approach works well even when the temperature is turned down to 0.0 (with maximal determinism during text generation), because the approach ensures a slightly different prompt from the one that failed.

With these changes, I was able to raise the performance from <5% to on the order of 70-90% pretty quickly.

I suppose in addition to "keep thinking; a simpler solution will come", the wider lesson that was reinforced for me here is that popular ideas can easily suck up more than their share of oxygen in the public conversation (even good ideas like prompt engineering!). "Don't stop with what's popular; make sure everyone is looking at the real behaviors" is my takeaway for myself.

Your very own Star Trek Computer: Making sense of unexpected pipenv behavior with an LLM

Story time!

So, recently, I’m playing in someone else’s codebase, and I need to use their code to create a new Docker image that I can use for myself elsewhere. Their thin base image doesn’t have all my usual ML dependencies, so I need to extend the image. When I try, it blows up with an “Unknown compiler” message. With some legwork I learn that sklearn depends on scipy, and scipy needs to be compiled from source (and has for years now), and that thin base image doesn’t have the compiler. So I move to a non-default base image, and all is well.

Then I notice something I can’t explain, which I never would have noticed if I hadn’t been in this codebase. I had set Python to 3.9 using pyenv. I’m using pipenv --three to create a Python 3 Pipfile and then pipenv install to add some dependencies to it. (That’s all wrapped in a build script, because that’s how the codebase works, but that’s all the build script is doing.) But when I cat the Pipfile, the environment I just created is using Python 3.10, not 3.9! My mind is blown. I’ve never really used pipenv before… but I had a deeply rooted expectation that pipenv --three would use python3 --version for building the Pipfile. And clearly, it did not.

I read the pipenv docs and don’t see anything obviously relevant.

I do some internet searches. No answer.

I figure the explanation has to be obvious to anyone who actually has some pipenv experience – so I ask around. None of my usual suspects have pipenv experience either.

Then I have my epiphany. ChatGPT was just recently officially corporately blessed. This is the perfect question to use with ChatGPT – I can learn about pipenv, and I can use this to demonstrate how ChatGPT might help developers at the next engineering department sprint demo. (A good number of our developers are using Python professionally for the first time, in a codebase that extensively uses pipenv, and have not yet tried any LLMs – I like the odds of a sprint demo on this topic actually being useful for the audience.) So, I spend a few minutes composing a very careful message with an SSCCE and everything:

I need help making sense of my interactions with pipenv. I think I am creating a 
Python 3.9 environment and installing packages into it, and then freezing that as 
a Pipfile. But when I check the Pipfile, it shows 3.10 as my frozen environment.

Here's my shell interactions:

pipenv global 3.9
pipenv --three
python --version # output: 'Python 3.9.13'
pipenv install $pkg
cat Pipfile | grep python_version # output: 'python_version = "3.10"'

How does pipenv decide which version of Python to use? Be super succinct.

ChatGPT gives me a lovely (but still verbose) response, whose first and penultimate lines are entirely correct and ultimately all I need: “When you run pipenv --three, pipenv creates a new virtual environment using the latest version of Python that you have installed on your system. To create a virtual environment using Python 3.9, you can use pipenv --python 3.9 instead of pipenv --three.” Ah hah!

I ask it a bunch of related pipenv sense-making questions to get myself smarter on pipenv. ChatGPT gets most of my questions beautifully, perfectly, verifiably right. Then, harkening back to my “sklearn requires scipy requires a compiler” trouble from earlier, I ask a trickier (pure pip) question: “can I pip install precompiled binaries of scikit-learn and its upstreams instead of building from source?” ChatGPT’s answer is so wrong it’s painful: “Yes, there are precompiled binary versions of scikit-learn that you can install instead of building from source. You can try installing the precompiled binary version of scikit-learn by running the following command: pip install -U scikit-learn”. No, sorry, definitely not. In no Python world can you get a precompiled binary for all dependencies by upgrading a downstream user library written in Python. That’s a farcical statement.

And so I shared the story and take-aways at this sprint demo this week:

  • pyenv and pipenv work together… mostly.
  • pipenv will ignore your pyenv!
  • You get your very own Star Trek Computer now… it’s better than a rubber duck, and it’s the kindest, least judgmental coworker ever. (But it’s not your actual coworker; don’t send confidential or sensitive information.)
  • Robots are fallible too.