Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Propose adding a file format for custom pipelines #109

Merged
merged 1 commit into from
Jul 12, 2024
Merged

Propose adding a file format for custom pipelines #109

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 22 additions & 2 deletions .spellcheck-en-custom.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@ Akash
AMDGPU
Anil
arge
args
arXiv
backend
backends
Expand All @@ -20,6 +21,7 @@ Colab
compositional
Conda
config
configs
Containerfile
cpp
cuBLAS
Expand All @@ -31,6 +33,7 @@ dataset
DCO
Dependabot
dev
disambiguating
ditaa
docstring
dr
Expand All @@ -42,23 +45,29 @@ eval
Excalidraw
exfiltrate
exfiltrating
Filesystem
filesystem
Finetuning
formedness
freeform
GFX
GGUF
GGUFs
GiB
Gmail
gpu
Guang
hardcode
hardcoded
hipBLAS
ilab
impactful
Inferencing
instantiation
instructlab
io
ISA
iters
Jie
JIT
JSON
Expand All @@ -78,31 +87,36 @@ Markdownlint
md
Mergify
Merlinite
merlinite
mimimum
Miniforge
Mixtral
mixtral
MLX
mlx
MMLU
Nakamura
num
NVidia
Nvidia
Ollama
ollama
OpenStax
orchestrator
ots
Params
Pareja
PEFT
Pereira
PlantUML
PLOS
PNG
POC
Podman
podman
PR's
pre
preprint
PR's
pyenv
PyPI
PyTorch
Expand All @@ -117,6 +131,7 @@ Ren
repo
ROCm
RTX
runtime
RX
SaaS
safetensors
Expand All @@ -129,6 +144,8 @@ Shellcheck
Shivchander
Signoff
Sigstore
specifiying
src
Srivastava
Standup
subcommand
Expand All @@ -138,14 +155,16 @@ Sudalairaj
Taj
tatsu
TBD
templating
Tesla
th
tl
TODO
tox
traigers
triager
triager's
Triagers
triager's
triagers
UI
ui
Expand All @@ -165,3 +184,4 @@ XT
XTX
Xu
YAML
yaml
258 changes: 258 additions & 0 deletions docs/sdg/sdg-flow-yaml.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,258 @@
# SDG API - Add a file format for defining custom Flows

## Problem Statement

The `instructlab/sdg` library is introducing more extensive data generation pipelines. To enable customization, we should allow users of the library to provide a configuration file which defines a custom pipeline or extends an existing pipeline.

In terms of the API constructs, a Pipeline is created from a sequence of “block configurations” which express how to instantiate and invoke the individual steps (aka blocks) in the pipeline. A Flow construct serves as a template from which a sequence of block configs can be generated.

## Objective

- Library users can specify a custom flow using a well-defined file format.
- Library users can either use a custom flow standalone, or combine a custom flow with existing flows.
- The file format and library can evolve substantially without making breaking changes.
- Incompatible changes can be introduced while retaining support for existing custom flows for a deprecation period.

## Proposal

### Existing API Review

The current `Pipeline` API allows instantiation with a list of `Block` configurations.
These configurations could come from one or many sources. In its simplest form:

```python
pipeline = Pipeline(block_configs)
```

or if you had two separate lists of block configurations to append together:

```python
pipeline = Pipeline(block_configs1 + block_configs2)
```

### API Additions

We will add an API that instantiates a `Pipeline` object from a YAML file:

```python
pipeline = Pipeline.from_file(ctx, 'mycustomflow.yaml')
```

The YAML file format will mirror the API and look like this:

```yaml
version: 1.0
blocks:
- name: gen_knowledge
type: LLMBlock
config: # LLMBlock constructor kwargs
output_cols: ["question", "response"]
gen_kwargs: # kwargs for block.generate()
max_tokens": 2048,
drop_duplicates: ["question"]
- name: filter_faithfulness
type: FilterByValueBlock
config:
filter_column: judgment
filter_value: YES
operation: eq
drop_columns: ["judgment", "explanation"]
```

## Versioning

A mandatory `version` field in the YAML file expresses major and minor versions (e.g., 1.0, 1.1, 2.0).

Compatibility rules

1. If the major version of the YAML file is higher than the parser can handle, the parser should reject the file.
2. If the minor version of the YAML file is higher than the highest version the parser is aware of, the parser should read the file but ignore any unrecognized content.
3. If the file’s version is lower than the parser version, the parser should provide default values for any configuration introduced in later versions.

Example parsing logic:

```python
def parse_custom_flow(content):
version = content['version']
major, minor = map(int, version.split('.'))

if major > PARSER_MAJOR:
raise IncompatibleVersionError("The custom flow file format is from a future major version.")
elif major <= PARSER_MAJOR and minor > PARSER_MINOR:
logger.warning("The custom flow file may have new features that will be ignored.")
```

### Pipeline Context

The following runtime parameters will no longer be part of the pipeline configuration definition and instead available to blocks via a `PipelineContext` object:

- client - an OpenAI completions API client for talking to the teacher model via the serving backend (i.e. llama-cpp or vLLM)
- model_family - e.g. mixtral or merlinite
- model_id - a path name for the specific teacher model being used
- num_instructions_to_generate - how many samples to generate

For now, we assume there is no need to do any sort of templating in the custom pipelines based on these runtime parameters.

### Model Prompts

Based on whether model_family is mixtral or merlinite, a different prompt is used with the teacher model

```python
_MODEL_PROMPT_MIXTRAL = "<s> [INST] {prompt} [/INST]"
_MODEL_PROMPT_MERLINITE = "'<|system|>\nYou are an AI language model developed by IBM Research. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior.\n<|user|>\n{prompt}\n<|assistant|>\n'"
```

For now, we assume that the `LLMBlock` class will choose the appropriate model prompt based on the family and that there is no need to specify a custom prompt.

### Prompt Config Files

Every LLMBlock references a separate prompt config file, and presumably a custom pipeline will provide custom prompt configs too.

These prompt config files are quite simple YAML files - they contain a single object with system, introduction, principles, examples, and generation keys. See e.g. src/instructlab/sdg/configs/skills/freeform_questions.yaml

We will continue to use these config files unchanged, and custom files can be specified with an absolute path. Relative paths are assumed to be relative to the Python package e.g. `configs/skills/...`.

### Model Serving

Custom pipelines may have more unique model serving requirements. Instead of serving just one model, we may need to launch the model server with a model and an additional model with adapter. vLLM, for example, can host both a model and a model+adapter under two different model IDs.

The pipeline author needs some way of disambiguating between these multiple models - i.e. the definition of each `LLMBlock` needs to specify a particular model.

Right now the `Pipeline` constructor takes two relevant parameters - the OpenAI client instance, and the model ID for the default model. It's important to note that this model ID is defined by the user at runtime, and it may not match the model IDs that the pipeline author used.

The use cases will be:

1. Most LLMBlock definitions will use the default teacher model - and we can make the semantic that if the pipeline author doesn't specify a model in an `LLMBlock`, the default in `PipelineContext.model_id` is used.
2. In cases where a model+adapter is to be served, the pipeline author should choose a descriptive model ID using `block.gen_kwargs.model_id` and the user should ensure that this is the model ID that is served.

For example, a pipeline author might define:

```yaml
version: "1.0"
blocks:
- name: gen_questions
type: LLMBlock
config:
config_path: configs/skills/freeform_questions.yaml
add_num_samples: True
gen_kwargs:
model_id: mycustomadapter
output_cols:
- question
drop_duplicates:
- question
```

and the user will be required to define a serving configuration like:

```bash
--lora-modules=mycustomadapter=path/to/my_custom_adapter
```

### Re-use of Built-in Pipelines

A custom pipeline may want to extend an existing built-in pipeline. In that
case, a new block type, `ImportBlock`, may be used to import the blocks from
another configuration file.

```yaml
version: "1.0"
blocks:
- name: import_from_full
type: ImportBlock
path: configs/full/synth_freeform_skills.yaml
- name: custom_post_processing_block
type: LLMBlock
...
```

### CLI Integration

As of the current version of `ilab`, it supports `simple` and `full` as parameters to `--pipeline` to select one of the two types of built-in pipelines included in the library.

Once we have support for loading custom pipelines, we need a way for these to be specified with the CLI. We believe the most common case for custom pipelines is for them to extend the `full` pipeline and, as such, we should support extending existing pipelines with a custom pipeline rather than simply specifiying a single pipeline.

Here is a proposed CLI UX for this:

> `ilab data generate`

Use the default pipeline, `simple`.
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this default should be achieved as ilab data generate --pipeline PATH_TO_THE_SIMPLE_WORKFLOW_YAML for the purpose of unification.
we can provide alias like simple or full to the built-in YAML files to make users' like easy.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, simple and full should be thought of as aliases

But it is a bit more complex, and it had occurred to me we'll need to get into this as part of the CLI integration

Currently, the aliases look more like this:

full:
    knowledge:
      - MMLU_BENCH_FLOW
      - SYNTH_KNOWLEDGE_FLOW
    freeform_skills:
      - SYNTH_FREEFORM_SKILLS_FLOW
    grounded_skills:
      - SYNTH_GROUNDED_SKILLS_FLOW
simple:
    knowledge:
      - SIMPLE_KNOWLEDGE_FLOW
    freeform_skills:
      - SIMPLE_FREEFORM_SKILLS_FLOW
    grounded_skills:
      - SIMPLE_GROUNDED_SKILLS_FLOW

So we need to be able to express a (knowledge, freeform_skills, grounded_skills) tuple really ...

Copy link
Contributor Author

@markmc markmc Jul 11, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok, I think this issue warrants a significant change to the design

I order to specify a pipeline, you need to specify a chain of blocks for knowledge, freeform_skills, and grounded_skills

I think we can see from the POC that keeping the chains for knowledge, freeform_skills, and grounded_skills in separate files aids clarity

So I think we need something like this

src/
  instructlab/
    sdg/
      pipelines/
        simple/
          knowledge.yaml
          freeform_skills.yaml
          grounded_skills.yaml
        full/
          knowledge.yaml   # also contains the current contents of mmlu_bench.yaml
          freeform_skills.yaml
          grounded_skills.yaml

and a custom pipeline would be a directory with these 3 YAML files

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this could be confusing what a pipeline is ... I would rather say there are actually 3 pipelines under simple/full and we route them differently in our particular frontend. this semantics should not be enforced

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Right, "full" or "simple" is a set of pipelines. I agree with that.

But likewise, when using custom pipelines, you should be providing a set of custom pipelines

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

on Slack, we talked about add the idea of a PipelineRouter which I guess would have an interface like this:

class PipelineRouter:
    def __init__(self, ...);
        ...

    def generate(self, dataset) -> Dataset:
       ...

(hmm, that's the same interface as a Pipeline which suggests a base class that Pipeline and Router inherit from)

and there would be two variants:

class DirectRouter(Router):
    def __init__(self, pipeline_yaml):
        self.pipeline = load_pipeline(pipeline_yaml)

    def generate(dataset):
        return self.pipeline.generate(dataset)

and the use case for this would be if you are developing a pipeline against a known data set?

But the more common use case is covered by this variant that understands how our taxonomy repo is structured:

class TaxonomyRouter(Router):
    def __init__(self, freeform_skills_pipeline_yaml, grounded_skills_pipeline_yaml, knowledge_pipeline_yaml):
       self.freeform_skills_pipeline = load_pipeline(freeform_skills_pipeline_yaml)
       ....

   def generate(samples):
       if samples[0].get("document"):
         pipeline = self.knowledge_pipeline
      elif samples[0].get("seed_context"):
         pipeline = self.grounded_skills_pipeline
      else:
          pipeline = self.freeform_skills_pipeline
      return pipeline.generate(samples)

"full" and "simple" would be an alias for a set of yaml files

(the above is just a sketch, it would evolve quite a bit as you hack on it)

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

And the point of all of this is "full" isn't a pipeline, but instead its some logic that understands the taxonomy repo and looks at the data set before picking from one of 3 pipelines

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

OK, I think this is making sense to me.

From the ilab CLI perspective, it only cares about the taxonomy variant. It can still just talk about a "pipeline" as a user concept, but the target is either:

  • an alias that sets up a TaxonomyRouter with 3 built-in pipelines
  • a path to a directory that has 3 pipeline yamls in it, and we initialize the TaxonomyRouter with that

altered syntax is over here: https://github.com/instructlab/dev-docs/pull/109/files#r1674565115

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, the class definitions and methods themselves make sense to me as well.
Abstraction-wise, one thing I've been wondering though is whether routing is a thing that Pipeline should be responsible for, or it's a higher level concept should route pipelines (in our case, it could be the SDG).
Also as per Mark's word: routing sounds like "an object that has a set of pipelines, and knows which pipeline to use for different input datasets", which also doesn't sound like another pipeline but something conveys the details about the actual synthetic data generation process (a bunch of pipelines applied to a set of files from taxonomy, in our case).
PS: open to deal with this later as the design above already unblocks us I believe.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Out of all of this, I think the really important details are:

  • ilab data generate --pipeline can take a path to a directory. I updated the doc text to reflect that part already.

We expect the contents of that directory to have 3 specific files: knowledge.yaml, grounded_skills.yaml, and freeform_skills.yaml. I will add that in a moment.

There are also some really good points about how we can implement this in terms of abstractions. I am comfortable leaving that part out of the doc and see how it shakes out in the implementation. I have a feeling the final details will get sorted as we see what it looks like in implementation, but knowing that it's inspired by our discussion here. Let me know if anyone feels diffrently.


> `ilab data generate --pipeline full`

Use the built-in `full` pipeline.

> `ilab data generate --pipeline path/to/custom_pipeline_directory/`

Use a custom pipeline configuration. The custom pipeline may include references to the built-in flows to be used in conjunction with custom ones, but those details are contained within the yaml files in the custom directory.

### File and Directory Structure

The existing contents of `default_flows.py` will become these files in the source tree:

```text
src/
instructlab/
sdg/
pipelines/
simple/
knowledge.yaml
freeform_skills.yaml
grounded_skills.yaml
full/
knowledge.yaml # also contains the current contents of mmlu_bench.yaml
freeform_skills.yaml
grounded_skills.yaml
```

When the `--pipeline` option to `ilab data generate` is used to point to a
custom directory, we will assume that the same 3 files are present. All three
files will be loaded and used according to the type of taxonomy additions
present when running `ilab data generate`.

### Future CLI Improvements

A possible improvement would be to have a well-defined place on the filesystem where custom pipeline configs can be automatically loaded and included as options to the `--pipeline` parameter.

For example, if the config format included new parameters, `name: full-extended` and `extends: full`, and the CLI discovered and loaded it automatically, we could support `--pipeline full-extended` without needing the additional `--pipeline-extend` option.

`/usr/share/instructlab/sdg/` is a proposed location for this as a place for a distribution of InstructLab to include pre-defined custom pipelines, at least for Linux. See the [Filesystem Hierarchy Standard](https://refspecs.linuxfoundation.org/FHS_3.0/fhs/ch04s11.html) for more details on why this path is appropriate for this use-case.

It would also make sense to support a configuration directory for user's own custom pipeline configurations. Assuming there is a base config directory, these could go in a `sdg` subdirectory. There is a separate proposal that discusses a proposed configuration location: <https://github.com/instructlab/dev-docs/pull/104>. Note this is separate from the distribution-provided, read-only pipelines discussed above with a different location.

If we have a location with pipeline examples then a nice to have would be to have a `ilab data generate --list-pipelines`.

## Alternative Approaches

Alternatives already considered and discarded are listed below.

### No Custom Flows

It would be preferable to not support custom flows, especially so early in the project because:

- We will need an extensive API to support this customization, and we will need to be careful about making incompatible changes to that API once it has been adopted.
- We would learn more about the pipelines that users are creating if they were added to the library.

This approach was discarded because of strong demand from downstream users to define custom flows to encapsulate proprietary pipeline configuration..

### Custom Flows as Code

If we have an API for creating flows, users could define these custom flows in Python rather than with a configuration file format.

This approach was discarded because of a desire by downstream users to separate reusable logic from proprietary pipeline configuration.

The initial version of the initial SDG library API design (#98) proposed using YAML files and this was changed to Python code based on this feedback:

> Does this need to be a yaml file?
>
> or is it actually a Python dict passed to the library?
>
> I actually think it would be a nice simplification to not worry about config files at all, and from the library perspective, assume configuration is passed in via data structures.
>
> How that config is constructed could be a problem of the library consumer. Maybe they hardcode it. maybe they allow a subset to be considered. Some could be driven by CLI args, for example.

Since adopting YAML may now appear contradictory to that feedback, it is useful to understand how the feedback relates to this new design:

1. The feedback assumes that YAML will be used for custom pipelines, but wonders whether it would be better to implement that in the CLI instead of the library.
2. Not called out is that at the time it was unclear whether custom pipeline definitions would also need to include custom model serving configuration - if so, the model serving configuration would not belong in the SDG library. It is now better understood that no model serving configuration needs to be included in the pipeline definitions. (See above)
3. The POC implementation of this format makes it clear - in a way that wasn't clear from an API design - that using the YAML format within the library is an improvement.