[Feature] Allow targeting multiples of sequential targets

[aayush7511]

SUMMARY:
Closes #2481

TEST PLAN:
"please outline how the changes were tested"

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a significant enhancement to the model tracing functionality by enabling the grouping of multiple sequential targets into a single subgraph. This new capability provides users with finer control over the granularity of subgraphs, allowing for optimization of memory usage and execution speed based on specific model architectures and hardware constraints.

Highlights

• New Configuration Option: Introduced a new configuration option, sequential_targets_per_subgraph, to control the number of sequential targets grouped into a single subgraph during model tracing.
• Core Tracing Logic Update: Modified core tracing utilities (trace_subgraphs, topological_partition) to accept and utilize this new parameter, enabling more flexible subgraph generation based on user-defined grouping.
• Pipeline and Debug Script Integration: Integrated the new subgraph grouping capability into the sequential pipeline and the tracing debug script, allowing users to leverage this feature for optimized model processing.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces a new sequential_targets_per_subgraph parameter, allowing users to control the number of sequential targets grouped into each subgraph during model tracing. The changes involve adding this parameter to DatasetArguments, trace_subgraphs, and topological_partition functions, and modifying the partitioning logic accordingly. Review comments highlight a critical TypeError bug where the DatasetArguments field object is passed instead of its value, a potential ZeroDivisionError if targets_per_subgraph is non-positive, and several style and formatting inconsistencies (docstring, comments, function call alignment, line length) that need to be addressed for improved code quality and readability.

[gemini-code-assist]

This is passing the field object from the DatasetArguments dataclass instead of the integer value from the parsed command-line arguments. This will cause a TypeError at runtime. You should use args.targets_per_subgraph to get the value provided by the user.

Suggested change

	targets_per_subgraph=DatasetArguments.sequential_targets_per_subgraph
	targets_per_subgraph=args.targets_per_subgraph

[gemini-code-assist]

The docstring has an extra blank line after the new parameter, which is inconsistent with the rest of the docstring formatting in this file.

Suggested change

	:param targets_per_subgraph: number of targets to include per subgraph
	:param targets_per_subgraph: number of targets to include per subgraph

[gemini-code-assist]

This comment doesn't follow PEP 8 spacing guidelines (a space is required after '#'). Also, the comment could be more descriptive of what targets_seen tracks.

Suggested change

	targets_seen = 0 # global counter
	targets_seen = 0 # number of targets encountered so far

[gemini-code-assist]

This if statement has a style issue: the condition is unnecessarily wrapped in parentheses. According to PEP 8, parentheses are not needed for if conditions.

Suggested change

	if(targets_seen % targets_per_subgraph == 0):
	if targets_seen % targets_per_subgraph == 0:

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[kylesayrs]

Looks great! Please consider adding some tests to verify that each subgraph has the submodules you expect (see Subgraph.submodules).

[kylesayrs]

Suggested change

	targets_seen = 0 # number of targets encountered so far
	targets_seen = 0 # number of targets encountered so far

[kylesayrs]

Suggested change

	if(targets_per_subgraph <= 0):
	if targets_per_subgraph <= 0:

[aayush7511]

Hi @kylesayrs ,

To test if each submodule has the necessary targets, I'm generating a subgraph using trace_subgraph. Currently trace_subgraph requires model.config in one of its lines to run.

src/llmcompressor/pipelines/sequential/helpers.py

with contextlib.ExitStack() as stack:
        # calibration context
        stack.enter_context(calibration_forward_context(model))
        stack.enter_context(HooksMixin.disable_hooks())

        # flags useful for tracing
        stack.enter_context(patch_attr(model.config, "_attn_implementation", "eager"))
        stack.enter_context(patch_attr(torch.compiler, "_is_compiling_flag", True))

        # autowrap forwards
        stack.enter_context(autowrap_forwards(ancestors, ignore))

With the way the current tests are setup, the models are simple torch.nn.module layers with no config. What model would you recommend using or should I create a model with a dummy config file.

tests/llmcompressor/pipelines/sequential/test_helpers.py

class DummyModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.seq = torch.nn.Sequential(torch.nn.Linear(10, 20), torch.nn.ReLU())
        self.fc = torch.nn.Linear(20, 5)

    def forward(self, x):
        x = self.seq(x)
        return self.fc(x)

class DummyModelMultipleSequentialLayers(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.layer1 = torch.nn.Linear(10, 10)
        self.layer2 = torch.nn.Linear(10, 10)
        self.layer3 = torch.nn.Linear(10, 10)
        self.layer4 = torch.nn.Linear(10, 10)
        self.layer5 = torch.nn.Linear(10, 10)
        self.layer6 = torch.nn.Linear(10, 10)
        
    
    def forward(self, x):
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        x = self.layer5(x)
        x = self.layer6(x)
        return x

[kylesayrs]

@aayush7511 I recommend using Qwen/Qwen3-0.6B on the meta device, similar to these tests.

[brian-dellabetta]

Thanks for the contribution! lgtm pending open comments

[mergify]

The quality checks have failed. Please run make style and make quality under
the root directory to adddress the lint failures. You will need to install the
dev optional install to get the required linting packages:
https://github.com/vllm-project/llm-compressor/blob/main/CONTRIBUTING.md