feat(inspect): render composite pipeline structure (pipeline-led)#903
Open
timenick wants to merge 8 commits into
Open
feat(inspect): render composite pipeline structure (pipeline-led)#903timenick wants to merge 8 commits into
timenick wants to merge 8 commits into
Conversation
Composite models (encoder-decoder seq2seq, etc.) previously inspected as a single non-runnable component (e.g. a bart decoder shown as Task: text2text-generation with a dangling encoder_hidden_states input). Inspect now surfaces the composite structure: - Model Information: the Task row shows the pipeline tasks the model_type serves (e.g. 'summarization · table-question-answering [composite]'), with an Export row for the component -> export-task breakdown. - A dedicated 'Composite Pipeline' panel lists the components and how to build one. - JSON: 'task' stays the granular machine task (unchanged contract); additive 'pipeline_tasks' + 'composite' fields. Gated on the resolver's detected composite (TaskResolution.composite), so only auto-detected composites render the view -- CLIP inspected for feature-extraction and other single-component exports are unchanged. pipeline_tasks are derived from the live COMPOSITE_MODEL_REGISTRY (new loader.composite_pipeline_tasks), architecture-agnostic. New API: loader.composite_pipeline_tasks; inspect.CompositeInfo + resolve_composite_info.
DingmaomaoBJTU
left a comment
DingmaomaoBJTU
left a comment
Collaborator
There was a problem hiding this comment.
Good overall architecture — the pipeline-led composite view is clearly scoped (gated on resolver detection, not registry presence alone) and the additive JSON design is clean. I found a few correctness issues worth fixing before merge.
Address review on #903: resolve_composite_info now returns None when the detected composite yields no pipeline tasks for the model_type (registry divergence), instead of building a CompositeInfo with empty pipeline_tasks that renders a broken '[composite]' Task row. Add a regression test pinning the guard.
DingmaomaoBJTU
left a comment
DingmaomaoBJTU
left a comment
Collaborator
There was a problem hiding this comment.
Follow-up review after commit 90f24b8.
Resolved: Finding #1 (downstream): the pipeline_tasks=[] guard was added in resolve_composite_info — CompositeInfo with empty pipeline_tasks no longer reaches the formatter. Finding #5: the registry-divergence monkeypatch test was added.
Still open: Three issues remain — the falsy None/{} conflation in the same guard, the issubclass registry inconsistency that is the root cause of finding #2, and the formatter's own lack of a guard (finding #3). The side-effect import warning (#4) is also unaddressed.
…tighten guards - register_composite_model now requires a WinMLCompositeModel subclass, making the registry invariant explicit at its single write point; the (now redundant) issubclass filters in composite_pipeline_tasks and _composite_components_for_task are removed so all readers uniformly trust the registry (matching resolve_composite) - resolve_composite_info gates on `detected_components is None` (not falsy) so a genuine empty-breakdown composite is not conflated with the un-set case - output_table guards both composite branches on pipeline_tasks so a directly constructed CompositeInfo with empty pipeline_tasks cannot render a bare [composite] Task row - tests: registry invariant, empty-components case, formatter empty-pipeline guard
…accessor - Add _composite_registry(): the single registry-load trigger the three readers (resolve_composite / composite_pipeline_tasks / _composite_components_for_task) now share. It raises RuntimeError when COMPOSITE_MODEL_REGISTRY is empty, so a moved/renamed registration fails loudly instead of silently returning []/None and disabling the composite feature unnoticed. - Fix a mypy break the main merge surfaced: #896 re-tightened the _composite_components_for_task annotation to type[WinMLCompositeModel] while the prior commit had removed that import; re-add it under TYPE_CHECKING. - Test: the accessor raises loudly when the registry is empty.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
winml inspectnow surfaces the composite pipeline structure for composite models (encoder-decoder seq2seq, etc.). Previously a composite was shown as its single exported component — e.g. a bart decoder rendered asTask: text2text-generationwith a danglingencoder_hidden_statesinput ("non-runnable half"), with no hint that the model serves higher-level pipelines.Before
After (pipeline-led)
plus a dedicated Composite Pipeline panel listing the components and how to build a specific pipeline.
Behavior
TaskResolution.composite), so only models whose resolved task bridges to a composite show the view. A CLIP inspected forfeature-extraction(a runnable single encoder) and all non-composites are unchanged.COMPOSITE_MODEL_REGISTRY(newloader.composite_pipeline_tasks) — no hardcoded model/task names. bart →summarization · table-question-answering, marian →translation, qwen3 →text-generation, etc.taskstays the granular machine task (text2text-generation);pipeline_tasks+compositeare additive.API
loader.composite_pipeline_tasks(model_type) -> list[str]inspect.CompositeInfo,inspect.resolve_composite_info(...)Notes
--task/--model-classyields no composite view (pinscomposite=None) — the pipeline-led view is intentionally scoped to auto-detection.blip(infersimage-text-to-text, registers underimage-to-text) is not reached by the seq2seq detection bridge today; out of scope here.