Add VibeVoice Realtime TTS export - #728
Open
justinchuby wants to merge 2 commits into
Open
Conversation
Add source-faithful staged ONNX export for the pinned Microsoft VibeVoice Realtime checkpoint, with explicit cache state, host orchestration metadata, processor contracts, and graph/parity coverage. Reject unsupported ORT GenAI configuration rather than advertising incomplete runtime support. Co-authored-by: Copilot App <[email protected]> Signed-off-by: Justin Chu <[email protected]>
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
Performance Comparison
|
![github-code-quality[bot]](https://avatars.githubusercontent.com/u/9919?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
🔵 Needs a closer look
It introduces a large new multi-stage model/task/config + exporter behavior surface area where correctness depends on nuanced runtime and source-parity constraints that warrants final human review.
Pull request overview
Adds a source-faithful, host-orchestrated multi-stage Mobius export pipeline for VibeVoice Realtime TTS (microsoft/VibeVoice-Realtime-0.5B), wiring it into the registry/build system and ensuring runtime exporters fail-closed (ORT GenAI) or emit advisory-only metadata (ONNX GenAI), with extensive source-parity and contract tests.
Changes:
- Introduces
VibeVoiceStreaming*config/model/task to export six explicit ONNX stages (embedding, LM backbone, TTS backbone, speech connector, diffusion head, audio decoder) with explicit KV + conv-cache I/O. - Registers the new architecture/model_type, pins the first Transformers config probe to an immutable revision, and annotates source revision metadata where applicable.
- Adds L1–L3 test coverage (graph contracts, cache/metadata contracts, CUDA fp16 GQA lowering expectations, source parity + continuation), plus CI requirements and testdata case metadata.
File summaries
| File | Description |
|---|---|
| tests/weight_alignment_test.py | Excludes vibevoice_streaming from initializer-name identity mapping assumptions. |
| tests/model_coverage_test.py | Marks architecture alias as covered by the new staged VibeVoice Realtime tests. |
| tests/build_graph/speech_test.py | Adds stage expectations for the new vibevoice-streaming-tts task. |
| tests/_test_configs.py | Adds tiny-config overrides for vibevoice_streaming graph-build coverage. |
| testdata/cases/audio/vibevoice-realtime-0.5b.yaml | Adds pinned-case metadata explaining why golden L4/L5 isn’t asserted for host-orchestrated Realtime. |
| src/mobius/tasks/_vibevoice.py | Refactors conv-cache helpers to shared cache utilities. |
| src/mobius/tasks/_vibevoice_streaming.py | New task exporting six Realtime stages and embedding host-protocol metadata in graph properties. |
| src/mobius/tasks/_cache_utils.py | Adds shared helpers for causal-conv cache inputs/outputs. |
| src/mobius/tasks/init.py | Exposes and registers VibeVoiceStreamingTask under vibevoice-streaming-tts. |
| src/mobius/models/vibevoice.py | Extends tokenizer decoder to optionally preserve ratio order (needed by Realtime decoder). |
| src/mobius/models/vibevoice_streaming.py | New staged Realtime model with strict checkpoint weight routing and explicit mask/cache behavior. |
| src/mobius/models/vibevoice_streaming_test.py | New L1–L3 tests: stage graph contracts, cache I/O, metadata, CUDA fp16 GQA lowering, and source parity. |
| src/mobius/models/init.py | Exports VibeVoiceStreamingForConditionalGeneration. |
| src/mobius/integrations/transformers/_builder.py | Pins first config probe for Realtime model ID and tags source revision metadata for vibevoice_streaming. |
| src/mobius/integrations/transformers/_builder_test.py | Tests that “no revision provided” still pins the first config probe for Realtime. |
| src/mobius/integrations/ort_genai/auto_export.py | Fails closed for vibevoice_streaming with an explicit host-orchestration error. |
| src/mobius/integrations/onnx_genai/auto_export.py | Emits advisory-only component contract metadata for vibevoice_streaming. |
| src/mobius/_registry.py | Registers vibevoice_streaming and its architecture-discriminating alias, including pinned model_id/revision. |
| src/mobius/_configs/vibevoice.py | Adds streaming tokenizer/diffusion/config dataclasses, parsing + validation, and torch_dtype-to-dtype resolution. |
| src/mobius/_configs/init.py | Exports the new streaming config types from the configs package. |
| requirements/ci/vibevoice.txt | Pins the Microsoft VibeVoice source package commit for parity tests (installed in CI with --no-deps). |
Review details
- Files reviewed: 21/21 changed files
- Comments generated: 1
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
🏗️ Architecture Diff
No architecture changes detected. ✅ Legend: ⚪ No change · 🔵 Minor (attrs/inits) · 🟡 Moderate (nodes added/removed) · 🔴 Major (interface changed) |
Keep the specialized TTS backbone invocation signature-compatible, record its architecture alias as source-parity covered, and apply the repository formatter. Co-authored-by: Copilot App <[email protected]> Signed-off-by: Justin Chu <[email protected]>
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
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
microsoft/VibeVoice-Realtime-0.5Bat6bce5f06044837fe6d2c5d7a71a84f0416bd57e4: embedding, the 4-layer lower LM backbone, 20-layer TTS backbone, speech connector, diffusion velocity head, and causal audio decoder.79e516a3e20b599f137c9da03410a2a0b473b63b, adds source parity/continuation coverage, full-config validation, processor-contract coverage, exhaustive weight routing, revision-pin protection, and honest runtime metadata.Validation
9.536743e-7; FP32 CPU 5.248908 s / RTF 39.3668.Runtime boundaries