Skip to content

Support official VibeVoice 1.5B checkpoint - #726

Open
justinchuby wants to merge 3 commits into
mainfrom
justinchuby-support-official-vibevoice-tts
Open

Support official VibeVoice 1.5B checkpoint#726
justinchuby wants to merge 3 commits into
mainfrom
justinchuby-support-official-vibevoice-tts

Conversation

@justinchuby

Copy link
Copy Markdown
Member

Summary

  • make microsoft/VibeVoice-1.5B@c00898d a pinned direct build target while retaining its official weights and model identity
  • resolve only the required Transformers-native config and processor assets from vibevoice/VibeVoice-1.5B-hf@edc39f8
  • reject distinct unimplemented official VibeVoice realtime, ASR, and tokenizer repositories instead of routing them as TTS

Validation

  • focused VibeVoice builder, CLI, graph, parity, and weight-alignment tests
  • complete 1,204-key official-to-native safetensors index check
  • full non-integration regression suite and formatter
  • official no-weight CLI build producing all eight ONNX stages

Resolve the official legacy checkpoint against pinned Transformers-native config and processor sidecars while retaining official weights and provenance. Reject distinct unimplemented VibeVoice collection architectures explicitly.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
justinchuby requested review from a team and a lite review from Copilot September 4, 2026 16:35
@github-actions

github-actions Bot commented Sep 4, 2026

Copy link
Copy Markdown

🏗️ Architecture Diff

Comparing 8c38c06d7b888b

Model Sub-model Changes Status

No architecture changes detected.


Legend: ⚪ No change · 🔵 Minor (attrs/inits) · 🟡 Moderate (nodes added/removed) · 🔴 Major (interface changed)

@github-actions

github-actions Bot commented Sep 4, 2026

Copy link
Copy Markdown

Performance Comparison

Comparing 8c38c06d7b888b

Model Metric Baseline Current Delta
bert (feature-extraction) model_size_bytes 359 KB 359 KB +0.0%
bert (feature-extraction) num_nodes 68 68 +0.0%
falcon model_size_bytes 364 KB 364 KB +0.0%
falcon num_nodes 66 66 +0.0%
gemma2 model_size_bytes 428 KB 428 KB +0.0%
gemma2 num_nodes 105 105 +0.0%
gpt2 model_size_bytes 324 KB 324 KB +0.0%
gpt2 num_nodes 54 54 +0.0%
llama model_size_bytes 425 KB 425 KB +0.0%
llama num_nodes 60 60 +0.0%
llama (static-cache) model_size_bytes 425 KB 425 KB +0.0%
llama (static-cache) num_nodes 56 56 +0.0%
mamba (ssm-text-generation) model_size_bytes 296 KB 296 KB +0.0%
mamba (ssm-text-generation) num_nodes 94 94 +0.0%
phi3 model_size_bytes 421 KB 421 KB +0.0%
phi3 num_nodes 58 58 +0.0%
phi3 (static-cache) model_size_bytes 421 KB 421 KB +0.0%
phi3 (static-cache) num_nodes 54 54 +0.0%
qwen2 model_size_bytes 425 KB 425 KB +0.0%
qwen2 num_nodes 60 60 +0.0%
qwen2 (static-cache) model_size_bytes 425 KB 425 KB +0.0%
qwen2 (static-cache) num_nodes 56 56 +0.0%
qwen3_5_moe (hybrid-text-generation) model_size_bytes 506 KB 506 KB +0.0%
qwen3_5_moe (hybrid-text-generation) num_nodes 265 265 +0.0%
qwen3_5_text (hybrid-text-generation) model_size_bytes 458 KB 458 KB +0.0%
qwen3_5_text (hybrid-text-generation) num_nodes 127 127 +0.0%
qwen3_5_vl (hybrid-qwen-vl) model_size_bytes 977 KB 977 KB +0.0%
qwen3_5_vl (hybrid-qwen-vl) num_nodes 450 450 +0.0%
t5 (seq2seq) model_size_bytes 836 KB 836 KB +0.0%
t5 (seq2seq) num_nodes 176 176 +0.0%
whisper (speech-to-text) model_size_bytes 1008 KB 1008 KB +0.0%
whisper (speech-to-text) num_nodes 128 128 +0.0%

No performance regressions.

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

🟢 Approval recommended

The change is well-scoped, heavily tested across builder/CLI/weight-alignment paths, and only minor nits remain.

Pull request overview

Adds first-class support for the official microsoft/VibeVoice-1.5B checkpoint while preserving official checkpoint identity/weights and pinning only the required Transformers-native sidecar assets (config/processor) from the HF conversion mirror. This integrates the new provenance model into both the Transformers builder and the CLI runtime-asset selection so exports remain reproducible and fail-closed for unimplemented VibeVoice collection entries.

Changes:

  • Introduces a pinned VibeVoice source resolver + official→native weight key conversion path, and threads it through the Transformers build flow (config probing, metadata, and weight preprocessing).
  • Updates CLI/runtime config emission to optionally download processor assets from the pinned sidecar source without changing the requested checkpoint provenance.
  • Expands unit/integration/golden coverage to validate identity preservation, sidecar pinning behavior, unsupported-ID rejection, and official index key mapping.
File summaries
File Description
tests/weight_alignment_test.py Adds an integration test ensuring the official safetensors index keys map 1:1 onto the pinned HF-native conversion layout.
tests/vibevoice_golden_test.py Routes config/processor resolution through the pinned sidecar source for golden/parity runs while building from the official checkpoint target.
tests/cli_test.py Verifies CLI build preserves requested model identity and that runtime assets can be sourced from the pinned processor sidecar.
src/mobius/models/vibevoice.py Adds VibeVoiceSources, resolves official vs executable sources, and implements official weight key conversion + collision protection.
src/mobius/models/vibevoice_test.py Adds unit tests for source resolution, revision pinning, unsupported model rejection, and key conversion correctness.
src/mobius/integrations/transformers/_builder.py Uses resolve_vibevoice_sources to separate checkpoint provenance from executable sidecars and passes layout info into preprocess_weights.
src/mobius/integrations/transformers/_builder_test.py Adds regression tests ensuring official VibeVoice uses pinned sidecars while downloading official weights; validates fail-closed behavior.
src/mobius/main.py Pins the early Diffusers probe for official VibeVoice and introduces _runtime_asset_source to select processor downloads from metadata.
docs/model-catalog.md Updates catalog entry and documents the official checkpoint + pinned sidecar behavior and explicit rejection of unimplemented collection IDs.
Review details
  • Files reviewed: 9/9 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.

Comment thread tests/weight_alignment_test.py Outdated
justinchuby and others added 2 commits September 4, 2026 10:01
Reference the follow-up issue for replacing the temporary community sidecar dependency when Microsoft publishes pinned HF-native sidecars.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: 000b46be-a743-47cb-a063-dc3e16cc03d4
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants