0.1.7

Zonos2 TTS ComfyUI v0.1.7

Fixed

• Corrected the mixed FP8 Hugging Face repository:
  • drbaph/ZONOS-FP8 → drbaph/ZONOS2-FP8
• Fixed automatic FP8 model and asset downloads.
• Updated the FP8 model-loader dropdown entry.
• Corrected FP8 links in the English and Chinese READMEs.
• Corrected the project metadata and Hugging Face badge.
• Updated the example workflow JSON.
• Updated the workflow embedded inside the example PNG without changing its rendered image.

Verification

• 35 automated tests passing.
• Workflow JSON and embedded PNG workflow metadata validated.

0.1.6

Zonos2 TTS ComfyUI v0.1.6

Fixed

• Removed redundant ComfyUI memory-manager resume calls during voice cloning.
• Removed the second full bundle resume during speaker embedding extraction.
• Cached model-loader executions no longer unnecessarily resume the active bundle.
• Fixed duplicate Loaded Zonos2Model, Zonos2DAC, and Zonos2SpeakerEncoder log messages.
• Load messages now appear only when a model is newly registered.

Improved

• Main model, DAC, and speaker encoder patchers are resumed through one batched ComfyUI memory-management call.
• Preserved correct ComfyUI/AIMDO usage and residency tracking.
• Added regression coverage for batched resumes and accurate load logging.

Notes

• No ZONOS2 generation math, MoE routing, attention, or FP8 execution behavior was changed.
• A complete ComfyUI restart is required after updating.

Verification

• 35 automated tests passing.

0.1.5

Zonos2 TTS ComfyUI v0.1.5

Added

• Added mixed FP8 E4M3 checkpoint support.
• Added the drbaph/ZONOS-FP8 model catalog entry.
• Added automatic download support for zonos2-fp8-mixed.safetensors.
• Added FP8 metadata, policy, tensor-shape, and runtime validation.
• Added native ComfyUI quantized tensor integration.

FP8 Policy

• MoE expert gate/up (w13) projections use FP8 E4M3.
• Expert down (w2), attention, LM head, routers, embeddings, norms, and other sensitive paths remain BF16.
• FP8 checkpoints support dtype: auto and dtype: bf16.
• Unsupported FP16 execution is rejected with a clear error.

AIMDO

• Added actual-size-aware static and DynamicVRAM selection for FP8 models.
• Mixed FP8 automatically selects AIMDO VBAR below approximately 12.78 GiB total VRAM when DynamicVRAM is enabled.
• FP8 w13 and BF16 w2 expert projections are independently pageable.
• Routed expert projections are loaded into VRAM on demand.
• VBAR residency, page faults, and eviction use real AIMDO operations.
• Systems with sufficient VRAM continue using the faster static loading path.

Downloads

• Main checkpoints, DAC assets, and speaker encoder assets are checked independently.
• Existing complete asset directories are not downloaded again.
• A missing selected checkpoint is still downloaded when shared assets already exist.
• BF16 and FP8 presets download from their respective Hugging Face repositories.

Validation

• Retired all-layer FP8 checkpoint layouts are rejected.
• Malformed 3D expert tensors are rejected before inference.
• Clear compatibility errors are provided for unsupported FP8 checkpoints.
• Added native and loader regression coverage.

Documentation

• Updated English and Chinese documentation.
• Added FP8 installation, memory usage, dtype, AIMDO, model structure, and troubleshooting information.
• Added the mixed FP8 Hugging Face badge and model link.

Verification

• 33 automated tests passing.
• Static FP8 generation verified.
• Real AIMDO VBAR generation verified with all expert w13 and w2 projections registered.

0.1.4

Zonos2 TTS ComfyUI v0.1.4

Highlights

• Added adaptive ComfyUI/AIMDO memory management.
• Added genuine AIMDO VBAR paging for low-VRAM GPUs.
• Improved model loading, unloading, visualization, and lifecycle handling.

AIMDO and VRAM Management

• BF16 main model size is estimated at approximately 14.324 GiB.
• A 3 GiB runtime reserve creates an automatic VBAR cutoff of approximately 17.324 GiB total VRAM.
• 8 GiB, 12 GiB, and 16 GiB GPUs use the real AIMDO CoreModelPatcher and VBAR path when DynamicVRAM is enabled.
• Larger GPUs use the faster static CUDA path when the model fits with the runtime reserve.
• Dynamic loading pages only selected MoE experts into VRAM.
• VBAR allocation, residency, page faults, and eviction are genuine AIMDO operations.
• DAC and speaker encoder remain managed as smaller static modules.
• Automatic path selection is based on total VRAM rather than currently free VRAM.

Model Lifecycle

• Integrated the main model, DAC, and speaker encoder with ComfyUI model management.
• Reuses an existing bundle when model, dtype, and attention settings are unchanged.
• Changing model, dtype, or attention now performs a complete hard unload.
• Improved cleanup of tensors, AIMDO registrations, references, and accelerator caches.
• Restored accurate tensor visibility in ComfyUI Memory Visualization.

Performance

• Optimized single-token MoE decoding to execute only routed experts.
• Avoided scanning all experts during each autoregressive generation step.
• Preserved direct CUDA execution on GPUs with sufficient VRAM.
• Retained file-backed expert weights for dynamic low-VRAM execution.

Fixes

• Progress bars now use the checkpoint's real tensor count.
• Fixed resampler caching so it does not retain the speaker encoder indefinitely.
• Cached resamplers now follow speaker encoder device changes.
• Improved speaker encoder registration with ComfyUI model management.
• Improved model unloading and reloading after configuration changes.

Documentation

• Expanded English and Chinese AIMDO memory-management documentation.
• Documented static and dynamic loading behavior.
• Documented the 14.324 GiB model estimate, 3 GiB reserve, and 17.324 GiB cutoff.
• Added the complete supported-language tier table.
• Added the official Zyphra ZONOS2 blog badge.
• Expanded VRAM and out-of-memory troubleshooting guidance.

Testing

• Added runtime-management, AIMDO selection, lifecycle, progress, resampler, and MoE module tests.
• All 24 tests pass.

0.1.3

ZONOS2 TTS ComfyUI v0.1.3

Performance

• Added optimized single-token MoE expert dispatch.
• Autoregressive generation now runs only the selected expert instead of scanning all 16 experts.
• Measured approximately 1.6x–2x faster token generation on an RTX 5090.
• Generated audio tokens remain identical to the previous implementation.
• Multi-token prompt processing retains the original grouped dispatch path.

Voice Cloning

• Changed clean_speaker_background default to false, matching upstream ZONOS2.
• Improved reference-audio and accurate-mode tooltips.
• Added clearer guidance about voice identity, accent, cadence, emotion, and prosody limitations.
• Expanded troubleshooting recommendations for improving clone similarity and accent retention.

Documentation

• Updated English and Chinese documentation.
• Updated version badges to 0.1.3.
• Documented the optimized MoE decoding path.
• Added detailed reference-audio and sampling guidance.

Testing

• Added top-1 and top-2 MoE dispatch equivalence tests.
• Added regression tests for upstream-compatible clone defaults.
• Verified optimized and original paths produce identical production-model audio tokens.
• All 12 automated tests pass.

0.1.2

ZONOS2 TTS ComfyUI v0.1.2

Changed

• Added uv support to install.py.
  • Uses uv pip install --python <active ComfyUI Python> when available.
  • Falls back to python -m pip.
• Added uv installation instructions to both READMEs.
• Added all workflow and ZONOS2 images to the Chinese README.
• Moved the citation outside the expandable sections.
• Updated package, README badge, and workflow metadata to 0.1.2.

Added

• Added an informational requirements.txt.
  • Dependencies are commented out intentionally.
  • pyproject.toml and install.py remain authoritative.
• Added tests covering uv and pip installer behavior.

Removed

• Removed TODO.md.

Validation

• All 9 automated tests pass.
• Verified both README files contain the same static images and animated ZONOS2 GIF.