[PR #3335] [MERGED] Add ComfyUI-Qwen-TTS-5x (transformers 5.x support) to node_db/new #3205

New issue

Closed

opened 2026-10-28 21:16:04 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-28 21:16:04 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/ComfyUI-Manager/pull/3335
Author: @jesse890423
Created: 10/2/2026
Status: ✅ Merged
Merged: 10/11/2026
Merged by: @ltdrdata

Base: main ← Head: add-qwen3-tts-5x

📝 Commits (2)

6140aa0 Add ComfyUI-Qwen-TTS-5x to node_db/new
84cf403 Use the repo name as the entry id and spell out the coexistence rule

📊 Changes

1 file changed (+12 additions, -0 deletions)

View changed files

📝 node_db/new/custom-node-list.json (+12 -0)

📄 Description

Add ComfyUI-Qwen-TTS-5x to node_db/new

What

Registers jesse890423/ComfyUI-Qwen-TTS-5x,
a fork of flybirdxx/ComfyUI-Qwen-TTS
v1.0.7 adapted for the transformers 5.x environment.

Why a fork is needed

Upstream v1.0.7 declares transformers>=4.57.0,<5.0.0, so it does not claim 5.x
support. Running it on transformers 5.14.1 fails in five places, one of them
silently:

#	Symptom on transformers 5.x
1	`AttributeError` on load: `pad_token_id` is no longer copied onto `PretrainedConfig`
2	`ROPE_INIT_FUNCTIONS["default"]` removed in 5.14 (mapping it to `"linear"` would be wrong)
3	`create_causal_mask` signature change (`input_embeds` -> `inputs_embeds`, `cache_position` dropped)
4	Crash: `mat1 and mat2 shapes cannot be multiplied (1x55296 and 2048x2048)`
5	Silent failure: `inv_freq` is left as uninitialised meta memory, so audio is produced but the voice is completely wrong, with no exception and no warning

Fix 5 is worth calling out because it raises nothing. The model just emits
meaningless acoustic frames, which is easy to misdiagnose as a bad reference
recording.

The fork also fixes the progress bar, which never moved (four separate causes),
and a deprecation shim that printed the same line 59 times per model load.

Verification

Every fix was verified by measurement rather than inspection.

ComfyUI        0.37.0
transformers   5.14.1
torch          2.13.0+cu130
plugin         ComfyUI-Qwen-TTS v1.0.7

RoPE output after the inv_freq fix matches the closed form exactly:

pos 0 -> cos=[1.0, 1.0, 1.0]                sin=[0.0, 0.0, 0.0]
pos 1 -> cos=[0.540302, 0.692504, ...]     sin=[0.841471, 0.721414, ...]

Compatibility

Folder: custom_nodes/ComfyUI-Qwen-TTS-5x. The name field in
pyproject.toml is deliberately kept equal to the repository name so the
folder users see matches the repository they cloned.
Node class names are unchanged (FB_Qwen3TTS*), so existing workflows keep
working. nodename_pattern is ^FB_Qwen3TTS.
Do not install both. Because the class names are identical, enabling this
fork and upstream at the same time makes the nodes clash. Users should disable
or remove one of them (rename its folder to <name>.disabled, or use Manager's
Disable button).
transformers pin widened to >=4.57.0. The previous upper bound would make an
installer downgrade a 5.x environment, which would undo every fix here. The code
supports both 4.x and 5.x.
No model weights are bundled. Weights come from Qwen3-TTS and stay under the
Qwen3-TTS License Agreement.

License

Apache-2.0 for the plugin code. LICENSE (Apache-2.0 text) and NOTICE (upstream
attributions) are included in the repository. The fork complies with clause
4.1(a): the README states that the modifications are not endorsed by the original
right-holder.

Entry

{
  "author": "jesse890423",
  "title": "ComfyUI-Qwen-TTS-5x",
  "id": "comfyui-qwen-tts-5x",
  "reference": "https://github.com/jesse890423/ComfyUI-Qwen-TTS-5x",
  "nodename_pattern": "^FB_Qwen3TTS",
  "files": [
    "https://github.com/jesse890423/ComfyUI-Qwen-TTS-5x"
  ],
  "install_type": "git-clone",
  "description": "Speech synthesis, voice cloning and voice design nodes for Qwen3-TTS, patched for transformers 5.x ..."
}

Checklist

Entry added to node_db/new/custom-node-list.json
install_type is git-clone
Repository URL is reachable and public
LICENSE present (Apache-2.0)
Fork relationship and attribution documented in README and NOTICE

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/Comfy-Org/ComfyUI-Manager/pull/3335 **Author:** [@jesse890423](https://github.com/jesse890423) **Created:** 10/2/2026 **Status:** ✅ Merged **Merged:** 10/11/2026 **Merged by:** [@ltdrdata](https://github.com/ltdrdata) **Base:** `main` ← **Head:** `add-qwen3-tts-5x` --- ### 📝 Commits (2) - [`6140aa0`](https://github.com/Comfy-Org/ComfyUI-Manager/commit/6140aa0db5251b361970b15acb99edd34cdd7c70) Add ComfyUI-Qwen-TTS-5x to node_db/new - [`84cf403`](https://github.com/Comfy-Org/ComfyUI-Manager/commit/84cf40314e92740634bc213be47cef01faec5d9c) Use the repo name as the entry id and spell out the coexistence rule ### 📊 Changes **1 file changed** (+12 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `node_db/new/custom-node-list.json` (+12 -0) </details> ### 📄 Description # Add ComfyUI-Qwen-TTS-5x to node_db/new ## What Registers [jesse890423/ComfyUI-Qwen-TTS-5x](https://github.com/jesse890423/ComfyUI-Qwen-TTS-5x), a fork of [flybirdxx/ComfyUI-Qwen-TTS](https://github.com/flybirdxx/ComfyUI-Qwen-TTS) v1.0.7 adapted for the transformers 5.x environment. ## Why a fork is needed Upstream v1.0.7 declares `transformers>=4.57.0,<5.0.0`, so it does not claim 5.x support. Running it on transformers 5.14.1 fails in five places, one of them silently: | # | Symptom on transformers 5.x | |---|---| | 1 | `AttributeError` on load: `pad_token_id` is no longer copied onto `PretrainedConfig` | | 2 | `ROPE_INIT_FUNCTIONS["default"]` removed in 5.14 (mapping it to `"linear"` would be wrong) | | 3 | `create_causal_mask` signature change (`input_embeds` -> `inputs_embeds`, `cache_position` dropped) | | 4 | Crash: `mat1 and mat2 shapes cannot be multiplied (1x55296 and 2048x2048)` | | 5 | **Silent failure**: `inv_freq` is left as uninitialised meta memory, so audio is produced but the voice is completely wrong, with no exception and no warning | Fix 5 is worth calling out because it raises nothing. The model just emits meaningless acoustic frames, which is easy to misdiagnose as a bad reference recording. The fork also fixes the progress bar, which never moved (four separate causes), and a deprecation shim that printed the same line 59 times per model load. ## Verification Every fix was verified by measurement rather than inspection. ``` ComfyUI 0.37.0 transformers 5.14.1 torch 2.13.0+cu130 plugin ComfyUI-Qwen-TTS v1.0.7 ``` RoPE output after the `inv_freq` fix matches the closed form exactly: ``` pos 0 -> cos=[1.0, 1.0, 1.0] sin=[0.0, 0.0, 0.0] pos 1 -> cos=[0.540302, 0.692504, ...] sin=[0.841471, 0.721414, ...] ``` ## Compatibility - **Folder**: `custom_nodes/ComfyUI-Qwen-TTS-5x`. The `name` field in `pyproject.toml` is deliberately kept equal to the repository name so the folder users see matches the repository they cloned. - **Node class names are unchanged** (`FB_Qwen3TTS*`), so existing workflows keep working. `nodename_pattern` is `^FB_Qwen3TTS`. - **Do not install both.** Because the class names are identical, enabling this fork and upstream at the same time makes the nodes clash. Users should disable or remove one of them (rename its folder to `<name>.disabled`, or use Manager's Disable button). - `transformers` pin widened to `>=4.57.0`. The previous upper bound would make an installer downgrade a 5.x environment, which would undo every fix here. The code supports both 4.x and 5.x. - No model weights are bundled. Weights come from Qwen3-TTS and stay under the Qwen3-TTS License Agreement. ## License Apache-2.0 for the plugin code. `LICENSE` (Apache-2.0 text) and `NOTICE` (upstream attributions) are included in the repository. The fork complies with clause 4.1(a): the README states that the modifications are not endorsed by the original right-holder. ## Entry ```json { "author": "jesse890423", "title": "ComfyUI-Qwen-TTS-5x", "id": "comfyui-qwen-tts-5x", "reference": "https://github.com/jesse890423/ComfyUI-Qwen-TTS-5x", "nodename_pattern": "^FB_Qwen3TTS", "files": [ "https://github.com/jesse890423/ComfyUI-Qwen-TTS-5x" ], "install_type": "git-clone", "description": "Speech synthesis, voice cloning and voice design nodes for Qwen3-TTS, patched for transformers 5.x ..." } ``` ## Checklist - [x] Entry added to `node_db/new/custom-node-list.json` - [x] `install_type` is `git-clone` - [x] Repository URL is reachable and public - [x] `LICENSE` present (Apache-2.0) - [x] Fork relationship and attribution documented in README and NOTICE --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-28 21:16:04 +00:00