fix(voxcpm1): fix webui download failure and Yue lang mis-triggering issues - #424
Merged
0xShug0 merged 2 commits intoSep 3, 2026
Merged
Conversation
added 2 commits
September 3, 2026 02:56
Package id voxcpm1_0.5b_q8_0 contains '.' which is rejected by app/server/model_installer.cpp:valid_package_id() (only [A-Za-z0-9_-]). The WebUI install path (webui/native/src/lib/api.ts -> POST /v1/ui/models/install) forwards the id verbatim, so the native C++ PackageManager never receives the request and download from audio-cpp/audio.cpp-gguf appears to fail, while tools/model_manager_v2.py (no dot check) succeeds. Rename to voxcpm1_0_5b_q8_0 to match repo naming convention (qwen3_tts_0_6b, soprano_1_1_80m) and unblock validation: - model_specs/voxcpm1.json: recommended_package + packages[0].id - webui/configs/models_catalog.json: download_id for voxcpm1 - docs/community_models/voxcpm1.md: CLI example and package table Remote file name voxcpm-0.5b-q8_0-audiovae-f16.gguf unchanged; GGUF already embeds config/tokenizer sidecars, so single-file huggingface_snapshot install remains valid.
Port audio8_tts OpenCC fix (0eec2be) to voxcpm1 to avoid Cantonese mis-trigger on Traditional input without yue language. Uses shared engine::text::chinese_variant (TSCharacters 3222) and keeps Traditional only for yue/cantonese/zh-HK/zh-MO. - src/community_models/voxcpm1/session.cpp: add extract_request_language() (text_input.language -> voice.style.language -> language/lang option), maybe_convert_traditional_to_simplified_opt for prompt_text/reference_text and TTS text before chunking (offline) and before generate_streaming (streaming) - Mirrors src/community_models/audio8_tts/session.cpp:477,504 Fixes voxcpm1 rendering Traditional "發財" as Cantonese when no language is specified; yue/cantonese/zh-HK/zh-MO preserves Traditional.
Owner
|
@jasonchen31 Good catch! PR merged. |
CryptVenture
added a commit
to CryptVenture/audio.cpp
that referenced
this pull request
Sep 3, 2026
60 of the 85 catalog entries name a download_id that is not a packages[].id, a
path no package installs into, or both: 46 unresolvable ids and 14 paths that
are not the package's target_directory. They resolve only through a
compatibility shim in catalog.ts that strips precision suffixes and guesses;
where the guess fails the entry offers no install button at all.
Each entry now names the package that entry already meant and the directory
that package installs into:
- download_id is an exact packages[].id from the entry's own family. Where
the old id was a stem ("omnivoice", "chatterbox"), it becomes the family's
recommended package. Where it named a variant ("qwen3_tts_1_7b_custom_
voice", "dots_tts_mf_q8_0"), the variant is preserved and only the
spelling is corrected to the published id.
- path is models/<target_directory> of that package. 40 entries pointed at
pre-GGUF directories that no package writes ("models/chatterbox" ->
"models/Chatterbox-GGUF"), and several pointed at a single .gguf file
inside the directory, which only resolves when that exact filename ships.
- Three entries resolved to safetensors packages the native manager cannot
install, so they had no install choice; they now use their family's
recommended GGUF package.
No entry is added, removed or retasked here, and display names, tasks, modes
and min_vram_gb are untouched. The voxcpm1 download_id fixed in 0xShug0#424 is left
as it is.
Validation:
- Entries whose download_id is not a packages[].id, or whose path is not
that package's target_directory: 60 before, 0 after.
- No path is targeted by more than one download_id, so a catalog install
leaves one package per directory.
- tools/check_loader_catalog_sync.py reports loaders, model_specs and
model_manager_v2 in sync.
- Directory paths load exactly as the file paths they replace. Same server,
same input, audiosr registered twice:
models/AudioSR-GGUF 200
models/AudioSR-GGUF/audiosr-basic-f32.gguf 200
identical response fields from both.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ATa5YkLUPMDPRL7w1gCo9p
This was referenced Sep 3, 2026
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
audio-cpp/audio.cpp-ggufandRoot cause
model_specs/voxcpm1.json:251usedvoxcpm1_0.5b_q8_0containing.which is rejected byapp/server/model_installer.cpp:49 valid_package_id()(only[A-Za-z0-9_-], 128 chars).webui/native/src/lib/api.ts:72->POST /v1/ui/models/installforwards the id verbatim, sosrc/framework/package_manager/manager.cpp:490 PackageManager::install()is never reached.tools/model_manager_v2.py:148has no dot check, so CLI succeeded - appears as WebUI-only failure.voxcpm1was the only package inmodel_specs/*.jsonwith..voxcpm1lacked Traditional->Simplified normalization. Tokenizer treats many Traditional codepoints as Cantonese-correlated, so發財withoutlanguage=yuerendered with Yue voice. Same bug previously fixed inaudio8_tts:0eec2be.Changes
3407c74 - rename
voxcpm1_0.5b_q8_0->voxcpm1_0_5b_q8_0(aligns withqwen3_tts_0_6b,soprano_1_1_80m)model_specs/voxcpm1.json:229recommended_packagemodel_specs/voxcpm1.json:251packages[0].idwebui/configs/models_catalog.json:20download_idforvoxcpm1docs/community_models/voxcpm1.md:24,153CLI example + package tableRemote file unchanged:
VoxCPM1-GGUF/voxcpm-0.5b-q8_0-audiovae-f16.gguf(847888032bytes,audiocpp.embedded_files/audiocpp.model_spec.jsonverified viaRangefetch,audiocpp.embedded_files.namescontainsconfig.json/tokenizer.json). GGUF already embeds sidecars, so single-filehuggingface_snapshotremains valid.bc36c6d - port
audio8_tts:0eec2betovoxcpm1(cherry-pickedfix_tokenzier/audio.cpp:2d6101df)src/community_models/voxcpm1/session.cpp:4+ #include "engine/framework/text/chinese_variant.h"src/community_models/voxcpm1/session.cpp:148addextract_request_language()(text_input.language->voice.style.language->language/langoption) - mirrorsaudio8_tts:313src/community_models/voxcpm1/session.cpp:281offline:language = extract_request_language(request),maybe_convert = maybe_convert_traditional_to_simplified_opt(text, language), convertprompt_text/reference_text+converted_request.text_input.textbeforechunk_text_request(so word-budget operates on converted text)src/community_models/voxcpm1/session.cpp:392,465streaming: sameextract+ convertprompt_textandstreaming_textbeforegenerate_streamingyue/cantonese/zh-HK/zh-MO, otherwise converts via sharedengine::text::chinese_variant(TSCharacters3222,src/framework/text/chinese_variant.cpp+chinese_variant_data.inc, built viaCMakeLists.txt:380already inmain)References
src/community_models/audio8_tts/session.cpp:477,504,include/engine/framework/text/chinese_variant.h:1,src/framework/text/chinese_variant.cpp:1