Instructions to use xumo/reader-speech-models with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use xumo/reader-speech-models with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf xumo/reader-speech-models:Q4_K_M # Run inference directly in the terminal: llama cli -hf xumo/reader-speech-models:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf xumo/reader-speech-models:Q4_K_M # Run inference directly in the terminal: llama cli -hf xumo/reader-speech-models:Q4_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf xumo/reader-speech-models:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama cli -hf xumo/reader-speech-models:Q4_K_M
Use Docker
docker model run hf.co/xumo/reader-speech-models:Q4_K_M
- Llama new
- LM Studio
- Jan
- Ollama
How to use xumo/reader-speech-models with Ollama:
ollama run hf.co/xumo/reader-speech-models:Q4_K_M
- Unsloth Desktop
- Docker Model Runner
How to use xumo/reader-speech-models with Docker Model Runner:
docker model run hf.co/xumo/reader-speech-models:Q4_K_M
- Lemonade
How to use xumo/reader-speech-models with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull xumo/reader-speech-models:Q4_K_M
Run and chat with the model
lemonade run user.reader-speech-models-Q4_K_M
List all available models
lemonade list
- Atomic Chat
Publish exact-rebuilt Reader ASR variants and forced aligner with portable provenance
Browse files- .gitattributes +8 -0
- LICENSE.llama.cpp +21 -0
- Qwen3-ASR-0.6B/Qwen3-ASR-0.6B-F16.gguf +3 -0
- Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-F16.gguf +3 -0
- Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q4_K_M.gguf +3 -0
- Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q8_0.gguf +3 -0
- Qwen3-ForcedAligner-0.6B/Qwen3-ForcedAligner-0.6B-Q8_0.gguf +3 -0
- README.md +40 -2
- asr-conversion.lock.json +250 -0
- asr-rebuild.json +74 -0
- conversion/convert-reader-aligner-projector.py +77 -0
- conversion/convert-reader-asr-gguf.py +97 -0
- legacy-asr-conversion.patch +246 -0
- legacy-asr-converter.lock.json +12 -0
- models.lock.json +173 -5
- projectors/mmproj-ASR-0.6B-reader-compat.gguf +3 -0
- projectors/mmproj-ASR-1.7B-reader-compat.gguf +3 -0
- projectors/mmproj-ForcedAligner-reader-compat.gguf +3 -0
.gitattributes
CHANGED
|
@@ -37,3 +37,11 @@ qwen-tts/mmproj-f16.gguf filter=lfs diff=lfs merge=lfs -text
|
|
| 37 |
qwen-tts/talker-f16.gguf filter=lfs diff=lfs merge=lfs -text
|
| 38 |
voices/george.wav filter=lfs diff=lfs merge=lfs -text
|
| 39 |
voices/heart.wav filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
qwen-tts/talker-f16.gguf filter=lfs diff=lfs merge=lfs -text
|
| 38 |
voices/george.wav filter=lfs diff=lfs merge=lfs -text
|
| 39 |
voices/heart.wav filter=lfs diff=lfs merge=lfs -text
|
| 40 |
+
Qwen3-ASR-0.6B/Qwen3-ASR-0.6B-F16.gguf filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-F16.gguf filter=lfs diff=lfs merge=lfs -text
|
| 42 |
+
Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
|
| 43 |
+
Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
|
| 44 |
+
Qwen3-ForcedAligner-0.6B/Qwen3-ForcedAligner-0.6B-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
|
| 45 |
+
projectors/mmproj-ASR-0.6B-reader-compat.gguf filter=lfs diff=lfs merge=lfs -text
|
| 46 |
+
projectors/mmproj-ASR-1.7B-reader-compat.gguf filter=lfs diff=lfs merge=lfs -text
|
| 47 |
+
projectors/mmproj-ForcedAligner-reader-compat.gguf filter=lfs diff=lfs merge=lfs -text
|
LICENSE.llama.cpp
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
MIT License
|
| 2 |
+
|
| 3 |
+
Copyright (c) 2023-2026 The ggml authors
|
| 4 |
+
|
| 5 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 6 |
+
of this software and associated documentation files (the "Software"), to deal
|
| 7 |
+
in the Software without restriction, including without limitation the rights
|
| 8 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 9 |
+
copies of the Software, and to permit persons to whom the Software is
|
| 10 |
+
furnished to do so, subject to the following conditions:
|
| 11 |
+
|
| 12 |
+
The above copyright notice and this permission notice shall be included in all
|
| 13 |
+
copies or substantial portions of the Software.
|
| 14 |
+
|
| 15 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 16 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 17 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 18 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 19 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 20 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 21 |
+
SOFTWARE.
|
Qwen3-ASR-0.6B/Qwen3-ASR-0.6B-F16.gguf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:59cbfd307c21c15a494daef712c9333c19d8e4259f37db4217fd295d508faeb3
|
| 3 |
+
size 1509344064
|
Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-F16.gguf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4257be1bee1f24c2b8e6c2ef7c1c9e0345e17ec805d0ea48e00fae26ef58a52c
|
| 3 |
+
size 4069675840
|
Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q4_K_M.gguf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8225f44c74bee7bb9ee3f1d5abace9054b46354429e257496a2b7fff266f5b7a
|
| 3 |
+
size 1282435904
|
Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q8_0.gguf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dd7dd905b8b29d1fe59ba769ef46f8915f8599592859b9463a7251b466b51454
|
| 3 |
+
size 2165035840
|
Qwen3-ForcedAligner-0.6B/Qwen3-ForcedAligner-0.6B-Q8_0.gguf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0a9b58e7a61eca2acf902794dbe8973860d42cccca49235530b2f2c5f290ed76
|
| 3 |
+
size 639585920
|
README.md
CHANGED
|
@@ -1,6 +1,9 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
base_model:
|
|
|
|
|
|
|
|
|
|
| 4 |
- Qwen/Qwen3-TTS-12Hz-0.6B-Base
|
| 5 |
library_name: gguf
|
| 6 |
tags:
|
|
@@ -14,6 +17,14 @@ accepted by Reader; `models.lock.json` pins size and SHA-256. Bind downloads to
|
|
| 14 |
full publication commit, never a moving main branch.
|
| 15 |
|
| 16 |
## Included artifacts
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
- `qwen3-tts-0.6b-base-f16`
|
| 18 |
- `synthetic-heart`
|
| 19 |
- `synthetic-george`
|
|
@@ -22,12 +33,12 @@ full publication commit, never a moving main branch.
|
|
| 22 |
|
| 23 |
Qwen Base revision: `5d83992436eae1d760afd27aff78a71d676296fc`.
|
| 24 |
Converter: ggml-org/llama.cpp `67107ac34f62882d8b53ea41ad68010ea635b9be`.
|
| 25 |
-
Reader packaging source commit: `
|
| 26 |
The complete recipe and reconstruction evidence are included in this model repo.
|
| 27 |
Pinned public converter:
|
| 28 |
https://github.com/ggml-org/llama.cpp/blob/67107ac34f62882d8b53ea41ad68010ea635b9be/convert_hf_to_gguf.py
|
| 29 |
|
| 30 |
-
The
|
| 31 |
`reader-model.json` preserves the original source-file receipt; raw safetensors
|
| 32 |
listed in that receipt are source inputs and are not required for inference.
|
| 33 |
Rebuilding is a separate operation; consult `conversion.lock.json` and `rebuild.json`.
|
|
@@ -60,3 +71,30 @@ Kokoro reference source: https://huggingface.co/onnx-community/Kokoro-82M-v1.0-O
|
|
| 60 |
Upstream model cards mark Apache-2.0. See LICENSE. Reader changes: conversion to
|
| 61 |
F16 GGUF and explicit packaging/identity metadata. The publication commit is not
|
| 62 |
a model retraining or a new quality claim.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
base_model:
|
| 4 |
+
- Qwen/Qwen3-ASR-0.6B
|
| 5 |
+
- Qwen/Qwen3-ASR-1.7B
|
| 6 |
+
- Qwen/Qwen3-ForcedAligner-0.6B
|
| 7 |
- Qwen/Qwen3-TTS-12Hz-0.6B-Base
|
| 8 |
library_name: gguf
|
| 9 |
tags:
|
|
|
|
| 17 |
full publication commit, never a moving main branch.
|
| 18 |
|
| 19 |
## Included artifacts
|
| 20 |
+
- `asr-0.6b-f16`
|
| 21 |
+
- `asr-1.7b-f16`
|
| 22 |
+
- `asr-1.7b-q4-k-m`
|
| 23 |
+
- `asr-1.7b-q8-0`
|
| 24 |
+
- `aligner-q8-0`
|
| 25 |
+
- `mmproj-asr-0.6b-reader-compat`
|
| 26 |
+
- `mmproj-asr-1.7b-reader-compat`
|
| 27 |
+
- `mmproj-forcedaligner-reader-compat`
|
| 28 |
- `qwen3-tts-0.6b-base-f16`
|
| 29 |
- `synthetic-heart`
|
| 30 |
- `synthetic-george`
|
|
|
|
| 33 |
|
| 34 |
Qwen Base revision: `5d83992436eae1d760afd27aff78a71d676296fc`.
|
| 35 |
Converter: ggml-org/llama.cpp `67107ac34f62882d8b53ea41ad68010ea635b9be`.
|
| 36 |
+
Reader packaging source commit: `14d2029ba099c97ea4b2544bef0c58de4f11ed79` (may precede its source-repository push).
|
| 37 |
The complete recipe and reconstruction evidence are included in this model repo.
|
| 38 |
Pinned public converter:
|
| 39 |
https://github.com/ggml-org/llama.cpp/blob/67107ac34f62882d8b53ea41ad68010ea635b9be/convert_hf_to_gguf.py
|
| 40 |
|
| 41 |
+
The Qwen TTS GGUF pair includes the talker and projector/codec. The generated
|
| 42 |
`reader-model.json` preserves the original source-file receipt; raw safetensors
|
| 43 |
listed in that receipt are source inputs and are not required for inference.
|
| 44 |
Rebuilding is a separate operation; consult `conversion.lock.json` and `rebuild.json`.
|
|
|
|
| 71 |
Upstream model cards mark Apache-2.0. See LICENSE. Reader changes: conversion to
|
| 72 |
F16 GGUF and explicit packaging/identity metadata. The publication commit is not
|
| 73 |
a model retraining or a new quality claim.
|
| 74 |
+
|
| 75 |
+
## ASR/FA conversion provenance
|
| 76 |
+
|
| 77 |
+
The accepted text GGUF and Reader-compatible projectors were reconstructed exactly
|
| 78 |
+
from the official revisions and raw-file hashes in `asr-conversion.lock.json`.
|
| 79 |
+
`asr-rebuild.json` compares all eight outputs to their accepted SHA values.
|
| 80 |
+
|
| 81 |
+
To reconstruct: obtain public ggml-org/llama.cpp commit
|
| 82 |
+
`66d65ec29ba7c440cbc31b6f63b74a17b536ba65`, apply the included
|
| 83 |
+
`legacy-asr-conversion.patch`, and check the four source hashes in
|
| 84 |
+
`legacy-asr-converter.lock.json`. This freezes conversion-only sources; it does
|
| 85 |
+
not restore an old inference backend. Use the locked Python environment and the
|
| 86 |
+
patched source's gguf-py. Convert each official raw model to F16 text GGUF and an
|
| 87 |
+
F16 mmproj. The original model folder basename affects metadata and must match.
|
| 88 |
+
|
| 89 |
+
Build the CPU llama-quantize tool from current upstream commit
|
| 90 |
+
`67107ac34f62882d8b53ea41ad68010ea635b9be`. Quantize ASR 1.7B F16 to Q8_0/Q4_K_M
|
| 91 |
+
and FA F16 to Q8_0 with four threads. Apply the included standalone Reader
|
| 92 |
+
projector converter (`conversion/`) with `--encoder-compat`, using gguf-py from
|
| 93 |
+
that current pinned commit. No tensors are retrained; compare final SHA values
|
| 94 |
+
against the model lock before use.
|
| 95 |
+
|
| 96 |
+
The Reader-compatible mmproj files require Reader's encoder/aligner compatibility
|
| 97 |
+
adapter. Ordinary upstream projectors are not interchangeable with these accepted
|
| 98 |
+
files. See the per-artifact origins for source revisions and compatibility tags.
|
| 99 |
+
Converter code is covered separately by LICENSE.llama.cpp; model provenance and
|
| 100 |
+
Apache-2.0 attribution remain in LICENSE and the official Qwen model cards.
|
asr-conversion.lock.json
ADDED
|
@@ -0,0 +1,250 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schemaVersion": 1,
|
| 3 |
+
"python": "3.12.3",
|
| 4 |
+
"packages": {
|
| 5 |
+
"huggingface_hub": "0.36.2",
|
| 6 |
+
"Jinja2": "3.1.6",
|
| 7 |
+
"numpy": "2.4.2",
|
| 8 |
+
"packaging": "26.0",
|
| 9 |
+
"PyYAML": "6.0.3",
|
| 10 |
+
"regex": "2026.2.28",
|
| 11 |
+
"requests": "2.32.5",
|
| 12 |
+
"safetensors": "0.7.0",
|
| 13 |
+
"sentencepiece": "0.2.1",
|
| 14 |
+
"tokenizers": "0.22.2",
|
| 15 |
+
"torch": "2.10.0",
|
| 16 |
+
"tqdm": "4.67.3",
|
| 17 |
+
"transformers": "4.57.6"
|
| 18 |
+
},
|
| 19 |
+
"converterLockSha256": "5340b50a0f3aaa1687296e7b048706ff8cc300004c4252ea0df9933aa1802995",
|
| 20 |
+
"quantizerRevision": "67107ac34f62882d8b53ea41ad68010ea635b9be",
|
| 21 |
+
"models": [
|
| 22 |
+
{
|
| 23 |
+
"repository": "Qwen/Qwen3-ASR-0.6B",
|
| 24 |
+
"revision": "5eb144179a02acc5e5ba31e748d22b0cf3e303b0",
|
| 25 |
+
"sourceFiles": [
|
| 26 |
+
{
|
| 27 |
+
"path": ".gitattributes",
|
| 28 |
+
"size": 1519,
|
| 29 |
+
"sha256": "11ad7efa24975ee4b0c3c3a38ed18737f0658a5f75a0a96787b576a78a023361"
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"path": "README.md",
|
| 33 |
+
"size": 57456,
|
| 34 |
+
"sha256": "5058416891bc47a2051557765997e8c42f8eb78a0e33c3e775bd17d4b0ba4d50"
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"path": "chat_template.json",
|
| 38 |
+
"size": 1161,
|
| 39 |
+
"sha256": "75a8cfca24f00de72d796fbfed6858fc9614ef3dabd8696684cc3bc03a9c58ff"
|
| 40 |
+
},
|
| 41 |
+
{
|
| 42 |
+
"path": "config.json",
|
| 43 |
+
"size": 6193,
|
| 44 |
+
"sha256": "76d3ae4601ce939830b2517f4a6cadb86cc51316c3900af6b020b051c21a478c"
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"path": "generation_config.json",
|
| 48 |
+
"size": 142,
|
| 49 |
+
"sha256": "1da527824d81e07118facff437e03f2e24a23311e3bdeb2368973fe77e5f275c"
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"path": "merges.txt",
|
| 53 |
+
"size": 1671853,
|
| 54 |
+
"sha256": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5"
|
| 55 |
+
},
|
| 56 |
+
{
|
| 57 |
+
"path": "model.safetensors",
|
| 58 |
+
"size": 1876091704,
|
| 59 |
+
"sha256": "79d6cbd4c98c7bbffe9db2edac07f56cd6637d0d5944b27f6c2b8353840323ea"
|
| 60 |
+
},
|
| 61 |
+
{
|
| 62 |
+
"path": "preprocessor_config.json",
|
| 63 |
+
"size": 330,
|
| 64 |
+
"sha256": "45e120a4eda2c20c5d7f2ea9354e63536bf35e27aa573fb7cdf78017b378770d"
|
| 65 |
+
},
|
| 66 |
+
{
|
| 67 |
+
"path": "tokenizer_config.json",
|
| 68 |
+
"size": 12487,
|
| 69 |
+
"sha256": "4942d005604266809309cabc9f4e9cb89ce855d59b14681fdc0e1cc62ea26c4c"
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"path": "vocab.json",
|
| 73 |
+
"size": 2776833,
|
| 74 |
+
"sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
|
| 75 |
+
}
|
| 76 |
+
],
|
| 77 |
+
"projector": {
|
| 78 |
+
"sha256": "e25dc3f8eca8c21adaceab8e35cf158274728f3e8fe266a60869fbfd54956904",
|
| 79 |
+
"acceptedInputSha256": "e25dc3f8eca8c21adaceab8e35cf158274728f3e8fe266a60869fbfd54956904",
|
| 80 |
+
"exact": true
|
| 81 |
+
}
|
| 82 |
+
},
|
| 83 |
+
{
|
| 84 |
+
"repository": "Qwen/Qwen3-ASR-1.7B",
|
| 85 |
+
"revision": "7278e1e70fe206f11671096ffdd38061171dd6e5",
|
| 86 |
+
"sourceFiles": [
|
| 87 |
+
{
|
| 88 |
+
"path": ".gitattributes",
|
| 89 |
+
"size": 1519,
|
| 90 |
+
"sha256": "11ad7efa24975ee4b0c3c3a38ed18737f0658a5f75a0a96787b576a78a023361"
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"path": "README.md",
|
| 94 |
+
"size": 57456,
|
| 95 |
+
"sha256": "5058416891bc47a2051557765997e8c42f8eb78a0e33c3e775bd17d4b0ba4d50"
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"path": "chat_template.json",
|
| 99 |
+
"size": 1161,
|
| 100 |
+
"sha256": "75a8cfca24f00de72d796fbfed6858fc9614ef3dabd8696684cc3bc03a9c58ff"
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"path": "config.json",
|
| 104 |
+
"size": 6194,
|
| 105 |
+
"sha256": "2e74a751548b8ad7d7526d29365ad8144c345d8b412b1152d25dc6698452712f"
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"path": "generation_config.json",
|
| 109 |
+
"size": 142,
|
| 110 |
+
"sha256": "1da527824d81e07118facff437e03f2e24a23311e3bdeb2368973fe77e5f275c"
|
| 111 |
+
},
|
| 112 |
+
{
|
| 113 |
+
"path": "merges.txt",
|
| 114 |
+
"size": 1671853,
|
| 115 |
+
"sha256": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5"
|
| 116 |
+
},
|
| 117 |
+
{
|
| 118 |
+
"path": "model-00001-of-00002.safetensors",
|
| 119 |
+
"size": 4220320824,
|
| 120 |
+
"sha256": "a4cd1f1a04d90b757dc7f7dd26254e69a013b19e80efe590a83c6a3bde8608d6"
|
| 121 |
+
},
|
| 122 |
+
{
|
| 123 |
+
"path": "model-00002-of-00002.safetensors",
|
| 124 |
+
"size": 478200688,
|
| 125 |
+
"sha256": "6e0b9d9e09e2e0238e7ef3cc8a484ab387e91b90f1900bedf88bc92d7929ccfc"
|
| 126 |
+
},
|
| 127 |
+
{
|
| 128 |
+
"path": "model.safetensors.index.json",
|
| 129 |
+
"size": 64821,
|
| 130 |
+
"sha256": "f994739fe38e5210b9e3e8ce6c6307315e2ceac3cb630e7b7414d69dce520f60"
|
| 131 |
+
},
|
| 132 |
+
{
|
| 133 |
+
"path": "preprocessor_config.json",
|
| 134 |
+
"size": 330,
|
| 135 |
+
"sha256": "45e120a4eda2c20c5d7f2ea9354e63536bf35e27aa573fb7cdf78017b378770d"
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"path": "tokenizer_config.json",
|
| 139 |
+
"size": 12487,
|
| 140 |
+
"sha256": "4942d005604266809309cabc9f4e9cb89ce855d59b14681fdc0e1cc62ea26c4c"
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"path": "vocab.json",
|
| 144 |
+
"size": 2776833,
|
| 145 |
+
"sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
|
| 146 |
+
}
|
| 147 |
+
],
|
| 148 |
+
"projector": {
|
| 149 |
+
"sha256": "6955893dafaecf60cb7cfc871b5c8647d75ebbe83e8d88f1e69940c763e4254b",
|
| 150 |
+
"acceptedInputSha256": "6955893dafaecf60cb7cfc871b5c8647d75ebbe83e8d88f1e69940c763e4254b",
|
| 151 |
+
"exact": true
|
| 152 |
+
}
|
| 153 |
+
},
|
| 154 |
+
{
|
| 155 |
+
"repository": "Qwen/Qwen3-ForcedAligner-0.6B",
|
| 156 |
+
"revision": "c7cbfc2048c462b0d63a45797104fc9db3ad62b7",
|
| 157 |
+
"sourceFiles": [
|
| 158 |
+
{
|
| 159 |
+
"path": ".gitattributes",
|
| 160 |
+
"size": 1519,
|
| 161 |
+
"sha256": "11ad7efa24975ee4b0c3c3a38ed18737f0658a5f75a0a96787b576a78a023361"
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"path": "README.md",
|
| 165 |
+
"size": 57456,
|
| 166 |
+
"sha256": "5058416891bc47a2051557765997e8c42f8eb78a0e33c3e775bd17d4b0ba4d50"
|
| 167 |
+
},
|
| 168 |
+
{
|
| 169 |
+
"path": "chat_template.json",
|
| 170 |
+
"size": 1161,
|
| 171 |
+
"sha256": "75a8cfca24f00de72d796fbfed6858fc9614ef3dabd8696684cc3bc03a9c58ff"
|
| 172 |
+
},
|
| 173 |
+
{
|
| 174 |
+
"path": "config.json",
|
| 175 |
+
"size": 5982,
|
| 176 |
+
"sha256": "d616c65d46c4b90bdc651b0a0963ea932732241140f337f9bb6b0335a9c8ef09"
|
| 177 |
+
},
|
| 178 |
+
{
|
| 179 |
+
"path": "generation_config.json",
|
| 180 |
+
"size": 115,
|
| 181 |
+
"sha256": "948d089b23bca1d214e768d59c4438365665f52ec6d33678f4062206b3fbbb8c"
|
| 182 |
+
},
|
| 183 |
+
{
|
| 184 |
+
"path": "merges.txt",
|
| 185 |
+
"size": 1671853,
|
| 186 |
+
"sha256": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5"
|
| 187 |
+
},
|
| 188 |
+
{
|
| 189 |
+
"path": "model.safetensors",
|
| 190 |
+
"size": 1835544544,
|
| 191 |
+
"sha256": "47831d0e82f96b20e9034dba01a075ee06436654719f6a68289e49f1b65ce0e7"
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"path": "preprocessor_config.json",
|
| 195 |
+
"size": 330,
|
| 196 |
+
"sha256": "45e120a4eda2c20c5d7f2ea9354e63536bf35e27aa573fb7cdf78017b378770d"
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"path": "tokenizer_config.json",
|
| 200 |
+
"size": 12666,
|
| 201 |
+
"sha256": "3ab80063f8511deb9566e6ad438d17b7a6277fcffd52d92854112f19d36bd81c"
|
| 202 |
+
},
|
| 203 |
+
{
|
| 204 |
+
"path": "vocab.json",
|
| 205 |
+
"size": 2776833,
|
| 206 |
+
"sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
|
| 207 |
+
}
|
| 208 |
+
],
|
| 209 |
+
"projector": {
|
| 210 |
+
"sha256": "3c4bd01b0293cc9d394cd649e7e9b1210e52684bee5409e237075475975223ee",
|
| 211 |
+
"acceptedInputSha256": "3c4bd01b0293cc9d394cd649e7e9b1210e52684bee5409e237075475975223ee",
|
| 212 |
+
"exact": true
|
| 213 |
+
}
|
| 214 |
+
}
|
| 215 |
+
],
|
| 216 |
+
"outputs": {
|
| 217 |
+
"asr-0.6b-f16": {
|
| 218 |
+
"path": "Qwen3-ASR-0.6B/Qwen3-ASR-0.6B-F16.gguf",
|
| 219 |
+
"sha256": "59cbfd307c21c15a494daef712c9333c19d8e4259f37db4217fd295d508faeb3"
|
| 220 |
+
},
|
| 221 |
+
"asr-1.7b-f16": {
|
| 222 |
+
"path": "Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-F16.gguf",
|
| 223 |
+
"sha256": "4257be1bee1f24c2b8e6c2ef7c1c9e0345e17ec805d0ea48e00fae26ef58a52c"
|
| 224 |
+
},
|
| 225 |
+
"asr-1.7b-q4-k-m": {
|
| 226 |
+
"path": "Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q4_K_M.gguf",
|
| 227 |
+
"sha256": "8225f44c74bee7bb9ee3f1d5abace9054b46354429e257496a2b7fff266f5b7a"
|
| 228 |
+
},
|
| 229 |
+
"asr-1.7b-q8-0": {
|
| 230 |
+
"path": "Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q8_0.gguf",
|
| 231 |
+
"sha256": "dd7dd905b8b29d1fe59ba769ef46f8915f8599592859b9463a7251b466b51454"
|
| 232 |
+
},
|
| 233 |
+
"aligner-q8-0": {
|
| 234 |
+
"path": "Qwen3-ForcedAligner-0.6B/Qwen3-ForcedAligner-0.6B-Q8_0.gguf",
|
| 235 |
+
"sha256": "0a9b58e7a61eca2acf902794dbe8973860d42cccca49235530b2f2c5f290ed76"
|
| 236 |
+
},
|
| 237 |
+
"mmproj-asr-0.6b-reader-compat": {
|
| 238 |
+
"path": "projectors/mmproj-ASR-0.6B-reader-compat.gguf",
|
| 239 |
+
"sha256": "fffb83e0ab2731a8c8ca1741726d417a01b274a6cebc9bb53b84e0f899a6a533"
|
| 240 |
+
},
|
| 241 |
+
"mmproj-asr-1.7b-reader-compat": {
|
| 242 |
+
"path": "projectors/mmproj-ASR-1.7B-reader-compat.gguf",
|
| 243 |
+
"sha256": "b0a1c67a7bd6e0231dfddc013f0e7e1e3c9a702b4885f39063981aec503ba7fd"
|
| 244 |
+
},
|
| 245 |
+
"mmproj-forcedaligner-reader-compat": {
|
| 246 |
+
"path": "projectors/mmproj-ForcedAligner-reader-compat.gguf",
|
| 247 |
+
"sha256": "b4f0af1d7daeb9910a9970b90057aa1f038a2319da0452a6da574f4b34106f99"
|
| 248 |
+
}
|
| 249 |
+
}
|
| 250 |
+
}
|
asr-rebuild.json
ADDED
|
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schemaVersion": 1,
|
| 3 |
+
"recipeSha256": "91a534aad00af8a42d29c02526eb5ccb4257255d4bc7dbd9350cbcf1a5726acd",
|
| 4 |
+
"converterLockSha256": "5340b50a0f3aaa1687296e7b048706ff8cc300004c4252ea0df9933aa1802995",
|
| 5 |
+
"quantizerSha256": "959b57f358eca088304e643c04b7fde4c5b559ffb1ad4a4e51e0090c511912bd",
|
| 6 |
+
"outputs": {
|
| 7 |
+
"asr-0.6b-f16": {
|
| 8 |
+
"path": "Qwen3-ASR-0.6B/Qwen3-ASR-0.6B-F16.gguf",
|
| 9 |
+
"sha256": "59cbfd307c21c15a494daef712c9333c19d8e4259f37db4217fd295d508faeb3",
|
| 10 |
+
"acceptedSha256": "59cbfd307c21c15a494daef712c9333c19d8e4259f37db4217fd295d508faeb3",
|
| 11 |
+
"exact": true
|
| 12 |
+
},
|
| 13 |
+
"mmproj-asr-0.6b-reader-compat": {
|
| 14 |
+
"path": "projectors/mmproj-ASR-0.6B-reader-compat.gguf",
|
| 15 |
+
"sha256": "fffb83e0ab2731a8c8ca1741726d417a01b274a6cebc9bb53b84e0f899a6a533",
|
| 16 |
+
"acceptedSha256": "fffb83e0ab2731a8c8ca1741726d417a01b274a6cebc9bb53b84e0f899a6a533",
|
| 17 |
+
"exact": true
|
| 18 |
+
},
|
| 19 |
+
"asr-1.7b-f16": {
|
| 20 |
+
"path": "Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-F16.gguf",
|
| 21 |
+
"sha256": "4257be1bee1f24c2b8e6c2ef7c1c9e0345e17ec805d0ea48e00fae26ef58a52c",
|
| 22 |
+
"acceptedSha256": "4257be1bee1f24c2b8e6c2ef7c1c9e0345e17ec805d0ea48e00fae26ef58a52c",
|
| 23 |
+
"exact": true
|
| 24 |
+
},
|
| 25 |
+
"asr-1.7b-q8-0": {
|
| 26 |
+
"path": "Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q8_0.gguf",
|
| 27 |
+
"sha256": "dd7dd905b8b29d1fe59ba769ef46f8915f8599592859b9463a7251b466b51454",
|
| 28 |
+
"acceptedSha256": "dd7dd905b8b29d1fe59ba769ef46f8915f8599592859b9463a7251b466b51454",
|
| 29 |
+
"exact": true
|
| 30 |
+
},
|
| 31 |
+
"asr-1.7b-q4-k-m": {
|
| 32 |
+
"path": "Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q4_K_M.gguf",
|
| 33 |
+
"sha256": "8225f44c74bee7bb9ee3f1d5abace9054b46354429e257496a2b7fff266f5b7a",
|
| 34 |
+
"acceptedSha256": "8225f44c74bee7bb9ee3f1d5abace9054b46354429e257496a2b7fff266f5b7a",
|
| 35 |
+
"exact": true
|
| 36 |
+
},
|
| 37 |
+
"mmproj-asr-1.7b-reader-compat": {
|
| 38 |
+
"path": "projectors/mmproj-ASR-1.7B-reader-compat.gguf",
|
| 39 |
+
"sha256": "b0a1c67a7bd6e0231dfddc013f0e7e1e3c9a702b4885f39063981aec503ba7fd",
|
| 40 |
+
"acceptedSha256": "b0a1c67a7bd6e0231dfddc013f0e7e1e3c9a702b4885f39063981aec503ba7fd",
|
| 41 |
+
"exact": true
|
| 42 |
+
},
|
| 43 |
+
"aligner-q8-0": {
|
| 44 |
+
"path": "Qwen3-ForcedAligner-0.6B/Qwen3-ForcedAligner-0.6B-Q8_0.gguf",
|
| 45 |
+
"sha256": "0a9b58e7a61eca2acf902794dbe8973860d42cccca49235530b2f2c5f290ed76",
|
| 46 |
+
"acceptedSha256": "0a9b58e7a61eca2acf902794dbe8973860d42cccca49235530b2f2c5f290ed76",
|
| 47 |
+
"exact": true
|
| 48 |
+
},
|
| 49 |
+
"mmproj-forcedaligner-reader-compat": {
|
| 50 |
+
"path": "projectors/mmproj-ForcedAligner-reader-compat.gguf",
|
| 51 |
+
"sha256": "b4f0af1d7daeb9910a9970b90057aa1f038a2319da0452a6da574f4b34106f99",
|
| 52 |
+
"acceptedSha256": "b4f0af1d7daeb9910a9970b90057aa1f038a2319da0452a6da574f4b34106f99",
|
| 53 |
+
"exact": true
|
| 54 |
+
}
|
| 55 |
+
},
|
| 56 |
+
"rawInputsVerified": [
|
| 57 |
+
{
|
| 58 |
+
"repository": "Qwen/Qwen3-ASR-0.6B",
|
| 59 |
+
"revision": "5eb144179a02acc5e5ba31e748d22b0cf3e303b0",
|
| 60 |
+
"files": 10
|
| 61 |
+
},
|
| 62 |
+
{
|
| 63 |
+
"repository": "Qwen/Qwen3-ASR-1.7B",
|
| 64 |
+
"revision": "7278e1e70fe206f11671096ffdd38061171dd6e5",
|
| 65 |
+
"files": 12
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"repository": "Qwen/Qwen3-ForcedAligner-0.6B",
|
| 69 |
+
"revision": "c7cbfc2048c462b0d63a45797104fc9db3ad62b7",
|
| 70 |
+
"files": 10
|
| 71 |
+
}
|
| 72 |
+
],
|
| 73 |
+
"allOutputsExact": true
|
| 74 |
+
}
|
conversion/convert-reader-aligner-projector.py
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Repackage the legacy ForcedAligner projector for pinned upstream mtmd.
|
| 3 |
+
|
| 4 |
+
Requires numpy and the pinned upstream gguf-py on PYTHONPATH. Renames tensors
|
| 5 |
+
and compatible shape metadata; never requantizes or changes tensor data bytes.
|
| 6 |
+
"""
|
| 7 |
+
import argparse
|
| 8 |
+
import hashlib
|
| 9 |
+
import json
|
| 10 |
+
from pathlib import Path
|
| 11 |
+
|
| 12 |
+
import gguf
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
def sha256(path):
|
| 16 |
+
with path.open("rb") as source:
|
| 17 |
+
return hashlib.file_digest(source, "sha256").hexdigest()
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def convert(source, output, encoder_compat=False):
|
| 21 |
+
if output.exists() or output.with_suffix(".conversion.json").exists():
|
| 22 |
+
raise FileExistsError("choose a fresh output path")
|
| 23 |
+
reader = gguf.GGUFReader(source)
|
| 24 |
+
if "clip.audio.classify_num" not in reader.fields or not any(t.name == "a.classifier.weight" for t in reader.tensors):
|
| 25 |
+
raise ValueError("not a legacy ForcedAligner projector")
|
| 26 |
+
writer = gguf.GGUFWriter(output, "clip")
|
| 27 |
+
if encoder_compat:
|
| 28 |
+
writer.add_bool("clip.audio.reader_aligner_compat", True)
|
| 29 |
+
for key, field in reader.fields.items():
|
| 30 |
+
if key.startswith("GGUF.") or key == "general.architecture":
|
| 31 |
+
continue
|
| 32 |
+
new_key = "clip.audio.projector_type" if key == "clip.projector_type" else key
|
| 33 |
+
writer.add_key_value(new_key, field.contents(), field.types[0],
|
| 34 |
+
field.types[1] if len(field.types) > 1 else None)
|
| 35 |
+
tensor_hashes, tensor_types, tensor_shapes = {}, {}, {}
|
| 36 |
+
for tensor in reader.tensors:
|
| 37 |
+
name = tensor.name.replace("a.conv1d.", "a.conv2d.").replace("a.pre_encode.out.", "a.conv_out.")
|
| 38 |
+
data = tensor.data
|
| 39 |
+
if name.startswith("a.conv2d.") and name.endswith(".bias"):
|
| 40 |
+
data = data.reshape(-1, 1, 1)
|
| 41 |
+
writer.add_tensor(name, data, raw_dtype=tensor.tensor_type)
|
| 42 |
+
tensor_hashes[name] = hashlib.sha256(tensor.data.tobytes()).hexdigest()
|
| 43 |
+
tensor_types[name] = tensor.tensor_type
|
| 44 |
+
# GGUF dimensions are stored in the reverse order of NumPy dimensions.
|
| 45 |
+
tensor_shapes[name] = list(data.shape[::-1]) if name.startswith("a.conv2d.") and name.endswith(".bias") else tensor.shape.tolist()
|
| 46 |
+
writer.write_header_to_file()
|
| 47 |
+
writer.write_kv_data_to_file()
|
| 48 |
+
writer.write_tensors_to_file()
|
| 49 |
+
writer.close()
|
| 50 |
+
actual = gguf.GGUFReader(output)
|
| 51 |
+
if len(actual.tensors) != len(tensor_hashes) or any(
|
| 52 |
+
t.name not in tensor_hashes or
|
| 53 |
+
hashlib.sha256(t.data.tobytes()).hexdigest() != tensor_hashes[t.name] or
|
| 54 |
+
t.tensor_type != tensor_types[t.name] or t.shape.tolist() != tensor_shapes[t.name]
|
| 55 |
+
for t in actual.tensors):
|
| 56 |
+
raise RuntimeError("tensor byte/type/shape identity check failed")
|
| 57 |
+
output.with_suffix(".conversion.json").write_text(json.dumps({
|
| 58 |
+
"inputSha256": sha256(source), "outputSha256": sha256(output),
|
| 59 |
+
"upstreamCommit": "67107ac34f62882d8b53ea41ad68010ea635b9be",
|
| 60 |
+
"unchangedTensorBytes": len(tensor_hashes), "tensorSha256": tensor_hashes,
|
| 61 |
+
"readerAlignerEncoderCompatibility": encoder_compat,
|
| 62 |
+
}, indent=2) + "\n")
|
| 63 |
+
print(f"PASS: {len(tensor_hashes)} tensors retain identical data bytes")
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
def main():
|
| 67 |
+
parser = argparse.ArgumentParser(description=__doc__)
|
| 68 |
+
parser.add_argument("--input", type=Path, required=True)
|
| 69 |
+
parser.add_argument("--output", type=Path, required=True)
|
| 70 |
+
parser.add_argument("--encoder-compat", action="store_true",
|
| 71 |
+
help="Opt into Reader legacy alignment arithmetic (encoder and CUDA decoder)")
|
| 72 |
+
args = parser.parse_args()
|
| 73 |
+
convert(args.input, args.output, args.encoder_compat)
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
if __name__ == "__main__":
|
| 77 |
+
main()
|
conversion/convert-reader-asr-gguf.py
ADDED
|
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Repackage existing Qwen3-ASR GGUF variants for the pinned upstream runtime.
|
| 3 |
+
|
| 4 |
+
No requantization. Every tensor payload, quantization type and resulting shape
|
| 5 |
+
is checked after writing. Requires numpy and pinned upstream gguf-py.
|
| 6 |
+
"""
|
| 7 |
+
import argparse
|
| 8 |
+
import hashlib
|
| 9 |
+
import json
|
| 10 |
+
from pathlib import Path
|
| 11 |
+
import gguf
|
| 12 |
+
|
| 13 |
+
UPSTREAM = '67107ac34f62882d8b53ea41ad68010ea635b9be'
|
| 14 |
+
CHAT = r"{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
def sha(path):
|
| 18 |
+
with Path(path).open('rb') as source:
|
| 19 |
+
return hashlib.file_digest(source, 'sha256').hexdigest()
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
def convert(source, output, encoder_compat=False):
|
| 23 |
+
if output.exists() or output.with_suffix('.conversion.json').exists():
|
| 24 |
+
raise FileExistsError('choose a fresh output and conversion-record path')
|
| 25 |
+
reader = gguf.GGUFReader(source)
|
| 26 |
+
architecture = reader.fields['general.architecture'].contents()
|
| 27 |
+
projector = architecture == 'clip'
|
| 28 |
+
if encoder_compat and not projector:
|
| 29 |
+
raise ValueError('encoder compatibility applies to the projector; reuse the original text GGUF')
|
| 30 |
+
if projector:
|
| 31 |
+
if any(t.name == 'a.classifier.weight' for t in reader.tensors):
|
| 32 |
+
raise ValueError('ForcedAligner requires its own converter and compatibility policy')
|
| 33 |
+
if reader.fields.get('clip.projector_type') is None or reader.fields['clip.projector_type'].contents() != 'qwen3a':
|
| 34 |
+
raise ValueError('expected legacy Qwen3-ASR audio projector')
|
| 35 |
+
layers = reader.fields['clip.audio.block_count'].contents()
|
| 36 |
+
config = (layers, reader.fields['clip.audio.embedding_length'].contents(),
|
| 37 |
+
reader.fields['clip.audio.projection_dim'].contents())
|
| 38 |
+
if config not in [(18, 896, 1024), (24, 1024, 2048)] or len(reader.tensors) != 14 + 16 * layers:
|
| 39 |
+
raise ValueError('unsupported ASR projector tensor inventory')
|
| 40 |
+
elif architecture != 'qwen3' or len(reader.tensors) != 311 or reader.fields['qwen3.rope.dimension_sections'].contents() != [24, 20, 20, 0]:
|
| 41 |
+
raise ValueError('expected legacy 311-tensor Qwen3-ASR text model')
|
| 42 |
+
if not projector:
|
| 43 |
+
tokens = reader.fields['tokenizer.ggml.tokens'].contents()
|
| 44 |
+
if len(tokens) <= 151645 or tokens[151645] != '<|im_end|>' or '<asr_text>' not in tokens:
|
| 45 |
+
raise ValueError('unexpected Qwen3-ASR vocabulary')
|
| 46 |
+
output.parent.mkdir(parents=True, exist_ok=True)
|
| 47 |
+
writer = gguf.GGUFWriter(output, 'clip' if projector else 'qwen3vl')
|
| 48 |
+
if encoder_compat:
|
| 49 |
+
if reader.fields['clip.audio.n_window'].contents() != 50 or reader.fields['clip.audio.n_window_infer'].contents() != 800:
|
| 50 |
+
raise ValueError('unsupported legacy ASR encoder windows')
|
| 51 |
+
writer.add_bool('clip.audio.reader_asr_compat', True)
|
| 52 |
+
writer.add_uint32('clip.audio.n_window', 50)
|
| 53 |
+
writer.add_uint32('clip.audio.n_window_infer', 800)
|
| 54 |
+
skip = {'GGUF.version', 'GGUF.tensor_count', 'GGUF.kv_count', 'general.architecture'}
|
| 55 |
+
if projector:
|
| 56 |
+
skip.update(['clip.audio.n_window', 'clip.audio.n_window_infer'])
|
| 57 |
+
else:
|
| 58 |
+
skip.update(['tokenizer.chat_template', 'tokenizer.ggml.bos_token_id', 'tokenizer.ggml.clean_spaces'])
|
| 59 |
+
writer.add_string('tokenizer.chat_template', CHAT)
|
| 60 |
+
writer.add_uint32('tokenizer.ggml.bos_token_id', 151645)
|
| 61 |
+
writer.add_bool('tokenizer.ggml.clean_spaces', False)
|
| 62 |
+
writer.add_uint32('qwen3vl.n_deepstack_layers', 0)
|
| 63 |
+
for key, field in reader.fields.items():
|
| 64 |
+
if key in skip or key.startswith('GGUF.'):
|
| 65 |
+
continue
|
| 66 |
+
new = ('clip.audio.projector_type' if key == 'clip.projector_type' else key) if projector else key.replace('qwen3.', 'qwen3vl.', 1) if key.startswith('qwen3.') else key
|
| 67 |
+
writer.add_key_value(new, field.contents(), field.types[0], field.types[1] if len(field.types)>1 else None)
|
| 68 |
+
identities = {}
|
| 69 |
+
for tensor in reader.tensors:
|
| 70 |
+
name = tensor.name.replace('a.conv1d.', 'a.conv2d.').replace('a.pre_encode.out.', 'a.conv_out.') if projector else tensor.name
|
| 71 |
+
data = tensor.data
|
| 72 |
+
shape = tensor.shape.tolist()
|
| 73 |
+
if projector and name.startswith('a.conv2d.') and name.endswith('.bias'):
|
| 74 |
+
data = data.reshape(-1, 1, 1)
|
| 75 |
+
shape = [1, 1, int(tensor.data.size)]
|
| 76 |
+
writer.add_tensor(name, data, raw_dtype=tensor.tensor_type)
|
| 77 |
+
identities[name] = {'sha256': hashlib.sha256(tensor.data.tobytes()).hexdigest(), 'type': int(tensor.tensor_type), 'shape': shape}
|
| 78 |
+
writer.write_header_to_file(); writer.write_kv_data_to_file(); writer.write_tensors_to_file(); writer.close()
|
| 79 |
+
actual = gguf.GGUFReader(output)
|
| 80 |
+
if len(actual.tensors) != len(identities):
|
| 81 |
+
raise RuntimeError('converted tensor inventory changed')
|
| 82 |
+
for tensor in actual.tensors:
|
| 83 |
+
expected = identities[tensor.name]
|
| 84 |
+
if tensor.shape.tolist() != expected['shape'] or int(tensor.tensor_type) != expected['type'] or hashlib.sha256(tensor.data.tobytes()).hexdigest() != expected['sha256']:
|
| 85 |
+
raise RuntimeError(f'tensor bytes, quantization or shape changed: {tensor.name}')
|
| 86 |
+
result = {'schemaVersion':1, 'upstreamCommit':UPSTREAM, 'inputSha256':sha(source), 'outputSha256':sha(output), 'source':str(source.resolve()), 'output':str(output.resolve()), 'projector':projector, 'readerAsrEncoderCompatibility':encoder_compat, 'unchangedTensorBytes':len(identities), 'tensorIdentities':identities}
|
| 87 |
+
output.with_suffix('.conversion.json').write_text(json.dumps(result,indent=2)+'\n')
|
| 88 |
+
print(f'PASS {output.name}: {len(identities)} tensor payloads and quantization types unchanged',flush=True)
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
if __name__ == '__main__':
|
| 92 |
+
p = argparse.ArgumentParser(description=__doc__)
|
| 93 |
+
p.add_argument('--input', type=Path, required=True)
|
| 94 |
+
p.add_argument('--output', type=Path, required=True)
|
| 95 |
+
p.add_argument('--encoder-compat', action='store_true', help='Restore legacy ASR audio encoding; pair with the original text GGUF')
|
| 96 |
+
a = p.parse_args()
|
| 97 |
+
convert(a.input.resolve(strict=True), a.output.resolve(), a.encoder_compat)
|
legacy-asr-conversion.patch
ADDED
|
@@ -0,0 +1,246 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
|
| 2 |
+
index 09544173981afb424ce54e6463da0db003d72935..48a97a3b8737541d46297c16754e3371754ad8b1 100755
|
| 3 |
+
--- a/convert_hf_to_gguf.py
|
| 4 |
+
+++ b/convert_hf_to_gguf.py
|
| 5 |
+
@@ -4718,6 +4718,50 @@ class Qwen3VLTextModel(Qwen3Model):
|
| 6 |
+
yield from super().modify_tensors(data_torch, name, bid)
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
+@ModelBase.register("Qwen3ASRForConditionalGeneration", "Qwen3ASRThinkerForConditionalGeneration")
|
| 10 |
+
+class Qwen3ASRTextModel(Qwen3Model):
|
| 11 |
+
+ model_arch = gguf.MODEL_ARCH.QWEN3
|
| 12 |
+
+
|
| 13 |
+
+ def set_vocab(self):
|
| 14 |
+
+ try:
|
| 15 |
+
+ self._set_vocab_qwen()
|
| 16 |
+
+ except Exception:
|
| 17 |
+
+ # Qwen3-ASR uses Qwen2Tokenizer (BPE), not tiktoken
|
| 18 |
+
+ self._set_vocab_gpt2()
|
| 19 |
+
+
|
| 20 |
+
+ # Qwen3-ASR's tokenizer.json is absent, so SpecialVocab can't resolve
|
| 21 |
+
+ # the EOS token. Explicitly set it from tokenizer_config.json to prevent
|
| 22 |
+
+ # llama.cpp from falling back to the GPT-2 default (token 11 = ',').
|
| 23 |
+
+ tokenizer_config = self.dir_model / "tokenizer_config.json"
|
| 24 |
+
+ if tokenizer_config.exists():
|
| 25 |
+
+ with open(tokenizer_config, "r", encoding="utf-8") as f:
|
| 26 |
+
+ tc = json.load(f)
|
| 27 |
+
+ eos_text = tc.get("eos_token")
|
| 28 |
+
+ if eos_text:
|
| 29 |
+
+ for tid, info in tc.get("added_tokens_decoder", {}).items():
|
| 30 |
+
+ if info.get("content") == eos_text:
|
| 31 |
+
+ self.gguf_writer.add_eos_token_id(int(tid))
|
| 32 |
+
+ break
|
| 33 |
+
+
|
| 34 |
+
+ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
|
| 35 |
+
+ # Strip thinker. prefix (Qwen3-ASR wraps all tensors under thinker.)
|
| 36 |
+
+ if name.startswith("thinker."):
|
| 37 |
+
+ name = name.removeprefix("thinker.")
|
| 38 |
+
+
|
| 39 |
+
+ # Skip audio tensors - they go in the mmproj file
|
| 40 |
+
+ if name.startswith("audio_tower."):
|
| 41 |
+
+ return
|
| 42 |
+
+
|
| 43 |
+
+ # ForcedAligner: lm_head has shape [classify_num, hidden_size] instead of
|
| 44 |
+
+ # [vocab_size, hidden_size]. Skip it so the tied weight fallback is used.
|
| 45 |
+
+ if name == "lm_head.weight":
|
| 46 |
+
+ vocab_size = self.hparams.get("vocab_size", 0)
|
| 47 |
+
+ if data_torch.shape[0] != vocab_size:
|
| 48 |
+
+ return
|
| 49 |
+
+
|
| 50 |
+
+ yield from super().modify_tensors(data_torch, name, bid)
|
| 51 |
+
+
|
| 52 |
+
+
|
| 53 |
+
@ModelBase.register("Qwen3VLMoeForConditionalGeneration")
|
| 54 |
+
class Qwen3VLMoeTextModel(Qwen3MoeModel):
|
| 55 |
+
model_arch = gguf.MODEL_ARCH.QWEN3VLMOE
|
| 56 |
+
@@ -10314,6 +10358,77 @@ class AudioFlamingo3WhisperEncoderModel(WhisperEncoderModel):
|
| 57 |
+
return super().tensor_force_quant(name, new_name, bid, n_dims)
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
+@ModelBase.register("Qwen3ASRForConditionalGeneration", "Qwen3ASRThinkerForConditionalGeneration")
|
| 61 |
+
+class Qwen3ASREncoderModel(MmprojModel):
|
| 62 |
+
+ has_vision_encoder = False
|
| 63 |
+
+ has_audio_encoder = True
|
| 64 |
+
+
|
| 65 |
+
+ def __init__(self, *args, **kwargs):
|
| 66 |
+
+ super().__init__(*args, **kwargs)
|
| 67 |
+
+ if "d_model" in self.hparams:
|
| 68 |
+
+ self.hparams["hidden_size"] = self.hparams["d_model"]
|
| 69 |
+
+ self.hparams["intermediate_size"] = self.hparams["encoder_ffn_dim"]
|
| 70 |
+
+ self.hparams["num_attention_heads"] = self.hparams["encoder_attention_heads"]
|
| 71 |
+
+
|
| 72 |
+
+ def set_gguf_parameters(self):
|
| 73 |
+
+ super().set_gguf_parameters()
|
| 74 |
+
+ self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN3A)
|
| 75 |
+
+ self.gguf_writer.add_audio_num_mel_bins(self.hparams["num_mel_bins"])
|
| 76 |
+
+ self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
|
| 77 |
+
+ self.gguf_writer.add_audio_n_window(self.hparams["n_window"])
|
| 78 |
+
+ self.gguf_writer.add_audio_n_window_infer(self.hparams["n_window_infer"])
|
| 79 |
+
+
|
| 80 |
+
+ # ForcedAligner metadata (optional — only present for forced aligner models)
|
| 81 |
+
+ thinker_cfg = self.global_config.get("thinker_config", {})
|
| 82 |
+
+ classify_num = thinker_cfg.get("classify_num")
|
| 83 |
+
+ if classify_num is not None:
|
| 84 |
+
+ self.gguf_writer.add_audio_classify_num(classify_num)
|
| 85 |
+
+ timestamp_segment_time = self.global_config.get("timestamp_segment_time", 80.0)
|
| 86 |
+
+ self.gguf_writer.add_audio_timestamp_segment_time(timestamp_segment_time)
|
| 87 |
+
+
|
| 88 |
+
+ def get_audio_config(self) -> dict[str, Any] | None:
|
| 89 |
+
+ return self.global_config.get("thinker_config", {}).get("audio_config")
|
| 90 |
+
+
|
| 91 |
+
+ def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
|
| 92 |
+
+ # Generate sinusoidal position embeddings (buffer, not in safetensors)
|
| 93 |
+
+ assert self.hparams_audio is not None
|
| 94 |
+
+ max_timescale = 10000
|
| 95 |
+
+ length = self.hparams_audio.get("max_source_positions", 1500)
|
| 96 |
+
+ channels = self.hparams_audio["hidden_size"]
|
| 97 |
+
+ log_timescale_increment = np.log(max_timescale) / (channels // 2 - 1)
|
| 98 |
+
+ inv_timescales = torch.exp(-log_timescale_increment * torch.arange(channels // 2).float())
|
| 99 |
+
+ scaled_time = torch.arange(length)[:, np.newaxis] * inv_timescales[np.newaxis, :]
|
| 100 |
+
+ pos_embd = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1).to(dtype=torch.float32)
|
| 101 |
+
+ yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_ENC_EMBD_POS), pos_embd)
|
| 102 |
+
+
|
| 103 |
+
+ def tensor_force_quant(self, name, new_name, bid, n_dims):
|
| 104 |
+
+ if ".conv2d" in name and ".weight" in name:
|
| 105 |
+
+ return gguf.GGMLQuantizationType.F16
|
| 106 |
+
+ return super().tensor_force_quant(name, new_name, bid, n_dims)
|
| 107 |
+
+
|
| 108 |
+
+ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
|
| 109 |
+
+ # Strip thinker. prefix
|
| 110 |
+
+ if name.startswith("thinker."):
|
| 111 |
+
+ name = name.removeprefix("thinker.")
|
| 112 |
+
+
|
| 113 |
+
+ # ForcedAligner: include lm_head as classifier tensor
|
| 114 |
+
+ if name == "lm_head.weight":
|
| 115 |
+
+ thinker_cfg = self.global_config.get("thinker_config", {})
|
| 116 |
+
+ if thinker_cfg.get("classify_num") is not None:
|
| 117 |
+
+ yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_CLASSIFIER), data_torch)
|
| 118 |
+
+ return
|
| 119 |
+
+
|
| 120 |
+
+ # Skip text model tensors
|
| 121 |
+
+ if name.startswith("model."):
|
| 122 |
+
+ return
|
| 123 |
+
+
|
| 124 |
+
+ # Handle Conv2D bias shape for ggml compatibility
|
| 125 |
+
+ if "conv2d" in name and "bias" in name:
|
| 126 |
+
+ data_torch = data_torch.reshape(1, -1, 1, 1)
|
| 127 |
+
+
|
| 128 |
+
+ yield from super().modify_tensors(data_torch, name, bid)
|
| 129 |
+
+
|
| 130 |
+
+
|
| 131 |
+
@ModelBase.register("FalconH1ForCausalLM")
|
| 132 |
+
class FalconH1Model(Mamba2Model):
|
| 133 |
+
model_arch = gguf.MODEL_ARCH.FALCON_H1
|
| 134 |
+
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
|
| 135 |
+
index 839c6e787fc0db9db0eb5ff94b9894dd04bea961..38081453dacd64d6507a0db9e49c0da5e721eb41 100644
|
| 136 |
+
--- a/gguf-py/gguf/constants.py
|
| 137 |
+
+++ b/gguf-py/gguf/constants.py
|
| 138 |
+
@@ -332,6 +332,12 @@ class Keys:
|
| 139 |
+
HEAD_COUNT = "clip.audio.attention.head_count"
|
| 140 |
+
LAYERNORM_EPS = "clip.audio.attention.layer_norm_epsilon"
|
| 141 |
+
|
| 142 |
+
+ N_WINDOW = "clip.audio.n_window"
|
| 143 |
+
+ N_WINDOW_INFER = "clip.audio.n_window_infer"
|
| 144 |
+
+
|
| 145 |
+
+ CLASSIFY_NUM = "clip.audio.classify_num"
|
| 146 |
+
+ TIMESTAMP_SEGMENT_TIME = "clip.audio.timestamp_segment_time"
|
| 147 |
+
+
|
| 148 |
+
class Projector:
|
| 149 |
+
STACK_FACTOR = "clip.audio.projector.stack_factor"
|
| 150 |
+
|
| 151 |
+
@@ -799,6 +805,7 @@ class MODEL_TENSOR(IntEnum):
|
| 152 |
+
A_ENC_CONV_NORM = auto() # SSM conv
|
| 153 |
+
A_ENC_CONV_PW1 = auto()
|
| 154 |
+
A_ENC_CONV_PW2 = auto()
|
| 155 |
+
+ A_CLASSIFIER = auto() # forced aligner classifier head
|
| 156 |
+
|
| 157 |
+
|
| 158 |
+
MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
|
| 159 |
+
@@ -1234,6 +1241,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
|
| 160 |
+
MODEL_TENSOR.A_ENC_CONV_NORM: "a.blk.{bid}.conv_norm",
|
| 161 |
+
MODEL_TENSOR.A_ENC_CONV_PW1: "a.blk.{bid}.conv_pw1",
|
| 162 |
+
MODEL_TENSOR.A_ENC_CONV_PW2: "a.blk.{bid}.conv_pw2",
|
| 163 |
+
+ MODEL_TENSOR.A_CLASSIFIER: "a.classifier",
|
| 164 |
+
# NextN/MTP
|
| 165 |
+
MODEL_TENSOR.NEXTN_EH_PROJ: "blk.{bid}.nextn.eh_proj",
|
| 166 |
+
MODEL_TENSOR.NEXTN_EMBED_TOKENS: "blk.{bid}.nextn.embed_tokens",
|
| 167 |
+
@@ -1350,6 +1358,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
|
| 168 |
+
MODEL_TENSOR.A_MM_SOFT_EMB_NORM,
|
| 169 |
+
MODEL_TENSOR.A_MM_EMBEDDING,
|
| 170 |
+
MODEL_TENSOR.A_MM_HARD_EMB_NORM,
|
| 171 |
+
+ MODEL_TENSOR.A_CLASSIFIER,
|
| 172 |
+
],
|
| 173 |
+
MODEL_ARCH.LLAMA: [
|
| 174 |
+
MODEL_TENSOR.TOKEN_EMBD,
|
| 175 |
+
@@ -3890,6 +3899,7 @@ class VisionProjectorType:
|
| 176 |
+
JANUS_PRO = "janus_pro"
|
| 177 |
+
LFM2A = "lfm2a" # audio
|
| 178 |
+
MUSIC_FLAMINGO = "musicflamingo" # audio
|
| 179 |
+
+ QWEN3A = "qwen3a" # audio
|
| 180 |
+
GLM4V = "glm4v"
|
| 181 |
+
YOUTUVL = "youtuvl"
|
| 182 |
+
NEMOTRON_V2_VL = "nemotron_v2_vl"
|
| 183 |
+
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
|
| 184 |
+
index 9ee3ac9e8ffe7df7df4e2a9fd81823fc928c92ce..4bc1de5902bad563d50df18e45328f5355f33cb1 100644
|
| 185 |
+
--- a/gguf-py/gguf/gguf_writer.py
|
| 186 |
+
+++ b/gguf-py/gguf/gguf_writer.py
|
| 187 |
+
@@ -1224,9 +1224,21 @@ class GGUFWriter:
|
| 188 |
+
def add_audio_num_mel_bins(self, value: int) -> None:
|
| 189 |
+
self.add_uint32(Keys.ClipAudio.NUM_MEL_BINS, value)
|
| 190 |
+
|
| 191 |
+
+ def add_audio_n_window(self, value: int) -> None:
|
| 192 |
+
+ self.add_uint32(Keys.ClipAudio.N_WINDOW, value)
|
| 193 |
+
+
|
| 194 |
+
+ def add_audio_n_window_infer(self, value: int) -> None:
|
| 195 |
+
+ self.add_uint32(Keys.ClipAudio.N_WINDOW_INFER, value)
|
| 196 |
+
+
|
| 197 |
+
def add_audio_stack_factor(self, value: int) -> None:
|
| 198 |
+
self.add_uint32(Keys.ClipAudio.Projector.STACK_FACTOR, value)
|
| 199 |
+
|
| 200 |
+
+ def add_audio_classify_num(self, value: int) -> None:
|
| 201 |
+
+ self.add_uint32(Keys.ClipAudio.CLASSIFY_NUM, value)
|
| 202 |
+
+
|
| 203 |
+
+ def add_audio_timestamp_segment_time(self, value: float) -> None:
|
| 204 |
+
+ self.add_float32(Keys.ClipAudio.TIMESTAMP_SEGMENT_TIME, value)
|
| 205 |
+
+
|
| 206 |
+
def add_xielu_alpha_p(self, values: Sequence[float]):
|
| 207 |
+
self.add_array(Keys.xIELU.ALPHA_P, values)
|
| 208 |
+
|
| 209 |
+
diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
|
| 210 |
+
index e575610900e7557100b8c5b1344ab28c90f27822..8b4d78555d9bba5d84a120f46180ee57d2411b00 100644
|
| 211 |
+
--- a/gguf-py/gguf/tensor_mapping.py
|
| 212 |
+
+++ b/gguf-py/gguf/tensor_mapping.py
|
| 213 |
+
@@ -1686,6 +1686,7 @@ class TensorNameMap:
|
| 214 |
+
MODEL_TENSOR.A_ENC_EMBD_POS: (
|
| 215 |
+
"audio_tower.embed_positions", # ultravox
|
| 216 |
+
"audio_embedding.embedding", # lfm2
|
| 217 |
+
+ "audio_tower.positional_embedding", # qwen3-asr
|
| 218 |
+
),
|
| 219 |
+
|
| 220 |
+
MODEL_TENSOR.A_ENC_EMBD_NORM: (
|
| 221 |
+
@@ -1700,6 +1701,7 @@ class TensorNameMap:
|
| 222 |
+
"audio_tower.conv{bid}", # ultravox
|
| 223 |
+
"conformer.pre_encode.conv.{bid}", # lfm2
|
| 224 |
+
"model.audio_tower.subsample_conv_projection.conv_{bid}.conv", # gemma3n
|
| 225 |
+
+ "audio_tower.conv2d{bid}", # qwen3-asr
|
| 226 |
+
),
|
| 227 |
+
|
| 228 |
+
MODEL_TENSOR.A_ENC_CONV1D_NORM: (
|
| 229 |
+
@@ -1823,6 +1825,7 @@ class TensorNameMap:
|
| 230 |
+
MODEL_TENSOR.A_ENC_OUT: (
|
| 231 |
+
"conformer.pre_encode.out", # lfm2
|
| 232 |
+
"model.audio_tower.subsample_conv_projection.input_proj_linear", # gemma3n
|
| 233 |
+
+ "audio_tower.conv_out", # qwen3-asr
|
| 234 |
+
),
|
| 235 |
+
|
| 236 |
+
# note: some tensors below has "audio." pseudo-prefix, to prevent conflicts with vision tensors
|
| 237 |
+
@@ -1830,7 +1833,8 @@ class TensorNameMap:
|
| 238 |
+
|
| 239 |
+
MODEL_TENSOR.A_MMPROJ: (
|
| 240 |
+
"audio.multi_modal_projector.linear_{bid}", # ultravox
|
| 241 |
+
- "audio_adapter.model.{bid}" # lfm2
|
| 242 |
+
+ "audio_adapter.model.{bid}", # lfm2
|
| 243 |
+
+ "audio_tower.proj{bid}", # qwen3-asr
|
| 244 |
+
),
|
| 245 |
+
|
| 246 |
+
MODEL_TENSOR.A_MMPROJ_FC: (
|
legacy-asr-converter.lock.json
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schemaVersion": 1,
|
| 3 |
+
"purpose": "Frozen conversion-only provenance; no legacy inference runtime",
|
| 4 |
+
"baseRevision": "66d65ec29ba7c440cbc31b6f63b74a17b536ba65",
|
| 5 |
+
"patchSha256": "3c8bc44e4515508bb028c0702e6145b0ddf2e63ebd8aae5a4ecacda4fae88e25",
|
| 6 |
+
"files": {
|
| 7 |
+
"convert_hf_to_gguf.py": "6d8f9de4d76861449896102da059799440c69c7319d00bd99f9c7bb201193feb",
|
| 8 |
+
"gguf-py/gguf/constants.py": "fa6b88414512c6a581db82db397649d1bfcca0ab35435dfc095d69e75647695d",
|
| 9 |
+
"gguf-py/gguf/gguf_writer.py": "42e0afc47e5c013a59a632373f76ec51138d3a53ce24f9b9080cf1e92049f55e",
|
| 10 |
+
"gguf-py/gguf/tensor_mapping.py": "26d54809b9a4c0160e4f827ed36bb16c91ac9d6a7c2beb53862b1dbb6e6912cf"
|
| 11 |
+
}
|
| 12 |
+
}
|
models.lock.json
CHANGED
|
@@ -3,6 +3,164 @@
|
|
| 3 |
"release": "reader-models-2026-11-07",
|
| 4 |
"upstreamRevision": "67107ac34f62882d8b53ea41ad68010ea635b9be",
|
| 5 |
"artifacts": [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
{
|
| 7 |
"id": "qwen3-tts-0.6b-base-f16",
|
| 8 |
"origin": {
|
|
@@ -18,19 +176,25 @@
|
|
| 18 |
"path": "qwen-tts/talker-f16.gguf",
|
| 19 |
"size": 1210844480,
|
| 20 |
"sha256": "bd53cb41d56329b4eacd95b4d888294e3d1484d31498c2d15a0e505fdfdd2733",
|
| 21 |
-
"urls": [
|
|
|
|
|
|
|
| 22 |
},
|
| 23 |
{
|
| 24 |
"path": "qwen-tts/mmproj-f16.gguf",
|
| 25 |
"size": 620839104,
|
| 26 |
"sha256": "04fd9028d206a8e9356b54be6817a41bd1faa36f851745d4710ab6ee733041f5",
|
| 27 |
-
"urls": [
|
|
|
|
|
|
|
| 28 |
},
|
| 29 |
{
|
| 30 |
"path": "qwen-tts/reader-model.json",
|
| 31 |
"size": 2179,
|
| 32 |
"sha256": "b048d81860d58c9aea644b95bfdb089e5b16e56da983544c5030011c521e45ee",
|
| 33 |
-
"urls": [
|
|
|
|
|
|
|
| 34 |
"content": "{\n \"schemaVersion\": 1,\n \"repository\": \"Qwen/Qwen3-TTS-12Hz-0.6B-Base\",\n \"revision\": \"5d83992436eae1d760afd27aff78a71d676296fc\",\n \"files\": [\n {\n \"path\": \".gitattributes\",\n \"size\": 1519,\n \"sha256\": \"11ad7efa24975ee4b0c3c3a38ed18737f0658a5f75a0a96787b576a78a023361\"\n },\n {\n \"path\": \"README.md\",\n \"size\": 3640,\n \"sha256\": \"181187b6057906bd960bc7f938d0b7a16652509776a0d52c4885b4ae5ccda0ea\"\n },\n {\n \"path\": \"config.json\",\n \"size\": 4494,\n \"sha256\": \"2e714c787c8edb98b05432685cddb634add2de4d4e645f653d68251ef72ba011\"\n },\n {\n \"path\": \"generation_config.json\",\n \"size\": 245,\n \"sha256\": \"f1b90b4513f3b34c62851049e2492d7b4c5940daf1276f89c82b8ef04127f3aa\"\n },\n {\n \"path\": \"merges.txt\",\n \"size\": 1671839,\n \"sha256\": \"599bab54075088774b1733fde865d5bd747cbcc7a547c5bc12610e874e26f5e3\"\n },\n {\n \"path\": \"model.safetensors\",\n \"size\": 1829344272,\n \"sha256\": \"180b3b10eb1c9f1b4db7806d5475bae3071c0243c299d49926bab1da3b6946f6\"\n },\n {\n \"path\": \"preprocessor_config.json\",\n \"size\": 127,\n \"sha256\": \"efdde1022ea9d76928bf7a9cd53139138f5ba2e466e837f08f6105ab1af1c119\"\n },\n {\n \"path\": \"speech_tokenizer/config.json\",\n \"size\": 2336,\n \"sha256\": \"ee65bb901c876664ab8707c487157aa1a6ee57c65969b28fb5ec9dc211e68167\"\n },\n {\n \"path\": \"speech_tokenizer/configuration.json\",\n \"size\": 76,\n \"sha256\": \"6bc26d64eb5024b4d1dab5a52371958b429256d6c9d59787f1f5294a54e0cebd\"\n },\n {\n \"path\": \"speech_tokenizer/model.safetensors\",\n \"size\": 682293092,\n \"sha256\": \"836b7b357f5ea43e889936a3709af68dfe3751881acefe4ecf0dbd30ba571258\"\n },\n {\n \"path\": \"speech_tokenizer/preprocessor_config.json\",\n \"size\": 234,\n \"sha256\": \"fcb3805e597e786d4067706e602f6688524640f8d3396790e2e09b5942fcbdfb\"\n },\n {\n \"path\": \"tokenizer_config.json\",\n \"size\": 7344,\n \"sha256\": \"dc3c31c3bdaedd5016382bb3cbe07323026775ad51f5a4fb564505992ae4a670\"\n },\n {\n \"path\": \"vocab.json\",\n \"size\": 2776833,\n \"sha256\": \"ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910\"\n }\n ]\n}\n"
|
| 35 |
}
|
| 36 |
]
|
|
@@ -51,7 +215,9 @@
|
|
| 51 |
"path": "voices/heart.wav",
|
| 52 |
"size": 266444,
|
| 53 |
"sha256": "c69f19c5b9e2e46d3e0adc4d3851a169c2ddfe0e389e8e439059da0a19d3e02e",
|
| 54 |
-
"urls": [
|
|
|
|
|
|
|
| 55 |
}
|
| 56 |
]
|
| 57 |
},
|
|
@@ -71,7 +237,9 @@
|
|
| 71 |
"path": "voices/george.wav",
|
| 72 |
"size": 356444,
|
| 73 |
"sha256": "f0da705809cc48f21883bf5ef028cd848c543a3689db64a7a723e34b511507f0",
|
| 74 |
-
"urls": [
|
|
|
|
|
|
|
| 75 |
}
|
| 76 |
]
|
| 77 |
}
|
|
|
|
| 3 |
"release": "reader-models-2026-11-07",
|
| 4 |
"upstreamRevision": "67107ac34f62882d8b53ea41ad68010ea635b9be",
|
| 5 |
"artifacts": [
|
| 6 |
+
{
|
| 7 |
+
"id": "asr-0.6b-f16",
|
| 8 |
+
"origin": {
|
| 9 |
+
"state": "verified-conversion",
|
| 10 |
+
"repository": "Qwen/Qwen3-ASR-0.6B",
|
| 11 |
+
"revision": "5eb144179a02acc5e5ba31e748d22b0cf3e303b0",
|
| 12 |
+
"license": "Apache-2.0",
|
| 13 |
+
"recipeSha256": "91a534aad00af8a42d29c02526eb5ccb4257255d4bc7dbd9350cbcf1a5726acd"
|
| 14 |
+
},
|
| 15 |
+
"publishable": true,
|
| 16 |
+
"files": [
|
| 17 |
+
{
|
| 18 |
+
"path": "Qwen3-ASR-0.6B/Qwen3-ASR-0.6B-F16.gguf",
|
| 19 |
+
"size": 1509344064,
|
| 20 |
+
"sha256": "59cbfd307c21c15a494daef712c9333c19d8e4259f37db4217fd295d508faeb3",
|
| 21 |
+
"urls": []
|
| 22 |
+
}
|
| 23 |
+
]
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"id": "asr-1.7b-f16",
|
| 27 |
+
"origin": {
|
| 28 |
+
"state": "verified-conversion",
|
| 29 |
+
"repository": "Qwen/Qwen3-ASR-1.7B",
|
| 30 |
+
"revision": "7278e1e70fe206f11671096ffdd38061171dd6e5",
|
| 31 |
+
"license": "Apache-2.0",
|
| 32 |
+
"recipeSha256": "91a534aad00af8a42d29c02526eb5ccb4257255d4bc7dbd9350cbcf1a5726acd"
|
| 33 |
+
},
|
| 34 |
+
"publishable": true,
|
| 35 |
+
"files": [
|
| 36 |
+
{
|
| 37 |
+
"path": "Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-F16.gguf",
|
| 38 |
+
"size": 4069675840,
|
| 39 |
+
"sha256": "4257be1bee1f24c2b8e6c2ef7c1c9e0345e17ec805d0ea48e00fae26ef58a52c",
|
| 40 |
+
"urls": []
|
| 41 |
+
}
|
| 42 |
+
]
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"id": "asr-1.7b-q4-k-m",
|
| 46 |
+
"origin": {
|
| 47 |
+
"state": "verified-conversion",
|
| 48 |
+
"repository": "Qwen/Qwen3-ASR-1.7B",
|
| 49 |
+
"revision": "7278e1e70fe206f11671096ffdd38061171dd6e5",
|
| 50 |
+
"license": "Apache-2.0",
|
| 51 |
+
"recipeSha256": "91a534aad00af8a42d29c02526eb5ccb4257255d4bc7dbd9350cbcf1a5726acd"
|
| 52 |
+
},
|
| 53 |
+
"publishable": true,
|
| 54 |
+
"files": [
|
| 55 |
+
{
|
| 56 |
+
"path": "Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q4_K_M.gguf",
|
| 57 |
+
"size": 1282435904,
|
| 58 |
+
"sha256": "8225f44c74bee7bb9ee3f1d5abace9054b46354429e257496a2b7fff266f5b7a",
|
| 59 |
+
"urls": []
|
| 60 |
+
}
|
| 61 |
+
]
|
| 62 |
+
},
|
| 63 |
+
{
|
| 64 |
+
"id": "asr-1.7b-q8-0",
|
| 65 |
+
"origin": {
|
| 66 |
+
"state": "verified-conversion",
|
| 67 |
+
"repository": "Qwen/Qwen3-ASR-1.7B",
|
| 68 |
+
"revision": "7278e1e70fe206f11671096ffdd38061171dd6e5",
|
| 69 |
+
"license": "Apache-2.0",
|
| 70 |
+
"recipeSha256": "91a534aad00af8a42d29c02526eb5ccb4257255d4bc7dbd9350cbcf1a5726acd"
|
| 71 |
+
},
|
| 72 |
+
"publishable": true,
|
| 73 |
+
"files": [
|
| 74 |
+
{
|
| 75 |
+
"path": "Qwen3-ASR-1.7B/Qwen3-ASR-1.7B-Q8_0.gguf",
|
| 76 |
+
"size": 2165035840,
|
| 77 |
+
"sha256": "dd7dd905b8b29d1fe59ba769ef46f8915f8599592859b9463a7251b466b51454",
|
| 78 |
+
"urls": []
|
| 79 |
+
}
|
| 80 |
+
]
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"id": "aligner-q8-0",
|
| 84 |
+
"origin": {
|
| 85 |
+
"state": "verified-conversion",
|
| 86 |
+
"repository": "Qwen/Qwen3-ForcedAligner-0.6B",
|
| 87 |
+
"revision": "c7cbfc2048c462b0d63a45797104fc9db3ad62b7",
|
| 88 |
+
"license": "Apache-2.0",
|
| 89 |
+
"recipeSha256": "91a534aad00af8a42d29c02526eb5ccb4257255d4bc7dbd9350cbcf1a5726acd"
|
| 90 |
+
},
|
| 91 |
+
"publishable": true,
|
| 92 |
+
"files": [
|
| 93 |
+
{
|
| 94 |
+
"path": "Qwen3-ForcedAligner-0.6B/Qwen3-ForcedAligner-0.6B-Q8_0.gguf",
|
| 95 |
+
"size": 639585920,
|
| 96 |
+
"sha256": "0a9b58e7a61eca2acf902794dbe8973860d42cccca49235530b2f2c5f290ed76",
|
| 97 |
+
"urls": []
|
| 98 |
+
}
|
| 99 |
+
]
|
| 100 |
+
},
|
| 101 |
+
{
|
| 102 |
+
"id": "mmproj-asr-0.6b-reader-compat",
|
| 103 |
+
"origin": {
|
| 104 |
+
"state": "verified-conversion",
|
| 105 |
+
"compatibility": "reader-legacy-encoder",
|
| 106 |
+
"converter": "scripts/convert-reader-asr-gguf.py",
|
| 107 |
+
"repository": "Qwen/Qwen3-ASR-0.6B",
|
| 108 |
+
"revision": "5eb144179a02acc5e5ba31e748d22b0cf3e303b0",
|
| 109 |
+
"license": "Apache-2.0",
|
| 110 |
+
"recipeSha256": "91a534aad00af8a42d29c02526eb5ccb4257255d4bc7dbd9350cbcf1a5726acd"
|
| 111 |
+
},
|
| 112 |
+
"publishable": true,
|
| 113 |
+
"files": [
|
| 114 |
+
{
|
| 115 |
+
"path": "projectors/mmproj-ASR-0.6B-reader-compat.gguf",
|
| 116 |
+
"size": 378575776,
|
| 117 |
+
"sha256": "fffb83e0ab2731a8c8ca1741726d417a01b274a6cebc9bb53b84e0f899a6a533",
|
| 118 |
+
"urls": []
|
| 119 |
+
}
|
| 120 |
+
]
|
| 121 |
+
},
|
| 122 |
+
{
|
| 123 |
+
"id": "mmproj-asr-1.7b-reader-compat",
|
| 124 |
+
"origin": {
|
| 125 |
+
"state": "verified-conversion",
|
| 126 |
+
"compatibility": "reader-legacy-encoder",
|
| 127 |
+
"converter": "scripts/convert-reader-asr-gguf.py",
|
| 128 |
+
"repository": "Qwen/Qwen3-ASR-1.7B",
|
| 129 |
+
"revision": "7278e1e70fe206f11671096ffdd38061171dd6e5",
|
| 130 |
+
"license": "Apache-2.0",
|
| 131 |
+
"recipeSha256": "91a534aad00af8a42d29c02526eb5ccb4257255d4bc7dbd9350cbcf1a5726acd"
|
| 132 |
+
},
|
| 133 |
+
"publishable": true,
|
| 134 |
+
"files": [
|
| 135 |
+
{
|
| 136 |
+
"path": "projectors/mmproj-ASR-1.7B-reader-compat.gguf",
|
| 137 |
+
"size": 641774208,
|
| 138 |
+
"sha256": "b0a1c67a7bd6e0231dfddc013f0e7e1e3c9a702b4885f39063981aec503ba7fd",
|
| 139 |
+
"urls": []
|
| 140 |
+
}
|
| 141 |
+
]
|
| 142 |
+
},
|
| 143 |
+
{
|
| 144 |
+
"id": "mmproj-forcedaligner-reader-compat",
|
| 145 |
+
"origin": {
|
| 146 |
+
"state": "verified-conversion",
|
| 147 |
+
"compatibility": "reader-legacy-encoder",
|
| 148 |
+
"converter": "scripts/convert-reader-aligner-projector.py",
|
| 149 |
+
"repository": "Qwen/Qwen3-ForcedAligner-0.6B",
|
| 150 |
+
"revision": "c7cbfc2048c462b0d63a45797104fc9db3ad62b7",
|
| 151 |
+
"license": "Apache-2.0",
|
| 152 |
+
"recipeSha256": "91a534aad00af8a42d29c02526eb5ccb4257255d4bc7dbd9350cbcf1a5726acd"
|
| 153 |
+
},
|
| 154 |
+
"publishable": true,
|
| 155 |
+
"files": [
|
| 156 |
+
{
|
| 157 |
+
"path": "projectors/mmproj-ForcedAligner-reader-compat.gguf",
|
| 158 |
+
"size": 649913056,
|
| 159 |
+
"sha256": "b4f0af1d7daeb9910a9970b90057aa1f038a2319da0452a6da574f4b34106f99",
|
| 160 |
+
"urls": []
|
| 161 |
+
}
|
| 162 |
+
]
|
| 163 |
+
},
|
| 164 |
{
|
| 165 |
"id": "qwen3-tts-0.6b-base-f16",
|
| 166 |
"origin": {
|
|
|
|
| 176 |
"path": "qwen-tts/talker-f16.gguf",
|
| 177 |
"size": 1210844480,
|
| 178 |
"sha256": "bd53cb41d56329b4eacd95b4d888294e3d1484d31498c2d15a0e505fdfdd2733",
|
| 179 |
+
"urls": [
|
| 180 |
+
"https://huggingface.co/xumo/reader-speech-models/resolve/76fee4e9d3b4c99b0710e1453df2a0c5820e5de6/qwen-tts/talker-f16.gguf"
|
| 181 |
+
]
|
| 182 |
},
|
| 183 |
{
|
| 184 |
"path": "qwen-tts/mmproj-f16.gguf",
|
| 185 |
"size": 620839104,
|
| 186 |
"sha256": "04fd9028d206a8e9356b54be6817a41bd1faa36f851745d4710ab6ee733041f5",
|
| 187 |
+
"urls": [
|
| 188 |
+
"https://huggingface.co/xumo/reader-speech-models/resolve/76fee4e9d3b4c99b0710e1453df2a0c5820e5de6/qwen-tts/mmproj-f16.gguf"
|
| 189 |
+
]
|
| 190 |
},
|
| 191 |
{
|
| 192 |
"path": "qwen-tts/reader-model.json",
|
| 193 |
"size": 2179,
|
| 194 |
"sha256": "b048d81860d58c9aea644b95bfdb089e5b16e56da983544c5030011c521e45ee",
|
| 195 |
+
"urls": [
|
| 196 |
+
"https://huggingface.co/xumo/reader-speech-models/resolve/76fee4e9d3b4c99b0710e1453df2a0c5820e5de6/qwen-tts/reader-model.json"
|
| 197 |
+
],
|
| 198 |
"content": "{\n \"schemaVersion\": 1,\n \"repository\": \"Qwen/Qwen3-TTS-12Hz-0.6B-Base\",\n \"revision\": \"5d83992436eae1d760afd27aff78a71d676296fc\",\n \"files\": [\n {\n \"path\": \".gitattributes\",\n \"size\": 1519,\n \"sha256\": \"11ad7efa24975ee4b0c3c3a38ed18737f0658a5f75a0a96787b576a78a023361\"\n },\n {\n \"path\": \"README.md\",\n \"size\": 3640,\n \"sha256\": \"181187b6057906bd960bc7f938d0b7a16652509776a0d52c4885b4ae5ccda0ea\"\n },\n {\n \"path\": \"config.json\",\n \"size\": 4494,\n \"sha256\": \"2e714c787c8edb98b05432685cddb634add2de4d4e645f653d68251ef72ba011\"\n },\n {\n \"path\": \"generation_config.json\",\n \"size\": 245,\n \"sha256\": \"f1b90b4513f3b34c62851049e2492d7b4c5940daf1276f89c82b8ef04127f3aa\"\n },\n {\n \"path\": \"merges.txt\",\n \"size\": 1671839,\n \"sha256\": \"599bab54075088774b1733fde865d5bd747cbcc7a547c5bc12610e874e26f5e3\"\n },\n {\n \"path\": \"model.safetensors\",\n \"size\": 1829344272,\n \"sha256\": \"180b3b10eb1c9f1b4db7806d5475bae3071c0243c299d49926bab1da3b6946f6\"\n },\n {\n \"path\": \"preprocessor_config.json\",\n \"size\": 127,\n \"sha256\": \"efdde1022ea9d76928bf7a9cd53139138f5ba2e466e837f08f6105ab1af1c119\"\n },\n {\n \"path\": \"speech_tokenizer/config.json\",\n \"size\": 2336,\n \"sha256\": \"ee65bb901c876664ab8707c487157aa1a6ee57c65969b28fb5ec9dc211e68167\"\n },\n {\n \"path\": \"speech_tokenizer/configuration.json\",\n \"size\": 76,\n \"sha256\": \"6bc26d64eb5024b4d1dab5a52371958b429256d6c9d59787f1f5294a54e0cebd\"\n },\n {\n \"path\": \"speech_tokenizer/model.safetensors\",\n \"size\": 682293092,\n \"sha256\": \"836b7b357f5ea43e889936a3709af68dfe3751881acefe4ecf0dbd30ba571258\"\n },\n {\n \"path\": \"speech_tokenizer/preprocessor_config.json\",\n \"size\": 234,\n \"sha256\": \"fcb3805e597e786d4067706e602f6688524640f8d3396790e2e09b5942fcbdfb\"\n },\n {\n \"path\": \"tokenizer_config.json\",\n \"size\": 7344,\n \"sha256\": \"dc3c31c3bdaedd5016382bb3cbe07323026775ad51f5a4fb564505992ae4a670\"\n },\n {\n \"path\": \"vocab.json\",\n \"size\": 2776833,\n \"sha256\": \"ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910\"\n }\n ]\n}\n"
|
| 199 |
}
|
| 200 |
]
|
|
|
|
| 215 |
"path": "voices/heart.wav",
|
| 216 |
"size": 266444,
|
| 217 |
"sha256": "c69f19c5b9e2e46d3e0adc4d3851a169c2ddfe0e389e8e439059da0a19d3e02e",
|
| 218 |
+
"urls": [
|
| 219 |
+
"https://huggingface.co/xumo/reader-speech-models/resolve/76fee4e9d3b4c99b0710e1453df2a0c5820e5de6/voices/heart.wav"
|
| 220 |
+
]
|
| 221 |
}
|
| 222 |
]
|
| 223 |
},
|
|
|
|
| 237 |
"path": "voices/george.wav",
|
| 238 |
"size": 356444,
|
| 239 |
"sha256": "f0da705809cc48f21883bf5ef028cd848c543a3689db64a7a723e34b511507f0",
|
| 240 |
+
"urls": [
|
| 241 |
+
"https://huggingface.co/xumo/reader-speech-models/resolve/76fee4e9d3b4c99b0710e1453df2a0c5820e5de6/voices/george.wav"
|
| 242 |
+
]
|
| 243 |
}
|
| 244 |
]
|
| 245 |
}
|
projectors/mmproj-ASR-0.6B-reader-compat.gguf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fffb83e0ab2731a8c8ca1741726d417a01b274a6cebc9bb53b84e0f899a6a533
|
| 3 |
+
size 378575776
|
projectors/mmproj-ASR-1.7B-reader-compat.gguf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b0a1c67a7bd6e0231dfddc013f0e7e1e3c9a702b4885f39063981aec503ba7fd
|
| 3 |
+
size 641774208
|
projectors/mmproj-ForcedAligner-reader-compat.gguf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b4f0af1d7daeb9910a9970b90057aa1f038a2319da0452a6da574f4b34106f99
|
| 3 |
+
size 649913056
|