Add files using upload-large-folder tool
Browse files- .eval_results/open_asr_leaderboard.yaml +79 -0
- .gitattributes +2 -0
- README.md +291 -0
- added_tokens.json +29 -0
- chat_template.jinja +15 -0
- config.json +125 -0
- configuration_arkasr.py +47 -0
- figures/ark_asr_architecture.png +3 -0
- generation_config.json +7 -0
- merges.txt +0 -0
- model-00001-of-00002.safetensors +3 -0
- model-00002-of-00002.safetensors +3 -0
- model.safetensors.index.json +934 -0
- modeling_arkasr.py +263 -0
- modeling_audio.py +293 -0
- preprocessor_config.json +18 -0
- processing_arkasr.py +446 -0
- processor_config.json +9 -0
- special_tokens_map.json +23 -0
- tokenizer.json +3 -0
- tokenizer_config.json +239 -0
- vocab.json +0 -0
.eval_results/open_asr_leaderboard.yaml
ADDED
|
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
- dataset:
|
| 2 |
+
id: hf-audio/open-asr-leaderboard
|
| 3 |
+
task_id: mean_wer
|
| 4 |
+
value: 5.13
|
| 5 |
+
date: '2026-06-22'
|
| 6 |
+
source:
|
| 7 |
+
url: https://huggingface.co/datasets/hf-audio/open-asr-leaderboard
|
| 8 |
+
name: open-asr-leaderboard
|
| 9 |
+
user: hf-audio
|
| 10 |
+
|
| 11 |
+
- dataset:
|
| 12 |
+
id: hf-audio/open-asr-leaderboard
|
| 13 |
+
task_id: ami_wer
|
| 14 |
+
value: 8.91
|
| 15 |
+
date: '2026-06-22'
|
| 16 |
+
source:
|
| 17 |
+
url: https://huggingface.co/datasets/hf-audio/open-asr-leaderboard
|
| 18 |
+
name: open-asr-leaderboard
|
| 19 |
+
user: hf-audio
|
| 20 |
+
|
| 21 |
+
- dataset:
|
| 22 |
+
id: hf-audio/open-asr-leaderboard
|
| 23 |
+
task_id: earnings22_wer
|
| 24 |
+
value: 8.25
|
| 25 |
+
date: '2026-06-22'
|
| 26 |
+
source:
|
| 27 |
+
url: https://huggingface.co/datasets/hf-audio/open-asr-leaderboard
|
| 28 |
+
name: open-asr-leaderboard
|
| 29 |
+
user: hf-audio
|
| 30 |
+
|
| 31 |
+
- dataset:
|
| 32 |
+
id: hf-audio/open-asr-leaderboard
|
| 33 |
+
task_id: gigaspeech_wer
|
| 34 |
+
value: 7.30
|
| 35 |
+
date: '2026-06-22'
|
| 36 |
+
source:
|
| 37 |
+
url: https://huggingface.co/datasets/hf-audio/open-asr-leaderboard
|
| 38 |
+
name: open-asr-leaderboard
|
| 39 |
+
user: hf-audio
|
| 40 |
+
|
| 41 |
+
- dataset:
|
| 42 |
+
id: hf-audio/open-asr-leaderboard
|
| 43 |
+
task_id: librispeech_clean_wer
|
| 44 |
+
value: 1.09
|
| 45 |
+
date: '2026-06-22'
|
| 46 |
+
source:
|
| 47 |
+
url: https://huggingface.co/datasets/hf-audio/open-asr-leaderboard
|
| 48 |
+
name: open-asr-leaderboard
|
| 49 |
+
user: hf-audio
|
| 50 |
+
|
| 51 |
+
- dataset:
|
| 52 |
+
id: hf-audio/open-asr-leaderboard
|
| 53 |
+
task_id: librispeech_other_wer
|
| 54 |
+
value: 2.41
|
| 55 |
+
date: '2026-06-22'
|
| 56 |
+
source:
|
| 57 |
+
url: https://huggingface.co/datasets/hf-audio/open-asr-leaderboard
|
| 58 |
+
name: open-asr-leaderboard
|
| 59 |
+
user: hf-audio
|
| 60 |
+
|
| 61 |
+
- dataset:
|
| 62 |
+
id: hf-audio/open-asr-leaderboard
|
| 63 |
+
task_id: spgispeech_wer
|
| 64 |
+
value: 2.49
|
| 65 |
+
date: '2026-06-22'
|
| 66 |
+
source:
|
| 67 |
+
url: https://huggingface.co/datasets/hf-audio/open-asr-leaderboard
|
| 68 |
+
name: open-asr-leaderboard
|
| 69 |
+
user: hf-audio
|
| 70 |
+
|
| 71 |
+
- dataset:
|
| 72 |
+
id: hf-audio/open-asr-leaderboard
|
| 73 |
+
task_id: voxpopuli_wer
|
| 74 |
+
value: 5.48
|
| 75 |
+
date: '2026-06-22'
|
| 76 |
+
source:
|
| 77 |
+
url: https://huggingface.co/datasets/hf-audio/open-asr-leaderboard
|
| 78 |
+
name: open-asr-leaderboard
|
| 79 |
+
user: hf-audio
|
.gitattributes
CHANGED
|
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
figures/ark_asr_architecture.png filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,291 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
tags:
|
| 4 |
+
- automatic-speech-recognition
|
| 5 |
+
- speech
|
| 6 |
+
- audio
|
| 7 |
+
- transformers
|
| 8 |
+
- pytorch
|
| 9 |
+
- safetensors
|
| 10 |
+
- vllm
|
| 11 |
+
- ark-asr
|
| 12 |
+
pipeline_tag: automatic-speech-recognition
|
| 13 |
+
language:
|
| 14 |
+
- zh
|
| 15 |
+
- en
|
| 16 |
+
- de
|
| 17 |
+
- ja
|
| 18 |
+
- fr
|
| 19 |
+
- ko
|
| 20 |
+
- es
|
| 21 |
+
- pl
|
| 22 |
+
- it
|
| 23 |
+
- ro
|
| 24 |
+
- hu
|
| 25 |
+
- cs
|
| 26 |
+
- nl
|
| 27 |
+
- fi
|
| 28 |
+
- hr
|
| 29 |
+
- sk
|
| 30 |
+
- sl
|
| 31 |
+
- et
|
| 32 |
+
- lt
|
| 33 |
+
license: apache-2.0
|
| 34 |
+
repository: https://github.com/AutoArk/open-audio-opd
|
| 35 |
+
---
|
| 36 |
+
|
| 37 |
+
<div align="center">
|
| 38 |
+
|
| 39 |
+
# ARK-ASR-3B: State-of-the-Art Multilingual ASR with Online Policy Distillation
|
| 40 |
+
|
| 41 |
+
[](https://github.com/AutoArk/open-audio-opd)
|
| 42 |
+
[](https://arxiv.org/abs/2605.28139)
|
| 43 |
+
[](https://www.apache.org/licenses/LICENSE-2.0)
|
| 44 |
+
|
| 45 |
+
</div>
|
| 46 |
+
|
| 47 |
+
> **TL;DR** ARK-ASR-3B is an automatic speech recognition model trained with teacher-data adaptation and on-policy distillation. It achieves current state-of-the-art results on the Hugging Face Open ASR Leaderboard English short-form benchmark, with an average WER of **5.13%** across AMI, Earnings22, GigaSpeech, LibriSpeech, SPGISpeech, and VoxPopuli. The accompanying training, inference, and evaluation code is available at [AutoArk/open-audio-opd](https://github.com/AutoArk/open-audio-opd).
|
| 48 |
+
|
| 49 |
+
## Abstract
|
| 50 |
+
|
| 51 |
+
ARK-ASR is an audio ASR student model optimized with the **teacher-data adaptation + online policy distillation (TD + OPD)** recipe from `open-audio-opd`.
|
| 52 |
+
|
| 53 |
+
Instead of relying only on static supervised transcripts, OPD lets the student generate transcripts online and trains it against token-level teacher scores on the student's own generated behavior. This checkpoint is the 3B-scale ARK-ASR model trained with the TD + OPD recipe.
|
| 54 |
+
|
| 55 |
+
ARK-ASR currently supports Chinese, English, German, Japanese, French, Korean, Spanish, Polish, Italian, Romanian, Hungarian, Czech, Dutch, Finnish, Croatian, Slovak, Slovene, Estonian, and Lithuanian ASR.
|
| 56 |
+
|
| 57 |
+
## Supported Languages
|
| 58 |
+
|
| 59 |
+
Chinese, English, German, Japanese, French, Korean, Spanish, Polish, Italian, Romanian, Hungarian, Czech, Dutch, Finnish, Croatian, Slovak, Slovene, Estonian, and Lithuanian.
|
| 60 |
+
|
| 61 |
+
## Model Overview
|
| 62 |
+
|
| 63 |
+
<div align="center">
|
| 64 |
+
<img src="figures/ark_asr_architecture.png" width="95%" alt="ARK-ASR architecture"/>
|
| 65 |
+
<br>
|
| 66 |
+
<p><strong>Figure 1: ARK-ASR architecture.</strong> Audio is encoded by a Whisper-style encoder with RoPE, merged through an MLP adapter, and injected into a Qwen decoder by replacing audio placeholder token embeddings before transcript generation.</p>
|
| 67 |
+
</div>
|
| 68 |
+
|
| 69 |
+
- **Model size:** 3B-scale decoder LLM with a dedicated Whisper-style audio encoder and MLP adapter
|
| 70 |
+
- **Task:** automatic speech recognition
|
| 71 |
+
- **Architecture:** audio-capable autoregressive Transformers model with custom `arkasr` remote code
|
| 72 |
+
- **Checkpoint format:** `safetensors`
|
| 73 |
+
- **Sampling rate:** 16 kHz
|
| 74 |
+
- **Recommended inference code:** [`scripts/infer/ark_asr_transformers.py`](https://github.com/AutoArk/open-audio-opd/blob/main/scripts/infer/ark_asr_transformers.py)
|
| 75 |
+
- **vLLM serving:** [`scripts/vllm/ark_asr_vllm`](https://github.com/AutoArk/open-audio-opd/tree/master/scripts/vllm/ark_asr_vllm)
|
| 76 |
+
|
| 77 |
+
The model should be loaded with `trust_remote_code=True`. The official inference script handles the processor, tokenizer, audio prompt format, generation cleanup, and ASR token filtering.
|
| 78 |
+
|
| 79 |
+
## Performance
|
| 80 |
+
|
| 81 |
+
The following results are from the Hugging Face [Open ASR Leaderboard](https://huggingface.co/datasets/hf-audio/open-asr-leaderboard). Lower WER is better. ARK-ASR-3B reaches the current state of the art on this English short-form benchmark.
|
| 82 |
+
|
| 83 |
+
### English WER
|
| 84 |
+
|
| 85 |
+
| Model | AMI | Earnings22 | GigaSpeech | LS Clean | LS Other | SPGISpeech | VoxPopuli | Avg |
|
| 86 |
+
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
|
| 87 |
+
| ARK-ASR-3B | **8.91%** | **8.25%** | **7.30%** | **1.09%** | **2.41%** | **2.49%** | **5.48%** | **5.13%** |
|
| 88 |
+
| ARK-ASR-0.6B | 10.02% | 9.77% | 8.00% | 1.53% | 3.51% | 2.63% | 6.31% | 5.97% |
|
| 89 |
+
|
| 90 |
+
## Inference
|
| 91 |
+
|
| 92 |
+
Run ASR inference with Hugging Face Transformers:
|
| 93 |
+
|
| 94 |
+
```python
|
| 95 |
+
import torch
|
| 96 |
+
from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer
|
| 97 |
+
|
| 98 |
+
model_path = "AutoArk-AI/ARK-ASR-3B"
|
| 99 |
+
audio_path = "assets/libai.wav"
|
| 100 |
+
|
| 101 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 102 |
+
torch_dtype = torch.bfloat16 if device == "cuda" else torch.float32
|
| 103 |
+
|
| 104 |
+
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
|
| 105 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
| 106 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 107 |
+
model_path,
|
| 108 |
+
trust_remote_code=True,
|
| 109 |
+
torch_dtype=torch_dtype,
|
| 110 |
+
attn_implementation="sdpa",
|
| 111 |
+
).to(device)
|
| 112 |
+
model.eval()
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def build_bad_words_ids(tokenizer):
|
| 116 |
+
eos_ids = tokenizer.eos_token_id
|
| 117 |
+
keep_ids = {eos_ids} if isinstance(eos_ids, int) else set(eos_ids or [])
|
| 118 |
+
bad_ids = set(tokenizer.all_special_ids) - keep_ids
|
| 119 |
+
bad_ids.update(
|
| 120 |
+
token_id
|
| 121 |
+
for token, token_id in tokenizer.get_added_vocab().items()
|
| 122 |
+
if token.startswith("<") and token.endswith(">") and token_id not in keep_ids
|
| 123 |
+
)
|
| 124 |
+
return [[token_id] for token_id in sorted(bad_ids)]
|
| 125 |
+
|
| 126 |
+
conversation = [
|
| 127 |
+
{
|
| 128 |
+
"role": "user",
|
| 129 |
+
"content": [
|
| 130 |
+
{"type": "audio", "path": audio_path},
|
| 131 |
+
{"type": "text", "text": "Please transcribe this audio."},
|
| 132 |
+
],
|
| 133 |
+
}
|
| 134 |
+
]
|
| 135 |
+
|
| 136 |
+
inputs = processor.apply_chat_template(
|
| 137 |
+
conversation,
|
| 138 |
+
add_generation_prompt=True,
|
| 139 |
+
return_tensors="pt",
|
| 140 |
+
sampling_rate=16000,
|
| 141 |
+
audio_padding="longest",
|
| 142 |
+
text_kwargs={"padding": "longest"},
|
| 143 |
+
audio_max_length=30 * 16000,
|
| 144 |
+
)
|
| 145 |
+
inputs = inputs.to(device)
|
| 146 |
+
if "audios" in inputs:
|
| 147 |
+
inputs["audios"] = inputs["audios"].to(dtype=torch_dtype)
|
| 148 |
+
|
| 149 |
+
bad_words_ids = build_bad_words_ids(tokenizer)
|
| 150 |
+
with torch.inference_mode():
|
| 151 |
+
outputs = model.generate(
|
| 152 |
+
**inputs,
|
| 153 |
+
do_sample=False,
|
| 154 |
+
max_new_tokens=256,
|
| 155 |
+
pad_token_id=tokenizer.pad_token_id,
|
| 156 |
+
eos_token_id=tokenizer.eos_token_id,
|
| 157 |
+
bad_words_ids=bad_words_ids,
|
| 158 |
+
)
|
| 159 |
+
decoded_outputs = tokenizer.batch_decode(
|
| 160 |
+
outputs[:, inputs.input_ids.shape[1] :],
|
| 161 |
+
skip_special_tokens=True,
|
| 162 |
+
)
|
| 163 |
+
print(decoded_outputs)
|
| 164 |
+
```
|
| 165 |
+
|
| 166 |
+
For batch JSONL inference, use the open-source inference code:
|
| 167 |
+
|
| 168 |
+
```bash
|
| 169 |
+
git clone https://github.com/AutoArk/open-audio-opd
|
| 170 |
+
cd open-audio-opd
|
| 171 |
+
pip install -e .
|
| 172 |
+
```
|
| 173 |
+
|
| 174 |
+
The input JSONL should contain one ASR sample per line:
|
| 175 |
+
|
| 176 |
+
```json
|
| 177 |
+
{"audio":"/path/to/audio.wav","text":"","task":"asr","begin_time":-1,"end_time":-1}
|
| 178 |
+
```
|
| 179 |
+
|
| 180 |
+
```bash
|
| 181 |
+
python scripts/infer/ark_asr_transformers.py \
|
| 182 |
+
--input /path/to/input.jsonl \
|
| 183 |
+
--output runs/infer/predictions.jsonl \
|
| 184 |
+
--model_path AutoArk-AI/ARK-ASR-3B \
|
| 185 |
+
--processor_path AutoArk-AI/ARK-ASR-3B \
|
| 186 |
+
--batch_size 40 \
|
| 187 |
+
--dtype bfloat16 \
|
| 188 |
+
--attn_impl sdpa
|
| 189 |
+
```
|
| 190 |
+
|
| 191 |
+
The output JSONL preserves input metadata and adds:
|
| 192 |
+
|
| 193 |
+
- `pred_text`: cleaned prediction text for downstream evaluation
|
| 194 |
+
- `pred_text_raw`: raw decoded generation before cleanup
|
| 195 |
+
|
| 196 |
+
## vLLM Online Serving
|
| 197 |
+
|
| 198 |
+
ARK-ASR can also be deployed as a vLLM-backed online ASR service with the
|
| 199 |
+
adapter in
|
| 200 |
+
[`scripts/vllm/ark_asr_vllm`](https://github.com/AutoArk/open-audio-opd/tree/master/scripts/vllm/ark_asr_vllm).
|
| 201 |
+
The service exposes both a compact `/asr` endpoint and an OpenAI-style
|
| 202 |
+
`/v1/audio/transcriptions` endpoint.
|
| 203 |
+
|
| 204 |
+
Clone and install the serving code:
|
| 205 |
+
|
| 206 |
+
```bash
|
| 207 |
+
git clone https://github.com/AutoArk/open-audio-opd
|
| 208 |
+
cd open-audio-opd
|
| 209 |
+
pip install -e ".[vllm]"
|
| 210 |
+
```
|
| 211 |
+
|
| 212 |
+
Start the service:
|
| 213 |
+
|
| 214 |
+
```bash
|
| 215 |
+
MODEL=AutoArk-AI/ARK-ASR-3B \
|
| 216 |
+
GPU=0 \
|
| 217 |
+
PORT=8025 \
|
| 218 |
+
scripts/vllm/deploy_ark_asr_vllm_service.sh start
|
| 219 |
+
```
|
| 220 |
+
|
| 221 |
+
Check the service:
|
| 222 |
+
|
| 223 |
+
```bash
|
| 224 |
+
scripts/vllm/deploy_ark_asr_vllm_service.sh status
|
| 225 |
+
curl -sS http://127.0.0.1:8025/health
|
| 226 |
+
curl -sS http://127.0.0.1:8025/token-mask
|
| 227 |
+
```
|
| 228 |
+
|
| 229 |
+
Run one transcription request:
|
| 230 |
+
|
| 231 |
+
```bash
|
| 232 |
+
curl -sS -X POST http://127.0.0.1:8025/asr \
|
| 233 |
+
-F file=@/path/to/audio.wav \
|
| 234 |
+
-F max_new_tokens=256
|
| 235 |
+
```
|
| 236 |
+
|
| 237 |
+
OpenAI-style transcription endpoint:
|
| 238 |
+
|
| 239 |
+
```bash
|
| 240 |
+
curl -sS -X POST http://127.0.0.1:8025/v1/audio/transcriptions \
|
| 241 |
+
-F file=@/path/to/audio.wav \
|
| 242 |
+
-F model=ark-asr
|
| 243 |
+
```
|
| 244 |
+
|
| 245 |
+
Stop the service:
|
| 246 |
+
|
| 247 |
+
```bash
|
| 248 |
+
scripts/vllm/deploy_ark_asr_vllm_service.sh stop
|
| 249 |
+
```
|
| 250 |
+
|
| 251 |
+
The vLLM adapter registers the custom `arkasr` model, loads the local
|
| 252 |
+
processor/tokenizer with `trust_remote_code=True`, applies generation-time
|
| 253 |
+
token masking for non-ASR control tokens, and keeps `<|im_end|>` as the stop
|
| 254 |
+
token. Service logs and PID files are written under `runs/vllm/`.
|
| 255 |
+
|
| 256 |
+
## Evaluation
|
| 257 |
+
|
| 258 |
+
The repository also includes a J/WER evaluation entrypoint:
|
| 259 |
+
|
| 260 |
+
```bash
|
| 261 |
+
python scripts/eval/eval_jwer_ark_asr_transformers.py \
|
| 262 |
+
--input /path/to/test.jsonl \
|
| 263 |
+
--output runs/eval/result.jsonl \
|
| 264 |
+
--model_path AutoArk-AI/ARK-ASR-3B \
|
| 265 |
+
--processor_path AutoArk-AI/ARK-ASR-3B \
|
| 266 |
+
--batch_size 40 \
|
| 267 |
+
--dtype bfloat16 \
|
| 268 |
+
--attn_impl sdpa
|
| 269 |
+
```
|
| 270 |
+
|
| 271 |
+
No evaluation audio or dataset files are bundled with this model repository.
|
| 272 |
+
|
| 273 |
+
## Acknowledgements
|
| 274 |
+
|
| 275 |
+
The training code is based on [THUNLP/OPD](https://github.com/thunlp/OPD/) and [verl](https://github.com/volcengine/verl). The OPD recipe uses a stronger ASR teacher to score online student rollouts.
|
| 276 |
+
|
| 277 |
+
## Citation
|
| 278 |
+
|
| 279 |
+
If you find ARK-ASR or open-audio-opd useful, please cite:
|
| 280 |
+
|
| 281 |
+
```bibtex
|
| 282 |
+
@misc{lin2026dataefficientopd,
|
| 283 |
+
title={Data-Efficient On-Policy Distillation for Automatic Speech Recognition},
|
| 284 |
+
author={Lin, Yu and Wang, Yiming and Cai, Runyuan and Zeng, Xiaodong},
|
| 285 |
+
year={2026},
|
| 286 |
+
eprint={2605.28139},
|
| 287 |
+
archivePrefix={arXiv},
|
| 288 |
+
primaryClass={cs.AI},
|
| 289 |
+
url={https://arxiv.org/abs/2605.28139}
|
| 290 |
+
}
|
| 291 |
+
```
|
added_tokens.json
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"</tool_call>": 151658,
|
| 3 |
+
"<tool_call>": 151657,
|
| 4 |
+
"<|assistant|>": 151668,
|
| 5 |
+
"<|audio|>": 151663,
|
| 6 |
+
"<|begin_of_audio|>": 151666,
|
| 7 |
+
"<|end_content|>": 151651,
|
| 8 |
+
"<|end_global_token|>": 151649,
|
| 9 |
+
"<|end_of_audio|>": 151667,
|
| 10 |
+
"<|endoftext|>": 151643,
|
| 11 |
+
"<|file_sep|>": 151664,
|
| 12 |
+
"<|fim_middle|>": 151660,
|
| 13 |
+
"<|fim_pad|>": 151662,
|
| 14 |
+
"<|fim_prefix|>": 151659,
|
| 15 |
+
"<|fim_suffix|>": 151661,
|
| 16 |
+
"<|im_end|>": 151645,
|
| 17 |
+
"<|im_start|>": 151644,
|
| 18 |
+
"<|image_pad|>": 151655,
|
| 19 |
+
"<|object_ref_end|>": 151647,
|
| 20 |
+
"<|object_ref_start|>": 151646,
|
| 21 |
+
"<|start_content|>": 151650,
|
| 22 |
+
"<|start_global_token|>": 151648,
|
| 23 |
+
"<|system|>": 151669,
|
| 24 |
+
"<|user|>": 151665,
|
| 25 |
+
"<|video_pad|>": 151656,
|
| 26 |
+
"<|vision_end|>": 151653,
|
| 27 |
+
"<|vision_pad|>": 151654,
|
| 28 |
+
"<|vision_start|>": 151652
|
| 29 |
+
}
|
chat_template.jinja
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{% for message in messages %}
|
| 2 |
+
{% if message.role == "system" %}
|
| 3 |
+
<|system|>
|
| 4 |
+
{{ message.content }}
|
| 5 |
+
{% elif message.role == "user" %}
|
| 6 |
+
<|user|>
|
| 7 |
+
{{ message.content }}
|
| 8 |
+
{% elif message.role == "assistant" %}
|
| 9 |
+
<|assistant|>
|
| 10 |
+
{{ message.content }}
|
| 11 |
+
{% endif %}
|
| 12 |
+
{% endfor %}
|
| 13 |
+
{% if add_generation_prompt %}
|
| 14 |
+
<|assistant|>
|
| 15 |
+
{% endif %}
|
config.json
ADDED
|
@@ -0,0 +1,125 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"adapter_type": "mlp",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"ArkasrForConditionalGeneration"
|
| 5 |
+
],
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"audio_token_id": 151663,
|
| 8 |
+
"auto_map": {
|
| 9 |
+
"AutoConfig": "configuration_arkasr.ArkasrConfig",
|
| 10 |
+
"AutoModelForCausalLM": "modeling_arkasr.ArkasrForConditionalGeneration"
|
| 11 |
+
},
|
| 12 |
+
"bos_token_id": 151643,
|
| 13 |
+
"dtype": "bfloat16",
|
| 14 |
+
"eos_token_id": 151645,
|
| 15 |
+
"hidden_act": "silu",
|
| 16 |
+
"hidden_size": 2048,
|
| 17 |
+
"initializer_range": 0.02,
|
| 18 |
+
"intermediate_size": 11008,
|
| 19 |
+
"layer_types": [
|
| 20 |
+
"full_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"full_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"full_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"full_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"full_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"full_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"full_attention",
|
| 43 |
+
"full_attention",
|
| 44 |
+
"full_attention",
|
| 45 |
+
"full_attention",
|
| 46 |
+
"full_attention",
|
| 47 |
+
"full_attention",
|
| 48 |
+
"full_attention",
|
| 49 |
+
"full_attention",
|
| 50 |
+
"full_attention",
|
| 51 |
+
"full_attention",
|
| 52 |
+
"full_attention",
|
| 53 |
+
"full_attention",
|
| 54 |
+
"full_attention",
|
| 55 |
+
"full_attention"
|
| 56 |
+
],
|
| 57 |
+
"max_position_embeddings": 32768,
|
| 58 |
+
"max_whisper_length": 1500,
|
| 59 |
+
"max_window_layers": 70,
|
| 60 |
+
"merge_factor": 4,
|
| 61 |
+
"mlp_adapter_act": "gelu",
|
| 62 |
+
"model_type": "arkasr",
|
| 63 |
+
"num_attention_heads": 16,
|
| 64 |
+
"num_hidden_layers": 36,
|
| 65 |
+
"num_key_value_heads": 2,
|
| 66 |
+
"pad_token_id": 151643,
|
| 67 |
+
"rms_norm_eps": 1e-06,
|
| 68 |
+
"rope_scaling": null,
|
| 69 |
+
"rope_theta": 1000000.0,
|
| 70 |
+
"sliding_window": null,
|
| 71 |
+
"spec_aug": false,
|
| 72 |
+
"tie_word_embeddings": true,
|
| 73 |
+
"transformers_version": "4.57.6",
|
| 74 |
+
"use_cache": true,
|
| 75 |
+
"use_mrope": false,
|
| 76 |
+
"use_rope": true,
|
| 77 |
+
"use_sliding_window": false,
|
| 78 |
+
"vocab_size": 151936,
|
| 79 |
+
"whisper_config": {
|
| 80 |
+
"activation_dropout": 0.0,
|
| 81 |
+
"activation_function": "gelu",
|
| 82 |
+
"apply_spec_augment": false,
|
| 83 |
+
"architectures": [
|
| 84 |
+
"WhisperForConditionalGeneration"
|
| 85 |
+
],
|
| 86 |
+
"attention_dropout": 0.0,
|
| 87 |
+
"begin_suppress_tokens": [
|
| 88 |
+
220,
|
| 89 |
+
50257
|
| 90 |
+
],
|
| 91 |
+
"bos_token_id": 50257,
|
| 92 |
+
"classifier_proj_size": 256,
|
| 93 |
+
"d_model": 1280,
|
| 94 |
+
"decoder_attention_heads": 20,
|
| 95 |
+
"decoder_ffn_dim": 5120,
|
| 96 |
+
"decoder_layerdrop": 0.0,
|
| 97 |
+
"decoder_layers": 32,
|
| 98 |
+
"decoder_start_token_id": 50258,
|
| 99 |
+
"dropout": 0.0,
|
| 100 |
+
"dtype": "bfloat16",
|
| 101 |
+
"encoder_attention_heads": 20,
|
| 102 |
+
"encoder_ffn_dim": 5120,
|
| 103 |
+
"encoder_layerdrop": 0.0,
|
| 104 |
+
"encoder_layers": 32,
|
| 105 |
+
"eos_token_id": 50257,
|
| 106 |
+
"init_std": 0.02,
|
| 107 |
+
"mask_feature_length": 10,
|
| 108 |
+
"mask_feature_min_masks": 0,
|
| 109 |
+
"mask_feature_prob": 0.0,
|
| 110 |
+
"mask_time_length": 10,
|
| 111 |
+
"mask_time_min_masks": 2,
|
| 112 |
+
"mask_time_prob": 0.05,
|
| 113 |
+
"max_length": 448,
|
| 114 |
+
"max_source_positions": 1500,
|
| 115 |
+
"max_target_positions": 448,
|
| 116 |
+
"median_filter_width": 7,
|
| 117 |
+
"model_type": "whisper",
|
| 118 |
+
"num_hidden_layers": 32,
|
| 119 |
+
"num_mel_bins": 128,
|
| 120 |
+
"scale_embedding": false,
|
| 121 |
+
"use_cache": true,
|
| 122 |
+
"use_weighted_layer_sum": false,
|
| 123 |
+
"vocab_size": 51866
|
| 124 |
+
}
|
| 125 |
+
}
|
configuration_arkasr.py
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Any, Dict, Optional, Union
|
| 2 |
+
from transformers import Qwen2Config, WhisperConfig
|
| 3 |
+
|
| 4 |
+
|
| 5 |
+
class ArkasrConfig(Qwen2Config):
|
| 6 |
+
model_type = "arkasr"
|
| 7 |
+
is_composition = True
|
| 8 |
+
|
| 9 |
+
def __init__(
|
| 10 |
+
self,
|
| 11 |
+
whisper_config: Optional[Union[Dict[str, Any], WhisperConfig]] = None,
|
| 12 |
+
adapter_type: str = "mlp",
|
| 13 |
+
merge_factor: int = 4,
|
| 14 |
+
spec_aug: bool = False,
|
| 15 |
+
use_rope: bool = True,
|
| 16 |
+
max_whisper_length: int = 1500,
|
| 17 |
+
mlp_adapter_act: str = "gelu",
|
| 18 |
+
**kwargs, # 👈 所有 Qwen2Config 的参数都从这里进来
|
| 19 |
+
):
|
| 20 |
+
# === 1️⃣ 关键点：初始化 Qwen2Config（LM 部分）===
|
| 21 |
+
# 这里会吃掉：
|
| 22 |
+
# vocab_size / hidden_size / num_hidden_layers / rope_scaling / ...
|
| 23 |
+
super().__init__(**kwargs)
|
| 24 |
+
|
| 25 |
+
# === 2️⃣ Whisper 子配置 ===
|
| 26 |
+
if isinstance(whisper_config, dict):
|
| 27 |
+
self.whisper_config = WhisperConfig(**whisper_config)
|
| 28 |
+
elif isinstance(whisper_config, WhisperConfig):
|
| 29 |
+
self.whisper_config = whisper_config
|
| 30 |
+
else:
|
| 31 |
+
self.whisper_config = WhisperConfig()
|
| 32 |
+
|
| 33 |
+
# === 3️⃣ ArkASR 自己的参数 ===
|
| 34 |
+
self.adapter_type = adapter_type
|
| 35 |
+
self.merge_factor = int(merge_factor)
|
| 36 |
+
self.spec_aug = bool(spec_aug)
|
| 37 |
+
self.use_rope = bool(use_rope)
|
| 38 |
+
self.max_whisper_length = int(max_whisper_length)
|
| 39 |
+
self.mlp_adapter_act = mlp_adapter_act
|
| 40 |
+
|
| 41 |
+
def to_dict(self):
|
| 42 |
+
output = super().to_dict()
|
| 43 |
+
output["whisper_config"] = self.whisper_config.to_dict()
|
| 44 |
+
return output
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
__all__ = ["ArkasrConfig"]
|
figures/ark_asr_architecture.png
ADDED
|
Git LFS Details
|
generation_config.json
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"bos_token_id": 151643,
|
| 4 |
+
"eos_token_id": 151645,
|
| 5 |
+
"pad_token_id": 151643,
|
| 6 |
+
"transformers_version": "4.57.6"
|
| 7 |
+
}
|
merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
model-00001-of-00002.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b0f93d36c42c3487d95532c7d54952c7810bf629ee2a2265ed4e2f9e988e5342
|
| 3 |
+
size 4996098432
|
model-00002-of-00002.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:db6d540d2b230e322efc17176d7f67ae44a25705bd85c5caccdf5d8c426c072c
|
| 3 |
+
size 3130890232
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,934 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_parameters": 4063438848,
|
| 4 |
+
"total_size": 8126877696
|
| 5 |
+
},
|
| 6 |
+
"weight_map": {
|
| 7 |
+
"audio_encoder.adapting.0.bias": "model-00001-of-00002.safetensors",
|
| 8 |
+
"audio_encoder.adapting.0.weight": "model-00002-of-00002.safetensors",
|
| 9 |
+
"audio_encoder.adapting.2.bias": "model-00001-of-00002.safetensors",
|
| 10 |
+
"audio_encoder.adapting.2.weight": "model-00001-of-00002.safetensors",
|
| 11 |
+
"audio_encoder.layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 12 |
+
"audio_encoder.layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 13 |
+
"audio_encoder.whisper.conv1.bias": "model-00002-of-00002.safetensors",
|
| 14 |
+
"audio_encoder.whisper.conv1.weight": "model-00001-of-00002.safetensors",
|
| 15 |
+
"audio_encoder.whisper.conv2.bias": "model-00002-of-00002.safetensors",
|
| 16 |
+
"audio_encoder.whisper.conv2.weight": "model-00002-of-00002.safetensors",
|
| 17 |
+
"audio_encoder.whisper.embed_positions.weight": "model-00001-of-00002.safetensors",
|
| 18 |
+
"audio_encoder.whisper.layers.0.fc1.bias": "model-00001-of-00002.safetensors",
|
| 19 |
+
"audio_encoder.whisper.layers.0.fc1.weight": "model-00001-of-00002.safetensors",
|
| 20 |
+
"audio_encoder.whisper.layers.0.fc2.bias": "model-00001-of-00002.safetensors",
|
| 21 |
+
"audio_encoder.whisper.layers.0.fc2.weight": "model-00002-of-00002.safetensors",
|
| 22 |
+
"audio_encoder.whisper.layers.0.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 23 |
+
"audio_encoder.whisper.layers.0.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 24 |
+
"audio_encoder.whisper.layers.0.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 25 |
+
"audio_encoder.whisper.layers.0.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 26 |
+
"audio_encoder.whisper.layers.0.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 27 |
+
"audio_encoder.whisper.layers.0.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 28 |
+
"audio_encoder.whisper.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 29 |
+
"audio_encoder.whisper.layers.0.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 30 |
+
"audio_encoder.whisper.layers.0.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 31 |
+
"audio_encoder.whisper.layers.0.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 32 |
+
"audio_encoder.whisper.layers.0.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 33 |
+
"audio_encoder.whisper.layers.1.fc1.bias": "model-00001-of-00002.safetensors",
|
| 34 |
+
"audio_encoder.whisper.layers.1.fc1.weight": "model-00002-of-00002.safetensors",
|
| 35 |
+
"audio_encoder.whisper.layers.1.fc2.bias": "model-00002-of-00002.safetensors",
|
| 36 |
+
"audio_encoder.whisper.layers.1.fc2.weight": "model-00002-of-00002.safetensors",
|
| 37 |
+
"audio_encoder.whisper.layers.1.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 38 |
+
"audio_encoder.whisper.layers.1.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 39 |
+
"audio_encoder.whisper.layers.1.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 40 |
+
"audio_encoder.whisper.layers.1.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 41 |
+
"audio_encoder.whisper.layers.1.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 42 |
+
"audio_encoder.whisper.layers.1.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 43 |
+
"audio_encoder.whisper.layers.1.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 44 |
+
"audio_encoder.whisper.layers.1.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 45 |
+
"audio_encoder.whisper.layers.1.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 46 |
+
"audio_encoder.whisper.layers.1.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 47 |
+
"audio_encoder.whisper.layers.1.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 48 |
+
"audio_encoder.whisper.layers.10.fc1.bias": "model-00002-of-00002.safetensors",
|
| 49 |
+
"audio_encoder.whisper.layers.10.fc1.weight": "model-00001-of-00002.safetensors",
|
| 50 |
+
"audio_encoder.whisper.layers.10.fc2.bias": "model-00002-of-00002.safetensors",
|
| 51 |
+
"audio_encoder.whisper.layers.10.fc2.weight": "model-00001-of-00002.safetensors",
|
| 52 |
+
"audio_encoder.whisper.layers.10.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 53 |
+
"audio_encoder.whisper.layers.10.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 54 |
+
"audio_encoder.whisper.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 55 |
+
"audio_encoder.whisper.layers.10.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 56 |
+
"audio_encoder.whisper.layers.10.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 57 |
+
"audio_encoder.whisper.layers.10.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 58 |
+
"audio_encoder.whisper.layers.10.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 59 |
+
"audio_encoder.whisper.layers.10.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 60 |
+
"audio_encoder.whisper.layers.10.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 61 |
+
"audio_encoder.whisper.layers.10.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 62 |
+
"audio_encoder.whisper.layers.10.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 63 |
+
"audio_encoder.whisper.layers.11.fc1.bias": "model-00001-of-00002.safetensors",
|
| 64 |
+
"audio_encoder.whisper.layers.11.fc1.weight": "model-00001-of-00002.safetensors",
|
| 65 |
+
"audio_encoder.whisper.layers.11.fc2.bias": "model-00002-of-00002.safetensors",
|
| 66 |
+
"audio_encoder.whisper.layers.11.fc2.weight": "model-00002-of-00002.safetensors",
|
| 67 |
+
"audio_encoder.whisper.layers.11.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 68 |
+
"audio_encoder.whisper.layers.11.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 69 |
+
"audio_encoder.whisper.layers.11.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 70 |
+
"audio_encoder.whisper.layers.11.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 71 |
+
"audio_encoder.whisper.layers.11.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 72 |
+
"audio_encoder.whisper.layers.11.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 73 |
+
"audio_encoder.whisper.layers.11.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 74 |
+
"audio_encoder.whisper.layers.11.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 75 |
+
"audio_encoder.whisper.layers.11.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 76 |
+
"audio_encoder.whisper.layers.11.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 77 |
+
"audio_encoder.whisper.layers.11.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 78 |
+
"audio_encoder.whisper.layers.12.fc1.bias": "model-00002-of-00002.safetensors",
|
| 79 |
+
"audio_encoder.whisper.layers.12.fc1.weight": "model-00001-of-00002.safetensors",
|
| 80 |
+
"audio_encoder.whisper.layers.12.fc2.bias": "model-00002-of-00002.safetensors",
|
| 81 |
+
"audio_encoder.whisper.layers.12.fc2.weight": "model-00001-of-00002.safetensors",
|
| 82 |
+
"audio_encoder.whisper.layers.12.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 83 |
+
"audio_encoder.whisper.layers.12.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 84 |
+
"audio_encoder.whisper.layers.12.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 85 |
+
"audio_encoder.whisper.layers.12.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 86 |
+
"audio_encoder.whisper.layers.12.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 87 |
+
"audio_encoder.whisper.layers.12.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 88 |
+
"audio_encoder.whisper.layers.12.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 89 |
+
"audio_encoder.whisper.layers.12.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 90 |
+
"audio_encoder.whisper.layers.12.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 91 |
+
"audio_encoder.whisper.layers.12.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 92 |
+
"audio_encoder.whisper.layers.12.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 93 |
+
"audio_encoder.whisper.layers.13.fc1.bias": "model-00001-of-00002.safetensors",
|
| 94 |
+
"audio_encoder.whisper.layers.13.fc1.weight": "model-00002-of-00002.safetensors",
|
| 95 |
+
"audio_encoder.whisper.layers.13.fc2.bias": "model-00001-of-00002.safetensors",
|
| 96 |
+
"audio_encoder.whisper.layers.13.fc2.weight": "model-00001-of-00002.safetensors",
|
| 97 |
+
"audio_encoder.whisper.layers.13.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 98 |
+
"audio_encoder.whisper.layers.13.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 99 |
+
"audio_encoder.whisper.layers.13.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 100 |
+
"audio_encoder.whisper.layers.13.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 101 |
+
"audio_encoder.whisper.layers.13.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 102 |
+
"audio_encoder.whisper.layers.13.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 103 |
+
"audio_encoder.whisper.layers.13.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 104 |
+
"audio_encoder.whisper.layers.13.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 105 |
+
"audio_encoder.whisper.layers.13.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 106 |
+
"audio_encoder.whisper.layers.13.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 107 |
+
"audio_encoder.whisper.layers.13.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 108 |
+
"audio_encoder.whisper.layers.14.fc1.bias": "model-00002-of-00002.safetensors",
|
| 109 |
+
"audio_encoder.whisper.layers.14.fc1.weight": "model-00002-of-00002.safetensors",
|
| 110 |
+
"audio_encoder.whisper.layers.14.fc2.bias": "model-00001-of-00002.safetensors",
|
| 111 |
+
"audio_encoder.whisper.layers.14.fc2.weight": "model-00001-of-00002.safetensors",
|
| 112 |
+
"audio_encoder.whisper.layers.14.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 113 |
+
"audio_encoder.whisper.layers.14.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 114 |
+
"audio_encoder.whisper.layers.14.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 115 |
+
"audio_encoder.whisper.layers.14.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 116 |
+
"audio_encoder.whisper.layers.14.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 117 |
+
"audio_encoder.whisper.layers.14.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 118 |
+
"audio_encoder.whisper.layers.14.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 119 |
+
"audio_encoder.whisper.layers.14.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 120 |
+
"audio_encoder.whisper.layers.14.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 121 |
+
"audio_encoder.whisper.layers.14.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 122 |
+
"audio_encoder.whisper.layers.14.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 123 |
+
"audio_encoder.whisper.layers.15.fc1.bias": "model-00002-of-00002.safetensors",
|
| 124 |
+
"audio_encoder.whisper.layers.15.fc1.weight": "model-00001-of-00002.safetensors",
|
| 125 |
+
"audio_encoder.whisper.layers.15.fc2.bias": "model-00001-of-00002.safetensors",
|
| 126 |
+
"audio_encoder.whisper.layers.15.fc2.weight": "model-00001-of-00002.safetensors",
|
| 127 |
+
"audio_encoder.whisper.layers.15.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 128 |
+
"audio_encoder.whisper.layers.15.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 129 |
+
"audio_encoder.whisper.layers.15.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 130 |
+
"audio_encoder.whisper.layers.15.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 131 |
+
"audio_encoder.whisper.layers.15.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 132 |
+
"audio_encoder.whisper.layers.15.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 133 |
+
"audio_encoder.whisper.layers.15.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 134 |
+
"audio_encoder.whisper.layers.15.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 135 |
+
"audio_encoder.whisper.layers.15.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 136 |
+
"audio_encoder.whisper.layers.15.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 137 |
+
"audio_encoder.whisper.layers.15.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 138 |
+
"audio_encoder.whisper.layers.16.fc1.bias": "model-00001-of-00002.safetensors",
|
| 139 |
+
"audio_encoder.whisper.layers.16.fc1.weight": "model-00001-of-00002.safetensors",
|
| 140 |
+
"audio_encoder.whisper.layers.16.fc2.bias": "model-00001-of-00002.safetensors",
|
| 141 |
+
"audio_encoder.whisper.layers.16.fc2.weight": "model-00001-of-00002.safetensors",
|
| 142 |
+
"audio_encoder.whisper.layers.16.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 143 |
+
"audio_encoder.whisper.layers.16.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 144 |
+
"audio_encoder.whisper.layers.16.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 145 |
+
"audio_encoder.whisper.layers.16.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 146 |
+
"audio_encoder.whisper.layers.16.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 147 |
+
"audio_encoder.whisper.layers.16.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 148 |
+
"audio_encoder.whisper.layers.16.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 149 |
+
"audio_encoder.whisper.layers.16.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 150 |
+
"audio_encoder.whisper.layers.16.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 151 |
+
"audio_encoder.whisper.layers.16.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 152 |
+
"audio_encoder.whisper.layers.16.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 153 |
+
"audio_encoder.whisper.layers.17.fc1.bias": "model-00002-of-00002.safetensors",
|
| 154 |
+
"audio_encoder.whisper.layers.17.fc1.weight": "model-00002-of-00002.safetensors",
|
| 155 |
+
"audio_encoder.whisper.layers.17.fc2.bias": "model-00002-of-00002.safetensors",
|
| 156 |
+
"audio_encoder.whisper.layers.17.fc2.weight": "model-00001-of-00002.safetensors",
|
| 157 |
+
"audio_encoder.whisper.layers.17.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 158 |
+
"audio_encoder.whisper.layers.17.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 159 |
+
"audio_encoder.whisper.layers.17.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 160 |
+
"audio_encoder.whisper.layers.17.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 161 |
+
"audio_encoder.whisper.layers.17.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 162 |
+
"audio_encoder.whisper.layers.17.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 163 |
+
"audio_encoder.whisper.layers.17.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 164 |
+
"audio_encoder.whisper.layers.17.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 165 |
+
"audio_encoder.whisper.layers.17.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 166 |
+
"audio_encoder.whisper.layers.17.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 167 |
+
"audio_encoder.whisper.layers.17.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 168 |
+
"audio_encoder.whisper.layers.18.fc1.bias": "model-00001-of-00002.safetensors",
|
| 169 |
+
"audio_encoder.whisper.layers.18.fc1.weight": "model-00001-of-00002.safetensors",
|
| 170 |
+
"audio_encoder.whisper.layers.18.fc2.bias": "model-00001-of-00002.safetensors",
|
| 171 |
+
"audio_encoder.whisper.layers.18.fc2.weight": "model-00001-of-00002.safetensors",
|
| 172 |
+
"audio_encoder.whisper.layers.18.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 173 |
+
"audio_encoder.whisper.layers.18.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 174 |
+
"audio_encoder.whisper.layers.18.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 175 |
+
"audio_encoder.whisper.layers.18.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 176 |
+
"audio_encoder.whisper.layers.18.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 177 |
+
"audio_encoder.whisper.layers.18.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 178 |
+
"audio_encoder.whisper.layers.18.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 179 |
+
"audio_encoder.whisper.layers.18.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 180 |
+
"audio_encoder.whisper.layers.18.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 181 |
+
"audio_encoder.whisper.layers.18.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 182 |
+
"audio_encoder.whisper.layers.18.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 183 |
+
"audio_encoder.whisper.layers.19.fc1.bias": "model-00002-of-00002.safetensors",
|
| 184 |
+
"audio_encoder.whisper.layers.19.fc1.weight": "model-00002-of-00002.safetensors",
|
| 185 |
+
"audio_encoder.whisper.layers.19.fc2.bias": "model-00002-of-00002.safetensors",
|
| 186 |
+
"audio_encoder.whisper.layers.19.fc2.weight": "model-00001-of-00002.safetensors",
|
| 187 |
+
"audio_encoder.whisper.layers.19.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 188 |
+
"audio_encoder.whisper.layers.19.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 189 |
+
"audio_encoder.whisper.layers.19.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 190 |
+
"audio_encoder.whisper.layers.19.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 191 |
+
"audio_encoder.whisper.layers.19.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 192 |
+
"audio_encoder.whisper.layers.19.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 193 |
+
"audio_encoder.whisper.layers.19.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 194 |
+
"audio_encoder.whisper.layers.19.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 195 |
+
"audio_encoder.whisper.layers.19.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 196 |
+
"audio_encoder.whisper.layers.19.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 197 |
+
"audio_encoder.whisper.layers.19.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 198 |
+
"audio_encoder.whisper.layers.2.fc1.bias": "model-00002-of-00002.safetensors",
|
| 199 |
+
"audio_encoder.whisper.layers.2.fc1.weight": "model-00002-of-00002.safetensors",
|
| 200 |
+
"audio_encoder.whisper.layers.2.fc2.bias": "model-00001-of-00002.safetensors",
|
| 201 |
+
"audio_encoder.whisper.layers.2.fc2.weight": "model-00002-of-00002.safetensors",
|
| 202 |
+
"audio_encoder.whisper.layers.2.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 203 |
+
"audio_encoder.whisper.layers.2.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 204 |
+
"audio_encoder.whisper.layers.2.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 205 |
+
"audio_encoder.whisper.layers.2.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 206 |
+
"audio_encoder.whisper.layers.2.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 207 |
+
"audio_encoder.whisper.layers.2.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 208 |
+
"audio_encoder.whisper.layers.2.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 209 |
+
"audio_encoder.whisper.layers.2.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 210 |
+
"audio_encoder.whisper.layers.2.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 211 |
+
"audio_encoder.whisper.layers.2.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 212 |
+
"audio_encoder.whisper.layers.2.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 213 |
+
"audio_encoder.whisper.layers.20.fc1.bias": "model-00002-of-00002.safetensors",
|
| 214 |
+
"audio_encoder.whisper.layers.20.fc1.weight": "model-00002-of-00002.safetensors",
|
| 215 |
+
"audio_encoder.whisper.layers.20.fc2.bias": "model-00002-of-00002.safetensors",
|
| 216 |
+
"audio_encoder.whisper.layers.20.fc2.weight": "model-00002-of-00002.safetensors",
|
| 217 |
+
"audio_encoder.whisper.layers.20.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 218 |
+
"audio_encoder.whisper.layers.20.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 219 |
+
"audio_encoder.whisper.layers.20.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 220 |
+
"audio_encoder.whisper.layers.20.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 221 |
+
"audio_encoder.whisper.layers.20.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 222 |
+
"audio_encoder.whisper.layers.20.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 223 |
+
"audio_encoder.whisper.layers.20.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 224 |
+
"audio_encoder.whisper.layers.20.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 225 |
+
"audio_encoder.whisper.layers.20.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 226 |
+
"audio_encoder.whisper.layers.20.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 227 |
+
"audio_encoder.whisper.layers.20.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 228 |
+
"audio_encoder.whisper.layers.21.fc1.bias": "model-00001-of-00002.safetensors",
|
| 229 |
+
"audio_encoder.whisper.layers.21.fc1.weight": "model-00001-of-00002.safetensors",
|
| 230 |
+
"audio_encoder.whisper.layers.21.fc2.bias": "model-00002-of-00002.safetensors",
|
| 231 |
+
"audio_encoder.whisper.layers.21.fc2.weight": "model-00001-of-00002.safetensors",
|
| 232 |
+
"audio_encoder.whisper.layers.21.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 233 |
+
"audio_encoder.whisper.layers.21.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 234 |
+
"audio_encoder.whisper.layers.21.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 235 |
+
"audio_encoder.whisper.layers.21.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 236 |
+
"audio_encoder.whisper.layers.21.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 237 |
+
"audio_encoder.whisper.layers.21.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 238 |
+
"audio_encoder.whisper.layers.21.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 239 |
+
"audio_encoder.whisper.layers.21.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 240 |
+
"audio_encoder.whisper.layers.21.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 241 |
+
"audio_encoder.whisper.layers.21.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 242 |
+
"audio_encoder.whisper.layers.21.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 243 |
+
"audio_encoder.whisper.layers.22.fc1.bias": "model-00002-of-00002.safetensors",
|
| 244 |
+
"audio_encoder.whisper.layers.22.fc1.weight": "model-00001-of-00002.safetensors",
|
| 245 |
+
"audio_encoder.whisper.layers.22.fc2.bias": "model-00001-of-00002.safetensors",
|
| 246 |
+
"audio_encoder.whisper.layers.22.fc2.weight": "model-00002-of-00002.safetensors",
|
| 247 |
+
"audio_encoder.whisper.layers.22.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 248 |
+
"audio_encoder.whisper.layers.22.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 249 |
+
"audio_encoder.whisper.layers.22.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 250 |
+
"audio_encoder.whisper.layers.22.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 251 |
+
"audio_encoder.whisper.layers.22.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 252 |
+
"audio_encoder.whisper.layers.22.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 253 |
+
"audio_encoder.whisper.layers.22.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 254 |
+
"audio_encoder.whisper.layers.22.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 255 |
+
"audio_encoder.whisper.layers.22.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 256 |
+
"audio_encoder.whisper.layers.22.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 257 |
+
"audio_encoder.whisper.layers.22.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 258 |
+
"audio_encoder.whisper.layers.23.fc1.bias": "model-00002-of-00002.safetensors",
|
| 259 |
+
"audio_encoder.whisper.layers.23.fc1.weight": "model-00001-of-00002.safetensors",
|
| 260 |
+
"audio_encoder.whisper.layers.23.fc2.bias": "model-00001-of-00002.safetensors",
|
| 261 |
+
"audio_encoder.whisper.layers.23.fc2.weight": "model-00001-of-00002.safetensors",
|
| 262 |
+
"audio_encoder.whisper.layers.23.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 263 |
+
"audio_encoder.whisper.layers.23.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 264 |
+
"audio_encoder.whisper.layers.23.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 265 |
+
"audio_encoder.whisper.layers.23.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 266 |
+
"audio_encoder.whisper.layers.23.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 267 |
+
"audio_encoder.whisper.layers.23.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 268 |
+
"audio_encoder.whisper.layers.23.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 269 |
+
"audio_encoder.whisper.layers.23.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 270 |
+
"audio_encoder.whisper.layers.23.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 271 |
+
"audio_encoder.whisper.layers.23.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 272 |
+
"audio_encoder.whisper.layers.23.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 273 |
+
"audio_encoder.whisper.layers.24.fc1.bias": "model-00002-of-00002.safetensors",
|
| 274 |
+
"audio_encoder.whisper.layers.24.fc1.weight": "model-00001-of-00002.safetensors",
|
| 275 |
+
"audio_encoder.whisper.layers.24.fc2.bias": "model-00001-of-00002.safetensors",
|
| 276 |
+
"audio_encoder.whisper.layers.24.fc2.weight": "model-00001-of-00002.safetensors",
|
| 277 |
+
"audio_encoder.whisper.layers.24.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 278 |
+
"audio_encoder.whisper.layers.24.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 279 |
+
"audio_encoder.whisper.layers.24.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 280 |
+
"audio_encoder.whisper.layers.24.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 281 |
+
"audio_encoder.whisper.layers.24.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 282 |
+
"audio_encoder.whisper.layers.24.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 283 |
+
"audio_encoder.whisper.layers.24.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 284 |
+
"audio_encoder.whisper.layers.24.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 285 |
+
"audio_encoder.whisper.layers.24.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 286 |
+
"audio_encoder.whisper.layers.24.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 287 |
+
"audio_encoder.whisper.layers.24.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 288 |
+
"audio_encoder.whisper.layers.25.fc1.bias": "model-00002-of-00002.safetensors",
|
| 289 |
+
"audio_encoder.whisper.layers.25.fc1.weight": "model-00002-of-00002.safetensors",
|
| 290 |
+
"audio_encoder.whisper.layers.25.fc2.bias": "model-00001-of-00002.safetensors",
|
| 291 |
+
"audio_encoder.whisper.layers.25.fc2.weight": "model-00001-of-00002.safetensors",
|
| 292 |
+
"audio_encoder.whisper.layers.25.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 293 |
+
"audio_encoder.whisper.layers.25.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 294 |
+
"audio_encoder.whisper.layers.25.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 295 |
+
"audio_encoder.whisper.layers.25.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 296 |
+
"audio_encoder.whisper.layers.25.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 297 |
+
"audio_encoder.whisper.layers.25.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 298 |
+
"audio_encoder.whisper.layers.25.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 299 |
+
"audio_encoder.whisper.layers.25.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 300 |
+
"audio_encoder.whisper.layers.25.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 301 |
+
"audio_encoder.whisper.layers.25.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 302 |
+
"audio_encoder.whisper.layers.25.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 303 |
+
"audio_encoder.whisper.layers.26.fc1.bias": "model-00001-of-00002.safetensors",
|
| 304 |
+
"audio_encoder.whisper.layers.26.fc1.weight": "model-00001-of-00002.safetensors",
|
| 305 |
+
"audio_encoder.whisper.layers.26.fc2.bias": "model-00001-of-00002.safetensors",
|
| 306 |
+
"audio_encoder.whisper.layers.26.fc2.weight": "model-00001-of-00002.safetensors",
|
| 307 |
+
"audio_encoder.whisper.layers.26.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 308 |
+
"audio_encoder.whisper.layers.26.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 309 |
+
"audio_encoder.whisper.layers.26.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 310 |
+
"audio_encoder.whisper.layers.26.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 311 |
+
"audio_encoder.whisper.layers.26.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 312 |
+
"audio_encoder.whisper.layers.26.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 313 |
+
"audio_encoder.whisper.layers.26.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 314 |
+
"audio_encoder.whisper.layers.26.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 315 |
+
"audio_encoder.whisper.layers.26.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 316 |
+
"audio_encoder.whisper.layers.26.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 317 |
+
"audio_encoder.whisper.layers.26.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 318 |
+
"audio_encoder.whisper.layers.27.fc1.bias": "model-00002-of-00002.safetensors",
|
| 319 |
+
"audio_encoder.whisper.layers.27.fc1.weight": "model-00002-of-00002.safetensors",
|
| 320 |
+
"audio_encoder.whisper.layers.27.fc2.bias": "model-00001-of-00002.safetensors",
|
| 321 |
+
"audio_encoder.whisper.layers.27.fc2.weight": "model-00002-of-00002.safetensors",
|
| 322 |
+
"audio_encoder.whisper.layers.27.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 323 |
+
"audio_encoder.whisper.layers.27.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 324 |
+
"audio_encoder.whisper.layers.27.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 325 |
+
"audio_encoder.whisper.layers.27.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 326 |
+
"audio_encoder.whisper.layers.27.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 327 |
+
"audio_encoder.whisper.layers.27.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 328 |
+
"audio_encoder.whisper.layers.27.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 329 |
+
"audio_encoder.whisper.layers.27.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 330 |
+
"audio_encoder.whisper.layers.27.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 331 |
+
"audio_encoder.whisper.layers.27.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 332 |
+
"audio_encoder.whisper.layers.27.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 333 |
+
"audio_encoder.whisper.layers.28.fc1.bias": "model-00001-of-00002.safetensors",
|
| 334 |
+
"audio_encoder.whisper.layers.28.fc1.weight": "model-00001-of-00002.safetensors",
|
| 335 |
+
"audio_encoder.whisper.layers.28.fc2.bias": "model-00001-of-00002.safetensors",
|
| 336 |
+
"audio_encoder.whisper.layers.28.fc2.weight": "model-00002-of-00002.safetensors",
|
| 337 |
+
"audio_encoder.whisper.layers.28.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 338 |
+
"audio_encoder.whisper.layers.28.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 339 |
+
"audio_encoder.whisper.layers.28.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 340 |
+
"audio_encoder.whisper.layers.28.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 341 |
+
"audio_encoder.whisper.layers.28.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 342 |
+
"audio_encoder.whisper.layers.28.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 343 |
+
"audio_encoder.whisper.layers.28.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 344 |
+
"audio_encoder.whisper.layers.28.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 345 |
+
"audio_encoder.whisper.layers.28.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 346 |
+
"audio_encoder.whisper.layers.28.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 347 |
+
"audio_encoder.whisper.layers.28.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 348 |
+
"audio_encoder.whisper.layers.29.fc1.bias": "model-00001-of-00002.safetensors",
|
| 349 |
+
"audio_encoder.whisper.layers.29.fc1.weight": "model-00002-of-00002.safetensors",
|
| 350 |
+
"audio_encoder.whisper.layers.29.fc2.bias": "model-00001-of-00002.safetensors",
|
| 351 |
+
"audio_encoder.whisper.layers.29.fc2.weight": "model-00002-of-00002.safetensors",
|
| 352 |
+
"audio_encoder.whisper.layers.29.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 353 |
+
"audio_encoder.whisper.layers.29.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 354 |
+
"audio_encoder.whisper.layers.29.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 355 |
+
"audio_encoder.whisper.layers.29.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 356 |
+
"audio_encoder.whisper.layers.29.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 357 |
+
"audio_encoder.whisper.layers.29.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 358 |
+
"audio_encoder.whisper.layers.29.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 359 |
+
"audio_encoder.whisper.layers.29.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 360 |
+
"audio_encoder.whisper.layers.29.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 361 |
+
"audio_encoder.whisper.layers.29.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 362 |
+
"audio_encoder.whisper.layers.29.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 363 |
+
"audio_encoder.whisper.layers.3.fc1.bias": "model-00002-of-00002.safetensors",
|
| 364 |
+
"audio_encoder.whisper.layers.3.fc1.weight": "model-00002-of-00002.safetensors",
|
| 365 |
+
"audio_encoder.whisper.layers.3.fc2.bias": "model-00001-of-00002.safetensors",
|
| 366 |
+
"audio_encoder.whisper.layers.3.fc2.weight": "model-00001-of-00002.safetensors",
|
| 367 |
+
"audio_encoder.whisper.layers.3.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 368 |
+
"audio_encoder.whisper.layers.3.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 369 |
+
"audio_encoder.whisper.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 370 |
+
"audio_encoder.whisper.layers.3.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 371 |
+
"audio_encoder.whisper.layers.3.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 372 |
+
"audio_encoder.whisper.layers.3.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 373 |
+
"audio_encoder.whisper.layers.3.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 374 |
+
"audio_encoder.whisper.layers.3.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 375 |
+
"audio_encoder.whisper.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 376 |
+
"audio_encoder.whisper.layers.3.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 377 |
+
"audio_encoder.whisper.layers.3.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 378 |
+
"audio_encoder.whisper.layers.30.fc1.bias": "model-00002-of-00002.safetensors",
|
| 379 |
+
"audio_encoder.whisper.layers.30.fc1.weight": "model-00001-of-00002.safetensors",
|
| 380 |
+
"audio_encoder.whisper.layers.30.fc2.bias": "model-00002-of-00002.safetensors",
|
| 381 |
+
"audio_encoder.whisper.layers.30.fc2.weight": "model-00002-of-00002.safetensors",
|
| 382 |
+
"audio_encoder.whisper.layers.30.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 383 |
+
"audio_encoder.whisper.layers.30.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 384 |
+
"audio_encoder.whisper.layers.30.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 385 |
+
"audio_encoder.whisper.layers.30.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 386 |
+
"audio_encoder.whisper.layers.30.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 387 |
+
"audio_encoder.whisper.layers.30.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 388 |
+
"audio_encoder.whisper.layers.30.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 389 |
+
"audio_encoder.whisper.layers.30.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 390 |
+
"audio_encoder.whisper.layers.30.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 391 |
+
"audio_encoder.whisper.layers.30.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 392 |
+
"audio_encoder.whisper.layers.30.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 393 |
+
"audio_encoder.whisper.layers.31.fc1.bias": "model-00002-of-00002.safetensors",
|
| 394 |
+
"audio_encoder.whisper.layers.31.fc1.weight": "model-00001-of-00002.safetensors",
|
| 395 |
+
"audio_encoder.whisper.layers.31.fc2.bias": "model-00002-of-00002.safetensors",
|
| 396 |
+
"audio_encoder.whisper.layers.31.fc2.weight": "model-00001-of-00002.safetensors",
|
| 397 |
+
"audio_encoder.whisper.layers.31.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 398 |
+
"audio_encoder.whisper.layers.31.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 399 |
+
"audio_encoder.whisper.layers.31.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 400 |
+
"audio_encoder.whisper.layers.31.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 401 |
+
"audio_encoder.whisper.layers.31.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 402 |
+
"audio_encoder.whisper.layers.31.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 403 |
+
"audio_encoder.whisper.layers.31.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 404 |
+
"audio_encoder.whisper.layers.31.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 405 |
+
"audio_encoder.whisper.layers.31.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 406 |
+
"audio_encoder.whisper.layers.31.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 407 |
+
"audio_encoder.whisper.layers.31.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 408 |
+
"audio_encoder.whisper.layers.4.fc1.bias": "model-00001-of-00002.safetensors",
|
| 409 |
+
"audio_encoder.whisper.layers.4.fc1.weight": "model-00001-of-00002.safetensors",
|
| 410 |
+
"audio_encoder.whisper.layers.4.fc2.bias": "model-00001-of-00002.safetensors",
|
| 411 |
+
"audio_encoder.whisper.layers.4.fc2.weight": "model-00002-of-00002.safetensors",
|
| 412 |
+
"audio_encoder.whisper.layers.4.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 413 |
+
"audio_encoder.whisper.layers.4.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 414 |
+
"audio_encoder.whisper.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 415 |
+
"audio_encoder.whisper.layers.4.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 416 |
+
"audio_encoder.whisper.layers.4.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 417 |
+
"audio_encoder.whisper.layers.4.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 418 |
+
"audio_encoder.whisper.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 419 |
+
"audio_encoder.whisper.layers.4.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 420 |
+
"audio_encoder.whisper.layers.4.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 421 |
+
"audio_encoder.whisper.layers.4.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 422 |
+
"audio_encoder.whisper.layers.4.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 423 |
+
"audio_encoder.whisper.layers.5.fc1.bias": "model-00002-of-00002.safetensors",
|
| 424 |
+
"audio_encoder.whisper.layers.5.fc1.weight": "model-00002-of-00002.safetensors",
|
| 425 |
+
"audio_encoder.whisper.layers.5.fc2.bias": "model-00001-of-00002.safetensors",
|
| 426 |
+
"audio_encoder.whisper.layers.5.fc2.weight": "model-00002-of-00002.safetensors",
|
| 427 |
+
"audio_encoder.whisper.layers.5.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 428 |
+
"audio_encoder.whisper.layers.5.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 429 |
+
"audio_encoder.whisper.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 430 |
+
"audio_encoder.whisper.layers.5.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 431 |
+
"audio_encoder.whisper.layers.5.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 432 |
+
"audio_encoder.whisper.layers.5.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 433 |
+
"audio_encoder.whisper.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 434 |
+
"audio_encoder.whisper.layers.5.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 435 |
+
"audio_encoder.whisper.layers.5.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 436 |
+
"audio_encoder.whisper.layers.5.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 437 |
+
"audio_encoder.whisper.layers.5.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 438 |
+
"audio_encoder.whisper.layers.6.fc1.bias": "model-00002-of-00002.safetensors",
|
| 439 |
+
"audio_encoder.whisper.layers.6.fc1.weight": "model-00001-of-00002.safetensors",
|
| 440 |
+
"audio_encoder.whisper.layers.6.fc2.bias": "model-00001-of-00002.safetensors",
|
| 441 |
+
"audio_encoder.whisper.layers.6.fc2.weight": "model-00002-of-00002.safetensors",
|
| 442 |
+
"audio_encoder.whisper.layers.6.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 443 |
+
"audio_encoder.whisper.layers.6.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 444 |
+
"audio_encoder.whisper.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 445 |
+
"audio_encoder.whisper.layers.6.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 446 |
+
"audio_encoder.whisper.layers.6.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 447 |
+
"audio_encoder.whisper.layers.6.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 448 |
+
"audio_encoder.whisper.layers.6.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 449 |
+
"audio_encoder.whisper.layers.6.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 450 |
+
"audio_encoder.whisper.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 451 |
+
"audio_encoder.whisper.layers.6.self_attn_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 452 |
+
"audio_encoder.whisper.layers.6.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 453 |
+
"audio_encoder.whisper.layers.7.fc1.bias": "model-00002-of-00002.safetensors",
|
| 454 |
+
"audio_encoder.whisper.layers.7.fc1.weight": "model-00001-of-00002.safetensors",
|
| 455 |
+
"audio_encoder.whisper.layers.7.fc2.bias": "model-00002-of-00002.safetensors",
|
| 456 |
+
"audio_encoder.whisper.layers.7.fc2.weight": "model-00001-of-00002.safetensors",
|
| 457 |
+
"audio_encoder.whisper.layers.7.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 458 |
+
"audio_encoder.whisper.layers.7.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 459 |
+
"audio_encoder.whisper.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 460 |
+
"audio_encoder.whisper.layers.7.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 461 |
+
"audio_encoder.whisper.layers.7.self_attn.out_proj.weight": "model-00002-of-00002.safetensors",
|
| 462 |
+
"audio_encoder.whisper.layers.7.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 463 |
+
"audio_encoder.whisper.layers.7.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 464 |
+
"audio_encoder.whisper.layers.7.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 465 |
+
"audio_encoder.whisper.layers.7.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 466 |
+
"audio_encoder.whisper.layers.7.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 467 |
+
"audio_encoder.whisper.layers.7.self_attn_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 468 |
+
"audio_encoder.whisper.layers.8.fc1.bias": "model-00001-of-00002.safetensors",
|
| 469 |
+
"audio_encoder.whisper.layers.8.fc1.weight": "model-00002-of-00002.safetensors",
|
| 470 |
+
"audio_encoder.whisper.layers.8.fc2.bias": "model-00001-of-00002.safetensors",
|
| 471 |
+
"audio_encoder.whisper.layers.8.fc2.weight": "model-00001-of-00002.safetensors",
|
| 472 |
+
"audio_encoder.whisper.layers.8.final_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 473 |
+
"audio_encoder.whisper.layers.8.final_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 474 |
+
"audio_encoder.whisper.layers.8.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 475 |
+
"audio_encoder.whisper.layers.8.self_attn.out_proj.bias": "model-00002-of-00002.safetensors",
|
| 476 |
+
"audio_encoder.whisper.layers.8.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 477 |
+
"audio_encoder.whisper.layers.8.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 478 |
+
"audio_encoder.whisper.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 479 |
+
"audio_encoder.whisper.layers.8.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 480 |
+
"audio_encoder.whisper.layers.8.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 481 |
+
"audio_encoder.whisper.layers.8.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 482 |
+
"audio_encoder.whisper.layers.8.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 483 |
+
"audio_encoder.whisper.layers.9.fc1.bias": "model-00001-of-00002.safetensors",
|
| 484 |
+
"audio_encoder.whisper.layers.9.fc1.weight": "model-00001-of-00002.safetensors",
|
| 485 |
+
"audio_encoder.whisper.layers.9.fc2.bias": "model-00001-of-00002.safetensors",
|
| 486 |
+
"audio_encoder.whisper.layers.9.fc2.weight": "model-00001-of-00002.safetensors",
|
| 487 |
+
"audio_encoder.whisper.layers.9.final_layer_norm.bias": "model-00001-of-00002.safetensors",
|
| 488 |
+
"audio_encoder.whisper.layers.9.final_layer_norm.weight": "model-00001-of-00002.safetensors",
|
| 489 |
+
"audio_encoder.whisper.layers.9.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 490 |
+
"audio_encoder.whisper.layers.9.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
|
| 491 |
+
"audio_encoder.whisper.layers.9.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
|
| 492 |
+
"audio_encoder.whisper.layers.9.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 493 |
+
"audio_encoder.whisper.layers.9.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 494 |
+
"audio_encoder.whisper.layers.9.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 495 |
+
"audio_encoder.whisper.layers.9.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 496 |
+
"audio_encoder.whisper.layers.9.self_attn_layer_norm.bias": "model-00002-of-00002.safetensors",
|
| 497 |
+
"audio_encoder.whisper.layers.9.self_attn_layer_norm.weight": "model-00002-of-00002.safetensors",
|
| 498 |
+
"lm_head.weight": "model-00001-of-00002.safetensors",
|
| 499 |
+
"model.embed_tokens.weight": "model-00001-of-00002.safetensors",
|
| 500 |
+
"model.layers.0.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 501 |
+
"model.layers.0.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 502 |
+
"model.layers.0.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 503 |
+
"model.layers.0.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 504 |
+
"model.layers.0.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 505 |
+
"model.layers.0.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 506 |
+
"model.layers.0.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 507 |
+
"model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 508 |
+
"model.layers.0.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 509 |
+
"model.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 510 |
+
"model.layers.0.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 511 |
+
"model.layers.0.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 512 |
+
"model.layers.1.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 513 |
+
"model.layers.1.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 514 |
+
"model.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 515 |
+
"model.layers.1.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 516 |
+
"model.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 517 |
+
"model.layers.1.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 518 |
+
"model.layers.1.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 519 |
+
"model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 520 |
+
"model.layers.1.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 521 |
+
"model.layers.1.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 522 |
+
"model.layers.1.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 523 |
+
"model.layers.1.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 524 |
+
"model.layers.10.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 525 |
+
"model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 526 |
+
"model.layers.10.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 527 |
+
"model.layers.10.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 528 |
+
"model.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 529 |
+
"model.layers.10.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 530 |
+
"model.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 531 |
+
"model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 532 |
+
"model.layers.10.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 533 |
+
"model.layers.10.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 534 |
+
"model.layers.10.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 535 |
+
"model.layers.10.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 536 |
+
"model.layers.11.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 537 |
+
"model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 538 |
+
"model.layers.11.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 539 |
+
"model.layers.11.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 540 |
+
"model.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 541 |
+
"model.layers.11.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 542 |
+
"model.layers.11.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 543 |
+
"model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 544 |
+
"model.layers.11.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 545 |
+
"model.layers.11.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 546 |
+
"model.layers.11.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 547 |
+
"model.layers.11.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 548 |
+
"model.layers.12.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 549 |
+
"model.layers.12.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 550 |
+
"model.layers.12.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 551 |
+
"model.layers.12.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 552 |
+
"model.layers.12.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 553 |
+
"model.layers.12.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 554 |
+
"model.layers.12.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 555 |
+
"model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 556 |
+
"model.layers.12.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 557 |
+
"model.layers.12.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 558 |
+
"model.layers.12.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 559 |
+
"model.layers.12.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 560 |
+
"model.layers.13.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 561 |
+
"model.layers.13.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 562 |
+
"model.layers.13.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 563 |
+
"model.layers.13.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 564 |
+
"model.layers.13.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 565 |
+
"model.layers.13.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 566 |
+
"model.layers.13.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 567 |
+
"model.layers.13.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 568 |
+
"model.layers.13.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 569 |
+
"model.layers.13.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 570 |
+
"model.layers.13.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 571 |
+
"model.layers.13.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 572 |
+
"model.layers.14.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 573 |
+
"model.layers.14.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 574 |
+
"model.layers.14.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 575 |
+
"model.layers.14.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 576 |
+
"model.layers.14.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 577 |
+
"model.layers.14.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 578 |
+
"model.layers.14.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 579 |
+
"model.layers.14.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 580 |
+
"model.layers.14.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 581 |
+
"model.layers.14.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 582 |
+
"model.layers.14.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 583 |
+
"model.layers.14.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 584 |
+
"model.layers.15.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 585 |
+
"model.layers.15.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 586 |
+
"model.layers.15.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 587 |
+
"model.layers.15.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 588 |
+
"model.layers.15.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 589 |
+
"model.layers.15.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 590 |
+
"model.layers.15.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 591 |
+
"model.layers.15.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 592 |
+
"model.layers.15.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 593 |
+
"model.layers.15.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 594 |
+
"model.layers.15.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 595 |
+
"model.layers.15.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 596 |
+
"model.layers.16.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 597 |
+
"model.layers.16.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 598 |
+
"model.layers.16.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 599 |
+
"model.layers.16.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 600 |
+
"model.layers.16.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 601 |
+
"model.layers.16.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 602 |
+
"model.layers.16.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 603 |
+
"model.layers.16.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 604 |
+
"model.layers.16.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 605 |
+
"model.layers.16.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 606 |
+
"model.layers.16.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 607 |
+
"model.layers.16.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 608 |
+
"model.layers.17.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 609 |
+
"model.layers.17.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 610 |
+
"model.layers.17.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 611 |
+
"model.layers.17.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 612 |
+
"model.layers.17.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 613 |
+
"model.layers.17.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 614 |
+
"model.layers.17.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 615 |
+
"model.layers.17.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 616 |
+
"model.layers.17.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 617 |
+
"model.layers.17.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 618 |
+
"model.layers.17.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 619 |
+
"model.layers.17.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 620 |
+
"model.layers.18.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 621 |
+
"model.layers.18.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 622 |
+
"model.layers.18.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 623 |
+
"model.layers.18.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 624 |
+
"model.layers.18.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 625 |
+
"model.layers.18.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 626 |
+
"model.layers.18.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 627 |
+
"model.layers.18.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 628 |
+
"model.layers.18.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 629 |
+
"model.layers.18.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 630 |
+
"model.layers.18.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 631 |
+
"model.layers.18.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 632 |
+
"model.layers.19.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 633 |
+
"model.layers.19.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 634 |
+
"model.layers.19.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 635 |
+
"model.layers.19.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 636 |
+
"model.layers.19.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 637 |
+
"model.layers.19.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 638 |
+
"model.layers.19.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 639 |
+
"model.layers.19.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 640 |
+
"model.layers.19.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 641 |
+
"model.layers.19.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 642 |
+
"model.layers.19.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 643 |
+
"model.layers.19.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 644 |
+
"model.layers.2.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 645 |
+
"model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 646 |
+
"model.layers.2.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 647 |
+
"model.layers.2.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 648 |
+
"model.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 649 |
+
"model.layers.2.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 650 |
+
"model.layers.2.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 651 |
+
"model.layers.2.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 652 |
+
"model.layers.2.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 653 |
+
"model.layers.2.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 654 |
+
"model.layers.2.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 655 |
+
"model.layers.2.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 656 |
+
"model.layers.20.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 657 |
+
"model.layers.20.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 658 |
+
"model.layers.20.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 659 |
+
"model.layers.20.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 660 |
+
"model.layers.20.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 661 |
+
"model.layers.20.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 662 |
+
"model.layers.20.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 663 |
+
"model.layers.20.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 664 |
+
"model.layers.20.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 665 |
+
"model.layers.20.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 666 |
+
"model.layers.20.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 667 |
+
"model.layers.20.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 668 |
+
"model.layers.21.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 669 |
+
"model.layers.21.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 670 |
+
"model.layers.21.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 671 |
+
"model.layers.21.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 672 |
+
"model.layers.21.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 673 |
+
"model.layers.21.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 674 |
+
"model.layers.21.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 675 |
+
"model.layers.21.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 676 |
+
"model.layers.21.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 677 |
+
"model.layers.21.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 678 |
+
"model.layers.21.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 679 |
+
"model.layers.21.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 680 |
+
"model.layers.22.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 681 |
+
"model.layers.22.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 682 |
+
"model.layers.22.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 683 |
+
"model.layers.22.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 684 |
+
"model.layers.22.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 685 |
+
"model.layers.22.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 686 |
+
"model.layers.22.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 687 |
+
"model.layers.22.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 688 |
+
"model.layers.22.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 689 |
+
"model.layers.22.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 690 |
+
"model.layers.22.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 691 |
+
"model.layers.22.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 692 |
+
"model.layers.23.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 693 |
+
"model.layers.23.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 694 |
+
"model.layers.23.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 695 |
+
"model.layers.23.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 696 |
+
"model.layers.23.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 697 |
+
"model.layers.23.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 698 |
+
"model.layers.23.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 699 |
+
"model.layers.23.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 700 |
+
"model.layers.23.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 701 |
+
"model.layers.23.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 702 |
+
"model.layers.23.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 703 |
+
"model.layers.23.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 704 |
+
"model.layers.24.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 705 |
+
"model.layers.24.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 706 |
+
"model.layers.24.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 707 |
+
"model.layers.24.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 708 |
+
"model.layers.24.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 709 |
+
"model.layers.24.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 710 |
+
"model.layers.24.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 711 |
+
"model.layers.24.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 712 |
+
"model.layers.24.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 713 |
+
"model.layers.24.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 714 |
+
"model.layers.24.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 715 |
+
"model.layers.24.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 716 |
+
"model.layers.25.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 717 |
+
"model.layers.25.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 718 |
+
"model.layers.25.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 719 |
+
"model.layers.25.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 720 |
+
"model.layers.25.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 721 |
+
"model.layers.25.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 722 |
+
"model.layers.25.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 723 |
+
"model.layers.25.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 724 |
+
"model.layers.25.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 725 |
+
"model.layers.25.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 726 |
+
"model.layers.25.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 727 |
+
"model.layers.25.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 728 |
+
"model.layers.26.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 729 |
+
"model.layers.26.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 730 |
+
"model.layers.26.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 731 |
+
"model.layers.26.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 732 |
+
"model.layers.26.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 733 |
+
"model.layers.26.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 734 |
+
"model.layers.26.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 735 |
+
"model.layers.26.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 736 |
+
"model.layers.26.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 737 |
+
"model.layers.26.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 738 |
+
"model.layers.26.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 739 |
+
"model.layers.26.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 740 |
+
"model.layers.27.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 741 |
+
"model.layers.27.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 742 |
+
"model.layers.27.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 743 |
+
"model.layers.27.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 744 |
+
"model.layers.27.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 745 |
+
"model.layers.27.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 746 |
+
"model.layers.27.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 747 |
+
"model.layers.27.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 748 |
+
"model.layers.27.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 749 |
+
"model.layers.27.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 750 |
+
"model.layers.27.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 751 |
+
"model.layers.27.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 752 |
+
"model.layers.28.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 753 |
+
"model.layers.28.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 754 |
+
"model.layers.28.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 755 |
+
"model.layers.28.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 756 |
+
"model.layers.28.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 757 |
+
"model.layers.28.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 758 |
+
"model.layers.28.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 759 |
+
"model.layers.28.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 760 |
+
"model.layers.28.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 761 |
+
"model.layers.28.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 762 |
+
"model.layers.28.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 763 |
+
"model.layers.28.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 764 |
+
"model.layers.29.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 765 |
+
"model.layers.29.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 766 |
+
"model.layers.29.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 767 |
+
"model.layers.29.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 768 |
+
"model.layers.29.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 769 |
+
"model.layers.29.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 770 |
+
"model.layers.29.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 771 |
+
"model.layers.29.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 772 |
+
"model.layers.29.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 773 |
+
"model.layers.29.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 774 |
+
"model.layers.29.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 775 |
+
"model.layers.29.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 776 |
+
"model.layers.3.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 777 |
+
"model.layers.3.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 778 |
+
"model.layers.3.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 779 |
+
"model.layers.3.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 780 |
+
"model.layers.3.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 781 |
+
"model.layers.3.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 782 |
+
"model.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 783 |
+
"model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 784 |
+
"model.layers.3.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 785 |
+
"model.layers.3.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 786 |
+
"model.layers.3.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 787 |
+
"model.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 788 |
+
"model.layers.30.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 789 |
+
"model.layers.30.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 790 |
+
"model.layers.30.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 791 |
+
"model.layers.30.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 792 |
+
"model.layers.30.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 793 |
+
"model.layers.30.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 794 |
+
"model.layers.30.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 795 |
+
"model.layers.30.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 796 |
+
"model.layers.30.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 797 |
+
"model.layers.30.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 798 |
+
"model.layers.30.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 799 |
+
"model.layers.30.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 800 |
+
"model.layers.31.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 801 |
+
"model.layers.31.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 802 |
+
"model.layers.31.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 803 |
+
"model.layers.31.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 804 |
+
"model.layers.31.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 805 |
+
"model.layers.31.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 806 |
+
"model.layers.31.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 807 |
+
"model.layers.31.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 808 |
+
"model.layers.31.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 809 |
+
"model.layers.31.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 810 |
+
"model.layers.31.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 811 |
+
"model.layers.31.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 812 |
+
"model.layers.32.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 813 |
+
"model.layers.32.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 814 |
+
"model.layers.32.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 815 |
+
"model.layers.32.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 816 |
+
"model.layers.32.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 817 |
+
"model.layers.32.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 818 |
+
"model.layers.32.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 819 |
+
"model.layers.32.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 820 |
+
"model.layers.32.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 821 |
+
"model.layers.32.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 822 |
+
"model.layers.32.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 823 |
+
"model.layers.32.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 824 |
+
"model.layers.33.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 825 |
+
"model.layers.33.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 826 |
+
"model.layers.33.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 827 |
+
"model.layers.33.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 828 |
+
"model.layers.33.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 829 |
+
"model.layers.33.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 830 |
+
"model.layers.33.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 831 |
+
"model.layers.33.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 832 |
+
"model.layers.33.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 833 |
+
"model.layers.33.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 834 |
+
"model.layers.33.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 835 |
+
"model.layers.33.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 836 |
+
"model.layers.34.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 837 |
+
"model.layers.34.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 838 |
+
"model.layers.34.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 839 |
+
"model.layers.34.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 840 |
+
"model.layers.34.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 841 |
+
"model.layers.34.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 842 |
+
"model.layers.34.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 843 |
+
"model.layers.34.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 844 |
+
"model.layers.34.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 845 |
+
"model.layers.34.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 846 |
+
"model.layers.34.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 847 |
+
"model.layers.34.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 848 |
+
"model.layers.35.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 849 |
+
"model.layers.35.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 850 |
+
"model.layers.35.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 851 |
+
"model.layers.35.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 852 |
+
"model.layers.35.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 853 |
+
"model.layers.35.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 854 |
+
"model.layers.35.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 855 |
+
"model.layers.35.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 856 |
+
"model.layers.35.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 857 |
+
"model.layers.35.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 858 |
+
"model.layers.35.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 859 |
+
"model.layers.35.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 860 |
+
"model.layers.4.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 861 |
+
"model.layers.4.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 862 |
+
"model.layers.4.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 863 |
+
"model.layers.4.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 864 |
+
"model.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 865 |
+
"model.layers.4.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 866 |
+
"model.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 867 |
+
"model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 868 |
+
"model.layers.4.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 869 |
+
"model.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 870 |
+
"model.layers.4.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 871 |
+
"model.layers.4.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 872 |
+
"model.layers.5.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 873 |
+
"model.layers.5.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 874 |
+
"model.layers.5.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 875 |
+
"model.layers.5.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 876 |
+
"model.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 877 |
+
"model.layers.5.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 878 |
+
"model.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 879 |
+
"model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 880 |
+
"model.layers.5.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 881 |
+
"model.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 882 |
+
"model.layers.5.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 883 |
+
"model.layers.5.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 884 |
+
"model.layers.6.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 885 |
+
"model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 886 |
+
"model.layers.6.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 887 |
+
"model.layers.6.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 888 |
+
"model.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 889 |
+
"model.layers.6.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 890 |
+
"model.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 891 |
+
"model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 892 |
+
"model.layers.6.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 893 |
+
"model.layers.6.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 894 |
+
"model.layers.6.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 895 |
+
"model.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 896 |
+
"model.layers.7.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 897 |
+
"model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 898 |
+
"model.layers.7.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 899 |
+
"model.layers.7.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 900 |
+
"model.layers.7.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 901 |
+
"model.layers.7.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 902 |
+
"model.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 903 |
+
"model.layers.7.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 904 |
+
"model.layers.7.self_attn.q_proj.bias": "model-00002-of-00002.safetensors",
|
| 905 |
+
"model.layers.7.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 906 |
+
"model.layers.7.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 907 |
+
"model.layers.7.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 908 |
+
"model.layers.8.input_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 909 |
+
"model.layers.8.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 910 |
+
"model.layers.8.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 911 |
+
"model.layers.8.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 912 |
+
"model.layers.8.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 913 |
+
"model.layers.8.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
|
| 914 |
+
"model.layers.8.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 915 |
+
"model.layers.8.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 916 |
+
"model.layers.8.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 917 |
+
"model.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 918 |
+
"model.layers.8.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
|
| 919 |
+
"model.layers.8.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 920 |
+
"model.layers.9.input_layernorm.weight": "model-00001-of-00002.safetensors",
|
| 921 |
+
"model.layers.9.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 922 |
+
"model.layers.9.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 923 |
+
"model.layers.9.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 924 |
+
"model.layers.9.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
|
| 925 |
+
"model.layers.9.self_attn.k_proj.bias": "model-00002-of-00002.safetensors",
|
| 926 |
+
"model.layers.9.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 927 |
+
"model.layers.9.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 928 |
+
"model.layers.9.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
|
| 929 |
+
"model.layers.9.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 930 |
+
"model.layers.9.self_attn.v_proj.bias": "model-00002-of-00002.safetensors",
|
| 931 |
+
"model.layers.9.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 932 |
+
"model.norm.weight": "model-00002-of-00002.safetensors"
|
| 933 |
+
}
|
| 934 |
+
}
|
modeling_arkasr.py
ADDED
|
@@ -0,0 +1,263 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from __future__ import annotations
|
| 2 |
+
|
| 3 |
+
from typing import Optional, List, Tuple, Union, Dict
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
from torch import Tensor, nn
|
| 7 |
+
from transformers import Qwen2ForCausalLM
|
| 8 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
| 9 |
+
|
| 10 |
+
from .configuration_arkasr import ArkasrConfig
|
| 11 |
+
from .modeling_audio import WhisperSpecialEncoder
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
class AudioMLPAdapter(nn.Module):
|
| 15 |
+
def __init__(self, config: ArkasrConfig):
|
| 16 |
+
super().__init__()
|
| 17 |
+
whisper_config = config.whisper_config
|
| 18 |
+
self.merge_factor = int(config.merge_factor)
|
| 19 |
+
|
| 20 |
+
# 音频编码器
|
| 21 |
+
self.whisper = WhisperSpecialEncoder(
|
| 22 |
+
whisper_config,
|
| 23 |
+
use_rope=getattr(config, "use_rope", False),
|
| 24 |
+
)
|
| 25 |
+
# 禁用 Whisper 自带 LayerNorm
|
| 26 |
+
self.whisper.layer_norm = nn.Identity()
|
| 27 |
+
self.layer_norm = nn.LayerNorm(whisper_config.hidden_size)
|
| 28 |
+
|
| 29 |
+
act_fn_map = {
|
| 30 |
+
"gelu": nn.GELU(),
|
| 31 |
+
"relu": nn.ReLU(),
|
| 32 |
+
"selu": nn.SELU(),
|
| 33 |
+
}
|
| 34 |
+
act = act_fn_map.get(getattr(config, "mlp_adapter_act", "gelu"), nn.GELU())
|
| 35 |
+
|
| 36 |
+
input_dim = whisper_config.hidden_size * self.merge_factor
|
| 37 |
+
output_dim = config.hidden_size
|
| 38 |
+
|
| 39 |
+
self.adapting = nn.Sequential(
|
| 40 |
+
nn.Linear(input_dim, output_dim * 2),
|
| 41 |
+
act,
|
| 42 |
+
nn.Linear(output_dim * 2, output_dim),
|
| 43 |
+
)
|
| 44 |
+
|
| 45 |
+
def forward(self, audios: Tensor) -> Tensor:
|
| 46 |
+
"""
|
| 47 |
+
Args:
|
| 48 |
+
audios: (B, mel, T) 或 (B, raw_len) —— 由 WhisperSpecialEncoder 决定
|
| 49 |
+
Returns:
|
| 50 |
+
adapted_features: (B, Seq_Audio, LLM_Hidden_Dim)
|
| 51 |
+
"""
|
| 52 |
+
bsz = audios.size(0)
|
| 53 |
+
|
| 54 |
+
encoded = self.whisper(audios)[0] # (B, T, D)
|
| 55 |
+
encoded = self.layer_norm(encoded)
|
| 56 |
+
|
| 57 |
+
seq_len = encoded.size(1)
|
| 58 |
+
if seq_len % self.merge_factor != 0:
|
| 59 |
+
target_len = (seq_len // self.merge_factor) * self.merge_factor
|
| 60 |
+
if target_len <= 0:
|
| 61 |
+
# 极短音频兜底：pad 到 merge_factor
|
| 62 |
+
target_len = self.merge_factor
|
| 63 |
+
if seq_len < target_len:
|
| 64 |
+
pad_len = target_len - seq_len
|
| 65 |
+
pad = encoded.new_zeros((bsz, pad_len, encoded.size(-1)))
|
| 66 |
+
encoded = torch.cat([encoded, pad], dim=1)
|
| 67 |
+
else:
|
| 68 |
+
encoded = encoded[:, :target_len, :]
|
| 69 |
+
|
| 70 |
+
encoded = encoded.reshape(bsz, -1, encoded.size(-1) * self.merge_factor)
|
| 71 |
+
adapted = self.adapting(encoded) # (B, T/k, hidden)
|
| 72 |
+
return adapted
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
class ArkasrForConditionalGeneration(Qwen2ForCausalLM):
|
| 76 |
+
config_class = ArkasrConfig
|
| 77 |
+
_no_split_modules = ["WhisperSpecialEncoder"]
|
| 78 |
+
|
| 79 |
+
def __init__(self, config: ArkasrConfig):
|
| 80 |
+
super().__init__(config)
|
| 81 |
+
self.audio_encoder = AudioMLPAdapter(config)
|
| 82 |
+
|
| 83 |
+
self.audio_token_id = getattr(config, "audio_token_id", None)
|
| 84 |
+
if self.audio_token_id is None:
|
| 85 |
+
raise ValueError("`audio_token_id` must be defined in config.")
|
| 86 |
+
|
| 87 |
+
@staticmethod
|
| 88 |
+
def _cache_seq_len(past_key_values) -> int:
|
| 89 |
+
if past_key_values is None:
|
| 90 |
+
return 0
|
| 91 |
+
if hasattr(past_key_values, "get_seq_length"):
|
| 92 |
+
try:
|
| 93 |
+
return int(past_key_values.get_seq_length())
|
| 94 |
+
except Exception:
|
| 95 |
+
return 0
|
| 96 |
+
try:
|
| 97 |
+
return int(past_key_values[0][0].shape[-2])
|
| 98 |
+
except Exception:
|
| 99 |
+
return 0
|
| 100 |
+
|
| 101 |
+
def _inject_audio_embeddings_batch_encode_then_loop_scatter(
|
| 102 |
+
self,
|
| 103 |
+
input_ids: torch.LongTensor, # (B, S)
|
| 104 |
+
inputs_embeds: torch.FloatTensor, # (B, S, H)
|
| 105 |
+
audios: Tensor, # (B, ...)
|
| 106 |
+
) -> torch.FloatTensor:
|
| 107 |
+
"""
|
| 108 |
+
先对「有 audio token 的样本」做一次 batch 音频编码，
|
| 109 |
+
然后 for-loop 把每个样本的 audio features 按 audio_token 位置写回 inputs_embeds。
|
| 110 |
+
|
| 111 |
+
好处：
|
| 112 |
+
- encoder 只跑一次（快）
|
| 113 |
+
- 写回按样本做，不会跨样本错位（稳）
|
| 114 |
+
- 碰到某行没有 audio_token：直接跳过（TTS 行无影响）
|
| 115 |
+
|
| 116 |
+
约束：
|
| 117 |
+
- 每条样本的 audio_token 数量 n_i 需要和 audio_encoder 输出的 Sa 对齐。
|
| 118 |
+
如果不对齐：这里采用截断/补零对齐到 n_i（不报错）。
|
| 119 |
+
"""
|
| 120 |
+
B, S = input_ids.shape
|
| 121 |
+
H = inputs_embeds.size(-1)
|
| 122 |
+
device = inputs_embeds.device
|
| 123 |
+
dtype = inputs_embeds.dtype
|
| 124 |
+
|
| 125 |
+
# 找到哪些样本需要注入
|
| 126 |
+
mask = (input_ids == self.audio_token_id) # (B, S)
|
| 127 |
+
per_counts = mask.sum(dim=1) # (B,)
|
| 128 |
+
need_idx = (per_counts > 0).nonzero(as_tuple=False).squeeze(1) # (K,)
|
| 129 |
+
|
| 130 |
+
if need_idx.numel() == 0:
|
| 131 |
+
return inputs_embeds
|
| 132 |
+
|
| 133 |
+
# 只编码需要注入的那部分音频（K, ...)
|
| 134 |
+
audios_sub = audios.index_select(0, need_idx)
|
| 135 |
+
feats_sub = self.audio_encoder(audios_sub) # (K, Sa, H)
|
| 136 |
+
|
| 137 |
+
# 写回：逐样本替换（写回操作本身几��不耗时）
|
| 138 |
+
feats_sub = feats_sub.to(device=device, dtype=dtype)
|
| 139 |
+
Sa = feats_sub.size(1)
|
| 140 |
+
|
| 141 |
+
# 逐个样本注入
|
| 142 |
+
for k in range(need_idx.numel()):
|
| 143 |
+
i = int(need_idx[k].item())
|
| 144 |
+
n_i = int(per_counts[i].item())
|
| 145 |
+
if n_i <= 0:
|
| 146 |
+
continue
|
| 147 |
+
|
| 148 |
+
feat_i = feats_sub[k] # (Sa, H)
|
| 149 |
+
|
| 150 |
+
# 对齐到该样本的 audio token 数 n_i
|
| 151 |
+
if Sa < n_i:
|
| 152 |
+
pad = feat_i.new_zeros((n_i - Sa, H))
|
| 153 |
+
feat_i_use = torch.cat([feat_i, pad], dim=0)
|
| 154 |
+
elif Sa > n_i:
|
| 155 |
+
feat_i_use = feat_i[:n_i]
|
| 156 |
+
else:
|
| 157 |
+
feat_i_use = feat_i
|
| 158 |
+
|
| 159 |
+
pos_i = mask[i].nonzero(as_tuple=False).squeeze(1) # (n_i,)
|
| 160 |
+
# 写回 embeddings
|
| 161 |
+
inputs_embeds[i, pos_i, :] = feat_i_use
|
| 162 |
+
|
| 163 |
+
return inputs_embeds
|
| 164 |
+
|
| 165 |
+
def forward(
|
| 166 |
+
self,
|
| 167 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 168 |
+
audios: Optional[Tensor] = None,
|
| 169 |
+
attention_mask: Optional[Tensor] = None,
|
| 170 |
+
position_ids: Optional[Tensor] = None,
|
| 171 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 172 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 173 |
+
use_cache: Optional[bool] = None,
|
| 174 |
+
labels: Optional[torch.LongTensor] = None,
|
| 175 |
+
output_attentions: Optional[bool] = None,
|
| 176 |
+
output_hidden_states: Optional[bool] = None,
|
| 177 |
+
logits_to_keep: int | torch.Tensor = 0,
|
| 178 |
+
**kwargs,
|
| 179 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
| 180 |
+
|
| 181 |
+
if inputs_embeds is None:
|
| 182 |
+
if input_ids is None:
|
| 183 |
+
raise ValueError("Either `input_ids` or `inputs_embeds` must be provided.")
|
| 184 |
+
inputs_embeds = self.model.embed_tokens(input_ids)
|
| 185 |
+
|
| 186 |
+
# 只在首步（past_len==0）注入，避免 generation 后续重复 encode
|
| 187 |
+
past_len = self._cache_seq_len(past_key_values)
|
| 188 |
+
if audios is not None and input_ids is not None and past_len == 0:
|
| 189 |
+
inputs_embeds = self._inject_audio_embeddings_batch_encode_then_loop_scatter(
|
| 190 |
+
input_ids=input_ids,
|
| 191 |
+
inputs_embeds=inputs_embeds,
|
| 192 |
+
audios=audios,
|
| 193 |
+
)
|
| 194 |
+
|
| 195 |
+
outputs = self.model(
|
| 196 |
+
input_ids=None,
|
| 197 |
+
attention_mask=attention_mask,
|
| 198 |
+
position_ids=position_ids,
|
| 199 |
+
past_key_values=past_key_values,
|
| 200 |
+
inputs_embeds=inputs_embeds,
|
| 201 |
+
use_cache=use_cache,
|
| 202 |
+
output_attentions=output_attentions,
|
| 203 |
+
output_hidden_states=output_hidden_states,
|
| 204 |
+
)
|
| 205 |
+
|
| 206 |
+
hidden_states = outputs[0]
|
| 207 |
+
|
| 208 |
+
# logits（避免重复算 lm_head）
|
| 209 |
+
if isinstance(logits_to_keep, int) and logits_to_keep > 0:
|
| 210 |
+
hidden_for_logits = hidden_states[:, -logits_to_keep:, :]
|
| 211 |
+
elif isinstance(logits_to_keep, torch.Tensor):
|
| 212 |
+
hidden_for_logits = hidden_states[:, logits_to_keep, :]
|
| 213 |
+
else:
|
| 214 |
+
hidden_for_logits = hidden_states
|
| 215 |
+
|
| 216 |
+
logits = self.lm_head(hidden_for_logits)
|
| 217 |
+
|
| 218 |
+
loss = None
|
| 219 |
+
if labels is not None:
|
| 220 |
+
loss = self.loss_function(
|
| 221 |
+
logits=logits,
|
| 222 |
+
labels=labels,
|
| 223 |
+
vocab_size=self.config.vocab_size,
|
| 224 |
+
**kwargs,
|
| 225 |
+
)
|
| 226 |
+
|
| 227 |
+
return CausalLMOutputWithPast(
|
| 228 |
+
loss=loss,
|
| 229 |
+
logits=logits,
|
| 230 |
+
past_key_values=outputs.past_key_values,
|
| 231 |
+
hidden_states=outputs.hidden_states,
|
| 232 |
+
attentions=outputs.attentions,
|
| 233 |
+
)
|
| 234 |
+
|
| 235 |
+
def prepare_inputs_for_generation(
|
| 236 |
+
self,
|
| 237 |
+
input_ids,
|
| 238 |
+
past_key_values=None,
|
| 239 |
+
attention_mask=None,
|
| 240 |
+
inputs_embeds=None,
|
| 241 |
+
**kwargs,
|
| 242 |
+
):
|
| 243 |
+
past_len = self._cache_seq_len(past_key_values)
|
| 244 |
+
if past_len > 0:
|
| 245 |
+
input_ids = input_ids[:, -1:]
|
| 246 |
+
|
| 247 |
+
model_inputs = {
|
| 248 |
+
"input_ids": input_ids,
|
| 249 |
+
"past_key_values": past_key_values,
|
| 250 |
+
"use_cache": kwargs.get("use_cache"),
|
| 251 |
+
"attention_mask": attention_mask,
|
| 252 |
+
# audios 透传；forward 内 past_len==0 才注入，所以后续 step 不会重复编码
|
| 253 |
+
"audios": kwargs.get("audios", None),
|
| 254 |
+
}
|
| 255 |
+
|
| 256 |
+
if inputs_embeds is not None and past_key_values is None:
|
| 257 |
+
model_inputs["inputs_embeds"] = inputs_embeds
|
| 258 |
+
del model_inputs["input_ids"]
|
| 259 |
+
|
| 260 |
+
return model_inputs
|
| 261 |
+
|
| 262 |
+
|
| 263 |
+
__all__ = ["ArkasrForConditionalGeneration", "AudioMLPAdapter"]
|
modeling_audio.py
ADDED
|
@@ -0,0 +1,293 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Any, Optional, Tuple, Union
|
| 2 |
+
|
| 3 |
+
import torch
|
| 4 |
+
from torch import Tensor, nn
|
| 5 |
+
from torch.nn.functional import scaled_dot_product_attention
|
| 6 |
+
from transformers import WhisperConfig
|
| 7 |
+
from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions
|
| 8 |
+
from transformers.models.whisper.modeling_whisper import WhisperEncoder, WhisperEncoderLayer
|
| 9 |
+
from transformers.utils import logging
|
| 10 |
+
|
| 11 |
+
logger = logging.get_logger(__name__)
|
| 12 |
+
|
| 13 |
+
# ==========================================
|
| 14 |
+
# 1. Rotary Embedding 核心组件
|
| 15 |
+
# ==========================================
|
| 16 |
+
|
| 17 |
+
class RotaryEmbedding(nn.Module):
|
| 18 |
+
def __init__(self, dim, rope_ratio=1):
|
| 19 |
+
super().__init__()
|
| 20 |
+
self.dim = dim
|
| 21 |
+
self.rope_ratio = rope_ratio
|
| 22 |
+
|
| 23 |
+
@torch.no_grad()
|
| 24 |
+
def get_emb(self, seq_len: int, dtype: torch.dtype, device: torch.device, base: int = 10000):
|
| 25 |
+
"""生成 RoPE 缓存"""
|
| 26 |
+
base = base * self.rope_ratio
|
| 27 |
+
# 计算频率 theta
|
| 28 |
+
inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, dtype=torch.float, device=device) / self.dim))
|
| 29 |
+
|
| 30 |
+
# 生成位置索引
|
| 31 |
+
t = torch.arange(seq_len, device=device, dtype=torch.float)
|
| 32 |
+
freqs = torch.outer(t, inv_freq) # [seq_len, dim/2]
|
| 33 |
+
|
| 34 |
+
# 构造 cos 和 sin 缓存
|
| 35 |
+
# 形状: [seq_len, dim/2, 2]
|
| 36 |
+
emb = torch.stack([torch.cos(freqs), torch.sin(freqs)], dim=-1)
|
| 37 |
+
|
| 38 |
+
if dtype in (torch.float16, torch.bfloat16):
|
| 39 |
+
emb = emb.to(dtype)
|
| 40 |
+
return emb
|
| 41 |
+
|
| 42 |
+
def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
|
| 43 |
+
"""
|
| 44 |
+
x: [batch, num_heads, seq_len, head_dim]
|
| 45 |
+
rope_cache: [1, seq_len, dim/2, 2]
|
| 46 |
+
"""
|
| 47 |
+
b, nh, sq, hd = x.shape
|
| 48 |
+
rot_dim = rope_cache.shape[-2] * 2
|
| 49 |
+
|
| 50 |
+
# 将 x 分为旋转部分和不旋转部分
|
| 51 |
+
x_rot, x_pass = x[..., :rot_dim], x[..., rot_dim:]
|
| 52 |
+
|
| 53 |
+
# 调整 x_rot 形状以匹配 rope_cache: [b, nh, sq, rot_dim/2, 2]
|
| 54 |
+
x_shaped = x_rot.reshape(b, nh, sq, rot_dim // 2, 2)
|
| 55 |
+
|
| 56 |
+
# 计算旋转: (a+bi)(c+di) = (ac-bd) + (ad+bc)i
|
| 57 |
+
cos = rope_cache[..., 0] # [1, sq, rot_dim/2]
|
| 58 |
+
sin = rope_cache[..., 1] # [1, sq, rot_dim/2]
|
| 59 |
+
|
| 60 |
+
# 增加 head 维度
|
| 61 |
+
cos = cos.unsqueeze(1) # [1, 1, sq, rot_dim/2]
|
| 62 |
+
sin = sin.unsqueeze(1) # [1, 1, sq, rot_dim/2]
|
| 63 |
+
|
| 64 |
+
x_out = torch.stack([
|
| 65 |
+
x_shaped[..., 0] * cos - x_shaped[..., 1] * sin,
|
| 66 |
+
x_shaped[..., 1] * cos + x_shaped[..., 0] * sin
|
| 67 |
+
], dim=-1)
|
| 68 |
+
|
| 69 |
+
x_out = x_out.flatten(3) # 合并最后两维到 rot_dim
|
| 70 |
+
return torch.cat([x_out, x_pass], dim=-1)
|
| 71 |
+
|
| 72 |
+
# ==========================================
|
| 73 |
+
# 2. 基于 SDPA 的 RoPE Attention
|
| 74 |
+
# ==========================================
|
| 75 |
+
|
| 76 |
+
class WhisperRoPESdpaAttention(nn.Module):
|
| 77 |
+
"""
|
| 78 |
+
使用 PyTorch 原生 scaled_dot_product_attention 替代 WhisperFlashAttention2。
|
| 79 |
+
"""
|
| 80 |
+
def __init__(self, config: WhisperConfig, embed_dim: int, num_heads: int, dropout: float = 0.0):
|
| 81 |
+
super().__init__()
|
| 82 |
+
self.config = config
|
| 83 |
+
self.embed_dim = embed_dim
|
| 84 |
+
self.num_heads = num_heads
|
| 85 |
+
self.dropout = dropout
|
| 86 |
+
self.head_dim = embed_dim // num_heads
|
| 87 |
+
|
| 88 |
+
# Whisper 标准投影层
|
| 89 |
+
self.q_proj = nn.Linear(embed_dim, embed_dim)
|
| 90 |
+
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False)
|
| 91 |
+
self.v_proj = nn.Linear(embed_dim, embed_dim)
|
| 92 |
+
self.out_proj = nn.Linear(embed_dim, embed_dim)
|
| 93 |
+
|
| 94 |
+
self.is_causal = False
|
| 95 |
+
|
| 96 |
+
def forward(
|
| 97 |
+
self,
|
| 98 |
+
hidden_states: torch.Tensor,
|
| 99 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 100 |
+
layer_head_mask: Optional[torch.Tensor] = None,
|
| 101 |
+
output_attentions: bool = False,
|
| 102 |
+
rotary_pos_emb: Optional[torch.Tensor] = None,
|
| 103 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], None]:
|
| 104 |
+
|
| 105 |
+
bsz, q_len, _ = hidden_states.size()
|
| 106 |
+
|
| 107 |
+
# 1. 投影映射
|
| 108 |
+
query_states = self.q_proj(hidden_states)
|
| 109 |
+
key_states = self.k_proj(hidden_states)
|
| 110 |
+
value_states = self.v_proj(hidden_states)
|
| 111 |
+
|
| 112 |
+
# 2. 变形为 [batch, heads, seq, dim] 并确保内存连续
|
| 113 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
| 114 |
+
key_states = key_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
| 115 |
+
value_states = value_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
| 116 |
+
|
| 117 |
+
# 3. 应用 RoPE
|
| 118 |
+
if rotary_pos_emb is not None:
|
| 119 |
+
query_states = apply_rotary_pos_emb(query_states, rotary_pos_emb)
|
| 120 |
+
key_states = apply_rotary_pos_emb(key_states, rotary_pos_emb)
|
| 121 |
+
|
| 122 |
+
# 4. 数据类型对齐 (处理 fp32 LayerNorm 带来的类型不匹配)
|
| 123 |
+
target_dtype = self.q_proj.weight.dtype
|
| 124 |
+
query_states = query_states.to(target_dtype)
|
| 125 |
+
key_states = key_states.to(target_dtype)
|
| 126 |
+
value_states = value_states.to(target_dtype)
|
| 127 |
+
|
| 128 |
+
# 5. SDPA 计算 (关键：不要手动乘以 scaling, SDPA 内部会自动处理)
|
| 129 |
+
# 注意: 如果传入了 4D attention_mask，SDPA 会正确应用它
|
| 130 |
+
attn_output = scaled_dot_product_attention(
|
| 131 |
+
query_states,
|
| 132 |
+
key_states,
|
| 133 |
+
value_states,
|
| 134 |
+
attn_mask=attention_mask,
|
| 135 |
+
dropout_p=self.dropout if self.training else 0.0,
|
| 136 |
+
is_causal=self.is_causal,
|
| 137 |
+
)
|
| 138 |
+
|
| 139 |
+
# 6. 恢复形状并输出投影
|
| 140 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 141 |
+
attn_output = attn_output.reshape(bsz, q_len, self.embed_dim)
|
| 142 |
+
attn_output = self.out_proj(attn_output)
|
| 143 |
+
|
| 144 |
+
return attn_output, None, None
|
| 145 |
+
|
| 146 |
+
# ==========================================
|
| 147 |
+
# 3. 封装好的 Encoder 层和 Encoder
|
| 148 |
+
# ==========================================
|
| 149 |
+
|
| 150 |
+
class WhisperSpecialEncoderLayer(WhisperEncoderLayer):
|
| 151 |
+
def __init__(self, config: WhisperConfig):
|
| 152 |
+
super().__init__(config)
|
| 153 |
+
# 替换 Self-Attention 为我们的 RoPE SDPA 版本
|
| 154 |
+
self.self_attn = WhisperRoPESdpaAttention(
|
| 155 |
+
config=config,
|
| 156 |
+
embed_dim=self.embed_dim,
|
| 157 |
+
num_heads=config.encoder_attention_heads,
|
| 158 |
+
dropout=config.attention_dropout,
|
| 159 |
+
)
|
| 160 |
+
|
| 161 |
+
def forward(
|
| 162 |
+
self,
|
| 163 |
+
hidden_states: torch.Tensor,
|
| 164 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 165 |
+
layer_head_mask: Optional[torch.Tensor] = None,
|
| 166 |
+
output_attentions: bool = False,
|
| 167 |
+
rotary_pos_emb: Optional[torch.Tensor] = None,
|
| 168 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 169 |
+
) -> Tuple[torch.Tensor, Any]:
|
| 170 |
+
|
| 171 |
+
residual = hidden_states
|
| 172 |
+
hidden_states = self.self_attn_layer_norm(hidden_states)
|
| 173 |
+
|
| 174 |
+
hidden_states, attn_weights, _ = self.self_attn(
|
| 175 |
+
hidden_states=hidden_states,
|
| 176 |
+
attention_mask=attention_mask,
|
| 177 |
+
layer_head_mask=layer_head_mask,
|
| 178 |
+
output_attentions=output_attentions,
|
| 179 |
+
rotary_pos_emb=rotary_pos_emb,
|
| 180 |
+
)
|
| 181 |
+
|
| 182 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
|
| 183 |
+
hidden_states = residual + hidden_states
|
| 184 |
+
|
| 185 |
+
residual = hidden_states
|
| 186 |
+
hidden_states = self.final_layer_norm(hidden_states)
|
| 187 |
+
hidden_states = self.activation_fn(self.fc1(hidden_states))
|
| 188 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
|
| 189 |
+
hidden_states = self.fc2(hidden_states)
|
| 190 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
|
| 191 |
+
hidden_states = residual + hidden_states
|
| 192 |
+
|
| 193 |
+
if hidden_states.dtype == torch.float16:
|
| 194 |
+
clamp_value = torch.finfo(hidden_states.dtype).max - 1000
|
| 195 |
+
hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
|
| 196 |
+
|
| 197 |
+
return (hidden_states, None) # 保持与 Whisper 接口一致的 tuple 长度
|
| 198 |
+
|
| 199 |
+
class WhisperSpecialEncoder(WhisperEncoder):
|
| 200 |
+
def __init__(self, config: WhisperConfig, use_rope=True, rope_ratio=1):
|
| 201 |
+
super().__init__(config)
|
| 202 |
+
self.use_rope = use_rope
|
| 203 |
+
# 覆盖父类的层列表
|
| 204 |
+
self.layers = nn.ModuleList(
|
| 205 |
+
[WhisperSpecialEncoderLayer(config) for _ in range(config.encoder_layers)]
|
| 206 |
+
)
|
| 207 |
+
|
| 208 |
+
if use_rope:
|
| 209 |
+
# 计算 RoPE 维度: 通常是 head_dim 的一部分
|
| 210 |
+
head_dim = config.d_model // config.encoder_attention_heads
|
| 211 |
+
self.rotary_embedding = RotaryEmbedding(head_dim // 2, rope_ratio)
|
| 212 |
+
|
| 213 |
+
def forward(
|
| 214 |
+
self,
|
| 215 |
+
input_features,
|
| 216 |
+
attention_mask=None,
|
| 217 |
+
head_mask=None,
|
| 218 |
+
output_attentions=None,
|
| 219 |
+
output_hidden_states=None,
|
| 220 |
+
return_dict=None,
|
| 221 |
+
position_ids=None,
|
| 222 |
+
):
|
| 223 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 224 |
+
output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 225 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 226 |
+
|
| 227 |
+
# Whisper 卷积特征提取
|
| 228 |
+
inputs_embeds = nn.functional.gelu(self.conv1(input_features))
|
| 229 |
+
inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))
|
| 230 |
+
inputs_embeds = inputs_embeds.permute(0, 2, 1) # [B, T_down, D]
|
| 231 |
+
|
| 232 |
+
if self.use_rope:
|
| 233 |
+
# 生成旋转编码缓存
|
| 234 |
+
rotary_embs = self.rotary_embedding.get_emb(
|
| 235 |
+
seq_len=inputs_embeds.shape[1],
|
| 236 |
+
dtype=inputs_embeds.dtype,
|
| 237 |
+
device=inputs_embeds.device
|
| 238 |
+
)
|
| 239 |
+
# 形状调整为 [1, seq_len, dim/2, 2] 以便广播
|
| 240 |
+
rotary_embs = rotary_embs.unsqueeze(0)
|
| 241 |
+
hidden_states = inputs_embeds
|
| 242 |
+
else:
|
| 243 |
+
rotary_embs = None
|
| 244 |
+
# 回退到绝对位置编码
|
| 245 |
+
embed_pos = self.embed_positions.weight[:inputs_embeds.shape[1]]
|
| 246 |
+
hidden_states = inputs_embeds + embed_pos
|
| 247 |
+
|
| 248 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
|
| 249 |
+
|
| 250 |
+
encoder_states = () if output_hidden_states else None
|
| 251 |
+
all_attentions = () if output_attentions else None
|
| 252 |
+
|
| 253 |
+
for idx, encoder_layer in enumerate(self.layers):
|
| 254 |
+
if output_hidden_states:
|
| 255 |
+
encoder_states = encoder_states + (hidden_states,)
|
| 256 |
+
|
| 257 |
+
if self.gradient_checkpointing and self.training:
|
| 258 |
+
layer_outputs = self._gradient_checkpointing_func(
|
| 259 |
+
encoder_layer.__call__,
|
| 260 |
+
hidden_states,
|
| 261 |
+
None, # attention_mask
|
| 262 |
+
(head_mask[idx] if head_mask is not None else None),
|
| 263 |
+
output_attentions,
|
| 264 |
+
rotary_embs,
|
| 265 |
+
position_ids,
|
| 266 |
+
)
|
| 267 |
+
else:
|
| 268 |
+
layer_outputs = encoder_layer(
|
| 269 |
+
hidden_states,
|
| 270 |
+
attention_mask=None,
|
| 271 |
+
layer_head_mask=(head_mask[idx] if head_mask is not None else None),
|
| 272 |
+
output_attentions=output_attentions,
|
| 273 |
+
rotary_pos_emb=rotary_embs,
|
| 274 |
+
position_ids=position_ids,
|
| 275 |
+
)
|
| 276 |
+
|
| 277 |
+
hidden_states = layer_outputs[0]
|
| 278 |
+
|
| 279 |
+
if output_attentions:
|
| 280 |
+
all_attentions = all_attentions + (layer_outputs[2],)
|
| 281 |
+
|
| 282 |
+
hidden_states = self.layer_norm(hidden_states)
|
| 283 |
+
if output_hidden_states:
|
| 284 |
+
encoder_states = encoder_states + (hidden_states,)
|
| 285 |
+
|
| 286 |
+
if not return_dict:
|
| 287 |
+
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
|
| 288 |
+
|
| 289 |
+
return BaseModelOutputWithPastAndCrossAttentions(
|
| 290 |
+
last_hidden_state=hidden_states,
|
| 291 |
+
hidden_states=encoder_states,
|
| 292 |
+
attentions=all_attentions,
|
| 293 |
+
)
|
preprocessor_config.json
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"auto_map": {
|
| 3 |
+
"AutoProcessor": "processing_arkasr.ArkasrProcessor"
|
| 4 |
+
},
|
| 5 |
+
"chunk_length": 30,
|
| 6 |
+
"dither": 0.0,
|
| 7 |
+
"feature_extractor_type": "WhisperFeatureExtractor",
|
| 8 |
+
"feature_size": 128,
|
| 9 |
+
"hop_length": 160,
|
| 10 |
+
"n_fft": 400,
|
| 11 |
+
"n_samples": 480000,
|
| 12 |
+
"nb_max_frames": 3000,
|
| 13 |
+
"padding_side": "right",
|
| 14 |
+
"padding_value": 0.0,
|
| 15 |
+
"processor_class": "ArkasrProcessor",
|
| 16 |
+
"return_attention_mask": false,
|
| 17 |
+
"sampling_rate": 16000
|
| 18 |
+
}
|
processing_arkasr.py
ADDED
|
@@ -0,0 +1,446 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
from __future__ import annotations
|
| 3 |
+
|
| 4 |
+
import base64
|
| 5 |
+
import io
|
| 6 |
+
import json
|
| 7 |
+
import os
|
| 8 |
+
from typing import Any, Dict, List, Optional, Union
|
| 9 |
+
|
| 10 |
+
import numpy as np
|
| 11 |
+
import torch
|
| 12 |
+
import librosa
|
| 13 |
+
import soundfile as sf # 显式引入 soundfile 以处理 BytesIO
|
| 14 |
+
|
| 15 |
+
from transformers import AutoTokenizer, WhisperFeatureExtractor
|
| 16 |
+
from transformers.feature_extraction_utils import BatchFeature
|
| 17 |
+
from transformers.processing_utils import ProcessorMixin
|
| 18 |
+
from transformers.utils import logging
|
| 19 |
+
|
| 20 |
+
logger = logging.get_logger(__name__)
|
| 21 |
+
|
| 22 |
+
_AUDIO_MARKER = "<<AUDIO_TOKENS>>"
|
| 23 |
+
|
| 24 |
+
def _normalize_dtype_name(name: str) -> str:
|
| 25 |
+
name = name.strip().lower()
|
| 26 |
+
alias = {
|
| 27 |
+
"fp16": "float16",
|
| 28 |
+
"float16": "float16",
|
| 29 |
+
"half": "float16",
|
| 30 |
+
"bf16": "bfloat16",
|
| 31 |
+
"bfloat16": "bfloat16",
|
| 32 |
+
"fp32": "float32",
|
| 33 |
+
"float32": "float32",
|
| 34 |
+
"float": "float32",
|
| 35 |
+
}
|
| 36 |
+
return alias.get(name, name)
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def _resolve_torch_dtype(x: Any, default: str = "float32") -> torch.dtype:
|
| 40 |
+
if isinstance(x, torch.dtype):
|
| 41 |
+
return x
|
| 42 |
+
if x is None:
|
| 43 |
+
x = default
|
| 44 |
+
if isinstance(x, str):
|
| 45 |
+
name = _normalize_dtype_name(x)
|
| 46 |
+
if not hasattr(torch, name):
|
| 47 |
+
raise ValueError(f"Unknown torch dtype string: {x} (normalized: {name})")
|
| 48 |
+
return getattr(torch, name)
|
| 49 |
+
raise TypeError(f"audio_dtype/audio_torch_dtype must be str or torch.dtype or None, got {type(x)}")
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
class ArkasrProcessor(ProcessorMixin):
|
| 53 |
+
attributes = ["feature_extractor", "tokenizer"]
|
| 54 |
+
valid_kwargs = ["merge_factor", "audio_token", "audio_dtype"]
|
| 55 |
+
feature_extractor_class = ("WhisperFeatureExtractor", "SequenceFeatureExtractor")
|
| 56 |
+
tokenizer_class = ("PreTrainedTokenizerFast", "PreTrainedTokenizer")
|
| 57 |
+
|
| 58 |
+
def __init__(
|
| 59 |
+
self,
|
| 60 |
+
feature_extractor,
|
| 61 |
+
tokenizer,
|
| 62 |
+
merge_factor: int = 4,
|
| 63 |
+
audio_token: str = "<|audio|>",
|
| 64 |
+
audio_dtype: str = "float32",
|
| 65 |
+
**kwargs,
|
| 66 |
+
):
|
| 67 |
+
super().__init__(feature_extractor, tokenizer)
|
| 68 |
+
self.merge_factor = int(merge_factor)
|
| 69 |
+
self.audio_token = str(audio_token)
|
| 70 |
+
self.audio_dtype = str(audio_dtype)
|
| 71 |
+
|
| 72 |
+
self.bos_audio_token = "<|begin_of_audio|>"
|
| 73 |
+
self.eos_audio_token = "<|end_of_audio|>"
|
| 74 |
+
self.user_token = "<|user|>"
|
| 75 |
+
self.assistant_token = "<|assistant|>"
|
| 76 |
+
self.assistant_end_token = "<|im_end|>"
|
| 77 |
+
|
| 78 |
+
@classmethod
|
| 79 |
+
def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs) -> "ArkasrProcessor":
|
| 80 |
+
trust_remote_code = bool(kwargs.pop("trust_remote_code", False))
|
| 81 |
+
passthrough_keys = {"cache_dir", "force_download", "local_files_only", "token", "revision", "subfolder"}
|
| 82 |
+
shared_kwargs = {k: kwargs[k] for k in list(kwargs.keys()) if k in passthrough_keys}
|
| 83 |
+
|
| 84 |
+
merge_factor = 4
|
| 85 |
+
audio_token = "<|audio|>"
|
| 86 |
+
audio_dtype = "float32"
|
| 87 |
+
tokenizer_cfg: Dict[str, Any] = {}
|
| 88 |
+
feat_cfg: Dict[str, Any] = {}
|
| 89 |
+
|
| 90 |
+
proc_cfg_path = os.path.join(pretrained_model_name_or_path, "processor_config.json")
|
| 91 |
+
if os.path.isfile(proc_cfg_path):
|
| 92 |
+
with open(proc_cfg_path, "r", encoding="utf-8") as f:
|
| 93 |
+
proc_cfg = json.load(f)
|
| 94 |
+
merge_factor = int(proc_cfg.get("merge_factor", merge_factor))
|
| 95 |
+
audio_token = str(proc_cfg.get("audio_token", audio_token))
|
| 96 |
+
audio_dtype = str(proc_cfg.get("audio_dtype", audio_dtype))
|
| 97 |
+
tokenizer_cfg = proc_cfg.get("tokenizer_config", {}) or {}
|
| 98 |
+
feat_cfg = proc_cfg.get("feature_extractor_config", {}) or {}
|
| 99 |
+
|
| 100 |
+
feature_extractor = WhisperFeatureExtractor.from_pretrained(pretrained_model_name_or_path, **shared_kwargs)
|
| 101 |
+
for k, v in feat_cfg.items():
|
| 102 |
+
if hasattr(feature_extractor, k):
|
| 103 |
+
try: setattr(feature_extractor, k, v)
|
| 104 |
+
except Exception: pass
|
| 105 |
+
|
| 106 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 107 |
+
pretrained_model_name_or_path, use_fast=True, trust_remote_code=trust_remote_code, **shared_kwargs
|
| 108 |
+
)
|
| 109 |
+
for k, v in tokenizer_cfg.items():
|
| 110 |
+
if hasattr(tokenizer, k):
|
| 111 |
+
try: setattr(tokenizer, k, v)
|
| 112 |
+
except Exception: pass
|
| 113 |
+
|
| 114 |
+
return cls(
|
| 115 |
+
feature_extractor=feature_extractor,
|
| 116 |
+
tokenizer=tokenizer,
|
| 117 |
+
merge_factor=merge_factor,
|
| 118 |
+
audio_token=audio_token,
|
| 119 |
+
audio_dtype=audio_dtype,
|
| 120 |
+
)
|
| 121 |
+
|
| 122 |
+
# =========================
|
| 123 |
+
# audio helpers (Modified)
|
| 124 |
+
# =========================
|
| 125 |
+
def _load_audio_file(self, path: str, sampling_rate: int = 16000, offset: float = 0.0, duration: Optional[float] = None) -> np.ndarray:
|
| 126 |
+
# librosa load 支持 offset 和 duration
|
| 127 |
+
# offset: start reading after this time (in seconds)
|
| 128 |
+
# duration: only load up to this much audio (in seconds)
|
| 129 |
+
audio_array, _ = librosa.load(path, sr=int(sampling_rate), mono=True, offset=offset, duration=duration)
|
| 130 |
+
return np.asarray(audio_array, dtype=np.float32)
|
| 131 |
+
|
| 132 |
+
def _strip_data_url_prefix(self, b64: str) -> str:
|
| 133 |
+
if "," in b64 and b64[:30].lower().startswith("data:"):
|
| 134 |
+
return b64.split(",", 1)[1]
|
| 135 |
+
return b64
|
| 136 |
+
|
| 137 |
+
def _load_audio_base64(self, b64: str, sampling_rate: int = 16000, offset: float = 0.0, duration: Optional[float] = None) -> np.ndarray:
|
| 138 |
+
b64 = self._strip_data_url_prefix(b64)
|
| 139 |
+
raw = base64.b64decode(b64)
|
| 140 |
+
bio = io.BytesIO(raw)
|
| 141 |
+
|
| 142 |
+
# 使用 librosa 加载 BytesIO 同样支持 offset 和 duration
|
| 143 |
+
try:
|
| 144 |
+
wav, _sr = librosa.load(bio, sr=int(sampling_rate), mono=True, offset=offset, duration=duration)
|
| 145 |
+
return np.asarray(wav, dtype=np.float32)
|
| 146 |
+
except Exception as e:
|
| 147 |
+
# Fallback (手动切片，比较慢)
|
| 148 |
+
try:
|
| 149 |
+
bio.seek(0)
|
| 150 |
+
data, sr = sf.read(bio, dtype="float32", always_2d=True)
|
| 151 |
+
wav = data.mean(axis=1)
|
| 152 |
+
if int(sr) != int(sampling_rate):
|
| 153 |
+
wav = librosa.resample(wav, orig_sr=int(sr), target_sr=int(sampling_rate))
|
| 154 |
+
|
| 155 |
+
start_sample = int(offset * sampling_rate)
|
| 156 |
+
end_sample = None
|
| 157 |
+
if duration is not None:
|
| 158 |
+
end_sample = start_sample + int(duration * sampling_rate)
|
| 159 |
+
|
| 160 |
+
return np.asarray(wav[start_sample:end_sample], dtype=np.float32)
|
| 161 |
+
except Exception as e2:
|
| 162 |
+
raise ValueError("Failed to decode base64 audio.") from e2
|
| 163 |
+
|
| 164 |
+
def calculate_audio_token_count(self, mel_frames: int) -> int:
|
| 165 |
+
downsampled = (int(mel_frames) + 1) // 2
|
| 166 |
+
merged = downsampled // max(self.merge_factor, 1)
|
| 167 |
+
return max(int(merged), 1)
|
| 168 |
+
|
| 169 |
+
def _build_templates_and_audios(
|
| 170 |
+
self,
|
| 171 |
+
conversations: List[List[dict]],
|
| 172 |
+
sampling_rate: int,
|
| 173 |
+
add_generation_prompt: bool,
|
| 174 |
+
) -> tuple[List[str], List[np.ndarray], List[int]]:
|
| 175 |
+
prompts_template: List[str] = []
|
| 176 |
+
audios_raw: List[np.ndarray] = []
|
| 177 |
+
prompt_audio_counts: List[int] = []
|
| 178 |
+
|
| 179 |
+
for conv in conversations:
|
| 180 |
+
conv_str = ""
|
| 181 |
+
last_role = None
|
| 182 |
+
audio_count_this_conv = 0
|
| 183 |
+
|
| 184 |
+
for msg in conv:
|
| 185 |
+
role = msg["role"]
|
| 186 |
+
last_role = role
|
| 187 |
+
content = msg["content"]
|
| 188 |
+
|
| 189 |
+
if role == "user": conv_str += f"{self.user_token}"
|
| 190 |
+
elif role == "assistant": conv_str += f"{self.assistant_token}"
|
| 191 |
+
else: conv_str += f"<|{role}|>"
|
| 192 |
+
|
| 193 |
+
if isinstance(content, str):
|
| 194 |
+
conv_str += f"{content}"
|
| 195 |
+
elif isinstance(content, list):
|
| 196 |
+
for part in content:
|
| 197 |
+
ptype = part.get("type")
|
| 198 |
+
if ptype == "audio":
|
| 199 |
+
# ------------------------------------------------------------
|
| 200 |
+
# 修改点：解析 begin_time 和 end_time
|
| 201 |
+
# ------------------------------------------------------------
|
| 202 |
+
begin_time = part.get("begin_time", -1)
|
| 203 |
+
end_time = part.get("end_time", -1)
|
| 204 |
+
|
| 205 |
+
offset = 0.0
|
| 206 |
+
duration = None
|
| 207 |
+
|
| 208 |
+
# 只有当 begin_time >= 0 且有效时才应用切片
|
| 209 |
+
if begin_time is not None and begin_time >= 0:
|
| 210 |
+
offset = float(begin_time)
|
| 211 |
+
if end_time is not None and end_time > begin_time:
|
| 212 |
+
duration = float(end_time) - float(begin_time)
|
| 213 |
+
|
| 214 |
+
audio_raw_this = None
|
| 215 |
+
if "array" in part:
|
| 216 |
+
arr = part["array"]
|
| 217 |
+
if isinstance(arr, torch.Tensor):
|
| 218 |
+
arr = arr.detach().cpu().numpy()
|
| 219 |
+
full_arr = np.asarray(arr, dtype=np.float32).reshape(-1)
|
| 220 |
+
|
| 221 |
+
# 针对 array 的切片
|
| 222 |
+
start_idx = int(offset * sampling_rate)
|
| 223 |
+
end_idx = None
|
| 224 |
+
if duration is not None:
|
| 225 |
+
end_idx = start_idx + int(duration * sampling_rate)
|
| 226 |
+
audio_raw_this = full_arr[start_idx:end_idx]
|
| 227 |
+
|
| 228 |
+
elif "path" in part:
|
| 229 |
+
audio_raw_this = self._load_audio_file(
|
| 230 |
+
part["path"],
|
| 231 |
+
sampling_rate=sampling_rate,
|
| 232 |
+
offset=offset,
|
| 233 |
+
duration=duration
|
| 234 |
+
)
|
| 235 |
+
elif "base64" in part:
|
| 236 |
+
audio_raw_this = self._load_audio_base64(
|
| 237 |
+
part["base64"],
|
| 238 |
+
sampling_rate=sampling_rate,
|
| 239 |
+
offset=offset,
|
| 240 |
+
duration=duration
|
| 241 |
+
)
|
| 242 |
+
else:
|
| 243 |
+
raise ValueError("Audio part must contain 'path' or 'array' or 'base64'.")
|
| 244 |
+
|
| 245 |
+
audios_raw.append(audio_raw_this)
|
| 246 |
+
audio_count_this_conv += 1
|
| 247 |
+
conv_str += f"{self.bos_audio_token}{_AUDIO_MARKER}{self.eos_audio_token}"
|
| 248 |
+
|
| 249 |
+
elif ptype == "text":
|
| 250 |
+
conv_str += f"{part.get('text', '')}"
|
| 251 |
+
else:
|
| 252 |
+
raise ValueError(f"Unknown content part type: {ptype}")
|
| 253 |
+
else:
|
| 254 |
+
raise ValueError(f"Unsupported message content type: {type(content)}")
|
| 255 |
+
|
| 256 |
+
if add_generation_prompt:
|
| 257 |
+
if last_role == "user": conv_str += f"{self.assistant_token}"
|
| 258 |
+
elif last_role == "assistant": conv_str += f"{self.assistant_end_token}"
|
| 259 |
+
else: conv_str += f"{self.assistant_token}"
|
| 260 |
+
|
| 261 |
+
prompts_template.append(conv_str)
|
| 262 |
+
prompt_audio_counts.append(audio_count_this_conv)
|
| 263 |
+
|
| 264 |
+
return prompts_template, audios_raw, prompt_audio_counts
|
| 265 |
+
|
| 266 |
+
def _calculate_audio_token_counts_per_sample(
|
| 267 |
+
self,
|
| 268 |
+
audios_raw: List[np.ndarray],
|
| 269 |
+
sampling_rate: int,
|
| 270 |
+
audio_max_length: Optional[int],
|
| 271 |
+
audio_pad_to_multiple_of: Optional[int],
|
| 272 |
+
) -> List[int]:
|
| 273 |
+
del sampling_rate, audio_pad_to_multiple_of
|
| 274 |
+
|
| 275 |
+
hop_length = int(getattr(self.feature_extractor, "hop_length", 160))
|
| 276 |
+
max_audio_samples = int(audio_max_length) if audio_max_length is not None else None
|
| 277 |
+
token_counts: List[int] = []
|
| 278 |
+
|
| 279 |
+
for audio_raw in audios_raw:
|
| 280 |
+
audio_np = np.asarray(audio_raw, dtype=np.float32).reshape(-1)
|
| 281 |
+
effective_len = int(audio_np.shape[0])
|
| 282 |
+
if max_audio_samples is not None:
|
| 283 |
+
effective_len = min(effective_len, max_audio_samples)
|
| 284 |
+
|
| 285 |
+
mel_frames = effective_len // max(hop_length, 1)
|
| 286 |
+
token_counts.append(self.calculate_audio_token_count(int(mel_frames)))
|
| 287 |
+
|
| 288 |
+
return token_counts
|
| 289 |
+
|
| 290 |
+
# =========================
|
| 291 |
+
# apply_chat_template
|
| 292 |
+
# =========================
|
| 293 |
+
def apply_chat_template(
|
| 294 |
+
self,
|
| 295 |
+
conversation: Union[List[dict], List[List[dict]]],
|
| 296 |
+
chat_template: Optional[str] = None,
|
| 297 |
+
add_generation_prompt: bool = True,
|
| 298 |
+
**kwargs,
|
| 299 |
+
) -> Union[BatchFeature, str, List[str]]:
|
| 300 |
+
if chat_template is not None:
|
| 301 |
+
logger.warning("chat_template argument is ignored.")
|
| 302 |
+
|
| 303 |
+
tokenize = kwargs.pop("tokenize", True)
|
| 304 |
+
return_tensors = kwargs.pop("return_tensors", "pt")
|
| 305 |
+
kwargs.pop("return_dict", None)
|
| 306 |
+
|
| 307 |
+
audio_torch_dtype = kwargs.pop("audio_torch_dtype", None)
|
| 308 |
+
audio_dtype_override = kwargs.pop("audio_dtype", None)
|
| 309 |
+
dtype_source = audio_torch_dtype if audio_torch_dtype is not None else audio_dtype_override
|
| 310 |
+
target_dtype = _resolve_torch_dtype(dtype_source, default=getattr(self, "audio_dtype", "float32"))
|
| 311 |
+
|
| 312 |
+
text_kwargs = dict(kwargs.pop("text_kwargs", {}) or {})
|
| 313 |
+
for k in ("padding", "truncation", "max_length", "add_special_tokens"):
|
| 314 |
+
if k in kwargs and k not in text_kwargs:
|
| 315 |
+
text_kwargs[k] = kwargs.pop(k)
|
| 316 |
+
|
| 317 |
+
sampling_rate = int(kwargs.pop("sampling_rate", 16000))
|
| 318 |
+
audio_padding = kwargs.pop("audio_padding", "longest")
|
| 319 |
+
audio_max_length = kwargs.pop("audio_max_length", None)
|
| 320 |
+
audio_pad_to_multiple_of = kwargs.pop("audio_pad_to_multiple_of", None)
|
| 321 |
+
|
| 322 |
+
if kwargs:
|
| 323 |
+
logger.warning(f"Ignored unused kwargs: {list(kwargs.keys())}")
|
| 324 |
+
|
| 325 |
+
if isinstance(conversation, list) and conversation and isinstance(conversation[0], dict):
|
| 326 |
+
conversations = [conversation]
|
| 327 |
+
is_single = True
|
| 328 |
+
else:
|
| 329 |
+
conversations = conversation
|
| 330 |
+
is_single = False
|
| 331 |
+
|
| 332 |
+
prompt_templates, audios_raw, prompt_audio_counts = self._build_templates_and_audios(
|
| 333 |
+
conversations=conversations,
|
| 334 |
+
sampling_rate=sampling_rate,
|
| 335 |
+
add_generation_prompt=add_generation_prompt,
|
| 336 |
+
)
|
| 337 |
+
|
| 338 |
+
input_features = None
|
| 339 |
+
audio_token_counts: List[int] = []
|
| 340 |
+
|
| 341 |
+
if len(audios_raw) > 0:
|
| 342 |
+
feat = self.feature_extractor(
|
| 343 |
+
audios_raw,
|
| 344 |
+
sampling_rate=sampling_rate,
|
| 345 |
+
return_tensors="np",
|
| 346 |
+
return_attention_mask=False,
|
| 347 |
+
padding=audio_padding,
|
| 348 |
+
max_length=audio_max_length,
|
| 349 |
+
pad_to_multiple_of=audio_pad_to_multiple_of,
|
| 350 |
+
)
|
| 351 |
+
input_features = feat["input_features"]
|
| 352 |
+
if not isinstance(input_features, np.ndarray):
|
| 353 |
+
input_features = np.asarray(input_features)
|
| 354 |
+
|
| 355 |
+
audio_token_counts = self._calculate_audio_token_counts_per_sample(
|
| 356 |
+
audios_raw=audios_raw,
|
| 357 |
+
sampling_rate=sampling_rate,
|
| 358 |
+
audio_max_length=audio_max_length,
|
| 359 |
+
audio_pad_to_multiple_of=audio_pad_to_multiple_of,
|
| 360 |
+
)
|
| 361 |
+
|
| 362 |
+
prompts: List[str] = []
|
| 363 |
+
audio_idx = 0
|
| 364 |
+
for prompt_template, audio_count in zip(prompt_templates, prompt_audio_counts):
|
| 365 |
+
prompt = prompt_template
|
| 366 |
+
for _ in range(audio_count):
|
| 367 |
+
if audio_idx >= len(audio_token_counts):
|
| 368 |
+
raise ValueError("Audio token count mismatch while building prompts.")
|
| 369 |
+
audio_tokens_str = "".join([self.audio_token] * audio_token_counts[audio_idx])
|
| 370 |
+
prompt = prompt.replace(_AUDIO_MARKER, audio_tokens_str, 1)
|
| 371 |
+
audio_idx += 1
|
| 372 |
+
if _AUDIO_MARKER in prompt:
|
| 373 |
+
raise ValueError("Unresolved audio marker remained in prompt.")
|
| 374 |
+
prompts.append(prompt)
|
| 375 |
+
|
| 376 |
+
if audio_idx != len(audio_token_counts):
|
| 377 |
+
raise ValueError("Unused audio token counts remained after prompt construction.")
|
| 378 |
+
|
| 379 |
+
if not tokenize:
|
| 380 |
+
return prompts[0] if is_single else prompts
|
| 381 |
+
|
| 382 |
+
text_kwargs.setdefault("padding", "longest")
|
| 383 |
+
text_kwargs.setdefault("add_special_tokens", False)
|
| 384 |
+
text_kwargs["return_tensors"] = return_tensors
|
| 385 |
+
|
| 386 |
+
enc = self.tokenizer(prompts, **text_kwargs)
|
| 387 |
+
data: Dict[str, Any] = dict(enc)
|
| 388 |
+
|
| 389 |
+
if input_features is not None:
|
| 390 |
+
data["audios"] = torch.tensor(input_features, dtype=target_dtype)
|
| 391 |
+
|
| 392 |
+
return BatchFeature(data=data, tensor_type=return_tensors)
|
| 393 |
+
|
| 394 |
+
# ... (其余 batch_decode, decode, __call__, model_input_names 保持不变) ...
|
| 395 |
+
def batch_decode(self, *args, **kwargs):
|
| 396 |
+
return self.tokenizer.batch_decode(*args, **kwargs)
|
| 397 |
+
|
| 398 |
+
def decode(self, *args, **kwargs):
|
| 399 |
+
return self.tokenizer.decode(*args, **kwargs)
|
| 400 |
+
|
| 401 |
+
def __call__(
|
| 402 |
+
self,
|
| 403 |
+
text: Union[str, List[str]],
|
| 404 |
+
audios: Union[np.ndarray, torch.Tensor, List[Union[np.ndarray, torch.Tensor]]],
|
| 405 |
+
sampling_rate: int = 16000,
|
| 406 |
+
return_tensors: str = "pt",
|
| 407 |
+
**tokenizer_kwargs,
|
| 408 |
+
) -> BatchFeature:
|
| 409 |
+
# 简化版实现，不包含时间切片逻辑，因为直接传入的是 audio array
|
| 410 |
+
audios_list = []
|
| 411 |
+
def flatten_audios(obj):
|
| 412 |
+
if isinstance(obj, (list, tuple)):
|
| 413 |
+
if len(obj) > 0 and isinstance(obj[0], (float, int)):
|
| 414 |
+
audios_list.append(obj)
|
| 415 |
+
else:
|
| 416 |
+
for item in obj: flatten_audios(item)
|
| 417 |
+
elif isinstance(obj, (np.ndarray, torch.Tensor)):
|
| 418 |
+
audios_list.append(obj)
|
| 419 |
+
flatten_audios(audios)
|
| 420 |
+
|
| 421 |
+
audios_np: List[np.ndarray] = []
|
| 422 |
+
for a in audios_list:
|
| 423 |
+
if isinstance(a, torch.Tensor): a = a.detach().cpu().numpy()
|
| 424 |
+
a = np.asarray(a, dtype=np.float32).reshape(-1)
|
| 425 |
+
audios_np.append(a)
|
| 426 |
+
|
| 427 |
+
input_features = None
|
| 428 |
+
if audios_np:
|
| 429 |
+
feat = self.feature_extractor(audios_np, sampling_rate=int(sampling_rate), return_tensors="np", return_attention_mask=False, padding="longest")
|
| 430 |
+
input_features = feat["input_features"]
|
| 431 |
+
if not isinstance(input_features, np.ndarray): input_features = np.asarray(input_features)
|
| 432 |
+
|
| 433 |
+
tokenizer_kwargs = dict(tokenizer_kwargs or {})
|
| 434 |
+
tokenizer_kwargs.setdefault("padding", "longest")
|
| 435 |
+
tokenizer_kwargs.setdefault("add_special_tokens", False)
|
| 436 |
+
tokenizer_kwargs["return_tensors"] = return_tensors
|
| 437 |
+
|
| 438 |
+
enc = self.tokenizer(text, **tokenizer_kwargs)
|
| 439 |
+
data: Dict[str, Any] = dict(enc)
|
| 440 |
+
if input_features is not None:
|
| 441 |
+
data["audios"] = torch.tensor(input_features, dtype=_resolve_torch_dtype(getattr(self, "audio_dtype", "float32")))
|
| 442 |
+
return BatchFeature(data=data, tensor_type=return_tensors)
|
| 443 |
+
|
| 444 |
+
@property
|
| 445 |
+
def model_input_names(self):
|
| 446 |
+
return ["input_ids", "attention_mask", "audios"]
|
processor_config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"audio_dtype": "bfloat16",
|
| 3 |
+
"audio_token": "<|audio|>",
|
| 4 |
+
"auto_map": {
|
| 5 |
+
"AutoProcessor": "processing_arkasr.ArkasrProcessor"
|
| 6 |
+
},
|
| 7 |
+
"merge_factor": 4,
|
| 8 |
+
"processor_class": "ArkasrProcessor"
|
| 9 |
+
}
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"additional_special_tokens": [
|
| 3 |
+
"<|user|>",
|
| 4 |
+
"<|begin_of_audio|>",
|
| 5 |
+
"<|end_of_audio|>",
|
| 6 |
+
"<|assistant|>",
|
| 7 |
+
"<|system|>"
|
| 8 |
+
],
|
| 9 |
+
"eos_token": {
|
| 10 |
+
"content": "<|im_end|>",
|
| 11 |
+
"lstrip": false,
|
| 12 |
+
"normalized": false,
|
| 13 |
+
"rstrip": false,
|
| 14 |
+
"single_word": false
|
| 15 |
+
},
|
| 16 |
+
"pad_token": {
|
| 17 |
+
"content": "<|endoftext|>",
|
| 18 |
+
"lstrip": false,
|
| 19 |
+
"normalized": false,
|
| 20 |
+
"rstrip": false,
|
| 21 |
+
"single_word": false
|
| 22 |
+
}
|
| 23 |
+
}
|
tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bc63972a406328b950c3dea5f64994846a95e444609657d440f0f2ecc4721b32
|
| 3 |
+
size 11422866
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,239 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": false,
|
| 3 |
+
"add_prefix_space": false,
|
| 4 |
+
"added_tokens_decoder": {
|
| 5 |
+
"151643": {
|
| 6 |
+
"content": "<|endoftext|>",
|
| 7 |
+
"lstrip": false,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false,
|
| 11 |
+
"special": true
|
| 12 |
+
},
|
| 13 |
+
"151644": {
|
| 14 |
+
"content": "<|im_start|>",
|
| 15 |
+
"lstrip": false,
|
| 16 |
+
"normalized": false,
|
| 17 |
+
"rstrip": false,
|
| 18 |
+
"single_word": false,
|
| 19 |
+
"special": true
|
| 20 |
+
},
|
| 21 |
+
"151645": {
|
| 22 |
+
"content": "<|im_end|>",
|
| 23 |
+
"lstrip": false,
|
| 24 |
+
"normalized": false,
|
| 25 |
+
"rstrip": false,
|
| 26 |
+
"single_word": false,
|
| 27 |
+
"special": true
|
| 28 |
+
},
|
| 29 |
+
"151646": {
|
| 30 |
+
"content": "<|object_ref_start|>",
|
| 31 |
+
"lstrip": false,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false,
|
| 35 |
+
"special": true
|
| 36 |
+
},
|
| 37 |
+
"151647": {
|
| 38 |
+
"content": "<|object_ref_end|>",
|
| 39 |
+
"lstrip": false,
|
| 40 |
+
"normalized": false,
|
| 41 |
+
"rstrip": false,
|
| 42 |
+
"single_word": false,
|
| 43 |
+
"special": true
|
| 44 |
+
},
|
| 45 |
+
"151648": {
|
| 46 |
+
"content": "<|start_global_token|>",
|
| 47 |
+
"lstrip": false,
|
| 48 |
+
"normalized": false,
|
| 49 |
+
"rstrip": false,
|
| 50 |
+
"single_word": false,
|
| 51 |
+
"special": true
|
| 52 |
+
},
|
| 53 |
+
"151649": {
|
| 54 |
+
"content": "<|end_global_token|>",
|
| 55 |
+
"lstrip": false,
|
| 56 |
+
"normalized": false,
|
| 57 |
+
"rstrip": false,
|
| 58 |
+
"single_word": false,
|
| 59 |
+
"special": true
|
| 60 |
+
},
|
| 61 |
+
"151650": {
|
| 62 |
+
"content": "<|start_content|>",
|
| 63 |
+
"lstrip": false,
|
| 64 |
+
"normalized": false,
|
| 65 |
+
"rstrip": false,
|
| 66 |
+
"single_word": false,
|
| 67 |
+
"special": true
|
| 68 |
+
},
|
| 69 |
+
"151651": {
|
| 70 |
+
"content": "<|end_content|>",
|
| 71 |
+
"lstrip": false,
|
| 72 |
+
"normalized": false,
|
| 73 |
+
"rstrip": false,
|
| 74 |
+
"single_word": false,
|
| 75 |
+
"special": true
|
| 76 |
+
},
|
| 77 |
+
"151652": {
|
| 78 |
+
"content": "<|vision_start|>",
|
| 79 |
+
"lstrip": false,
|
| 80 |
+
"normalized": false,
|
| 81 |
+
"rstrip": false,
|
| 82 |
+
"single_word": false,
|
| 83 |
+
"special": true
|
| 84 |
+
},
|
| 85 |
+
"151653": {
|
| 86 |
+
"content": "<|vision_end|>",
|
| 87 |
+
"lstrip": false,
|
| 88 |
+
"normalized": false,
|
| 89 |
+
"rstrip": false,
|
| 90 |
+
"single_word": false,
|
| 91 |
+
"special": true
|
| 92 |
+
},
|
| 93 |
+
"151654": {
|
| 94 |
+
"content": "<|vision_pad|>",
|
| 95 |
+
"lstrip": false,
|
| 96 |
+
"normalized": false,
|
| 97 |
+
"rstrip": false,
|
| 98 |
+
"single_word": false,
|
| 99 |
+
"special": true
|
| 100 |
+
},
|
| 101 |
+
"151655": {
|
| 102 |
+
"content": "<|image_pad|>",
|
| 103 |
+
"lstrip": false,
|
| 104 |
+
"normalized": false,
|
| 105 |
+
"rstrip": false,
|
| 106 |
+
"single_word": false,
|
| 107 |
+
"special": true
|
| 108 |
+
},
|
| 109 |
+
"151656": {
|
| 110 |
+
"content": "<|video_pad|>",
|
| 111 |
+
"lstrip": false,
|
| 112 |
+
"normalized": false,
|
| 113 |
+
"rstrip": false,
|
| 114 |
+
"single_word": false,
|
| 115 |
+
"special": true
|
| 116 |
+
},
|
| 117 |
+
"151657": {
|
| 118 |
+
"content": "<tool_call>",
|
| 119 |
+
"lstrip": false,
|
| 120 |
+
"normalized": false,
|
| 121 |
+
"rstrip": false,
|
| 122 |
+
"single_word": false,
|
| 123 |
+
"special": false
|
| 124 |
+
},
|
| 125 |
+
"151658": {
|
| 126 |
+
"content": "</tool_call>",
|
| 127 |
+
"lstrip": false,
|
| 128 |
+
"normalized": false,
|
| 129 |
+
"rstrip": false,
|
| 130 |
+
"single_word": false,
|
| 131 |
+
"special": false
|
| 132 |
+
},
|
| 133 |
+
"151659": {
|
| 134 |
+
"content": "<|fim_prefix|>",
|
| 135 |
+
"lstrip": false,
|
| 136 |
+
"normalized": false,
|
| 137 |
+
"rstrip": false,
|
| 138 |
+
"single_word": false,
|
| 139 |
+
"special": false
|
| 140 |
+
},
|
| 141 |
+
"151660": {
|
| 142 |
+
"content": "<|fim_middle|>",
|
| 143 |
+
"lstrip": false,
|
| 144 |
+
"normalized": false,
|
| 145 |
+
"rstrip": false,
|
| 146 |
+
"single_word": false,
|
| 147 |
+
"special": false
|
| 148 |
+
},
|
| 149 |
+
"151661": {
|
| 150 |
+
"content": "<|fim_suffix|>",
|
| 151 |
+
"lstrip": false,
|
| 152 |
+
"normalized": false,
|
| 153 |
+
"rstrip": false,
|
| 154 |
+
"single_word": false,
|
| 155 |
+
"special": false
|
| 156 |
+
},
|
| 157 |
+
"151662": {
|
| 158 |
+
"content": "<|fim_pad|>",
|
| 159 |
+
"lstrip": false,
|
| 160 |
+
"normalized": false,
|
| 161 |
+
"rstrip": false,
|
| 162 |
+
"single_word": false,
|
| 163 |
+
"special": false
|
| 164 |
+
},
|
| 165 |
+
"151663": {
|
| 166 |
+
"content": "<|audio|>",
|
| 167 |
+
"lstrip": false,
|
| 168 |
+
"normalized": false,
|
| 169 |
+
"rstrip": false,
|
| 170 |
+
"single_word": false,
|
| 171 |
+
"special": false
|
| 172 |
+
},
|
| 173 |
+
"151664": {
|
| 174 |
+
"content": "<|file_sep|>",
|
| 175 |
+
"lstrip": false,
|
| 176 |
+
"normalized": false,
|
| 177 |
+
"rstrip": false,
|
| 178 |
+
"single_word": false,
|
| 179 |
+
"special": false
|
| 180 |
+
},
|
| 181 |
+
"151665": {
|
| 182 |
+
"content": "<|user|>",
|
| 183 |
+
"lstrip": false,
|
| 184 |
+
"normalized": false,
|
| 185 |
+
"rstrip": false,
|
| 186 |
+
"single_word": false,
|
| 187 |
+
"special": true
|
| 188 |
+
},
|
| 189 |
+
"151666": {
|
| 190 |
+
"content": "<|begin_of_audio|>",
|
| 191 |
+
"lstrip": false,
|
| 192 |
+
"normalized": false,
|
| 193 |
+
"rstrip": false,
|
| 194 |
+
"single_word": false,
|
| 195 |
+
"special": true
|
| 196 |
+
},
|
| 197 |
+
"151667": {
|
| 198 |
+
"content": "<|end_of_audio|>",
|
| 199 |
+
"lstrip": false,
|
| 200 |
+
"normalized": false,
|
| 201 |
+
"rstrip": false,
|
| 202 |
+
"single_word": false,
|
| 203 |
+
"special": true
|
| 204 |
+
},
|
| 205 |
+
"151668": {
|
| 206 |
+
"content": "<|assistant|>",
|
| 207 |
+
"lstrip": false,
|
| 208 |
+
"normalized": false,
|
| 209 |
+
"rstrip": false,
|
| 210 |
+
"single_word": false,
|
| 211 |
+
"special": true
|
| 212 |
+
},
|
| 213 |
+
"151669": {
|
| 214 |
+
"content": "<|system|>",
|
| 215 |
+
"lstrip": false,
|
| 216 |
+
"normalized": false,
|
| 217 |
+
"rstrip": false,
|
| 218 |
+
"single_word": false,
|
| 219 |
+
"special": true
|
| 220 |
+
}
|
| 221 |
+
},
|
| 222 |
+
"additional_special_tokens": [
|
| 223 |
+
"<|user|>",
|
| 224 |
+
"<|begin_of_audio|>",
|
| 225 |
+
"<|end_of_audio|>",
|
| 226 |
+
"<|assistant|>",
|
| 227 |
+
"<|system|>"
|
| 228 |
+
],
|
| 229 |
+
"bos_token": null,
|
| 230 |
+
"clean_up_tokenization_spaces": false,
|
| 231 |
+
"eos_token": "<|im_end|>",
|
| 232 |
+
"errors": "replace",
|
| 233 |
+
"extra_special_tokens": {},
|
| 234 |
+
"model_max_length": 131072,
|
| 235 |
+
"pad_token": "<|endoftext|>",
|
| 236 |
+
"split_special_tokens": false,
|
| 237 |
+
"tokenizer_class": "Qwen2Tokenizer",
|
| 238 |
+
"unk_token": null
|
| 239 |
+
}
|
vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|