프로그램 사용/ai 프로그램2026. 5. 11. 14:58

텔레그램 봇 업데이트 / 그림 그리기

음성인식 - 그림 들어가면 그리기

 

 

그림 넣고 캡션에 그림 있으면 img2img로 작동시키기

저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

rag 관련 라이브러리, 방법 조사중  (0) 2026.05.12
stable diffusion 모델 변경(SD -> SDXL)  (0) 2026.05.12
litert-lm 와 gemma4-e2b mtp 일단 실패  (0) 2026.05.10
vLLM  (0) 2026.05.10
Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice 시도 실패  (0) 2026.05.10
Posted by 구차니
카테고리 없음2026. 5. 11. 06:33

둠 이터널 dlc / ancient god part 1

길 못 찾아서 찾아봄 ㅠㅠ

[링크 : https://youtu.be/AZJuH2eG3Ug?t=3220]

저작자표시 (새창열림)
Posted by 구차니
프로그램 사용/ai 프로그램2026. 5. 10. 18:53

litert-lm 와 gemma4-e2b mtp 일단 실패

흐음.. 신기한걸 알았는데.. 또 파이썬이네.

핸드폰에서 돌려놓기도 하는거 보면 제법 가볍긴 한다 보다.

 

 

D:\study\llm>pip install litert-lm
D:\study\llm>litert-lm
CLI tool for LiteRT-LM models.

Usage: litert-lm [OPTIONS] COMMAND [ARGS]...

Commands:
  benchmark  Benchmarks a LiteRT-LM model.
  delete     Deletes a model from the local storage.
  import     Imports a model from a local path or HuggingFace hub.
  list       Lists all imported LiteRT-LM models.
  rename     Renames a model.
  run        Runs a LiteRT-LM model interactively or with a single prompt.
  serve      Start a server with a Gemini or OpenAI compatible API (alpha feature)

Global options:
  --version   Show the version and exit.
  -h, --help  Show this message and exit.

D:\study\llm>litert-lm run --from-huggingface-repo=litert-community/gemma-4-E4B-it-litert-lm gemma-4-E4B-it.litertlm --backend=gpu  --enable-speculative-decoding=true --prompt="What is the capital of France?"
Downloading gemma-4-E4B-it.litertlm from litert-community/gemma-4-E4B-it-litert-lm...
gemma-4-E4B-it.litertlm:   0%|                                                 | 92.3k/3.66G [00:01<16:44:38, 60.7kB/s]
gemma-4-E4B-it.litertlm: 100%|████████████████████████████████████████████████████| 3.66G/3.66G [05:00<00:00, 12.2MB/s]
C:\Users\minimonk\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\file_download.py:143: UserWarning: `huggingface_hub` cache-system uses symlinks by default to efficiently store duplicated files but your machine does not support them in C:\Users\minimonk\.cache\huggingface\hub\models--litert-community--gemma-4-E4B-it-litert-lm. Caching files will still work but in a degraded version that might require more space on your disk. This warning can be disabled by setting the `HF_HUB_DISABLE_SYMLINKS_WARNING` environment variable. For more details, see https://huggingface.co/docs/huggingface_hub/how-to-cache#limitations.
To support symlinks on Windows, you either need to activate Developer Mode or to run Python as an administrator. In order to activate developer mode, see this article: https://docs.microsoft.com/en-us/windows/apps/get-started/enable-your-device-for-development
  warnings.warn(message)
The capital of France is **Paris**.

 

 

찾아보니 저장소의 파일명이 그런거였군.

[링크 : https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm/tree/main]

 

[링크 : https://huggingface.co/metricspace/gemma4-E2B-it-litert-128k-mtp/tree/main]

 

mtp 들어가면서 안되나?

D:\study\llm>litert-lm run --from-huggingface-repo=metricspace/gemma4-E2B-it-litert-128k-mtp model.litertlm --backend=gpu  --enable-speculative-decoding=true --prompt="What is the capital of France?"
Downloading model.litertlm from metricspace/gemma4-E2B-it-litert-128k-mtp...
E0000 00:00:1778407860.973266    8280 delegate_webgpu.cc:373] Failed to create litert::ml_drift::DelegateKernelLiteRt: RESOURCE_EXHAUSTED: Requested allocation size - 4294967296 bytes. Max allocation size for this GPU - 2147483648 bytes. Shape - {bhwdc, {1, 1, 8192, 1, 131072}}, data type - float32.
=== Source Location Trace: ===
third_party/ml_drift/common/task/tensor_desc.cc:1846
third_party/ml_drift/common/gpu_model_util.cc:232
third_party/ml_drift/common/gpu_model_util.cc:269
third_party/ml_drift/common/gpu_model_util.cc:432
third_party/odml/litert/ml_drift/delegate/delegate_kernel.cc:765
third_party/odml/litert/ml_drift/delegate/delegate_kernel.cc:695
third_party/odml/litert/ml_drift/delegate/delegate_kernel.cc:787
third_party/odml/litert/ml_drift/delegate/delegate_kernel.cc:284
third_party/odml/litert/ml_drift/delegate/delegate_kernel_litert.cc:167
ERROR: Failed to initialize kernel.
ERROR: Node number 223 (STABLEHLO_COMPOSITE) failed to prepare.
E0000 00:00:1778407862.768911    8280 engine.cc:491] Failed to create engine: INTERNAL: ERROR: [third_party/odml/litert_lm/runtime/executor/llm_litert_compiled_model_executor.cc:1928]
??ERROR: [./third_party/odml/litert/litert/cc/litert_compiled_model.h:1780]
=== Source Location Trace: ===
./third_party/odml/litert/litert/cc/litert_macros.h:538
third_party/odml/litert_lm/runtime/executor/llm_litert_compiled_model_executor_factory.cc:144
third_party/odml/litert_lm/runtime/core/engine_impl.cc:384
An error occurred
Traceback (most recent call last):
  File "C:\Users\minimonk\AppData\Local\Programs\Python\Python310\lib\site-packages\litert_lm_cli\model.py", line 255, in run_interactive
    engine_cm = litert_lm.Engine(
  File "C:\Users\ minimonk \AppData\Local\Programs\Python\Python310\lib\site-packages\litert_lm\engine.py", line 82, in __init__
    raise RuntimeError(
RuntimeError: Failed to create LiteRT-LM engine for C:\Users\ minimonk \.cache\huggingface\hub\models--metricspace--gemma4-E2B-it-litert-128k-mtp\snapshots\4dae3505f550397923c206eaa63be84f17ee43cb\model.litertlm

 

 

[링크 : https://github.com/google-ai-edge/LiteRT-LM]

[링크 : https://huggingface.co/metricspace/gemma4-E2B-it-litert-128k-mtp]

[링크 : https://pypi.org/project/litert-lm/]

[링크 : https://www.reddit.com/r/LocalLLaMA/comments/1somixt/practical_local_llm_on_android_gemma_4_via/?tl=ko]

저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

stable diffusion 모델 변경(SD -> SDXL)  (0) 2026.05.12
텔레그램 봇 업데이트 / 그림 그리기  (0) 2026.05.11
vLLM  (0) 2026.05.10
Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice 시도 실패  (0) 2026.05.10
supertone/supertonic3 시도  (0) 2026.05.10
Posted by 구차니
프로그램 사용/ai 프로그램2026. 5. 10. 18:34

vLLM

python 에서 돌아가는 녀석인 듯.

 

[링크 : https://vllm.ai/]

 

[링크 : https://github.com/vllm-project/vllm/releases]

저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

텔레그램 봇 업데이트 / 그림 그리기  (0) 2026.05.11
litert-lm 와 gemma4-e2b mtp 일단 실패  (0) 2026.05.10
Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice 시도 실패  (0) 2026.05.10
supertone/supertonic3 시도  (0) 2026.05.10
outetts 시도  (0) 2026.05.10
Posted by 구차니
프로그램 사용/ai 프로그램2026. 5. 10. 18:32

Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice 시도 실패

qwen 형님으로 모셔야 하나 ㅋㅋㅋ

 

D:\study\llm>pip install soundfile torch qwen_tts
D:\study\llm>python
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> import soundfile as sf
>>> from qwen_tts import Qwen3TTSModel

********
Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.
********

'sox' is not recognized as an internal or external command,
operable program or batch file.
SoX could not be found!

    If you do not have SoX, proceed here:
     - - - http://sox.sourceforge.net/ - - -

    If you do (or think that you should) have SoX, double-check your
    path variables.

>>>
>>> model = Qwen3TTSModel.from_pretrained(
...     "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
...     device_map="cuda:0",
...     dtype=torch.bfloat16,
...     attn_implementation="flash_attention_2",
... )
config.json: 4.91kB [00:00, 4.70MB/s]
C:\Users\minimonk\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\file_download.py:143: UserWarning: `huggingface_hub` cache-system uses symlinks by default to efficiently store duplicated files but your machine does not support them in C:\Users\minimonk\.cache\huggingface\hub\models--Qwen--Qwen3-TTS-12Hz-1.7B-CustomVoice. Caching files will still work but in a degraded version that might require more space on your disk. This warning can be disabled by setting the `HF_HUB_DISABLE_SYMLINKS_WARNING` environment variable. For more details, see https://huggingface.co/docs/huggingface_hub/how-to-cache#limitations.
To support symlinks on Windows, you either need to activate Developer Mode or to run Python as an administrator. In order to activate developer mode, see this article: https://docs.microsoft.com/en-us/windows/apps/get-started/enable-your-device-for-development
  warnings.warn(message)
model.safetensors:   0%|                                                                   | 0.00/3.83G [00:00<?, ?B/s]

model.safetensors: 100%|██████████████████████████████████████████████████████████| 3.83G/3.83G [04:45<00:00, 13.4MB/s]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Users\minimonk\AppData\Local\Programs\Python\Python310\lib\site-packages\qwen_tts\inference\qwen3_tts_model.py", line 112, in from_pretrained
    model = AutoModel.from_pretrained(pretrained_model_name_or_path, **kwargs)
  File "C:\ Users\minimonk\AppData \Local\Programs\Python\Python310\lib\site-packages\transformers\models\auto\auto_factory.py", line 604, in from_pretrained
    return model_class.from_pretrained(
  File "C:\ Users\minimonk\AppData \Local\Programs\Python\Python310\lib\site-packages\qwen_tts\core\models\modeling_qwen3_tts.py", line 1876, in from_pretrained
    model = super().from_pretrained(
  File "C:\ Users\minimonk\AppData \Local\Programs\Python\Python310\lib\site-packages\transformers\modeling_utils.py", line 277, in _wrapper
    return func(*args, **kwargs)
  File "C:\ Users\minimonk\AppData \Local\Programs\Python\Python310\lib\site-packages\transformers\modeling_utils.py", line 4971, in from_pretrained
    model = cls(config, *model_args, **model_kwargs)
  File "C:\ Users\minimonk\AppData \Local\Programs\Python\Python310\lib\site-packages\qwen_tts\core\models\modeling_qwen3_tts.py", line 1817, in __init__
    super().__init__(config)
  File "C:\ Users\minimonk\AppData \Local\Programs\Python\Python310\lib\site-packages\transformers\modeling_utils.py", line 2076, in __init__
    self.config._attn_implementation_internal = self._check_and_adjust_attn_implementation(
  File "C:\ Users\minimonk\AppData \Local\Programs\Python\Python310\lib\site-packages\transformers\modeling_utils.py", line 2686, in _check_and_adjust_attn_implementation
    applicable_attn_implementation = self.get_correct_attn_implementation(
  File "C:\ Users\minimonk\AppData \Local\Programs\Python\Python310\lib\site-packages\transformers\modeling_utils.py", line 2714, in get_correct_attn_implementation
    self._flash_attn_2_can_dispatch(is_init_check)
  File "C:\ Users\minimonk\AppData \Local\Programs\Python\Python310\lib\site-packages\transformers\modeling_utils.py", line 2422, in _flash_attn_2_can_dispatch
    raise ImportError(f"{preface} the package flash_attn seems to be not installed. {install_message}")
ImportError: FlashAttention2 has been toggled on, but it cannot be used due to the following error: the package flash_attn seems to be not installed. Please refer to the documentation of https://huggingface.co/docs/transformers/perf_infer_gpu_one#flashattention-2 to install Flash Attention 2.

 

에라이, 역시 리눅스 환경 기준으로 해야하나?

D:\study\llm>pip install flash_attn
Collecting flash_attn
  Using cached flash_attn-2.8.3.tar.gz (8.4 MB)
ERROR: Could not install packages due to an OSError: [Errno 2] No such file or directory: 'C:\\Users\\minimonk\\AppData\\Local\\Temp\\pip-install-gkk0v5su\\flash-attn_bdc9b907b4714d19aa80016a5ecbd8e6\\csrc/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp'
HINT: This error might have occurred since this system does not have Windows Long Path support enabled. You can find information on how to enable this at https://pip.pypa.io/warnings/enable-long-paths

 

화자와 언어가 달라도 될까 궁금하네

[링크 : https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice]

저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

litert-lm 와 gemma4-e2b mtp 일단 실패  (0) 2026.05.10
vLLM  (0) 2026.05.10
supertone/supertonic3 시도  (0) 2026.05.10
outetts 시도  (0) 2026.05.10
huggingface 에서 다운로드 받기(python)  (0) 2026.05.10
Posted by 구차니
프로그램 사용/ai 프로그램2026. 5. 10. 18:19

supertone/supertonic3 시도

알아서 받고 한글도 정말 잘 변환해준다.

잠시 검색해보니 한국 회사인것 같고. hybe 자회사로 게임같은데서 보이스 체인저로 유명한 듯?

라이센스는 좀 읽어 봐야겠지만 대충 번역기 돌려서 보니 SaaS 까지도 허용하는 것 같긴한데..

 

outetts 처럼 빌드는 필요없이 그냥 pip만으로 설치되니 good!

그리고 auto_download 하면 먼가 열심히 받고 알아서 한다.

D:\study\llm>pip install supertonic
D:\study\llm>python
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> from supertonic import TTS
>>> tts = TTS(auto_download=True)
Downloading (incomplete total...): 0.00B [00:00, ?B/s]                                                                 Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 26 files: 100%|███████████████████████████████████████████████████████████████| 26/26 [00:36<00:00,  1.40s/it]
Download complete: : 404MB [00:36, 19.4MB/s]                                                                           >>> style = tts.get_voice_style(voice_name="M1")
>>>
>>> text = "A gentle breeze moved through the open window while everyone listened to the story."
>>> wav, duration = tts.synthesize(text, voice_style=style, lang="en")
>>>
>>> tts.save_audio(wav, "output.wav")
>>> print(f"Generated {duration:.2f}s of audio")

>>> text = "안녕? 난 잼미니야 만나서 반가워"
>>> wav, duration = tts.synthesize(text, voice_style=style, lang="ko")
>>> tts.save_audio(wav, "output_ko.wav")

 

[링크 : https://huggingface.co/Supertone/supertonic-3]

[링크 : https://www.supertone.ai/ko]

저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

vLLM  (0) 2026.05.10
Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice 시도 실패  (0) 2026.05.10
outetts 시도  (0) 2026.05.10
huggingface 에서 다운로드 받기(python)  (0) 2026.05.10
stable diffusion 사용법  (0) 2026.05.09
Posted by 구차니
프로그램 사용/ai 프로그램2026. 5. 10. 17:36

outetts 시도

윈도우에서 하려고 했더니

step 1에서 바로 좌절. 먼가 그럼 미친듯이 깔지 말고 컴파일러 부터 확인하고 가라고!!! 버럭버럭!

D:\study\llm> pip install outetts
      *** CMake configuration failed
      [end of output]

  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for llama-cpp-python
Failed to build llama-cpp-python
error: failed-wheel-build-for-install

× Failed to build installable wheels for some pyproject.toml based projects
╰─> llama-cpp-python

D:\study\llm>

[링크 : https://github.com/edwko/OuteTTS?tab=readme-ov-file#installation]

[링크 : https://huggingface.co/unsloth/Llama-OuteTTS-1.0-1B]

 

 

Running the example
With both of the models generated, the LLM model and the voice decoder model,
we can run the example:

$ build/bin/llama-tts -m  ./models/outetts-0.2-0.5B-q8_0.gguf \
    -mv ./models/wavtokenizer-large-75-f16.gguf \
    -p "Hello world"
...
main: audio written to file 'output.wav'

[링크 : https://git.comtegra.pl/ajastrzebski/llama-cpp/-/tree/master/examples/tts]

[링크 : https://huggingface.co/OuteAI/OuteTTS-0.2-500M-GGUF/tree/main]

[링크 : https://huggingface.co/ggml-org/WavTokenizer/tree/main]

 

D:\study\llm\llama-b9093-bin-win-cuda-12.4-x64>llama-tts -m ..\OuteTTS-0.3-500M-Q8_0.gguf  -mv ..\WavTokenizer-Large-75-F16.gguf -p "hello i am sam. how are you?"
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 6143 MiB):
  Device 0: NVIDIA GeForce GTX 1060 6GB, compute capability 6.1, VMM: yes, VRAM: 6143 MiB
load_backend: loaded CUDA backend from D:\study\llm\llama-b9093-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from D:\study\llm\llama-b9093-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from D:\study\llm\llama-b9093-bin-win-cuda-12.4-x64\ggml-cpu-haswell.dll
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
common_params_fit_impl: getting device memory data for initial parameters:
common_memory_breakdown_print: | memory breakdown [MiB]   | total   free    self   model   context   compute    unaccounted |
common_memory_breakdown_print: |   - CUDA0 (GTX 1060 6GB) |  6143 = 5197 + ( 931 =   506 +      96 +     329) +          15 |
common_memory_breakdown_print: |   - Host                 |                  162 =   143 +       0 +      19                |
common_params_fit_impl: projected to use 931 MiB of device memory vs. 5197 MiB of free device memory
common_params_fit_impl: will leave 4265 >= 1024 MiB of free device memory, no changes needed
common_fit_params: successfully fit params to free device memory
common_fit_params: fitting params to free memory took 0.44 seconds
llama_model_loader: loaded meta data with 25 key-value pairs and 290 tensors from ..\OuteTTS-0.3-500M-Q8_0.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = OuteTTS 0.3 500M
llama_model_loader: - kv   3:                           general.basename str              = OuteTTS-0.3
llama_model_loader: - kv   4:                         general.size_label str              = 500M
llama_model_loader: - kv   5:                          qwen2.block_count u32              = 24
llama_model_loader: - kv   6:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv   7:                     qwen2.embedding_length u32              = 896
llama_model_loader: - kv   8:                  qwen2.feed_forward_length u32              = 4864
llama_model_loader: - kv   9:                 qwen2.attention.head_count u32              = 14
llama_model_loader: - kv  10:              qwen2.attention.head_count_kv u32              = 2
llama_model_loader: - kv  11:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,157696]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,157696]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 151644
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 151645
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  22:                    tokenizer.chat_template str              = outetts-0.3
llama_model_loader: - kv  23:               general.quantization_version u32              = 2
llama_model_loader: - kv  24:                          general.file_type u32              = 7
llama_model_loader: - type  f32:  121 tensors
llama_model_loader: - type q8_0:  169 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q8_0
print_info: file size   = 506.02 MiB (8.50 BPW)
llama_prepare_model_devices: using device CUDA0 (NVIDIA GeForce GTX 1060 6GB) (0000:01:00.0) - 5197 MiB free
load: 0 unused tokens
load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
load: printing all EOG tokens:
load:   - 128247 ('</s>')
load:   - 151643 ('<|endoftext|>')
load:   - 151645 ('<|im_end|>')
load:   - 151662 ('<|fim_pad|>')
load:   - 151663 ('<|repo_name|>')
load:   - 151664 ('<|file_sep|>')
load: special tokens cache size = 5152
load: token to piece cache size = 0.9712 MB
print_info: arch                  = qwen2
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 32768
print_info: n_embd                = 896
print_info: n_embd_inp            = 896
print_info: n_layer               = 24
print_info: n_head                = 14
print_info: n_head_kv             = 2
print_info: n_rot                 = 64
print_info: n_swa                 = 0
print_info: is_swa_any            = 0
print_info: n_embd_head_k         = 64
print_info: n_embd_head_v         = 64
print_info: n_gqa                 = 7
print_info: n_embd_k_gqa          = 128
print_info: n_embd_v_gqa          = 128
print_info: f_norm_eps            = 0.0e+00
print_info: f_norm_rms_eps        = 1.0e-06
print_info: f_clamp_kqv           = 0.0e+00
print_info: f_max_alibi_bias      = 0.0e+00
print_info: f_logit_scale         = 0.0e+00
print_info: f_attn_scale          = 0.0e+00
print_info: f_attn_value_scale    = 0.0000
print_info: n_ff                  = 4864
print_info: n_expert              = 0
print_info: n_expert_used         = 0
print_info: n_expert_groups       = 0
print_info: n_group_used          = 0
print_info: causal attn           = 1
print_info: pooling type          = -1
print_info: rope type             = 2
print_info: rope scaling          = linear
print_info: freq_base_train       = 1000000.0
print_info: freq_scale_train      = 1
print_info: n_ctx_orig_yarn       = 32768
print_info: rope_yarn_log_mul     = 0.0000
print_info: rope_finetuned        = unknown
print_info: model type            = 1B
print_info: model params          = 499.19 M
print_info: general.name          = OuteTTS 0.3 500M
print_info: vocab type            = BPE
print_info: n_vocab               = 157696
print_info: n_merges              = 151387
print_info: BOS token             = 151644 '<|im_start|>'
print_info: EOS token             = 151645 '<|im_end|>'
print_info: EOT token             = 151645 '<|im_end|>'
print_info: PAD token             = 151645 '<|im_end|>'
print_info: LF token              = 198 'Ċ'
print_info: FIM PRE token         = 151659 '<|fim_prefix|>'
print_info: FIM SUF token         = 151661 '<|fim_suffix|>'
print_info: FIM MID token         = 151660 '<|fim_middle|>'
print_info: FIM PAD token         = 151662 '<|fim_pad|>'
print_info: FIM REP token         = 151663 '<|repo_name|>'
print_info: FIM SEP token         = 151664 '<|file_sep|>'
print_info: EOG token             = 128247 '</s>'
print_info: EOG token             = 151643 '<|endoftext|>'
print_info: EOG token             = 151645 '<|im_end|>'
print_info: EOG token             = 151662 '<|fim_pad|>'
print_info: EOG token             = 151663 '<|repo_name|>'
print_info: EOG token             = 151664 '<|file_sep|>'
print_info: max token length      = 256
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 23 repeating layers to GPU
load_tensors: offloaded 25/25 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   143.17 MiB
load_tensors:        CUDA0 model buffer size =   506.07 MiB
..........................................................
common_init_result: added </s> logit bias = -inf
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
common_init_result: added <|fim_pad|> logit bias = -inf
common_init_result: added <|repo_name|> logit bias = -inf
common_init_result: added <|file_sep|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 8192
llama_context: n_ctx_seq     = 8192
llama_context: n_batch       = 8192
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = auto
llama_context: kv_unified    = false
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (8192) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.60 MiB
llama_kv_cache:      CUDA0 KV buffer size =    96.00 MiB
llama_kv_cache: size =   96.00 MiB (  8192 cells,  24 layers,  1/1 seqs), K (f16):   48.00 MiB, V (f16):   48.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 64
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 64
sched_reserve: reserving ...
sched_reserve: Flash Attention was auto, set to enabled
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve:      CUDA0 compute buffer size =   329.26 MiB
sched_reserve:  CUDA_Host compute buffer size =    19.51 MiB
sched_reserve: graph nodes  = 823
sched_reserve: graph splits = 2
sched_reserve: reserve took 9.05 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
common_params_fit_impl: getting device memory data for initial parameters:
common_memory_breakdown_print: | memory breakdown [MiB]   | total   free    self   model   context   compute    unaccounted |
common_memory_breakdown_print: |   - CUDA0 (GTX 1060 6GB) |  6143 = 4255 + ( 496 =   120 +       0 +     376) +        1392 |
common_memory_breakdown_print: |   - Host                 |                   36 =     4 +       0 +      32                |
common_params_fit_impl: projected to use 496 MiB of device memory vs. 4255 MiB of free device memory
common_params_fit_impl: will leave 3758 >= 1024 MiB of free device memory, no changes needed
common_fit_params: successfully fit params to free device memory
common_fit_params: fitting params to free memory took -0.78 seconds
llama_model_loader: loaded meta data with 25 key-value pairs and 161 tensors from ..\WavTokenizer-Large-75-F16.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = wavtokenizer-dec
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = WavTokenizer Large Speech 75token
llama_model_loader: - kv   3:                           general.finetune str              = speech-75token
llama_model_loader: - kv   4:                           general.basename str              = WavTokenizer
llama_model_loader: - kv   5:                         general.size_label str              = large
llama_model_loader: - kv   6:                            general.license str              = mit
llama_model_loader: - kv   7:               wavtokenizer-dec.block_count u32              = 12
llama_model_loader: - kv   8:            wavtokenizer-dec.context_length u32              = 8192
llama_model_loader: - kv   9:          wavtokenizer-dec.embedding_length u32              = 1282
llama_model_loader: - kv  10:      wavtokenizer-dec.attention.head_count u32              = 1
llama_model_loader: - kv  11: wavtokenizer-dec.attention.layer_norm_epsilon f32              = 0.000001
llama_model_loader: - kv  12:                          general.file_type u32              = 1
llama_model_loader: - kv  13:                wavtokenizer-dec.vocab_size u32              = 4096
llama_model_loader: - kv  14:           wavtokenizer-dec.features_length u32              = 512
llama_model_loader: - kv  15:       wavtokenizer-dec.feed_forward_length u32              = 2304
llama_model_loader: - kv  16: wavtokenizer-dec.attention.group_norm_epsilon f32              = 0.000001
llama_model_loader: - kv  17: wavtokenizer-dec.attention.group_norm_groups u32              = 32
llama_model_loader: - kv  18:   wavtokenizer-dec.posnet.embedding_length u32              = 768
llama_model_loader: - kv  19:        wavtokenizer-dec.posnet.block_count u32              = 6
llama_model_loader: - kv  20: wavtokenizer-dec.convnext.embedding_length u32              = 768
llama_model_loader: - kv  21:      wavtokenizer-dec.convnext.block_count u32              = 12
llama_model_loader: - kv  22:          wavtokenizer-dec.attention.causal bool             = false
llama_model_loader: - kv  23:                       tokenizer.ggml.model str              = none
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  110 tensors
llama_model_loader: - type  f16:   51 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = F16
print_info: file size   = 124.15 MiB (16.03 BPW)
llama_prepare_model_devices: using device CUDA0 (NVIDIA GeForce GTX 1060 6GB) (0000:01:00.0) - 4255 MiB free
load: adding 4096 dummy tokens
print_info: arch                  = wavtokenizer-dec
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 8192
print_info: n_embd                = 512
print_info: n_embd_inp            = 512
print_info: n_layer               = 12
print_info: n_head                = 1
print_info: n_head_kv             = 1
print_info: n_rot                 = 512
print_info: n_swa                 = 0
print_info: is_swa_any            = 0
print_info: n_embd_head_k         = 512
print_info: n_embd_head_v         = 512
print_info: n_gqa                 = 1
print_info: n_embd_k_gqa          = 512
print_info: n_embd_v_gqa          = 512
print_info: f_norm_eps            = 1.0e-06
print_info: f_norm_rms_eps        = 0.0e+00
print_info: f_clamp_kqv           = 0.0e+00
print_info: f_max_alibi_bias      = 0.0e+00
print_info: f_logit_scale         = 0.0e+00
print_info: f_attn_scale          = 0.0e+00
print_info: f_attn_value_scale    = 0.0000
print_info: n_ff                  = 2304
print_info: n_expert              = 0
print_info: n_expert_used         = 0
print_info: n_expert_groups       = 0
print_info: n_group_used          = 0
print_info: causal attn           = 0
print_info: pooling type          = -1
print_info: rope type             = -1
print_info: rope scaling          = linear
print_info: freq_base_train       = 10000.0
print_info: freq_scale_train      = 1
print_info: n_ctx_orig_yarn       = 8192
print_info: rope_yarn_log_mul     = 0.0000
print_info: rope_finetuned        = unknown
print_info: model type            = ?B
print_info: model params          = 64.98 M
print_info: general.name          = WavTokenizer Large Speech 75token
print_info: vocab type            = no vocab
print_info: n_vocab               = 4096
print_info: n_merges              = 0
print_info: max token length      = 0
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 11 repeating layers to GPU
load_tensors: offloaded 13/13 layers to GPU
load_tensors:   CPU_Mapped model buffer size =     4.00 MiB
load_tensors:        CUDA0 model buffer size =   120.15 MiB
.......................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 8192
llama_context: n_ctx_seq     = 8192
llama_context: n_batch       = 8192
llama_context: n_ubatch      = 8192
llama_context: causal_attn   = 0
llama_context: flash_attn    = auto
llama_context: kv_unified    = false
llama_context: freq_base     = 10000.0
llama_context: freq_scale    = 1
llama_context:  CUDA_Host  output buffer size =     0.02 MiB
sched_reserve: reserving ...
sched_reserve: Flash Attention was auto, set to enabled
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve:      CUDA0 compute buffer size =   376.00 MiB
sched_reserve:  CUDA_Host compute buffer size =    32.03 MiB
sched_reserve: graph nodes  = 401
sched_reserve: graph splits = 2
sched_reserve: reserve took 14.06 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
sampler seed: 0
sampler params:
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = -1
        top_k = 4, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900
sampler chain: logits -> top-k -> dist
main: loading done
main: constructing prompt ..
main: prompt: 'hello<|space|>i<|space|>am<|space|>sam<|space|>how<|space|>are<|space|>you'


main: llama tokens: 151667, 198, 1782, 155780, 151929, 152412, 152308, 152585, 152460, 153375, 156777, 198, 74455, 155808, 151799, 151873, 151863, 152446, 152372, 152204, 152728, 152229, 152470, 151970, 153413, 152419, 153334, 153289, 153374, 153199, 152040, 153260, 152721, 152680, 153297, 152419, 153248, 152400, 152691, 153368, 153437, 156777, 198, 1722, 155828, 152607, 152256, 152991, 152299, 152688, 153163, 153016, 152789, 153198, 152712, 151911, 153107, 152623, 152170, 152395, 152852, 152207, 152461, 153321, 153309, 151750, 152137, 153340, 152573, 152267, 153347, 151789, 152681, 153339, 151992, 152512, 151751, 152179, 153434, 153180, 152900, 153440, 152474, 153122, 153129, 151904, 152311, 156777, 198, 1499, 155791, 152276, 152454, 153354, 152544, 153204, 153272, 152708, 153433, 152319, 153226, 153043, 152325, 153267, 152622, 156777, 198, 4250, 155797, 153454, 153342, 151989, 152458, 153420, 152303, 152271, 152827, 153036, 153196, 151708, 153263, 152561, 153207, 152213, 152112, 153204, 151722, 152542, 156777, 198, 19789, 155796, 153353, 153182, 152345, 152471, 152477, 153014, 152002, 152191, 151734, 152312, 152810, 152237, 153224, 153169, 153224, 152244, 153387, 153404, 156777, 198, 16069, 155811, 152265, 151946, 151808, 152412, 152363, 152305, 153156, 152733, 152810, 153157, 152016, 152100, 152069, 153234, 152317, 152589, 152707, 153121, 153341, 152159, 152114, 153156, 153001, 153504, 153376, 152272, 152433, 152325, 151941, 156777, 198, 285, 155788, 152238, 152255, 153427, 152318, 153009, 152381, 152474, 152680, 152157, 153255, 152324, 151682, 156777, 198, 32955, 155804, 153490, 153419, 152364, 152405, 152682, 152206, 152078, 153369, 152725, 153193, 153027, 152946, 152488, 153070, 151883, 152890, 152489, 153144, 153375, 152358, 151685, 152494, 152117, 152740, 156777, 198, 37448, 480, 155840, 151902, 152720, 153377, 152027, 152378, 152821, 153207, 153459, 153028, 153068, 152507, 153255, 152158, 152921, 151958, 152609, 152748, 152822, 152286, 151714, 152730, 152377, 152353, 152470, 152606, 152162, 152186, 153071, 152244, 153118, 153375, 153018, 152712, 153098, 152976, 152336, 151843, 153202, 152297, 151736, 153380, 153502, 152702, 152115, 153181, 152735, 153277, 153457, 152393, 153112, 152595, 156777, 198, 19098, 155808, 152464, 153452, 152595, 153312, 151937, 151933, 153197, 152239, 153163, 152922, 153402, 152034, 152591, 153438, 152215, 151673, 152005, 151785, 152642, 151924, 153278, 151805, 151974, 153482, 152718, 152862, 153347, 156777, 198, 72, 155780, 151795, 152111, 152746, 152377, 153471, 152309, 156777, 198, 19016, 155788, 153181, 152271, 152190, 152842, 152224, 152701, 152939, 152536, 152091, 151815, 152733, 151672, 156777, 198, 14689, 155788, 152291, 152072, 152942, 151734, 153042, 153504, 152589, 153333, 151839, 151941, 153038, 153180, 156777, 198, 36996, 8303, 155832, 152231, 152256, 152835, 152801, 152985, 153400, 152393, 152818, 152765, 152249, 152600, 151699, 152302, 152752, 153018, 153009, 151992, 153054, 152847, 153354, 153228, 152662, 153355, 152532, 153393, 151782, 152458, 152048, 152757, 152428, 153195, 151906, 153006, 153178, 153250, 152331, 152284, 152780, 153138, 153319, 151980, 153142, 152418, 152228, 152733, 156777, 198, 9096, 155801, 151698, 153321, 152217, 153039, 152935, 153400, 152122, 152531, 153106, 152169, 152892, 152957, 151851, 152427, 152826, 152451, 151851, 152901, 152885, 152594, 153446, 153080, 156777, 198, 14689, 155795, 152658, 151700, 153321, 152450, 152530, 153191, 151673, 151690, 151698, 152714, 152846, 152981, 153171, 153384, 153364, 153188, 153246, 156777, 198, 1055, 155779, 151869, 152388, 152711, 153334, 151736, 156777, 198, 1782, 155780, 153483, 153240, 152241, 152558, 152697, 153046, 156777, 198, 5804, 1363, 155820, 152941, 152764, 152605, 153034, 153434, 153372, 153347, 151887, 152453, 152758, 152133, 152510, 152694, 152431, 152321, 153088, 152676, 152223, 152581, 152459, 152015, 152502, 153063, 152712, 153294, 153451, 153032, 152903, 152859, 152989, 151748, 152669, 152661, 152650, 152409, 151861, 156777, 198, 300, 7973, 155828, 153095, 152469, 152988, 152894, 151819, 152391, 153019, 152058, 153062, 153230, 151826, 152112, 152306, 152264, 152769, 153390, 152384, 152435, 152790, 153393, 152983, 152540, 152252, 152034, 153107, 152540, 151919, 151893, 152558, 152817, 152946, 152956, 152129, 152715, 153131, 153490, 151734, 152271, 152707, 151734, 153321, 152450, 156777, 198, 8088, 155792, 152452, 153497, 153353, 152679, 152533, 152382, 152374, 152611, 153341, 153163, 152285, 153411, 152495, 153141, 152320, 156777, 198, 1199, 155781, 151764, 152360, 153295, 152634, 153342, 152199, 152271, 156777, 198, 43366, 155799, 152308, 151682, 152889, 152016, 152385, 152629, 152495, 151826, 153321, 152958, 152180, 151886, 153432, 152922, 152128, 153024, 153040, 152593, 152287, 151677, 156777, 198, 53660, 155808, 151727, 152092, 152680, 153331, 151699, 152316, 152938, 152289, 152433, 153384, 151781, 153137, 153259, 152175, 153213, 152291, 151869, 152691, 152489, 151941, 152049, 152034, 153053, 152179, 153160, 151676, 153367, 156777, 198, 268, 4123, 480, 155821, 152350, 152173, 152536, 151991, 151960, 153144, 153013, 152358, 152234, 153135, 152291, 153235, 152143, 152583, 152402, 153483, 152678, 152192, 152533, 152946, 151797, 153103, 152310, 152293, 151825, 152548, 153442, 152109, 152659, 153325, 152781, 152570, 152957, 151752, 152265, 153381, 152515, 156777, 198, 437, 155787, 152957, 152659, 151975, 152709, 152402, 152836, 152174, 151792, 153409, 153327, 152990, 156777, 198, 275, 155781, 152520, 153038, 152067, 153273, 153185, 152265, 152974, 156777, 198, 94273, 155799, 152953, 152938, 153427, 152244, 151920, 153423, 152929, 152367, 153052, 152129, 152331, 152257, 152987, 152777, 153448, 152408, 151696, 152408, 152326, 152699, 156777, 198, 385, 16239, 155828, 152306, 152268, 153438, 153228, 152978, 152957, 153153, 153393, 152795, 152110, 152918, 152923, 152467, 152331, 153053, 153330, 151889, 153444, 152234, 152624, 151779, 152801, 152784, 152139, 152222, 152751, 152512, 153287, 153141, 153052, 151840, 152589, 152508, 153499, 152109, 152255, 151739, 152267, 152759, 153318, 153165, 153349, 156777,


<|im_start|>
<|text_start|>the<|space|>overall<|space|>package<|space|>from<|space|>just<|space|>two<|space|>people<|space|>is<|space|>pretty<|space|>remarkable<|space|>sure<|space|>i<|space|>have<|space|>some<|space|>critiques<|space|>about<|space|>some<|space|>of<|space|>the<|space|>gameplay<|space|>aspects<|space|>but<|space|>its<|space|>still<|space|>really<|space|>enjoyable<|space|>and<|space|>it<|space|>looks<|space|>lovely<|space|>hello<|space|>i<|space|>am<|space|>sam<|space|>how<|space|>are<|space|>you<|text_end|>
<|audio_start|>
the<|t_0.08|><|257|><|740|><|636|><|913|><|788|><|1703|><|space|>
overall<|t_0.36|><|127|><|201|><|191|><|774|><|700|><|532|><|1056|><|557|><|798|><|298|><|1741|><|747|><|1662|><|1617|><|1702|><|1527|><|368|><|1588|><|1049|><|1008|><|1625|><|747|><|1576|><|728|><|1019|><|1696|><|1765|><|space|>
package<|t_0.56|><|935|><|584|><|1319|><|627|><|1016|><|1491|><|1344|><|1117|><|1526|><|1040|><|239|><|1435|><|951|><|498|><|723|><|1180|><|535|><|789|><|1649|><|1637|><|78|><|465|><|1668|><|901|><|595|><|1675|><|117|><|1009|><|1667|><|320|><|840|><|79|><|507|><|1762|><|1508|><|1228|><|1768|><|802|><|1450|><|1457|><|232|><|639|><|space|>
from<|t_0.19|><|604|><|782|><|1682|><|872|><|1532|><|1600|><|1036|><|1761|><|647|><|1554|><|1371|><|653|><|1595|><|950|><|space|>
just<|t_0.25|><|1782|><|1670|><|317|><|786|><|1748|><|631|><|599|><|1155|><|1364|><|1524|><|36|><|1591|><|889|><|1535|><|541|><|440|><|1532|><|50|><|870|><|space|>
two<|t_0.24|><|1681|><|1510|><|673|><|799|><|805|><|1342|><|330|><|519|><|62|><|640|><|1138|><|565|><|1552|><|1497|><|1552|><|572|><|1715|><|1732|><|space|>
people<|t_0.39|><|593|><|274|><|136|><|740|><|691|><|633|><|1484|><|1061|><|1138|><|1485|><|344|><|428|><|397|><|1562|><|645|><|917|><|1035|><|1449|><|1669|><|487|><|442|><|1484|><|1329|><|1832|><|1704|><|600|><|761|><|653|><|269|><|space|>
is<|t_0.16|><|566|><|583|><|1755|><|646|><|1337|><|709|><|802|><|1008|><|485|><|1583|><|652|><|10|><|space|>
pretty<|t_0.32|><|1818|><|1747|><|692|><|733|><|1010|><|534|><|406|><|1697|><|1053|><|1521|><|1355|><|1274|><|816|><|1398|><|211|><|1218|><|817|><|1472|><|1703|><|686|><|13|><|822|><|445|><|1068|><|space|>
remarkable<|t_0.68|><|230|><|1048|><|1705|><|355|><|706|><|1149|><|1535|><|1787|><|1356|><|1396|><|835|><|1583|><|486|><|1249|><|286|><|937|><|1076|><|1150|><|614|><|42|><|1058|><|705|><|681|><|798|><|934|><|490|><|514|><|1399|><|572|><|1446|><|1703|><|1346|><|1040|><|1426|><|1304|><|664|><|171|><|1530|><|625|><|64|><|1708|><|1830|><|1030|><|443|><|1509|><|1063|><|1605|><|1785|><|721|><|1440|><|923|><|space|>
sure<|t_0.36|><|792|><|1780|><|923|><|1640|><|265|><|261|><|1525|><|567|><|1491|><|1250|><|1730|><|362|><|919|><|1766|><|543|><|1|><|333|><|113|><|970|><|252|><|1606|><|133|><|302|><|1810|><|1046|><|1190|><|1675|><|space|>
i<|t_0.08|><|123|><|439|><|1074|><|705|><|1799|><|637|><|space|>
have<|t_0.16|><|1509|><|599|><|518|><|1170|><|552|><|1029|><|1267|><|864|><|419|><|143|><|1061|><|0|><|space|>
some<|t_0.16|><|619|><|400|><|1270|><|62|><|1370|><|1832|><|917|><|1661|><|167|><|269|><|1366|><|1508|><|space|>
critiques<|t_0.60|><|559|><|584|><|1163|><|1129|><|1313|><|1728|><|721|><|1146|><|1093|><|577|><|928|><|27|><|630|><|1080|><|1346|><|1337|><|320|><|1382|><|1175|><|1682|><|1556|><|990|><|1683|><|860|><|1721|><|110|><|786|><|376|><|1085|><|756|><|1523|><|234|><|1334|><|1506|><|1578|><|659|><|612|><|1108|><|1466|><|1647|><|308|><|1470|><|746|><|556|><|1061|><|space|>
about<|t_0.29|><|26|><|1649|><|545|><|1367|><|1263|><|1728|><|450|><|859|><|1434|><|497|><|1220|><|1285|><|179|><|755|><|1154|><|779|><|179|><|1229|><|1213|><|922|><|1774|><|1408|><|space|>
some<|t_0.23|><|986|><|28|><|1649|><|778|><|858|><|1519|><|1|><|18|><|26|><|1042|><|1174|><|1309|><|1499|><|1712|><|1692|><|1516|><|1574|><|space|>
of<|t_0.07|><|197|><|716|><|1039|><|1662|><|64|><|space|>
the<|t_0.08|><|1811|><|1568|><|569|><|886|><|1025|><|1374|><|space|>
gameplay<|t_0.48|><|1269|><|1092|><|933|><|1362|><|1762|><|1700|><|1675|><|215|><|781|><|1086|><|461|><|838|><|1022|><|759|><|649|><|1416|><|1004|><|551|><|909|><|787|><|343|><|830|><|1391|><|1040|><|1622|><|1779|><|1360|><|1231|><|1187|><|1317|><|76|><|997|><|989|><|978|><|737|><|189|><|space|>
aspects<|t_0.56|><|1423|><|797|><|1316|><|1222|><|147|><|719|><|1347|><|386|><|1390|><|1558|><|154|><|440|><|634|><|592|><|1097|><|1718|><|712|><|763|><|1118|><|1721|><|1311|><|868|><|580|><|362|><|1435|><|868|><|247|><|221|><|886|><|1145|><|1274|><|1284|><|457|><|1043|><|1459|><|1818|><|62|><|599|><|1035|><|62|><|1649|><|778|><|space|>
but<|t_0.20|><|780|><|1825|><|1681|><|1007|><|861|><|710|><|702|><|939|><|1669|><|1491|><|613|><|1739|><|823|><|1469|><|648|><|space|>
its<|t_0.09|><|92|><|688|><|1623|><|962|><|1670|><|527|><|599|><|space|>
still<|t_0.27|><|636|><|10|><|1217|><|344|><|713|><|957|><|823|><|154|><|1649|><|1286|><|508|><|214|><|1760|><|1250|><|456|><|1352|><|1368|><|921|><|615|><|5|><|space|>
really<|t_0.36|><|55|><|420|><|1008|><|1659|><|27|><|644|><|1266|><|617|><|761|><|1712|><|109|><|1465|><|1587|><|503|><|1541|><|619|><|197|><|1019|><|817|><|269|><|377|><|362|><|1381|><|507|><|1488|><|4|><|1695|><|space|>
enjoyable<|t_0.49|><|678|><|501|><|864|><|319|><|288|><|1472|><|1341|><|686|><|562|><|1463|><|619|><|1563|><|471|><|911|><|730|><|1811|><|1006|><|520|><|861|><|1274|><|125|><|1431|><|638|><|621|><|153|><|876|><|1770|><|437|><|987|><|1653|><|1109|><|898|><|1285|><|80|><|593|><|1709|><|843|><|space|>
and<|t_0.15|><|1285|><|987|><|303|><|1037|><|730|><|1164|><|502|><|120|><|1737|><|1655|><|1318|><|space|>
it<|t_0.09|><|848|><|1366|><|395|><|1601|><|1513|><|593|><|1302|><|space|>
looks<|t_0.27|><|1281|><|1266|><|1755|><|572|><|248|><|1751|><|1257|><|695|><|1380|><|457|><|659|><|585|><|1315|><|1105|><|1776|><|736|><|24|><|736|><|654|><|1027|><|space|>
lovely<|t_0.56|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|1481|><|1721|><|1123|><|438|><|1246|><|1251|><|795|><|659|><|1381|><|1658|><|217|><|1772|><|562|><|952|><|107|><|1129|><|1112|><|467|><|550|><|1079|><|840|><|1615|><|1469|><|1380|><|168|><|917|><|836|><|1827|><|437|><|583|><|67|><|595|><|1087|><|1646|><|1493|><|1677|><|space|>main: prompt size: 871

main: time for prompt: 252.929 ms

000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000

main: time for decoder:       1412.620 ms
common_perf_print:    sampling time =      66.79 ms
common_perf_print:    samplers time =      26.02 ms /   199 tokens
common_perf_print:        load time =     655.95 ms
common_perf_print: prompt eval time =     234.97 ms /   871 tokens (    0.27 ms per token,  3706.90 tokens per second)
common_perf_print:        eval time =    1341.80 ms /   198 runs   (    6.78 ms per token,   147.56 tokens per second)
common_perf_print:       total time =    1075.97 ms /  1069 tokens
common_perf_print: unaccounted time =    -567.58 ms / -52.8 %      (total - sampling - prompt eval - eval) / (total)
common_perf_print:    graphs reused =        196
common_memory_breakdown_print: | memory breakdown [MiB]   | total   free    self   model   context   compute    unaccounted |
common_memory_breakdown_print: |   - CUDA0 (GTX 1060 6GB) |  6143 = 3733 + ( 931 =   506 +      96 +     329) +        1479 |
common_memory_breakdown_print: |   - Host                 |                  162 =   143 +       0 +      19                |

codes: '
hello<|t_0.96|><|865|><|1506|><|865|><|1419|><|1819|><|838|><|624|><|1251|><|899|><|954|><|1096|><|710|><|1152|><|1418|><|710|><|1301|><|1120|><|17|><|1456|><|1405|><|776|><|1668|><|1390|><|86|><|1292|><|1023|><|1683|><|1589|><|1092|><|1556|><|1479|><|1294|><|1292|><|805|><|1683|><|1430|><|900|><|1714|><|995|><|1294|><|1432|><|1007|><|1622|><|1120|><|861|><|1803|><|995|><|1092|><|1668|><|710|><|1433|><|933|><|670|><|32|><|1293|><|1251|><|1134|><|1701|><|1347|><|816|><|642|><|95|><|508|><|48|><|503|><|653|><|1707|><|1041|><|267|><|1817|><|248|><|1754|><|space|>
i<|t_0.28|><|73|><|642|><|169|><|614|><|983|><|169|><|843|><|443|><|1092|><|752|><|252|><|1378|><|1315|><|221|><|1448|><|1083|><|565|><|866|><|93|><|767|><|1697|><|space|>
am<|t_0.16|><|422|><|852|><|408|><|847|><|1007|><|550|><|874|><|673|><|191|><|127|><|220|><|716|><|space|>
sam<|t_0.43|><|775|><|487|><|646|><|519|><|493|><|1513|><|1|><|1166|><|640|><|556|><|0|><|1061|><|18|><|333|><|719|><|632|><|693|><|907|><|430|><|1312|><|1086|><|1098|><|1333|><|974|><|816|><|440|><|1755|><|1324|><|1534|><|662|><|1812|><|385|><|space|>
how<|t_0.20|><|1663|><|1028|><|1488|><|1314|><|1393|><|1723|><|1303|><|1497|><|951|><|1181|><|789|><|142|><|1475|><|66|><|297|><|space|>
are<|t_0.13|><|798|><|1803|><|562|><|123|><|756|><|968|><|381|><|890|><|1773|><|1039|><|space|>
you<|t_0.08|><|193|><|92|><|1221|><|1334|><|562|><|1415|>
<|audio_end|>
<|im_end|>'
main: codes size: 199
codes audio: '<|865|><|1506|><|865|><|1419|><|1819|><|838|><|624|><|1251|><|899|><|954|><|1096|><|710|><|1152|><|1418|><|710|><|1301|><|1120|><|17|><|1456|><|1405|><|776|><|1668|><|1390|><|86|><|1292|><|1023|><|1683|><|1589|><|1092|><|1556|><|1479|><|1294|><|1292|><|805|><|1683|><|1430|><|900|><|1714|><|995|><|1294|><|1432|><|1007|><|1622|><|1120|><|861|><|1803|><|995|><|1092|><|1668|><|710|><|1433|><|933|><|670|><|32|><|1293|><|1251|><|1134|><|1701|><|1347|><|816|><|642|><|95|><|508|><|48|><|503|><|653|><|1707|><|1041|><|267|><|1817|><|248|><|1754|><|73|><|642|><|169|><|614|><|983|><|169|><|843|><|443|><|1092|><|752|><|252|><|1378|><|1315|><|221|><|1448|><|1083|><|565|><|866|><|93|><|767|><|1697|><|422|><|852|><|408|><|847|><|1007|><|550|><|874|><|673|><|191|><|127|><|220|><|716|><|775|><|487|><|646|><|519|><|493|><|1513|><|1|><|1166|><|640|><|556|><|0|><|1061|><|18|><|333|><|719|><|632|><|693|><|907|><|430|><|1312|><|1086|><|1098|><|1333|><|974|><|816|><|440|><|1755|><|1324|><|1534|><|662|><|1812|><|385|><|1663|><|1028|><|1488|><|1314|><|1393|><|1723|><|1303|><|1497|><|951|><|1181|><|789|><|142|><|1475|><|66|><|297|><|798|><|1803|><|562|><|123|><|756|><|968|><|381|><|890|><|1773|><|1039|><|193|><|92|><|1221|><|1334|><|562|><|1415|>'
main: codes audio size: 168
main: time for vocoder:      220.671 ms
main: time for spectral ops: 850.860 ms
main: total time:            2737.218 ms
main: audio written to file 'output.wav'

 

영어는 잘되는데 한글은 잘 안되는 듯.

output.wav
0.10MB

 

GPT 통해서 영어로 "안녕? 난 잼미니야 만나서 반가워" 를 TTS에 유리하게 바꾸어 달라고 했는데

'안녕? 난' 은 날아가고 '잼미니야 맨나서 빵가워' 정도로 들린다.

D:\study\llm\llama-b9093-bin-win-cuda-12.4-x64>llama-tts -m ..\OuteTTS-0.3-500M-Q8_0.gguf  -mv ..\WavTokenizer-Large-75-F16.gguf -p "Annyoung? Nahn Jemmini-ya. Mannaseo bangawo."

 

hello_ko.wav
0.15MB

 

 

+

허깅페이스에서 타입에 아예 tts가 있었군.

[링크 : https://huggingface.co/models?pipeline_tag=text-to-speech]

저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice 시도 실패  (0) 2026.05.10
supertone/supertonic3 시도  (0) 2026.05.10
huggingface 에서 다운로드 받기(python)  (0) 2026.05.10
stable diffusion 사용법  (0) 2026.05.09
stable diffusion python service  (0) 2026.05.08
Posted by 구차니
프로그램 사용/ai 프로그램2026. 5. 10. 17:17

huggingface 에서 다운로드 받기(python)

git clone  하듯 받을수 있을것 같기도 한데..

아무튼 파이썬을 통해서 한번에 받는 방법

 

step 1.

hugging face 아래에 복사 아이콘 (like 2 왼쪽) 을 누른다.

[링크 : https://huggingface.co/lysandre/arxiv-nlp]

 

step 2.

pip 로 huggingface_hub 패키지를 설치한다.

 

step 3. 

python 실행해서 아래를 따라한다.

전체 리포지토리 다운로드하기
snapshot_download() 함수는 특정 버전의 전체 리포지토리를 다운로드합니다. 이 함수는 내부적으로 hf_hub_download() 함수를 사용하므로, 다운로드한 모든 파일은 로컬 디스크에 캐시되어 저장됩니다. 다운로드는 여러 파일을 동시에 받아오기 때문에 빠르게 진행됩니다.

전체 리포지토리를 다운로드하려면 repo_id와 repo_type을 인자로 넘겨주면 됩니다:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="lysandre/arxiv-nlp")

[링크 : https://huggingface.co/docs/huggingface_hub/ko/guides/download]

 

음.. 걍 하나씩 받을까 -_-

Download complete: : 2.52GB [03:13, 50.8MB/s]              'C:\\Users\\minimonk\\.cache\\huggingface\\hub\\models--unsloth--Llama-OuteTTS-1.0-1B\\snapshots\\52b901173c6cf817148fdaef981e52408332c3ca'
저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

supertone/supertonic3 시도  (0) 2026.05.10
outetts 시도  (0) 2026.05.10
stable diffusion 사용법  (0) 2026.05.09
stable diffusion python service  (0) 2026.05.08
opencode + qwen3.6 35b q2 사용 테스트  (0) 2026.05.08
Posted by 구차니
프로그램 사용/ai 프로그램2026. 5. 9. 12:51

stable diffusion 사용법

[링크 : https://selgyun.tistory.com/4]

 

(키워드) - 강화

[키워드] -약화

batch - 동시에 여러개 이미지 생성

[링크 : https://selgyun.tistory.com/5] txt2img

[링크 : https://selgyun.tistory.com/6] img2img

[링크 : https://selgyun.tistory.com/7] Lora - 모델타입?

 

+

2026.05.11

모델 받는 법, 종류

[링크 : https://healtable.tistory.com/7]

 

+

--ckpt CKPT model.ckpt Path to checkpoint of Stable Diffusion model; if specified, this checkpoint will be added to the list of checkpoints and loaded.
--ckpt-dir CKPT_DIR None Path to directory with Stable Diffusion checkpoints.

[링크 : https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings]

[링크 : https://www.reddit.com/r/StableDiffusion/comments/117p89v/how_can_i_choose_which_checkpoint_to_load_when/?tl=ko]

 

+

VAE stands for variational autoencoder
EMA and MSE. (Exponential Moving Average and Mean Square Error)

[링크 : https://stable-diffusion-art.com/automatic1111/]

[링크 : https://stable-diffusion-art.com/how-to-use-vae/]

[링크 : https://huggingface.co/stabilityai/sd-vae-ft-ema#visual]

 

지금 이거랑 비슷한 현상인데 봐야겠다.

[링크 : https://samablog.tistory.com/m/203]

 

+

이전 모델이 머였는지 모르겠는데, 새로 받은건 SDXL 계열이고.. 그래서 VAE를 automatic으로 해주라는데

어디서 설정을 하는건지 못 찾는 중..

Illustrious XL is an advanced Stable Diffusion XL (SD XL)-based model, developed by OnomaAI Research, optimized specifically for illustration and animation tasks. It is built upon the Kohaku XL-Beta - Revision 5 checkpoint, leveraging its robust foundation to deliver high-quality generative capabilities.

[링크 : https://civitai.com/models/795765/illustrious-xl]

 

gpt 에서 보면 VAE가 의심되긴한데 CLIP도 문제라지만..fp16 VAE(half-vae ?) 문제일수도 있다는데

일단 내가 한건 모델만 바꾼거라.. vae와 매칭이 안되는게 의심되긴 한다.

1. VAE가 모델과 맞지 않는 경우

가장 흔한 원인입니다.

예를 들어:

SD1.5 계열 모델
SDXL 계열 모델
애니 특화 모델
실사 특화 모델

각각 latent 분포가 조금 다릅니다.

그런데 다른 계열 VAE를 연결하면:

색감 이상
얼굴 붕괴
디테일 손실
프롬프트 반영 약화
채도 과다/부족
눈/손 이상

등이 발생합니다.

예:

SDXL 모델에 SD1.5 VAE 사용
Anything 계열에 실사 VAE 사용

→ “고양이”를 입력했는데 이상한 생물처럼 나오는 경우도 있습니다.


5. 실제로는 CLIP 문제인 경우도 많음

많은 사용자가 VAE 문제라고 생각하지만 실제론:

CLIP 인코딩
CFG scale
sampler
step 수
LoRA 충돌

문제인 경우도 매우 많습니다.

예:

CFG 너무 높음 → 과적합
Euler a → 프롬프트 드리프트
LoRA weight 과다 → 키워드 왜곡

그러면:

"1girl, red hair"

인데 갑자기:

배경 과다
캐릭터 여러 명
이상한 의상

등이 나옵니다.



6. fp16 VAE 문제

저사양 환경에서 흔합니다.

fp16 VAE는:

VRAM 절약
속도 향상

장점이 있지만:

검은 이미지
washed out
NaN
디테일 손실

이 발생하기도 합니다.

특히 AMD/구형 CUDA에서 자주 보입니다.

[링크 : https://chatgpt.com/share/6a01e096-dc48-83e9-8aa3-cfe937d3b9e6]

 

 

 

+

2026.05.12

ft-MSE가 내눈에는 좋아 보인다.

[링크 : https://huggingface.co/stabilityai/sd-vae-ft-mse/tree/main]

[링크 : https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main]

[링크 : https://huggingface.co/stabilityai/sd-vae-ft-ema/tree/main]

[링크 : https://huggingface.co/stabilityai/sd-vae-ft-ema-original/tree/main]

 

+

[링크 : https://www.reddit.com/r/StableDiffusion/comments/1594hbj/minimum_nonsquare_resolution_with_sdxl/]

 

저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

outetts 시도  (0) 2026.05.10
huggingface 에서 다운로드 받기(python)  (0) 2026.05.10
stable diffusion python service  (0) 2026.05.08
opencode + qwen3.6 35b q2 사용 테스트  (0) 2026.05.08
llama.cpp gemma4:e2b 실행시 에러  (0) 2026.05.08
Posted by 구차니
프로그램 사용/ai 프로그램2026. 5. 8. 16:21

stable diffusion python service

 

from diffusers import StableDiffusionPipeline

[링크 : https://www.reddit.com/r/Python/comments/10g5nay/use_python_to_build_a_free_stable_diffusion_app/?tl=ko]

[링크 : https://www.assemblyai.com/blog/build-a-free-stable-diffusion-app-with-a-gpu-backend]

 

매모리 터져나갈게 보이니 e4b는 gpu 0번에서

stable diffusion은 gpu 1번에서 돌려서 두개 연동해

img2img로 장난치거나 그림 그려 가 포함되면 txt2img로 돌려도 괜찮을듯

---

by gpt

api 모드를 활성화 해주고

./webui.sh --api

 

txt2img

import requests
import base64

url = "http://127.0.0.1:7860/sdapi/v1/txt2img"

payload = {
    "prompt": "masterpiece, ultra detailed, cyberpunk girl, neon city, rain",
    "negative_prompt": "low quality, blurry",
    "steps": 30,
    "width": 768,
    "height": 768,
    "cfg_scale": 7,
    "sampler_name": "DPM++ 2M Karras"
}

response = requests.post(url, json=payload)

result = response.json()

# base64 이미지 저장
image_data = base64.b64decode(result["images"][0])

with open("generated.png", "wb") as f:
    f.write(image_data)

print("generated.png 저장 완료")

 

오.. 된다.

 

img2img

import requests
import base64

# 입력 이미지 읽기
with open("input.png", "rb") as f:
    image_base64 = base64.b64encode(f.read()).decode()

url = "http://127.0.0.1:7860/sdapi/v1/img2img"

payload = {
    "init_images": [image_base64],

    "prompt": "cyberpunk style, neon lights, futuristic",
    "negative_prompt": "low quality, blurry",

    "denoising_strength": 0.55,

    "steps": 30,
    "cfg_scale": 7,
    "width": 768,
    "height": 768,
    "sampler_name": "DPM++ 2M Karras"
}

response = requests.post(url, json=payload)

result = response.json()

image_data = base64.b64decode(result["images"][0])

with open("modified.png", "wb") as f:
    f.write(image_data)

print("modified.png 저장 완료")

 

[링크 : https://chatgpt.com/share/69fd90a2-2bbc-83e9-8f04-6cecdddc1b41]

저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

huggingface 에서 다운로드 받기(python)  (0) 2026.05.10
stable diffusion 사용법  (0) 2026.05.09
opencode + qwen3.6 35b q2 사용 테스트  (0) 2026.05.08
llama.cpp gemma4:e2b 실행시 에러  (0) 2026.05.08
stable diffusion gpu 선택하기  (0) 2026.05.08
Posted by 구차니

티스토리툴바