AMD环境都安装了，但是还是用的CPU，是哪里没有配置对吗

[liufumingbaba]

[dlsandy]

麻烦提供下任务执行时的引擎日志

[188198]

麻烦提供下任务执行时的引擎日志

9070已经官网通过pip安装了rocm及其配套pytorch，无法调用gpu。Faster-Whisper-TransWithAI-ChickenRice这个倒可以调用GPU,可模型单调

[engine] 启动 U:\Transub-3.2.0-win\transub-engine\runtime\python.exe -m transub_engine serve --host 127.0.0.1 --port 8765
[transub-engine] ffmpeg: U:\Transub-3.2.0-win_internal\bin\ffmpeg.exe
[transub-engine] torch: preload failed (torch/lib not found)
[transub-engine] gpu: status=no_gpu name=- cublas12=False ct2Cuda=False ortGpuCuda=False pathInjected=1
INFO: Started server process [13228]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:8765 (Press CTRL+C to quit)
INFO: 127.0.0.1:54318 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:54320 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:54318 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:54320 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:54320 - "POST /v1/models/recommend HTTP/1.1" 200 OK
INFO: 127.0.0.1:54318 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:49755 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:49757 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:49755 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:49757 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:49757 - "POST /v1/models/recommend HTTP/1.1" 200 OK
INFO: 127.0.0.1:49755 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:49756 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:49756 - "POST /v1/compute/acquire HTTP/1.1" 200 OK
INFO: 127.0.0.1:49757 - "POST /v1/models/recommend HTTP/1.1" 200 OK
[engine] 开始批次 · 1 个文件 · ASR=qwen3-asr-1.7b-ja-anime-galgame · 档位=quality
[engine] 免费推理翻译 · 引擎内 llama（跳过外部适配器）
[engine] #1/1 启动 4k688.com@ADN-785.mp4 · 准备引擎任务…
[engine] #1/1 翻译后端 · 推理翻译：sakura-galtransl-14b-q6k）
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 创建引擎任务（qwen3-asr-1.7b-ja-anime-galgame）…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 准备原文…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 加载识别模型（qwen3-asr-1.7b-ja-anime-galgame）…
INFO: 127.0.0.1:49756 - "POST /v1/jobs HTTP/1.1" 200 OK
[engine] #1/1 启动 4k688.com@ADN-785.mp4 · 任务已启动
[engine] #1/1 处理中 4k688.com@ADN-785.mp4 · 准备音频…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 准备加载 qwen3-asr-1.7b-ja-anime-galgame（导入 Torch）…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 检查 NumPy / Numba 兼容性…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 检测 Torch/CUDA（子进程）…
WARNING transub_engine.asr.qwen3_asr: Qwen3-ASR: torch probe failed (code=1): Traceback (most recent call last):
File "", line 1, in
ModuleNotFoundError: No module named 'torch'
Traceback (most recent call last):
File "U:\Transub-3.2.0-win\transub-engine\runtime\Lib\site-packages\transub_engine\jobs.py", line 404, in run_job
result = execute_job(job.request, job=job)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.0-win\transub-engine\runtime\Lib\site-packages\transub_engine\jobs.py", line 1141, in execute_job
source_cues = run_asr(
^^^^^^^^
File "U:\Transub-3.2.0-win\transub-engine\runtime\Lib\site-packages\transub_engine\asr_init.py", line 125, in run_asr
cues = qwen3_asr_transcribe(
^^^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.0-win\transub-engine\runtime\Lib\site-packages\transub_engine\asr\qwen3_asr.py", line 1470, in qwen3_asr_transcribe
raise RuntimeError(
RuntimeError: Torch / CUDA 导入失败：torch/lib not found。请完全退出后重试，或在设置中「补齐 CUDA PyTorch」。
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 子进程导入失败 · 尝试本进程预加载 Torch…
[engine] #1/1 失败 4k688.com@ADN-785.mp4 · Torch / CUDA 导入失败：torch/lib not found。请完全退出后重试，或在设置中「补齐 CUDA PyTorch」。 · 已试 qwen3-asr-1.7b-ja-anime-galgame · 可：另选 ASR 模型
INFO transub_engine.runtime_release: release_gpu_memory reason=api freed=['whisper_cache:0', 'whisperseg_cache:0', 'sensevoice_cache:0', 'reazon_k2_cache:0', 'qwen3_asr_cache:0', 'parakeet_cache:0', 'cohere_asr_cache:0', 'wav2vec_align_cache:0', 'gc'] errors=["ctranslate2:No module named 'ctranslate2'"]
INFO: 127.0.0.1:49755 - "POST /v1/runtime/release-gpu HTTP/1.1" 200 OK
[engine] 批次结束 · 已释放引擎显存
[engine] 任务失败或已中断 · 已停止引擎与本地 LLM 进程
[engine] 进程退出 code=1 signal=
[engine] 启动 U:\Transub-3.2.0-win\transub-engine\runtime\python.exe -m transub_engine serve --host 127.0.0.1 --port 8765
[transub-engine] ffmpeg: U:\Transub-3.2.0-win_internal\bin\ffmpeg.exe
[transub-engine] torch: preload failed (torch/lib not found)
[transub-engine] gpu: status=no_gpu name=- cublas12=False ct2Cuda=False ortGpuCuda=False pathInjected=1
INFO: Started server process [14592]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:8765 (Press CTRL+C to quit)
INFO: 127.0.0.1:49559 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:49557 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:49557 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:49559 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:49559 - "POST /v1/models/recommend HTTP/1.1" 200 OK
INFO: 127.0.0.1:49557 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:49558 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:49558 - "POST /v1/compute/acquire HTTP/1.1" 200 OK
INFO: 127.0.0.1:49559 - "POST /v1/models/recommend HTTP/1.1" 200 OK
[engine] 开始批次 · 1 个文件 · ASR=qwen3-asr-1.7b-ja-anime-galgame · 档位=quality
[engine] 免费推理翻译 · 引擎内 llama（跳过外部适配器）
[engine] #1/1 启动 4k688.com@ADN-785.mp4 · 准备引擎任务…
[engine] #1/1 翻译后端 · 推理翻译：sakura-galtransl-14b-q6k）
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 创建引擎任务（qwen3-asr-1.7b-ja-anime-galgame）…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 准备原文…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 加载识别模型（qwen3-asr-1.7b-ja-anime-galgame）…
INFO: 127.0.0.1:49558 - "POST /v1/jobs HTTP/1.1" 200 OK
WARNING transub_engine.asr.qwen3_asr: Qwen3-ASR: torch probe failed (code=1): Traceback (most recent call last):
File "", line 1, in
ModuleNotFoundError: No module named 'torch'
Traceback (most recent call last):
File "U:\Transub-3.2.0-win\transub-engine\runtime\Lib\site-packages\transub_engine\jobs.py", line 404, in run_job
result = execute_job(job.request, job=job)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.0-win\transub-engine\runtime\Lib\site-packages\transub_engine\jobs.py", line 1141, in execute_job
source_cues = run_asr(
^^^^^^^^
File "U:\Transub-3.2.0-win\transub-engine\runtime\Lib\site-packages\transub_engine\asr_init.py", line 125, in run_asr
cues = qwen3_asr_transcribe(
^^^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.0-win\transub-engine\runtime\Lib\site-packages\transub_engine\asr\qwen3_asr.py", line 1470, in qwen3_asr_transcribe
raise RuntimeError(
RuntimeError: Torch / CUDA 导入失败：torch/lib not found。请完全退出后重试，或在设置中「补齐 CUDA PyTorch」。
[engine] #1/1 启动 4k688.com@ADN-785.mp4 · 任务已启动
[engine] #1/1 处理中 4k688.com@ADN-785.mp4 · 准备音频…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 准备加载 qwen3-asr-1.7b-ja-anime-galgame（导入 Torch）…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 检查 NumPy / Numba 兼容性…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 检测 Torch/CUDA（子进程）…
[engine] #1/1 加载识别模型 4k688.com@ADN-785.mp4 · 子进程导入失败 · 尝试本进程预加载 Torch…
[engine] #1/1 失败 4k688.com@ADN-785.mp4 · Torch / CUDA 导入失败：torch/lib not found。请完全退出后重试，或在设置中「补齐 CUDA PyTorch」。 · 已试 qwen3-asr-1.7b-ja-anime-galgame · 可：另选 ASR 模型
INFO transub_engine.runtime_release: release_gpu_memory reason=api freed=['whisper_cache:0', 'whisperseg_cache:0', 'sensevoice_cache:0', 'reazon_k2_cache:0', 'qwen3_asr_cache:0', 'parakeet_cache:0', 'cohere_asr_cache:0', 'wav2vec_align_cache:0', 'gc'] errors=["ctranslate2:No module named 'ctranslate2'"]
INFO: 127.0.0.1:49557 - "POST /v1/runtime/release-gpu HTTP/1.1" 200 OK
[engine] 批次结束 · 已释放引擎显存
[engine] 任务失败或已中断 · 已停止引擎与本地 LLM 进程
[engine] 进程退出 code=1 signal=

[dlsandy]

我已经根据这些问题重新做了针对性补丁，请下载：https://github.com/dlsandy/Transub/releases/tag/v3.2.1

[188198]

我已经根据这些问题重新做了针对性补丁，请下载：https://github.com/dlsandy/Transub/releases/tag/v3.2.1

还是不行，会出错停止

[13:53:46] ROCm PyTorch 已安装，正在重启引擎…
[13:54:09] 已把 ROCm PyTorch 装进引擎运行时（2.12.0+rocm10.0.0，AMD Radeon RX 9070）
[13:54:24] 已添加 1 个文件
[13:54:26] 正在执行「@MFYD-183.mp4」…
[13:54:27] 开始生成字幕 1 个文件…
[13:54:27] 引擎后端：Transub Engine · ASR=qwen3-asr-1.7b-ja-anime-galgame · 档位=quality
[13:54:27] 正在提交引擎任务（首次加载模型可能需数分钟）…
[13:54:27] [任务] 开始：引擎字幕任务
[13:54:29] 开始任务（翻译），共 1 个文件
[13:54:29] #1/1 开始处理：@MFYD-183.mp4 · 准备引擎任务…
[13:54:29] #1/1 加载识别模型：@MFYD-183.mp4 · 加载识别模型（qwen3-asr-1.7b-ja-anime-galgame）…
[13:54:29] #1/1 开始处理：@MFYD-183.mp4 · 任务已启动
[13:54:30] #1/1 加载识别模型：@MFYD-183.mp4 · 加载 Qwen3-ASR（VAD 帧时间戳）…
[13:56:51] #1/1 翻译中：@MFYD-183.mp4
[13:56:58] 失败：@MFYD-183.mp4 — Qwen3-ASR TEN failed 4/2466 frames (@MFYD-183.mp4): CUDA error: invalid argument
Search for `hipErrorInvalidValue' in https://rocm.docs.amd.com/projects/HIP/en/latest/index.html for more information.
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing AMD_SERIALIZE_KERNEL=3
Device-side assertion tracking was not enabled by user.
[13:56:58] 任务结束：成功 0 · 跳过 0 · 失败 1
[13:57:04] [任务] 结束

[188198]

我已经根据这些问题重新做了针对性补丁，请下载：https://github.com/dlsandy/Transub/releases/tag/v3.2.1

官网的是10对应2.13，7.14对应的2.12.软件里是10对应2.12，不知道有无关系

[dlsandy]

原始是你所使用的模型（qwen3-asr-1.7b-ja-anime-galgame）默认 beam=4 + 多帧 batch + bf16/SDPA，在 ROCm / RDNA4 上很容易触发该错误。
你可以下载下面这个补丁文件，并解压缩到软件目录下覆盖：
Transub-3.2.1-amd-rocm-qwen-hotfix-20260920.zip

[188198]

原始是你所使用的模型（qwen3-asr-1.7b-ja-anime-galgame）默认 beam=4 + 多帧 batch + bf16/SDPA，在 ROCm / RDNA4 上很容易触发该错误。 你可以下载下面这个补丁文件，并解压缩到软件目录下覆盖： Transub-3.2.1-amd-rocm-qwen-hotfix-20260920.zip

9070已经可以运行了。速度有点感人

[dlsandy]

原始是你所使用的模型（qwen3-asr-1.7b-ja-anime-galgame）默认 beam=4 + 多帧 batch + bf16/SDPA，在 ROCm / RDNA4 上很容易触发该错误。 你可以下载下面这个补丁文件，并解压缩到软件目录下覆盖： Transub-3.2.1-amd-rocm-qwen-hotfix-20260920.zip

9070已经可以运行了。速度有点感人

目前的处理完一部影片的时长大约需要多久呢？
以及你所使用的翻译模式和模型是哪个呢？

[188198]

原始是你所使用的模型（qwen3-asr-1.7b-ja-anime-galgame）默认 beam=4 + 多帧 batch + bf16/SDPA，在 ROCm / RDNA4 上很容易触发该错误。 你可以下载下面这个补丁文件，并解压缩到软件目录下覆盖： Transub-3.2.1-amd-rocm-qwen-hotfix-20260920.zip

9070已经可以运行了。速度有点感人

目前的处理完一部影片的时长大约需要多久呢？ 以及你所使用的翻译模式和模型是哪个呢？

如果使用Qwen3-ASR-1.7B速度较慢的话，可以尝试换成Anime Whisper。目前ROCm对Qwen3-ASR-1.7B的运行支持仍不是很好。

预设的软对白，推理，质量，时间大概跑了下，1小时时长的要近半小时的吧，9070负荷90左右了

[dlsandy]

原始是你所使用的模型（qwen3-asr-1.7b-ja-anime-galgame）默认 beam=4 + 多帧 batch + bf16/SDPA，在 ROCm / RDNA4 上很容易触发该错误。 你可以下载下面这个补丁文件，并解压缩到软件目录下覆盖： Transub-3.2.1-amd-rocm-qwen-hotfix-20260920.zip

9070已经可以运行了。速度有点感人

目前的处理完一部影片的时长大约需要多久呢？ 以及你所使用的翻译模式和模型是哪个呢？
如果使用Qwen3-ASR-1.7B速度较慢的话，可以尝试换成Anime Whisper。目前ROCm对Qwen3-ASR-1.7B的运行支持仍不是很好。

预设的软对白，推理，质量，时间大概跑了下，1小时时长的要近半小时的吧，9070负荷90左右了

你可以试下组合方案内的：软对白 · 推理 · 均衡 这一档（时间参考是2小时的影片处理时间约10分钟左右），预设里的质量档所使用的sakura-galtransl-14b-q6k模型参数量偏大9070跑起来可能比较困难。

[188198]

原始是你所使用的模型（qwen3-asr-1.7b-ja-anime-galgame）默认 beam=4 + 多帧 batch + bf16/SDPA，在 ROCm / RDNA4 上很容易触发该错误。 你可以下载下面这个补丁文件，并解压缩到软件目录下覆盖： Transub-3.2.1-amd-rocm-qwen-hotfix-20260920.zip

9070已经可以运行了。速度有点感人

目前的处理完一部影片的时长大约需要多久呢？ 以及你所使用的翻译模式和模型是哪个呢？
如果使用Qwen3-ASR-1.7B速度较慢的话，可以尝试换成Anime Whisper。目前ROCm对Qwen3-ASR-1.7B的运行支持仍不是很好。

预设的软对白，推理，质量，时间大概跑了下，1小时时长的要近半小时的吧，9070负荷90左右了

你可以试下组合方案内的：软对白 · 推理 · 均衡 这一档（时间参考是2小时的影片处理时间约10分钟左右），预设里的质量档所使用的sakura-galtransl-14b-q6k模型参数量偏大9070跑起来可能比较困难。

Anime Whisper支持rocm么？一直卡住，我还特意重新装的软件，重新下的mod。同样的视频千问3分多就开始了，5分钟的日文歌曲都一直卡住

[engine] 启动 U:\Transub-3.2.1-win\transub-engine\runtime\python.exe -m transub_engine serve --host 127.0.0.1 --port 8765
[transub-engine] ffmpeg: U:\Transub-3.2.1-win_internal\bin\ffmpeg.exe
[transub-engine] torch: preloaded 2.12.0+rocm10.0.0 cuda=None
[transub-engine] gpu: status=ready name=AMD Radeon RX 9070 cublas12=False ct2Cuda=False ortGpuCuda=False pathInjected=1 rocm=gfx1201 torchRocm=True
INFO: Started server process [6292]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:8765 (Press CTRL+C to quit)
INFO: 127.0.0.1:64544 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64542 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64543 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
[engine] 进程退出 code=1 signal=
[engine] 启动 U:\Transub-3.2.1-win\transub-engine\runtime\python.exe -m transub_engine serve --host 127.0.0.1 --port 8765
[transub-engine] ffmpeg: U:\Transub-3.2.1-win_internal\bin\ffmpeg.exe
[transub-engine] torch: preloaded 2.12.0+rocm10.0.0 cuda=None
[transub-engine] gpu: status=ready name=AMD Radeon RX 9070 cublas12=False ct2Cuda=False ortGpuCuda=False pathInjected=1 rocm=gfx1201 torchRocm=True
INFO: Started server process [16132]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:8765 (Press CTRL+C to quit)
INFO: 127.0.0.1:64558 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:64557 - "POST /v1/models/download HTTP/1.1" 200 OK
INFO: 127.0.0.1:64571 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:64557 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:64579 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64577 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64578 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:64579 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64577 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64578 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
[engine] 进程退出 code=1 signal=
[engine] 启动 U:\Transub-3.2.1-win\transub-engine\runtime\python.exe -m transub_engine serve --host 127.0.0.1 --port 8765
[transub-engine] ffmpeg: U:\Transub-3.2.1-win_internal\bin\ffmpeg.exe
[transub-engine] torch: preloaded 2.12.0+rocm10.0.0 cuda=None
[transub-engine] gpu: status=ready name=AMD Radeon RX 9070 cublas12=False ct2Cuda=False ortGpuCuda=False pathInjected=1 rocm=gfx1201 torchRocm=True
INFO: Started server process [16184]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:8765 (Press CTRL+C to quit)
INFO: 127.0.0.1:64594 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:64593 - "POST /v1/models/download HTTP/1.1" 200 OK
INFO: 127.0.0.1:63947 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:63948 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:63949 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:63947 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:63948 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:63946 - "POST /v1/models/recommend HTTP/1.1" 200 OK
INFO: 127.0.0.1:63949 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:63947 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:63948 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:63945 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:63947 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:63959 - "GET /v1/models HTTP/1.1" 200 OK
[engine] 进程退出 code=1 signal=
[engine] 启动 U:\Transub-3.2.1-win\transub-engine\runtime\python.exe -m transub_engine serve --host 127.0.0.1 --port 8765
[transub-engine] ffmpeg: U:\Transub-3.2.1-win_internal\bin\ffmpeg.exe
[transub-engine] torch: preloaded 2.12.0+rocm10.0.0 cuda=None
[transub-engine] gpu: status=ready name=AMD Radeon RX 9070 cublas12=False ct2Cuda=False ortGpuCuda=False pathInjected=2 rocm=gfx1201 torchRocm=True
INFO: Started server process [9424]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:8765 (Press CTRL+C to quit)
INFO: 127.0.0.1:63797 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:63796 - "POST /v1/models/download HTTP/1.1" 200 OK
INFO transub_engine.pip_mirror_util: install_local_wheel_files: U:\Transub-3.2.1-win\transub-engine\runtime\python.exe -u -m pip install --force-reinstall --no-deps --no-index --no-cache-dir --disable-pip-version-check C:\Users\Administrator\AppData\Local\TransubEngine\wheel-cache\transformers-4.57.3-py3-none-any.whl
INFO transub_engine.runtime_extras: runtime_extras pip: U:\Transub-3.2.1-win\transub-engine\runtime\python.exe -m pip install --upgrade --force-reinstall --prefer-binary --only-binary=numba,llvmlite --find-links U:\Transub-3.2.1-win\transub-engine\wheels -i https://repo.huaweicloud.com/repository/pypi/simple --trusted-host repo.huaweicloud.com huggingface-hub>=0.34.0,<1.0 numpy>=1.24.0,<2.5 tokenizers==0.22.2
INFO: 127.0.0.1:64762 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:64763 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:64179 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:64762 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:64763 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64179 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64762 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64763 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:64179 - "POST /v1/models/recommend HTTP/1.1" 200 OK
INFO: 127.0.0.1:63796 - "POST /v1/models/recommend HTTP/1.1" 200 OK
INFO: 127.0.0.1:64762 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:64761 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:58313 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:58311 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:58311 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:58313 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:58313 - "POST /v1/models/recommend HTTP/1.1" 200 OK
INFO: 127.0.0.1:58311 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:58312 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:58312 - "POST /v1/compute/acquire HTTP/1.1" 200 OK
INFO: 127.0.0.1:58312 - "POST /v1/models/recommend HTTP/1.1" 200 OK
[engine] 开始批次 · 1 个文件 · ASR=anime-whisper · 档位=balanced
[engine] 免费推理翻译 · 引擎内 llama（跳过外部适配器）
[engine] #1/1 启动 @MFYD-183.mp4 · 准备引擎任务…
[engine] #1/1 翻译后端 · 推理翻译：sakura-galtransl-7b-q6k）
[engine] #1/1 加载识别模型 @MFYD-183.mp4 · 创建引擎任务（anime-whisper）…
[engine] #1/1 加载识别模型 @MFYD-183.mp4 · 准备原文…
[engine] #1/1 加载识别模型 @MFYD-183.mp4 · 加载识别模型（anime-whisper）…
INFO: 127.0.0.1:58312 - "POST /v1/jobs HTTP/1.1" 200 OK
[engine] #1/1 启动 @MFYD-183.mp4 · 任务已启动
[engine] #1/1 处理中 @MFYD-183.mp4 · 准备音频…
INFO transub_engine.asr.whisper_fw: anime-whisper timing: island VAD short frames (align=firered)
INFO transub_engine.asr.whisper_fw: anime-whisper FireRed path start: @MFYD-183.mp4 | model=anime-whisper | duration=11517.3s | VAD owns timeline (knobs from job vad options)
INFO transub_engine.vad.firered_vad_seg: FireRedVAD collect start: @MFYD-183.mp4 | threshold=0.45 | max_speech=5.0s | min_speech=0.18s | min_sil=0.50s | pad=0.08s | extend=0.04s | gap=0.50s | gpu=True
INFO transub_engine.vad.ten_vad_seg: TEN VAD decode: ffmpeg | @MFYD-183.mp4 | max 11517s
INFO: 127.0.0.1:58324 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:58325 - "POST /v1/models/recommend HTTP/1.1" 200 OK
INFO: 127.0.0.1:58323 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO transub_engine.vad.ten_vad_seg: TEN VAD decode ok: ffmpeg | @MFYD-183.mp4 | 11517.3s @ 16000 Hz
INFO fireredvad.utils.param: #param of DetectModel <class 'fireredvad.core.detect_model.DetectModel'> is 588417 = 2.2 MB (float32)
WARNING fireredvad.vad: Too long input, split every 30000 frames
WARNING fireredvad.core.vad_postprocessor: Unexpected short speech segment, check vad_postprocessor.py
[engine] （上条重复 4 次）
INFO transub_engine.vad.firered_vad_seg: FireRedVAD collect done: @MFYD-183.mp4 | audio=11517.3s | raw=3061 | merged=258 | spoken=11013.9s (96%) | first=32.8-71.6s | last=11494.6-11496.6s
INFO transub_engine.asr.whisper_fw: anime-whisper TEN frames: @MFYD-183.mp4 | islands=258 -> frames=2331 | spoken=11014.8s | max_frame=5.00s | first=32.8-37.8s | last=11494.6-11496.6s
OMP: Error #15: Initializing libiomp5md.dll, but found libomp140.x86_64.dll already initialized.
OMP: Hint This means that multiple copies of the OpenMP runtime have been linked into the program. That is dangerous, since it can degrade performance or cause incorrect results. The best thing to do is to ensure that only a single OpenMP runtime is linked into the process, e.g. by avoiding static linking of the OpenMP runtime in any library. As an unsafe, unsupported, undocumented workaround you can set the environment variable KMP_DUPLICATE_LIB_OK=TRUE to allow the program to continue to execute, but that may cause crashes or silently produce incorrect results. For more information, please see http://www.intel.com/software/products/support/.
[engine] 进程退出 code=3 signal=
[engine] 已请求停止当前条目（其余任务继续）
[engine] #1/2 已取消 @MFYD-183.mp4 · 已取消
[engine] 本条已停止，准备继续下一项…
[engine] 启动 U:\Transub-3.2.1-win\transub-engine\runtime\python.exe -m transub_engine serve --host 127.0.0.1 --port 8765
[transub-engine] ffmpeg: U:\Transub-3.2.1-win_internal\bin\ffmpeg.exe
[transub-engine] torch: preloaded 2.12.0+rocm10.0.0 cuda=None
[transub-engine] gpu: status=ready name=AMD Radeon RX 9070 cublas12=False ct2Cuda=False ortGpuCuda=False pathInjected=2 rocm=gfx1201 torchRocm=True
INFO: Started server process [3656]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:8765 (Press CTRL+C to quit)
[engine] #2/2 启动 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv · 准备引擎任务…
[engine] #2/2 翻译后端 · 推理翻译：sakura-galtransl-7b-q6k）
[engine] #2/2 加载识别模型 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv · 创建引擎任务（anime-whisper）…
[engine] #2/2 加载识别模型 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv · 准备原文…
[engine] #2/2 加载识别模型 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv · 加载识别模型（anime-whisper）…
INFO: 127.0.0.1:56397 - "POST /v1/jobs HTTP/1.1" 200 OK
[engine] #2/2 启动 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv · 任务已启动
[engine] #2/2 处理中 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv · 准备音频…
INFO transub_engine.asr.whisper_fw: anime-whisper timing: island VAD short frames (align=firered)
INFO transub_engine.asr.whisper_fw: anime-whisper FireRed path start: 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv | model=anime-whisper | duration=311.1s | VAD owns timeline (knobs from job vad options)
INFO transub_engine.vad.firered_vad_seg: FireRedVAD collect start: 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv | threshold=0.45 | max_speech=5.0s | min_speech=0.18s | min_sil=0.50s | pad=0.08s | extend=0.04s | gap=0.50s | gpu=True
INFO transub_engine.vad.ten_vad_seg: TEN VAD decode: ffmpeg | 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv | max 311s
INFO transub_engine.vad.ten_vad_seg: TEN VAD decode ok: ffmpeg | 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv | 311.1s @ 16000 Hz
INFO fireredvad.utils.param: #param of DetectModel <class 'fireredvad.core.detect_model.DetectModel'> is 588417 = 2.2 MB (float32)
WARNING fireredvad.vad: Too long input, split every 30000 frames
INFO transub_engine.vad.firered_vad_seg: FireRedVAD collect done: 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv | audio=311.1s | raw=66 | merged=10 | spoken=232.6s (75%) | first=23.5-116.1s | last=296.5-298.5s
INFO transub_engine.asr.whisper_fw: anime-whisper TEN frames: 《大海》日语版 JAYWALK 原作者张雨生 - 1.1(Av4495691,P1).Flv | islands=10 -> frames=51 | spoken=232.6s | max_frame=5.00s | first=23.5-28.5s | last=296.5-298.5s
OMP: Error #15: Initializing libiomp5md.dll, but found libomp140.x86_64.dll already initialized.
OMP: Hint This means that multiple copies of the OpenMP runtime have been linked into the program. That is dangerous, since it can degrade performance or cause incorrect results. The best thing to do is to ensure that only a single OpenMP runtime is linked into the process, e.g. by avoiding static linking of the OpenMP runtime in any library. As an unsafe, unsupported, undocumented workaround you can set the environment variable KMP_DUPLICATE_LIB_OK=TRUE to allow the program to continue to execute, but that may cause crashes or silently produce incorrect results. For more information, please see http://www.intel.com/software/products/support/.
[engine] 进程退出 code=3 signal=

[dlsandy]

我会在今晚再次对AMD GPU支持方面做一次全面的检查。
明天会发布修改后的版本。

[188198]

我会在今晚再次对AMD GPU支持方面做一次全面的检查。 明天会发布修改后的版本。

好的，幸苦了

[188198]

原始是你所使用的模型（qwen3-asr-1.7b-ja-anime-galgame）默认 beam=4 + 多帧 batch + bf16/SDPA，在 ROCm / RDNA4 上很容易触发该错误。 你可以下载下面这个补丁文件，并解压缩到软件目录下覆盖： Transub-3.2.1-amd-rocm-qwen-hotfix-20260920.zip

9070已经可以运行了。速度有点感人

目前的处理完一部影片的时长大约需要多久呢？ 以及你所使用的翻译模式和模型是哪个呢？
如果使用Qwen3-ASR-1.7B速度较慢的话，可以尝试换成Anime Whisper。目前ROCm对Qwen3-ASR-1.7B的运行支持仍不是很好。

预设的软对白，推理，质量，时间大概跑了下，1小时时长的要近半小时的吧，9070负荷90左右了

你可以试下组合方案内的：软对白 · 推理 · 均衡 这一档（时间参考是2小时的影片处理时间约10分钟左右），预设里的质量档所使用的sakura-galtransl-14b-q6k模型参数量偏大9070跑起来可能比较困难。

3,22 Anime Whisper依旧不可用，千问可以。不过好像报错原因不一样了

[engine] 启动 U:\Transub-3.2.2-win\transub-engine\runtime\python.exe -m transub_engine serve --host 127.0.0.1 --port 8765
[transub-engine] ffmpeg: U:\Transub-3.2.2-win_internal\bin\ffmpeg.exe
[transub-engine] torch: preloaded 2.12.0+rocm10.0.0 cuda=None
[transub-engine] gpu: status=ready name=AMD Radeon RX 9070 cublas12=False ct2Cuda=False ortGpuCuda=False pathInjected=0 rocm=gfx1201 torchRocm=True
INFO: Started server process [14444]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:8765 (Press CTRL+C to quit)
INFO: 127.0.0.1:51849 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:51847 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:51848 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:53690 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:53688 - "GET /v1/models HTTP/1.1" 200 OK
INFO: 127.0.0.1:53688 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:53690 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:53690 - "POST /v1/models/recommend HTTP/1.1" 200 OK
INFO: 127.0.0.1:53688 - "GET /v1/runtime/audio-separate HTTP/1.1" 200 OK
INFO: 127.0.0.1:53689 - "GET /v1/runtime/gpu HTTP/1.1" 200 OK
INFO: 127.0.0.1:53689 - "POST /v1/compute/acquire HTTP/1.1" 200 OK
INFO: 127.0.0.1:53690 - "POST /v1/models/recommend HTTP/1.1" 200 OK
[engine] 开始批次 · 1 个文件 · ASR=anime-whisper · 档位=quality
[engine] 免费推理翻译 · 引擎内 llama（跳过外部适配器）
[engine] #1/1 启动 4k688.com@HUNTC-626.mp4 · 准备引擎任务…
[engine] #1/1 翻译后端 · 推理翻译：sakura-galtransl-7b-q6k）
[engine] #1/1 加载识别模型 4k688.com@HUNTC-626.mp4 · 创建引擎任务（anime-whisper）…
[engine] #1/1 加载识别模型 4k688.com@HUNTC-626.mp4 · 准备原文…
[engine] #1/1 加载识别模型 4k688.com@HUNTC-626.mp4 · 加载识别模型（anime-whisper）…
INFO: 127.0.0.1:53689 - "POST /v1/jobs HTTP/1.1" 200 OK
[engine] #1/1 启动 4k688.com@HUNTC-626.mp4 · 任务已启动
[engine] #1/1 处理中 4k688.com@HUNTC-626.mp4 · 准备音频…
[engine] #1/1 轻度降噪 4k688.com@HUNTC-626.mp4 · 轻度降噪…
Traceback (most recent call last):
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\transub_engine\asr\whisper_av_shim.py", line 186, in import_whisper_model_class
from faster_whisper import WhisperModel # type: ignore
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\transub_engine\device.py", line 926, in guarded_import
return real_import(name, globals, locals, fromlist, level)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\faster_whisper_init_.py", line 1, in
from faster_whisper.audio import decode_audio
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\transub_engine\device.py", line 926, in guarded_import
return real_import(name, globals, locals, fromlist, level)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\faster_whisper\audio.py", line 15, in
import av
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\transub_engine\device.py", line 926, in guarded_import
return real_import(name, globals, locals, fromlist, level)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ModuleNotFoundError: No module named 'av'
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\transub_engine\jobs.py", line 404, in run_job
result = execute_job(job.request, job=job)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\transub_engine\jobs.py", line 1141, in execute_job
source_cues = run_asr(
^^^^^^^^
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\transub_engine\asr_init.py", line 80, in run_asr
return whisper_transcribe(
^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\transub_engine\asr\whisper_fw.py", line 2265, in whisper_transcribe
WhisperModel, used_ffmpeg_audio = import_whisper_model_class()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "U:\Transub-3.2.2-win\transub-engine\runtime\Lib\site-packages\transub_engine\asr\whisper_av_shim.py", line 192, in import_whisper_model_class
raise RuntimeError(
RuntimeError: Whisper 需要 faster-whisper。请执行: pip install 'transub-engine[asr-whisper]' 原始错误: No module named 'av'
[engine] #1/1 失败 4k688.com@HUNTC-626.mp4 · Whisper 需要 faster-whisper。请执行: pip install 'transub-engine[asr-whisper]' 原始错误: No module named 'av' · 已试 anime-whisper · 可：打开引擎日志
INFO transub_engine.runtime_release: release_gpu_memory reason=api freed=['whisper_cache:0', 'whisperseg_cache:0', 'sensevoice_cache:0', 'reazon_k2_cache:0', 'qwen3_asr_cache:0', 'parakeet_cache:0', 'cohere_asr_cache:0', 'wav2vec_align_cache:0', 'gc', 'torch.cuda.empty_cache'] errors=["ctranslate2:No module named 'ctranslate2'"]
INFO: 127.0.0.1:53688 - "POST /v1/runtime/release-gpu HTTP/1.1" 200 OK
[engine] 批次结束 · 已释放引擎显存
[engine] 任务失败或已中断 · 已停止引擎与本地 LLM 进程
[engine] 进程退出 code=1 signal=

[dlsandy]

目前Anime Whisper官方不支持AMD GPU，你可以使用Qwen3-asr，其效果要优于Anime Whisper