Skip to content

Speakr is a personal, self-hosted web application designed for transcribing audio recordings

Python

Add PyTorch 2.6 weights_only fix documentation for WhisperX ASR

Murtaza Nasir2b32a3dd2805
diff --git a/README.md b/README.md
index 95cd617f..4922e18e 100644
--- a/README.md
+++ b/README.md
@@ -131,6 +131,8 @@ docker compose up -d
- **For voice profiles & speaker embeddings:** Use [WhisperX ASR Service](https://github.com/murtaza-nasir/whisperx-asr-service) (recommended)
- **For basic speaker diarization:** Use [OpenAI Whisper ASR Webservice](https://github.com/ahmetoner/whisper-asr-webservice)
+> **⚠️ PyTorch 2.6 Users:** If you encounter a "Weights only load failed" error with WhisperX, add `TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=true` to your ASR container's environment in docker-compose.yml. See [troubleshooting](https://murtaza-nasir.github.io/speakr/troubleshooting#pytorch-26-weights-loading-error-whisperx-asr-service) for details.
+
See [installation guide](https://murtaza-nasir.github.io/speakr/getting-started/installation#running-asr-service-for-speaker-diarization) for complete setup instructions.
**[View Full Installation Guide →](https://murtaza-nasir.github.io/speakr/getting-started)**
diff --git a/docs/getting-started/installation.md b/docs/getting-started/installation.md
index 5b070bca..5406055f 100644
--- a/docs/getting-started/installation.md
+++ b/docs/getting-started/installation.md
@@ -295,6 +295,9 @@ curl http://localhost:9000/health
See the [WhisperX ASR Service README](https://github.com/murtaza-nasir/whisperx-asr-service#readme) for detailed configuration options, troubleshooting, and performance tuning.
+!!! warning "PyTorch 2.6 Compatibility"
+ If you encounter a "Weights only load failed" error, add `TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=true` to your ASR container's environment variables in docker-compose.yml. See [troubleshooting](../troubleshooting.md#pytorch-26-weights-loading-error-whisperx-asr-service) for details.
+
#### Option 2: OpenAI Whisper ASR Webservice (Basic Diarization Only)
**Use this if you:**
@@ -336,6 +339,7 @@ services:
- COMPUTE_TYPE=float16
- BATCH_SIZE=16
- HF_TOKEN=your_huggingface_token_here
+ - TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=true # Required for PyTorch 2.6+
deploy:
resources:
reservations:
@@ -411,6 +415,7 @@ services:
- ASR_COMPUTE_TYPE=int8
- ASR_ENGINE=whisperx
- HF_TOKEN=your_huggingface_token_here
+ - TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=true # Required for PyTorch 2.6+
deploy:
resources:
reservations:
@@ -473,6 +478,7 @@ services:
- ASR_COMPUTE_TYPE=int8
- ASR_ENGINE=whisperx
- HF_TOKEN=your_huggingface_token_here
+ - TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=true # Required for PyTorch 2.6+
deploy:
resources:
reservations:
diff --git a/docs/index.md b/docs/index.md
index a44a63f3..c981deca 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -147,6 +147,14 @@ Learn more about [audio synchronization features](user-guide/transcripts.md#audi
## Latest Updates
+!!! warning "PyTorch 2.6 Compatibility Issue with WhisperX ASR"
+ If you're using the WhisperX ASR service and encounter a "Weights only load failed" error after a recent update, add this environment variable to your ASR container in docker-compose.yml:
+ ```yaml
+ environment:
+ - TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=true
+ ```
+ This is caused by a PyTorch 2.6 change. See [troubleshooting](troubleshooting.md#pytorch-26-weights-loading-error-whisperx-asr-service) for details.
+
!!! info "Version 0.6.5 - Separate Chat Model Configuration"
**New Feature** - Configure different AI models for chat vs background tasks
diff --git a/docs/troubleshooting.md b/docs/troubleshooting.md
index e3db3e29..ef43fda8 100644
--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@@ -122,6 +122,38 @@ Even with ASR enabled, you must explicitly request diarization when uploading or
After transcription, speakers appear as generic labels (SPEAKER_01, etc.). You must manually [identify speakers](user-guide/transcripts.md#speaker-identification) by clicking the labels and assigning names. Manage your [speaker library](user-guide/settings.md#speakers-management-tab) in account settings.
+### PyTorch 2.6 Weights Loading Error (WhisperX ASR Service)
+
+If you're getting an error like this when trying to transcribe:
+
+```
+Weights only load failed. This file can still be loaded...
+WeightsUnpickler error: Unsupported global: GLOBAL omegaconf.listconfig.ListConfig was not an allowed global by default.
+```
+
+This is caused by a breaking change in PyTorch 2.6 where the default value of `weights_only` in `torch.load` changed from `False` to `True`. The pyannote models used for voice activity detection in WhisperX are affected.
+
+**Solution**: Add this environment variable to your WhisperX ASR service container in your docker-compose.yml:
+
+```yaml
+whisper-asr:
+ image: learnedmachine/whisperx-asr-service:latest
+ environment:
+ - TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=true
+ # ... your other environment variables
+```
+
+!!! warning "Important"
+ This environment variable must be set directly in the docker-compose.yml file under the `environment` section of the ASR service container. Setting it in a `.env` file may not work reliably.
+
+After adding this variable, restart your ASR container:
+```bash
+docker compose down whisper-asr
+docker compose up -d whisper-asr
+```
+
+For more details, see the [WhisperX issue discussion](https://github.com/m-bain/whisperX/issues/1304).
+
### WhisperX Shows UNKNOWN_SPEAKER
If WhisperX only shows "UNKNOWN_SPEAKER" instead of numbered speakers (SPEAKER_00, SPEAKER_01, etc.), check these common issues: