Simon/Scriberr
1
0
Fork 0
mirror of https://github.com/rishikanthc/Scriberr.git synced 2026-03-03 02:57:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat: add RTX 5090 Blackwell GPU support (sm_120)

Browse Source 
Add support for NVIDIA RTX 50-series GPUs (Blackwell architecture) which
require CUDA 12.8+ and PyTorch cu128 wheels due to the new sm_120 compute
capability.

Changes:
- Add configurable PYTORCH_CUDA_VERSION environment variable to control
  PyTorch wheel version at runtime (cu126 for legacy, cu128 for Blackwell)
- Update all model adapters to use dynamic CUDA version instead of
  hardcoded cu126 URLs
- Update Dockerfile.cuda.12.9 for Blackwell with CUDA 12.9.1 base image,
  PYTORCH_CUDA_VERSION=cu128, and missing WHISPERX_ENV/yt-dlp
- Update Dockerfile.cuda with explicit PYTORCH_CUDA_VERSION=cu126
- Add docker-compose.blackwell.yml for pre-built Blackwell image
- Add docker-compose.build.blackwell.yml for local Blackwell builds
- Add GPU compatibility documentation to README

Fixes: rishikanthc/Scriberr#104
This commit is contained in:
Fran Fitzpatrick
2025-12-23 08:36:23 -06:00
committed by Rishikanth Chandrasekaran
parent 718cb74b70
commit 8f537548d4
11 changed files with 146 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
Dockerfile.cuda
View File

@@ -59,7 +59,9 @@ ENV PYTHONUNBUFFERED=1 \
PUID=1000 \
PGID=1000 \
NVIDIA_VISIBLE_DEVICES=all \
NVIDIA_DRIVER_CAPABILITIES=compute,utility
NVIDIA_DRIVER_CAPABILITIES=compute,utility \
# PyTorch CUDA wheel version: cu126 for legacy GPUs (GTX 10-series through RTX 40-series)
PYTORCH_CUDA_VERSION=cu126
WORKDIR /app

15
Dockerfile.cuda.12.9
View File

@@ -1,4 +1,6 @@
# Multi-stage build for Scriberr with CUDA support
# Multi-stage build for Scriberr with CUDA 12.9 support (Blackwell GPUs)
# This Dockerfile targets NVIDIA RTX 50-series (Blackwell architecture, sm_120)
# For legacy GPUs (GTX 10-series through RTX 40-series), use Dockerfile.cuda instead
# Builds React UI and Go server, then ships with NVIDIA CUDA runtime
########################
@@ -54,10 +56,14 @@ ENV PYTHONUNBUFFERED=1 \
PORT=8080 \
DATABASE_PATH=/app/data/scriberr.db \
UPLOAD_DIR=/app/data/uploads \
WHISPERX_ENV=/app/whisperx-env \
APP_ENV=production \
PUID=1000 \
PGID=1000 \
NVIDIA_VISIBLE_DEVICES=all \
NVIDIA_DRIVER_CAPABILITIES=compute,utility
NVIDIA_DRIVER_CAPABILITIES=compute,utility \
# PyTorch CUDA wheel version: cu128 for Blackwell GPUs (RTX 50-series, sm_120)
PYTORCH_CUDA_VERSION=cu128
WORKDIR /app
@@ -76,6 +82,11 @@ RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
&& chmod 755 /usr/local/bin/uv \
&& uv --version
# Install yt-dlp standalone binary (required by API handlers for YouTube URL processing)
RUN curl -L https://github.com/yt-dlp/yt-dlp/releases/latest/download/yt-dlp -o /usr/local/bin/yt-dlp \
&& chmod a+rx /usr/local/bin/yt-dlp \
&& yt-dlp --version
# Install Deno (JavaScript runtime required for yt-dlp YouTube downloads)
# YouTube now requires JS execution for video cipher decryption
# See: https://github.com/yt-dlp/yt-dlp/issues/14404

24
README.md
View File

@@ -275,6 +275,30 @@ volumes:
docker compose -f docker-compose.cuda.yml up -d
```
#### GPU Compatibility
Scriberr provides separate Docker images for different NVIDIA GPU generations due to CUDA/PyTorch compatibility requirements:
| GPU Generation | Compute Capability | Docker Image | Docker Compose File |
|:---|:---|:---|:---|
| GTX 10-series (Pascal) | sm_61 | `scriberr-cuda` | `docker-compose.cuda.yml` |
| RTX 20-series (Turing) | sm_75 | `scriberr-cuda` | `docker-compose.cuda.yml` |
| RTX 30-series (Ampere) | sm_86 | `scriberr-cuda` | `docker-compose.cuda.yml` |
| RTX 40-series (Ada Lovelace) | sm_89 | `scriberr-cuda` | `docker-compose.cuda.yml` |
| **RTX 50-series (Blackwell)** | sm_120 | `scriberr-cuda-blackwell` | `docker-compose.blackwell.yml` |
**RTX 50-series users (RTX 5080, 5090, etc.):** You must use the Blackwell-specific image. The standard CUDA image will not work due to PyTorch CUDA compatibility requirements. Use:
```bash
docker compose -f docker-compose.blackwell.yml up -d
```
Or for local builds:
```bash
docker compose -f docker-compose.build.blackwell.yml up -d
```
### App Startup
When you run Scriberr for the first time, it may take several minutes to start. This is normal!

34
docker-compose.blackwell.yml Normal file
View File

@@ -0,0 +1,34 @@
# Docker Compose for Scriberr with Blackwell GPU support (RTX 50-series)
# Uses pre-built image from GitHub Container Registry
# For legacy GPUs (GTX 10-series through RTX 40-series), use docker-compose.cuda.yml instead
version: "3.9"
services:
scriberr:
image: ghcr.io/rishikanthc/scriberr-cuda-blackwell:latest
ports:
- "8080:8080"
volumes:
- scriberr_data:/app/data
- env_data:/app/whisperx-env
restart: unless-stopped
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities:
- gpu
environment:
- NVIDIA_VISIBLE_DEVICES=all
- NVIDIA_DRIVER_CAPABILITIES=compute,utility
- PUID=${PUID:-1000}
- PGID=${PGID:-1000}
# Security: already set in container, but can be overridden
- APP_ENV=production
# CORS: comma-separated list of allowed origins for production
# - ALLOWED_ORIGINS=https://your-domain.com
volumes:
scriberr_data: {}
env_data: {}

37
docker-compose.build.blackwell.yml Normal file
View File

@@ -0,0 +1,37 @@
# Docker Compose for building Scriberr with Blackwell GPU support locally
# Uses Dockerfile.cuda.12.9 for RTX 50-series (Blackwell architecture, sm_120)
# For legacy GPUs (GTX 10-series through RTX 40-series), use docker-compose.build.cuda.yml instead
services:
scriberr:
build:
context: .
dockerfile: Dockerfile.cuda.12.9
image: scriberr:local-blackwell
container_name: scriberr-blackwell
ports:
- "8080:8080"
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities:
- gpu
environment:
- NVIDIA_VISIBLE_DEVICES=all
- NVIDIA_DRIVER_CAPABILITIES=compute,utility
- PUID=${PUID:-1000}
- PGID=${PGID:-1000}
# Security: already set in container, but can be overridden
- APP_ENV=production
# CORS: comma-separated list of allowed origins for production
# - ALLOWED_ORIGINS=https://your-domain.com
volumes:
- ./scriberr_data:/app/data
- ./env-data:/app/whisperx-env
restart: unless-stopped
volumes:
scriberr_data:
env-data:

16
internal/transcription/adapters/base_adapter.go
View File

@@ -27,6 +27,22 @@ var (
requestGroup singleflight.Group
)
// GetPyTorchCUDAVersion returns the PyTorch CUDA wheel version to use.
// This is configurable via the PYTORCH_CUDA_VERSION environment variable.
// Defaults to "cu126" for CUDA 12.6 (legacy GPUs: GTX 10-series through RTX 40-series).
// Set to "cu128" for CUDA 12.8 (Blackwell GPUs: RTX 50-series).
func GetPyTorchCUDAVersion() string {
if cudaVersion := os.Getenv("PYTORCH_CUDA_VERSION"); cudaVersion != "" {
return cudaVersion
}
return "cu126" // Default to CUDA 12.6 for legacy compatibility
}
// GetPyTorchWheelURL returns the full PyTorch wheel URL for the configured CUDA version.
func GetPyTorchWheelURL() string {
return fmt.Sprintf("https://download.pytorch.org/whl/%s", GetPyTorchCUDAVersion())
}
// CheckEnvironmentReady checks if a UV environment is ready with caching and singleflight
func CheckEnvironmentReady(envPath, importStatement string) bool {
cacheKey := fmt.Sprintf("%s:%s", envPath, importStatement)

8
internal/transcription/adapters/canary_adapter.go
View File

@@ -207,8 +207,8 @@ func (c *CanaryAdapter) setupCanaryEnvironment() error {
return nil
}
// Create pyproject.toml (same as Parakeet since they share environment)
pyprojectContent := `[project]
// Create pyproject.toml with configurable PyTorch CUDA version
pyprojectContent := fmt.Sprintf(`[project]
name = "parakeet-transcription"
version = "0.1.0"
description = "Audio transcription using NVIDIA Parakeet models"
@@ -241,14 +241,14 @@ triton = [
[[tool.uv.index]]
name = "pytorch"
url = "https://download.pytorch.org/whl/cu126"
url = "%s"
explicit = true
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true
`
`, GetPyTorchWheelURL())
if err := os.WriteFile(pyprojectPath, []byte(pyprojectContent), 0644); err != nil {
return fmt.Errorf("failed to write pyproject.toml: %w", err)
}

8
internal/transcription/adapters/parakeet_adapter.go
View File

@@ -179,8 +179,8 @@ func (p *ParakeetAdapter) setupParakeetEnvironment() error {
return fmt.Errorf("failed to create parakeet directory: %w", err)
}
// Create pyproject.toml
pyprojectContent := `[project]
// Create pyproject.toml with configurable PyTorch CUDA version
pyprojectContent := fmt.Sprintf(`[project]
name = "parakeet-transcription"
version = "0.1.0"
description = "Audio transcription using NVIDIA Parakeet models"
@@ -213,14 +213,14 @@ triton = [
[[tool.uv.index]]
name = "pytorch"
url = "https://download.pytorch.org/whl/cu126"
url = "%s"
explicit = true
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true
`
`, GetPyTorchWheelURL())
pyprojectPath := filepath.Join(p.envPath, "pyproject.toml")
if err := os.WriteFile(pyprojectPath, []byte(pyprojectContent), 0644); err != nil {
return fmt.Errorf("failed to write pyproject.toml: %w", err)

8
internal/transcription/adapters/pyannote_adapter.go
View File

@@ -216,10 +216,10 @@ func (p *PyAnnoteAdapter) setupPyAnnoteEnvironment() error {
return fmt.Errorf("failed to create pyannote directory: %w", err)
}
// Create pyproject.toml for PyAnnote
// Create pyproject.toml with configurable PyTorch CUDA version
// Note: We explicitly pin torch and torchaudio to 2.1.2 to ensure compatibility with pyannote.audio 3.1
// Newer versions of torchaudio (2.2+) removed AudioMetaData which causes crashes
pyprojectContent := `[project]
pyprojectContent := fmt.Sprintf(`[project]
name = "pyannote-diarization"
version = "0.1.0"
description = "Audio diarization using PyAnnote"
@@ -245,14 +245,14 @@ torchaudio = [
[[tool.uv.index]]
name = "pytorch"
url = "https://download.pytorch.org/whl/cu126"
url = "%s"
explicit = true
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true
`
`, GetPyTorchWheelURL())
pyprojectPath := filepath.Join(p.envPath, "pyproject.toml")
if err := os.WriteFile(pyprojectPath, []byte(pyprojectContent), 0644); err != nil {
return fmt.Errorf("failed to write pyproject.toml: %w", err)

8
internal/transcription/adapters/sortformer_adapter.go
View File

@@ -195,8 +195,8 @@ func (s *SortformerAdapter) setupSortformerEnvironment() error {
return fmt.Errorf("failed to create sortformer directory: %w", err)
}
// Create pyproject.toml (same as other NVIDIA models)
pyprojectContent := `[project]
// Create pyproject.toml with configurable PyTorch CUDA version
pyprojectContent := fmt.Sprintf(`[project]
name = "parakeet-transcription"
version = "0.1.0"
description = "Audio transcription using NVIDIA Parakeet models"
@@ -230,14 +230,14 @@ triton = [
[[tool.uv.index]]
name = "pytorch"
url = "https://download.pytorch.org/whl/cu126"
url = "%s"
explicit = true
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true
`
`, GetPyTorchWheelURL())
pyprojectPath := filepath.Join(s.envPath, "pyproject.toml")
if err := os.WriteFile(pyprojectPath, []byte(pyprojectContent), 0644); err != nil {
return fmt.Errorf("failed to write pyproject.toml: %w", err)

5
internal/transcription/adapters/whisperx_adapter.go
View File

@@ -359,9 +359,10 @@ func (w *WhisperXAdapter) updateWhisperXDependencies(whisperxPath string) error
"yt-dlp[default]",`)
}
// Pin PyTorch to CUDA 12.6
// Set PyTorch CUDA version based on environment configuration
// The repo already has the correct [tool.uv.sources] configuration, we just need to update the CUDA version
content = strings.ReplaceAll(content, "https://download.pytorch.org/whl/cu128", "https://download.pytorch.org/whl/cu126")
// This allows using cu126 for legacy GPUs (GTX 10-series through RTX 40-series) or cu128 for Blackwell (RTX 50-series)
content = strings.ReplaceAll(content, "https://download.pytorch.org/whl/cu128", GetPyTorchWheelURL())
if err := os.WriteFile(pyprojectPath, []byte(content), 0644); err != nil {
return fmt.Errorf("failed to write pyproject.toml: %w", err)