Skip to content

REFERENCE_HARDWARE.md

Latest commit

 

History

History
38 lines (29 loc) · 2.21 KB

REFERENCE_HARDWARE.md

File metadata and controls

38 lines (29 loc) · 2.21 KB

Hardware & Environment Compatibility Guide (RTX 5090)

This document summarizes the specific library versions required to run the Jarvis voice assistant on an NVIDIA RTX 5090 (Blackwell architecture).

Key Versions

Component Version Justification
Operating System Windows 10/11 (win32) Project environment.
Python 3.10.x Compatibility with chatterbox-tts and faster-whisper.
PyTorch 2.11.0.dev20260206+cu128 CRITICAL: RTX 5090 (sm_120) requires CUDA 12.8+ support found in Nightly builds. Stable 2.5/2.6 (CUDA 12.1/12.4) fails with "no kernel image".
CUDA 12.8 (via Torch Nightly) Required for Blackwell GPU architecture.
TorchAudio 2.11.0.dev20260206+cu128 Must match PyTorch version/index.
TorchVision 0.26.0.dev20260206+cu128 Must match PyTorch version/index.
NumPy 1.25.2 FORCED: chatterbox-tts requires < 1.26.0. Pipecat-ai prefers higher but works with this version.
Pillow 11.3.0 FORCED: pipecat-ai requires < 12.0.0.
Pipecat-ai 0.0.101 Core orchestration framework.

Dependency Conflict Resolution

  1. RTX 5090 "No Kernel Image" Error: Standard PyTorch releases do not yet include the Blackwell compute kernels. We switched to the Nightly Index (https://download.pytorch.org/whl/nightly/cu128) to resolve this.

  2. The NumPy Conflict: pipecat-ai tries to install NumPy 2.x, but chatterbox-tts (Resemble AI) has a strict check for NumPy 1.x (specifically < 1.26.0). We manually pinned 1.25.2 to satisfy both.

  3. Chatterbox strict Torch check: chatterbox-tts originally requested torch==2.6.0. Since the 5090 required the Nightly version (2.11.0.dev), we performed a --force-reinstall to bypass the strict equality check while maintaining system functionality.

Reinstallation Command

If the environment needs to be rebuilt, use this specific order:

# 1. Install GPU compute stack
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 --force-reinstall

# 2. Fix downstream conflicts
pip install "numpy<1.26.0,>=1.24.0" "pillow<12.0,>=11.1.0"