Skip to content

RTX 50xx (Blackwell) owners: field reports wanted — sm_120 support shipped in 0.4.4, needs real-hardware validation #12

Description

@Martossien

TL;DR: if you own an RTX 5060 Ti / 5070 / 5080 / 5090, one run + your logs would validate work that is already shipped.

Release 0.4.4 added native Blackwell (sm_120) support everywhere:

  • Docker images: CUDA 12.8.1 base + torch cu130, sm_120 SASS compiled into all three CUDA stacks (llama.cpp, audio.cpp, parakeet.cpp) — verified with cuobjdump on the built binaries. Requires NVIDIA driver ≥ 580.
  • Native install: install.sh now selects the torch cu130 wheel automatically when your driver reports CUDA ≥ 13 (drivers in the 12.8–12.9 range get cu126 plus an actionable warning to upgrade to ≥ 580).

What has NOT happened yet: a full run on real Blackwell silicon. Everything is proven at the binary level, nothing at runtime.

What a field report looks like

  1. Driver ≥ 580 (nvidia-smi shows CUDA 13.x).
  2. Either path:
  3. Run one meeting recording through the portal (any language, a few minutes of audio is enough).
  4. Paste here: your GPU model, driver version, which path you took, and whether the job completed — with logs if anything failed (especially the LLM stage and the STT stage).

A failure report is just as valuable as a success report — the last external tester's two bug reports shaped an entire release. Merci !

Metadata

Metadata

Assignees

No one assigned

    Labels

    help wantedExtra attention is needed

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions