nsaspy/zara
1
0
Fork 0
Code Issues 95 Pull requests 18 Projects Releases Packages Wiki Activity Actions

feat(stt): route AMD device tokens to the whisper.cpp Vulkan backend #272

Closed
nsaspy wants to merge 4 commits from rage/250-amd-stt-device-routing into master
pull from: rage/250-amd-stt-device-routing
merge into: nsaspy:master
Conversation 0 Commits 4 Files changed 8 +420 -33
nsaspy commented 2026-09-04 22:44:05 +00:00
Owner
Copy link

Closes #250

Problem

On an AMD-only machine (RX 5500 XT, RADV Vulkan; no CUDA, no ROCm userspace):

  • --wake --device rocm (default provider) rewrote the AMD token to CUDA, failed, and silently continued on CPU.
  • --wake --device vulkan crashed at argument parsing with ValueError: Unsupported STT device 'vulkan' despite being an advertised choice.

The flake already records the architecture: whisper.cpp is the AMD/Radeon STT path (Vulkan) — and whisperCppVulkan is packaged — but the CLI could never reach it from the AMD tokens.

Fix

  • amd|rocm|hip|vulkan normalize to vulkan for every provider; cuda keeps NVIDIA-only semantics; cuda on whisper-cpp still fails with actionable guidance.
  • Local non-whisper-cpp providers are rerouted to the whisper.cpp Vulkan backend with a visible stderr notice, before model resolution (so GGML files resolve).
  • Remote providers (openai, groq) are never rerouted; their device token stays inert.
  • --device help, config.py comment, and the dev-shell hint updated.

Evidence (RAGE run rage/250-amd-stt-device-routing, start 9554387)

  • RED: 18 deterministic failures encoding the new contract (tokens → cuda, missing routing helper, exact user ValueError).
  • GREEN after the minimal __main__.py change; t/test_dictation.py internal compat contract untouched and green (64/64 focused).
  • Exact head cbf2aad: scripts/test-all.sh 11/11 phases, nix flake check all checks, nix build produced zarathushtra-full.
  • Research/design artifacts: rage/250-amd-stt-device-routing-{research,design}.org.
Closes #250 ## Problem On an AMD-only machine (RX 5500 XT, RADV Vulkan; no CUDA, no ROCm userspace): - `--wake --device rocm` (default provider) rewrote the AMD token to **CUDA**, failed, and silently continued on CPU. - `--wake --device vulkan` crashed at argument parsing with `ValueError: Unsupported STT device 'vulkan'` despite being an advertised choice. The flake already records the architecture: *whisper.cpp is the AMD/Radeon STT path (Vulkan)* — and `whisperCppVulkan` is packaged — but the CLI could never reach it from the AMD tokens. ## Fix - `amd|rocm|hip|vulkan` normalize to `vulkan` for **every** provider; `cuda` keeps NVIDIA-only semantics; `cuda` on whisper-cpp still fails with actionable guidance. - Local non-whisper-cpp providers are rerouted to the whisper.cpp Vulkan backend with a visible stderr notice, before model resolution (so GGML files resolve). - Remote providers (`openai`, `groq`) are never rerouted; their device token stays inert. - `--device` help, `config.py` comment, and the dev-shell hint updated. ## Evidence (RAGE run `rage/250-amd-stt-device-routing`, start `9554387`) - RED: 18 deterministic failures encoding the new contract (tokens → cuda, missing routing helper, exact user `ValueError`). - GREEN after the minimal `__main__.py` change; `t/test_dictation.py` internal compat contract untouched and green (64/64 focused). - Exact head `cbf2aad`: `scripts/test-all.sh` 11/11 phases, `nix flake check` all checks, `nix build` produced `zarathushtra-full`. - Research/design artifacts: `rage/250-amd-stt-device-routing-{research,design}.org`.
--device amd/rocm/hip/vulkan request AMD GPU acceleration. They now
normalize to vulkan for every provider instead of lying into CUDA, and
local non-whisper-cpp providers are rerouted to the whisper.cpp Vulkan
backend with a visible notice instead of silently degrading to CPU or
crashing on 'vulkan'. cuda keeps NVIDIA semantics with the existing CPU
fallback; remote providers are never rerouted.

Consumed issue: #250. RAGE start: 9554387.
docs(rage): record 250 merge evidence
Some checks failed
CI / test (pull_request) Failing after 6s
Details
CI / android skeleton gate (pull_request) Failing after 6s
Details
CI / shared mic / Arch Linux (pull_request) Failing after 8m57s
Details
CI / shared mic / Ubuntu 24.04 (pull_request) Failing after 8m51s
Details
570beec07a
nsaspy closed this pull request 2026-09-04 23:09:25 +00:00
Some checks failed
CI / test (pull_request) Failing after 6s
Details
CI / android skeleton gate (pull_request) Failing after 6s
Details
CI / shared mic / Arch Linux (pull_request) Failing after 8m57s
Details
CI / shared mic / Ubuntu 24.04 (pull_request) Failing after 8m51s
Details

Pull request closed

Please reopen this pull request to perform a merge.
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels   
bug
Something isn't working

  
documentation
Improvements or additions to documentation

  
duplicate
This issue or pull request already exists

  
enhancement
New feature or request

  
good first issue
Good for newcomers

  
help wanted
Extra attention is needed

  
invalid
This doesn't seem right

  
question
Further information is requested

  
wontfix
This will not be worked on

No labels
bug
documentation
duplicate
enhancement
good first issue
help wanted
invalid
question
wontfix
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
nsaspy/zara!272
No description provided.