feat(voicestudio): force the GGUF TTS runtime onto Vulkan
CI / check (push) Waiting to run

PyTorch in the Nix env is CPU-only, so the bundled omnivoice.cpp
runtime (ggml Vulkan) is the only GPU path.

Put vulkan-tools on the backend PATH: the app finds the adapter by
running `vulkaninfo` and otherwise reports none, since NixOS has no
FHS library directories for its fallback probe.

Default OMNIVOICE_GGUF_BACKEND=vulkan so the runtime opens the GPU or
fails loudly instead of falling back to the CPU. Overridable.

Refs: nix-overlay-ltr
This commit is contained in:
2026-11-20 21:20:37 +03:00
parent fc657f4d55
commit 66c64572d2
3 changed files with 13 additions and 3 deletions
+1
View File
@@ -48,6 +48,7 @@
{"id":"nix-overlay-ka5","title":"Add Strata with ROCm 10 HIP support for gfx1151","description":"Package Niko1221/Strata using the existing TheRock ROCm 10 SDK, compile HIP kernels for gfx1151, document usage, and verify build and CLI.","status":"closed","priority":2,"issue_type":"feature","assignee":"minder","created_at":"2026-11-12T18:13:13.558104256Z","created_by":"minder","updated_at":"2026-11-12T18:28:20.466366213Z","closed_at":"2026-11-12T18:28:20.465072110Z","close_reason":"Completed: Strata 0.2.13 builds with TheRock ROCm 10.0 for gfx1151; installed command checks, formatting, flake evaluation, HTTP health/models/OpenAI/Anthropic mock APIs pass. Upstream HIP intrinsics numerical parity passed on host Radeon 8060S. Engine starts with package runtime libraries (unset inherited LD_LIBRARY_PATH pointing to GCC 15). Full model inference not exercised.","source_repo":"nix-overlay","compaction_level":0,"original_size":0} {"id":"nix-overlay-ka5","title":"Add Strata with ROCm 10 HIP support for gfx1151","description":"Package Niko1221/Strata using the existing TheRock ROCm 10 SDK, compile HIP kernels for gfx1151, document usage, and verify build and CLI.","status":"closed","priority":2,"issue_type":"feature","assignee":"minder","created_at":"2026-11-12T18:13:13.558104256Z","created_by":"minder","updated_at":"2026-11-12T18:28:20.466366213Z","closed_at":"2026-11-12T18:28:20.465072110Z","close_reason":"Completed: Strata 0.2.13 builds with TheRock ROCm 10.0 for gfx1151; installed command checks, formatting, flake evaluation, HTTP health/models/OpenAI/Anthropic mock APIs pass. Upstream HIP intrinsics numerical parity passed on host Radeon 8060S. Engine starts with package runtime libraries (unset inherited LD_LIBRARY_PATH pointing to GCC 15). Full model inference not exercised.","source_repo":"nix-overlay","compaction_level":0,"original_size":0}
{"id":"nix-overlay-kk7","title":"Add VoiceStudio package exposed through the overlay","description":"Track the flake input github:debpalash/VoiceStudio (v0.5.8) and packages/voicestudio, which re-exports upstream's flake package because our pinned nixpkgs has no electron_44-bin.","status":"closed","priority":2,"issue_type":"task","created_at":"2026-11-20T10:09:32.426719507Z","created_by":"minder","updated_at":"2026-11-20T10:09:35.769652770Z","closed_at":"2026-11-20T10:09:35.769131488Z","close_reason":"done","source_repo":"nix-overlay","compaction_level":0,"original_size":0} {"id":"nix-overlay-kk7","title":"Add VoiceStudio package exposed through the overlay","description":"Track the flake input github:debpalash/VoiceStudio (v0.5.8) and packages/voicestudio, which re-exports upstream's flake package because our pinned nixpkgs has no electron_44-bin.","status":"closed","priority":2,"issue_type":"task","created_at":"2026-11-20T10:09:32.426719507Z","created_by":"minder","updated_at":"2026-11-20T10:09:35.769652770Z","closed_at":"2026-11-20T10:09:35.769131488Z","close_reason":"done","source_repo":"nix-overlay","compaction_level":0,"original_size":0}
{"id":"nix-overlay-kzh","title":"Autoload a calibrated Kyojin HC sidecar from the model directory","description":"ms-s1-max / Qwen: use hc_gs_sidecar.safetensors by convention with a verified model-local checksum. Keep Yamz signature and unknown-file fallback, isolate persisted state by the sidecar selected, and qualify a real load of the 6.05 bpw file without a sidecar environment setting.","status":"closed","priority":2,"issue_type":"task","assignee":"minder","created_at":"2026-11-19T11:50:52.697232601Z","created_by":"minder","updated_at":"2026-11-19T11:58:32.264165296Z","closed_at":"2026-11-19T11:58:32.258339526Z","close_reason":"Canonical model-local HC calibration verified end to end","source_repo":"nix-overlay","compaction_level":0,"original_size":0,"comments":[{"id":32,"issue_id":"nix-overlay-kzh","author":"minder","text":"Scope packages/kyojin/package.nix and one upstream server patch. Install the verified sidecar under the conventional name in the local 6.05 model folder, leaving model-specific env out of the test; run detector tests and a native private service check.","created_at":"2026-11-19T11:51:18Z"},{"id":33,"issue_id":"nix-overlay-kzh","author":"minder","text":"Built .#kyojin and evaluated and built ms-s1-max with a local input override. The old and new packages started privately with the same 6.05 bpw directory, 8192 ctx and one slot, with no external sidecar path or cache tag. Old: 0/97 sidecar sites (published SHA mismatch). New: 96 sidecar, 1 plain, 0 rejected. Separate disk namespaces, same public model ID. 6/6 tool calls and 6/6 structured responses correct. Qwen source suite in the ROCm environment: 120 tests, OK (1 skipped). nixfmt --check passed. Evidence: /DATA/LLMs/benchmark_results/kyojin-hc-calibration-20261119/autoload.json","created_at":"2026-11-19T11:58:32Z"}]} {"id":"nix-overlay-kzh","title":"Autoload a calibrated Kyojin HC sidecar from the model directory","description":"ms-s1-max / Qwen: use hc_gs_sidecar.safetensors by convention with a verified model-local checksum. Keep Yamz signature and unknown-file fallback, isolate persisted state by the sidecar selected, and qualify a real load of the 6.05 bpw file without a sidecar environment setting.","status":"closed","priority":2,"issue_type":"task","assignee":"minder","created_at":"2026-11-19T11:50:52.697232601Z","created_by":"minder","updated_at":"2026-11-19T11:58:32.264165296Z","closed_at":"2026-11-19T11:58:32.258339526Z","close_reason":"Canonical model-local HC calibration verified end to end","source_repo":"nix-overlay","compaction_level":0,"original_size":0,"comments":[{"id":32,"issue_id":"nix-overlay-kzh","author":"minder","text":"Scope packages/kyojin/package.nix and one upstream server patch. Install the verified sidecar under the conventional name in the local 6.05 model folder, leaving model-specific env out of the test; run detector tests and a native private service check.","created_at":"2026-11-19T11:51:18Z"},{"id":33,"issue_id":"nix-overlay-kzh","author":"minder","text":"Built .#kyojin and evaluated and built ms-s1-max with a local input override. The old and new packages started privately with the same 6.05 bpw directory, 8192 ctx and one slot, with no external sidecar path or cache tag. Old: 0/97 sidecar sites (published SHA mismatch). New: 96 sidecar, 1 plain, 0 rejected. Separate disk namespaces, same public model ID. 6/6 tool calls and 6/6 structured responses correct. Qwen source suite in the ROCm environment: 120 tests, OK (1 skipped). nixfmt --check passed. Evidence: /DATA/LLMs/benchmark_results/kyojin-hc-calibration-20261119/autoload.json","created_at":"2026-11-19T11:58:32Z"}]}
{"id":"nix-overlay-ltr","title":"voicestudio: force the GGUF TTS runtime onto Vulkan","description":"Torch in the Nix env is CPU-only. Route the bundled omnivoice.cpp (ggml Vulkan) runtime to the GPU on NixOS: OMNIVOICE_GGUF_BACKEND=vulkan, and vulkan-tools on PATH so the app's vulkaninfo probe can see the adapter.","status":"closed","priority":2,"issue_type":"task","assignee":"minder","created_at":"2026-11-20T17:51:32.398813790Z","created_by":"minder","updated_at":"2026-11-20T18:20:37.126758327Z","closed_at":"2026-11-20T18:20:37.125560290Z","close_reason":"Completed: vulkan-tools on backend PATH + OMNIVOICE_GGUF_BACKEND=vulkan default; app API reports Vulkan0 (Intel UHD) verified via vulkaninfo; omnivoice.cpp TTS on Vulkan0 100s-class vs CPU, audio round-trips through ASR","source_repo":"nix-overlay","compaction_level":0,"original_size":0}
{"id":"nix-overlay-m5y","title":"Add CodeAF package (Agent-Field/CodeAF)","description":"Add https://github.com/Agent-Field/CodeAF as a new package under packages/codeaf with default.nix + package.nix, staged for blueprint discovery, README table updated, nix build verified.","status":"closed","priority":2,"issue_type":"feature","assignee":"Alexander Miroshnichenko","owner":"alex@millerson.name","created_at":"2026-10-06T18:07:38Z","created_by":"Alexander Miroshnichenko","updated_at":"2026-10-06T18:46:34Z","closed_at":"2026-10-06T18:46:34Z","close_reason":"Closed","source_repo":".","compaction_level":0,"original_size":0} {"id":"nix-overlay-m5y","title":"Add CodeAF package (Agent-Field/CodeAF)","description":"Add https://github.com/Agent-Field/CodeAF as a new package under packages/codeaf with default.nix + package.nix, staged for blueprint discovery, README table updated, nix build verified.","status":"closed","priority":2,"issue_type":"feature","assignee":"Alexander Miroshnichenko","owner":"alex@millerson.name","created_at":"2026-10-06T18:07:38Z","created_by":"Alexander Miroshnichenko","updated_at":"2026-10-06T18:46:34Z","closed_at":"2026-10-06T18:46:34Z","close_reason":"Closed","source_repo":".","compaction_level":0,"original_size":0}
{"id":"nix-overlay-nn2","title":"marmel: drop session-boot-trust-store.patch once upstream stops panicking on a missing CA store","description":"packages/marmel/session-boot-trust-store.patch forwards SSL_CERT_FILE into\nthe env_clear()d child in tests/test_session_boot.rs. It is a workaround,\nnot the fix: src/llm/client.rs default_http_client() calls unwrap_or_default()\non Result<Client, reqwest::Error>, and reqwest's Client::default() is\nClient::new(), which repeats the same build and .expect()s it, so the fallback\nalways panics. Reproduced independently of Nix on a stock busybox image with no\n/etc/ssl.\n\nUpstream issue filed 2026-10-16, still open:\nhttps://github.com/Na1w/marmel/issues/13\n(\"marmel panics at startup instead of reporting a missing CA trust store\")\n\nBlocked on upstream. Once upstream propagates the error instead of discarding\nit, drop the patch, bump marmel to the fixing commit, and rebuild.\n","status":"closed","priority":3,"issue_type":"task","owner":"alex@millerson.name","created_at":"2026-10-16T11:32:29Z","created_by":"Alexander Miroshnichenko","updated_at":"2026-11-09T07:03:29.680285424Z","closed_at":"2026-11-09T07:03:29.673967963Z","close_reason":"Upstream fixed in 62cb756 (#16, 2026-10-18); session-boot-trust-store.patch was already dropped in adbed5f3c074. Current pin afcc1dc is 31 commits ahead of the fix and the package carries no such patch.","external_ref":"https://github.com/Na1w/marmel/issues/13","source_repo":".","compaction_level":0,"original_size":0} {"id":"nix-overlay-nn2","title":"marmel: drop session-boot-trust-store.patch once upstream stops panicking on a missing CA store","description":"packages/marmel/session-boot-trust-store.patch forwards SSL_CERT_FILE into\nthe env_clear()d child in tests/test_session_boot.rs. It is a workaround,\nnot the fix: src/llm/client.rs default_http_client() calls unwrap_or_default()\non Result<Client, reqwest::Error>, and reqwest's Client::default() is\nClient::new(), which repeats the same build and .expect()s it, so the fallback\nalways panics. Reproduced independently of Nix on a stock busybox image with no\n/etc/ssl.\n\nUpstream issue filed 2026-10-16, still open:\nhttps://github.com/Na1w/marmel/issues/13\n(\"marmel panics at startup instead of reporting a missing CA trust store\")\n\nBlocked on upstream. Once upstream propagates the error instead of discarding\nit, drop the patch, bump marmel to the fixing commit, and rebuild.\n","status":"closed","priority":3,"issue_type":"task","owner":"alex@millerson.name","created_at":"2026-10-16T11:32:29Z","created_by":"Alexander Miroshnichenko","updated_at":"2026-11-09T07:03:29.680285424Z","closed_at":"2026-11-09T07:03:29.673967963Z","close_reason":"Upstream fixed in 62cb756 (#16, 2026-10-18); session-boot-trust-store.patch was already dropped in adbed5f3c074. Current pin afcc1dc is 31 commits ahead of the fix and the package carries no such patch.","external_ref":"https://github.com/Na1w/marmel/issues/13","source_repo":".","compaction_level":0,"original_size":0}
{"id":"nix-overlay-ocf","title":"agnix: ship agnix-lsp and agnix-mcp in the same package","description":"Extend packages/agnix to build -p agnix-cli -p agnix-lsp -p agnix-mcp, as upstream's release job does.","status":"closed","priority":3,"issue_type":"task","assignee":"minder","created_at":"2026-11-03T12:17:08.080999670Z","created_by":"minder","updated_at":"2026-11-03T12:33:21.279454904Z","closed_at":"2026-11-03T12:33:21.278951385Z","close_reason":"packages/agnix now builds -p agnix-cli -p agnix-lsp -p agnix-mcp; nix build .#agnix produces all three binaries; agnix-lsp answers LSP initialize with capabilities, agnix-mcp answers MCP initialize with serverInfo agnix 0.60.3","source_repo":"nix-overlay","compaction_level":0,"original_size":0} {"id":"nix-overlay-ocf","title":"agnix: ship agnix-lsp and agnix-mcp in the same package","description":"Extend packages/agnix to build -p agnix-cli -p agnix-lsp -p agnix-mcp, as upstream's release job does.","status":"closed","priority":3,"issue_type":"task","assignee":"minder","created_at":"2026-11-03T12:17:08.080999670Z","created_by":"minder","updated_at":"2026-11-03T12:33:21.279454904Z","closed_at":"2026-11-03T12:33:21.278951385Z","close_reason":"packages/agnix now builds -p agnix-cli -p agnix-lsp -p agnix-mcp; nix build .#agnix produces all three binaries; agnix-lsp answers LSP initialize with capabilities, agnix-mcp answers MCP initialize with serverInfo agnix 0.60.3","source_repo":"nix-overlay","compaction_level":0,"original_size":0}
+1 -1
View File
@@ -58,7 +58,7 @@ A custom Nix overlay and flake providing additional packages not found in upstre
| `relay-free-llm` | RESTful API to route user prompts to various AI model providers with automatic failover and intent-based routing | AI LLM Gateway | | `relay-free-llm` | RESTful API to route user prompts to various AI model providers with automatic failover and intent-based routing | AI LLM Gateway |
| `stakpak` | DevOps AI agent that generates infrastructure code, debugs Kubernetes, configures CI/CD, and automates deployments | AI Agents | | `stakpak` | DevOps AI agent that generates infrastructure code, debugs Kubernetes, configures CI/CD, and automates deployments | AI Agents |
| `traycer` | Open-source AI orchestration app for agentic coding | AI Coding Agents | | `traycer` | Open-source AI orchestration app for agentic coding | AI Coding Agents |
| `voicestudio` | Fully-local voice studio — cloning, design, dubbing, dictation, transcription and audiobook creation in 646 languages; Python backend built by Nix (CPU torch), no runtime download on first launch | AI Audio | | `voicestudio` | Fully-local voice studio — cloning, design, dubbing, dictation, transcription and audiobook creation in 646 languages; Python backend built by Nix (CPU torch, GGUF engine pinned to Vulkan), no runtime download on first launch | AI Audio |
## Usage ## Usage
+11 -2
View File
@@ -37,12 +37,21 @@ else
backend = upstream.writeShellApplication { backend = upstream.writeShellApplication {
name = "voicestudio-backend"; name = "voicestudio-backend";
# Opt-in engines build their own venvs with `uv venv --python 3.11`; ffmpeg and uv # Opt-in engines build their own venvs with `uv venv --python 3.11`; ffmpeg and uv
# are already on PATH from upstream's launcher. # are already on PATH from upstream's launcher. The app finds the GPU by running
runtimeInputs = [ upstream.python311 ]; # `vulkaninfo`; without it on PATH it reports no adapter and offers no GPU engine.
runtimeInputs = [
upstream.python311
upstream.vulkan-tools
];
text = '' text = ''
export LD_LIBRARY_PATH="${libraryPath}''${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" export LD_LIBRARY_PATH="${libraryPath}''${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
# Never fetch an interpreter: uv's own CPython cannot run here. # Never fetch an interpreter: uv's own CPython cannot run here.
export UV_PYTHON_DOWNLOADS=never UV_PYTHON_PREFERENCE=only-system export UV_PYTHON_DOWNLOADS=never UV_PYTHON_PREFERENCE=only-system
# Torch is CPU-only here, so the bundled omnivoice.cpp runtime (GGUF engine) is the
# only GPU path. Pin it to Vulkan rather than let ggml fall back to the CPU quietly;
# a missing driver then fails loudly. Override: OMNIVOICE_GGUF_BACKEND=auto|cpu, and
# OMNIVOICE_GGUF_DEVICE=<n> to pick another adapter.
export OMNIVOICE_GGUF_BACKEND="''${OMNIVOICE_GGUF_BACKEND:-vulkan}"
cd ${resources} cd ${resources}
# -s -B: no user site, and the sources live in the read-only store. # -s -B: no user site, and the sources live in the read-only store.