LlamaModelParams::default() sets n_gpu_layers=0, so even with --features
rocm the model ran entirely on CPU at ~19 t/s. Setting n_gpu_layers to a
large sentinel value asks llama.cpp to offload every layer the model
has; llama.cpp clamps to the real count (27 for Gemma 2 2B). Observed
throughput jumps from 19 t/s to 74 t/s on an RX 9070 once the ROCm
binary is also compiled for gfx1201 (see tooling commit).
Also adds server/sr-voice/.gitignore so locally-built binaries don't
sneak into the worktree. Release binaries ship out-of-tree per #850.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Critical fixes:
- Pause mechanism: workers now hold requests during pause instead of
dropping them. Queue and worker pool share the same AtomicBool flag
via VoiceQueue::paused_flag(). Submit() rejects while paused.
- Seed type: sr-voice accepts u64 seeds over IPC (explicit u32 truncation
for llama.cpp sampler, documented).
Warning fixes:
- HashMap → BTreeMap in cache.rs and worker.rs (D-010 determinism mandate).
Added Ord derives to CacheKey, ContentType, TellCategory.
- VoicePipe::generate() watchdog kills child after 120s timeout to prevent
indefinite blocking on read_line.
- VoiceCacheStore Drop impl calls save_all() on shutdown.
- trait-modifiers.ron: fixed 3 wrong trait names (Impulsive→Compassionate,
Methodical→Incurious, Stubborn→Ruthless) to match PersonalityTrait enum.
Suggestion fixes:
- Worker spawn: log error + reduce pool instead of panic on thread failure.
- on_battery(): added macOS detection via pmset.
- Epistemic markers: lowercased constants, removed redundant to_lowercase().
- cache.rs: documented non-atomic write tradeoff.
- queue.rs: reprioritize() bypasses pause check (it runs during pause).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Gemma 2 T&C compliance: exposed HTTP ports allow mods or external code
to reach the model, complicating license enforcement. Switch to piped
stdin/stdout (JSONL protocol) so the model is only reachable through
the game server's internal queue.
- worker.rs: VoicePipe owns Child + piped stdin/stdout, VoiceProcessConfig
replaces port-based config, workers spawn their own sr-voice child
- hardware.rs: remove VoiceInstanceManager (port/process lifecycle),
replace with evaluate_scaling() free function + HardwareProbe::voice_config()
- sr-voice: add --stdio flag to serve command, new stdio.rs JSONL mode
- Remove ureq dependency from server crate (no longer needed)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
jpmschweitzerandClaude Opus 4.6