Add protocol performance profiles

src/mka/application/meeting_service.py

@@ -16,6 +16,7 @@ import yaml
 
 from mka.application.config import AppSettings
 from mka.application.glossary import GlossaryRepository, render_glossary_terms
+from mka.application.performance import DEFAULT_PERFORMANCE_PROFILE, resolve_ollama_num_thread
 
 STAGES = ("preparing", "transcription", "diarization", "protocol_generation")
 
@@ -65,6 +66,7 @@ class ParticipantInput:
 class ProcessingOptions:
     diarization_enabled: bool = False
     audio_normalization: bool = True
+    performance_profile: str = DEFAULT_PERFORMANCE_PROFILE
 
 
 @dataclass(frozen=True)
@@ -229,6 +231,7 @@ class MeetingProcessingService:
                 "model": self.settings.protocol_model,
                 "ollama_endpoint": self.settings.ollama_endpoint,
                 "protocol_num_ctx": self.settings.protocol_num_ctx,
+                "protocol_num_thread": resolve_ollama_num_thread(options.performance_profile),
                 "protocol_safe_input_token_budget": (
                     self.settings.protocol_safe_input_token_budget
                 ),
@@ -357,6 +360,7 @@ class MeetingProcessingService:
         run_dir: Path,
         speaker_mappings: dict[str, str],
         progress_sink: Callable[[AppProgressEvent], None] | None = None,
+        performance_profile: str = DEFAULT_PERFORMANCE_PROFILE,
     ) -> ProcessingOutcome:
         """Persist confirmed mappings and regenerate only the direct protocol."""
         review = self.load_speaker_mapping_review(run_dir)
@@ -401,6 +405,7 @@ class MeetingProcessingService:
             model=self.settings.protocol_model,
             ollama_endpoint=self.settings.ollama_endpoint,
             protocol_num_ctx=self.settings.protocol_num_ctx,
+            protocol_num_thread=resolve_ollama_num_thread(performance_profile),
             protocol_safe_input_token_budget=(self.settings.protocol_safe_input_token_budget),
         )
         protocol_path = Path(result.protocol_path)

src/mka/application/performance.py

@@ -0,0 +1,23 @@
+"""Abstract performance profiles resolved to current backend runtime options."""
+
+from __future__ import annotations
+
+DEFAULT_PERFORMANCE_PROFILE = "auto"
+PERFORMANCE_PROFILES = ("auto", "fast", "efficient", "powersave")
+
+_OLLAMA_THREADS_BY_PROFILE = {
+    "fast": 16,
+    "efficient": 10,
+    "powersave": 4,
+}
+
+
+def resolve_ollama_num_thread(profile: str) -> int | None:
+    """Resolve a profile, leaving automatic thread selection to Ollama for Auto."""
+    if profile == "auto":
+        return None
+    try:
+        return _OLLAMA_THREADS_BY_PROFILE[profile]
+    except KeyError as exc:
+        allowed = ", ".join(PERFORMANCE_PROFILES)
+        raise ValueError(f"Unknown performance profile {profile!r}; expected: {allowed}.") from exc

src/mka/ui/streamlit_app.py

@@ -34,6 +34,7 @@ from mka.application.people_yaml import (
     export_people_yaml,
     import_people_yaml,
 )
+from mka.application.performance import DEFAULT_PERFORMANCE_PROFILE, PERFORMANCE_PROFILES
 from mka.application.run_inputs import (
     RunInputJsonError,
     RunInputState,
@@ -495,7 +496,13 @@ def _render_result() -> None:
     ):
         try:
             with st.spinner("Regenerating protocol without rerunning audio processing..."):
-                regenerated = service.regenerate_protocol(outcome.run_dir, selections)
+                regenerated = service.regenerate_protocol(
+                    outcome.run_dir,
+                    selections,
+                    performance_profile=st.session_state.get(
+                        "performance_profile", DEFAULT_PERFORMANCE_PROFILE
+                    ),
+                )
         except (OSError, RuntimeError, ValueError) as exc:
             st.error(f"Protocol regeneration failed: {exc}")
         else:
@@ -551,6 +558,13 @@ def main() -> None:
     participants = _render_participants()
 
     st.header("Processing options")
+    performance_profile = st.selectbox(
+        "Performance profile",
+        options=PERFORMANCE_PROFILES,
+        format_func=str.title,
+        key="performance_profile",
+        help="Controls protocol-generation performance using an abstract runtime profile.",
+    )
     audio_normalization = st.checkbox(
         "Audio normalization",
         value=True,
@@ -619,6 +633,7 @@ def main() -> None:
                     ProcessingOptions(
                         diarization_enabled=diarization_enabled,
                         audio_normalization=audio_normalization,
+                        performance_profile=performance_profile,
                     ),
                     progress_sink=callback,
                 )