fix: batch quest analysis on zerogpu

Sync GitHub commit 45ef859; split MiniCPM quest refresh into smaller ZeroGPU batches.

README.md

@@ -219,6 +219,7 @@ ADVISOR_ADAPTER_ID=build-small-hackathon/hackathon-advisor-minicpm5-lora
 ADVISOR_ADAPTER_REVISION=25de69bcde397e1bcdd852923b56a42f10222650
 ADVISOR_QUEST_ANALYZER_BACKEND=minicpm-transformers
 ADVISOR_QUEST_ADAPTER_ID=artifacts/quest-lora
+ADVISOR_QUEST_ANALYSIS_BATCH_SIZE=24
 ADVISOR_CACHE_DIR=/data/advisor-cache
 ADVISOR_REFRESH_EMBEDDING_TIMEOUT_SECONDS=1800
 ADVISOR_EMBEDDING_MODEL_REPO=ggml-org/embeddinggemma-300m-qat-q8_0-GGUF

app.py

@@ -63,6 +63,7 @@ MAX_AUDIO_UPLOAD_BYTES = 25 * 1024 * 1024
 AUDIO_UPLOAD_SUFFIXES = {".aac", ".aif", ".aiff", ".flac", ".m4a", ".mp3", ".oga", ".ogg", ".opus", ".wav", ".webm"}
 DEFAULT_HF_ORG = "build-small-hackathon"
 DEFAULT_REFRESH_EMBEDDING_TIMEOUT_SECONDS = 1800
+DEFAULT_QUEST_ANALYSIS_BATCH_SIZE = 24
 REFRESH_SUBPROCESS_LOG_TAIL_LINES = 80
 REFRESH_STAGE_LABELS = {
     "crawling": "Fetching public Spaces",
@@ -132,7 +133,6 @@ def _transcribe_voice(audio_path: str) -> dict[str, Any]:
     return voice_transcriber.transcribe(Path(audio_path)).to_dict()
 
 
-@gpu_task
 def _analyze_dashboard_quests(project_rows: list[dict[str, Any]]) -> dict[str, Any]:
     missing_evidence_keys = [
         str(item.get("id") or index)
@@ -145,6 +145,24 @@ def _analyze_dashboard_quests(project_rows: list[dict[str, Any]]) -> dict[str, A
             f"missing evidence keys for {len(missing_evidence_keys)} projects"
         )
     projects = [Project.from_dict(item) for item in project_rows]
+    matches_by_project: dict[str, list[dict[str, Any]]] = {}
+    source = "quest-analyzer"
+    batch_size = _quest_analysis_batch_size()
+    for start in range(0, len(project_rows), batch_size):
+        batch_rows = project_rows[start : start + batch_size]
+        result = _analyze_dashboard_quest_batch(batch_rows)
+        source = str(result["source"])
+        matches_by_project.update(result["matches_by_project"])
+    validated = validate_matches_by_project(matches_by_project, projects, source=source)
+    return {
+        "source": validated.source,
+        "matches_by_project": validated.matches_by_project,
+    }
+
+
+@gpu_task
+def _analyze_dashboard_quest_batch(project_rows: list[dict[str, Any]]) -> dict[str, Any]:
+    projects = [Project.from_dict(item) for item in project_rows]
     analyzer = create_quest_analyzer(device="cuda" if zero_gpu_enabled() else "local")
     matches = analyzer.analyze(projects)
     source = getattr(analyzer, "source", "quest-analyzer")
@@ -155,6 +173,16 @@ def _analyze_dashboard_quests(project_rows: list[dict[str, Any]]) -> dict[str, A
     }
 
 
+def _quest_analysis_batch_size() -> int:
+    raw = os.environ.get("ADVISOR_QUEST_ANALYSIS_BATCH_SIZE", "").strip()
+    if not raw:
+        return DEFAULT_QUEST_ANALYSIS_BATCH_SIZE
+    batch_size = int(raw)
+    if batch_size <= 0:
+        raise RuntimeError("ADVISOR_QUEST_ANALYSIS_BATCH_SIZE must be a positive integer.")
+    return batch_size
+
+
 def _refresh_public_state() -> dict[str, Any]:
     with _refresh_lock:
         return dict(_refresh_state)

tests/test_app.py

@@ -290,6 +290,48 @@ def test_dashboard_refresh_quest_analysis_uses_minicpm_analyzer(monkeypatch) ->
     assert quests == {"Off the Grid", "Field Notes"}
 
 
+def test_dashboard_refresh_quest_analysis_batches_minicpm(monkeypatch) -> None:
+    projects = [
+        Project(
+            id=f"build-small-hackathon/batched-{index}",
+            title=f"Batched {index}",
+            summary="Small local demo",
+            tags=("gradio",),
+            models=(),
+            datasets=(),
+            likes=0,
+            sdk="gradio",
+            license="mit",
+            created_at="2026-06-01T00:00:00+00:00",
+            last_modified="2026-06-08T00:00:00+00:00",
+            host=f"https://batched-{index}.hf.space",
+            url=f"https://huggingface.co/spaces/build-small-hackathon/batched-{index}",
+            readme_body="README evidence",
+            app_file_source="import gradio as gr",
+        )
+        for index in range(3)
+    ]
+    calls = []
+
+    class FakeMiniCPMAnalyzer:
+        source = "minicpm-json-quest-analyzer"
+
+        def analyze(self, batch):
+            calls.append([project.id for project in batch])
+            return {project.id: [] for project in batch}
+
+    monkeypatch.setenv("ADVISOR_QUEST_ANALYSIS_BATCH_SIZE", "2")
+    monkeypatch.setattr(app_module, "create_quest_analyzer", lambda device: FakeMiniCPMAnalyzer())
+
+    result = app_module._analyze_dashboard_quests([project.to_refresh_snapshot_dict() for project in projects])
+
+    assert calls == [
+        ["build-small-hackathon/batched-0", "build-small-hackathon/batched-1"],
+        ["build-small-hackathon/batched-2"],
+    ]
+    assert set(result["matches_by_project"]) == {project.id for project in projects}
+
+
 def test_dashboard_refresh_quest_analysis_requires_two_segment_snapshot() -> None:
     project = Project(
         id="build-small-hackathon/missing-evidence",