fix(perf): apply MVP threading hotfix

Configure numerical library and torch threading for H1, keep raster threading/timing instrumentation, remove CONCERTATION from forced masks after real PDF FP testing, and record coordination archive state.
2026-06-08 10:41:15 +02:00
parent 3249f3a337
commit 21a408a9e4
68 changed files with 2075 additions and 20 deletions
--- a/anonymizer_core_refactored_onnx.py
+++ b/anonymizer_core_refactored_onnx.py
@@ -19,9 +19,20 @@ import os
 import re
 import shutil
 import sys
-from concurrent.futures import ProcessPoolExecutor
+import time
+from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from datetime import datetime

+# --- H1 perf (D-19) : usage multi-cœur des libs numériques en EXE frozen ---
+# En PyInstaller frozen, OpenMP/MKL/BLAS tombent souvent à 1 thread (CPU ~12 %).
+# Ces variables sont lues par numpy/torch/onnxruntime à leur init : elles doivent
+# donc être posées AVANT l'import de pdfplumber/PIL (numpy transitif) ci-dessous.
+# setdefault : on n'écrase jamais un réglage explicite posé par l'utilisateur/admin.
+_n_cpu_threads = str(os.cpu_count() or 4)
+for _env in ("OMP_NUM_THREADS", "MKL_NUM_THREADS", "OPENBLAS_NUM_THREADS",
+             "NUMEXPR_NUM_THREADS", "VECLIB_MAXIMUM_THREADS"):
+    os.environ.setdefault(_env, _n_cpu_threads)
+
 log = logging.getLogger(__name__)
 from dataclasses import dataclass, field
 from pathlib import Path
@@ -64,6 +75,46 @@ except Exception:
    _doctr_ocr_predictor = None  # type: ignore
    _DOCTR_AVAILABLE = False

+_doctr_model_cache = None
+_TORCH_THREADS_CONFIGURED = False
+
+def _configure_torch_threads():
+    """Configure les threads PyTorch pour exploiter tous les cœurs en mode frozen.
+
+    En EXE PyInstaller, torch ne configure pas ses threads par défaut et reste
+    à 1 thread intra-op + 1 inter-op, ce qui limite l'OCR docTR et le NER
+    à ~12 % CPU sur une machine 8 threads.
+
+    Idempotent : appelable depuis l'OCR (doc scanné) comme depuis le NER (doc
+    natif sans OCR). `set_num_interop_threads` ne peut être posé qu'une seule
+    fois avant tout travail parallèle ; le flag évite un 2e appel qui lèverait.
+    """
+    global _TORCH_THREADS_CONFIGURED
+    if _TORCH_THREADS_CONFIGURED:
+        return
+    try:
+        import torch
+        n_cpus = os.cpu_count() or 4
+        torch.set_num_threads(n_cpus)
+        try:
+            torch.set_num_interop_threads(min(n_cpus, 8))
+        except Exception:
+            pass  # inter-op déjà figé par un travail torch antérieur — non bloquant
+        _TORCH_THREADS_CONFIGURED = True
+        log.info("torch threads config: intra=%d inter=%d (CPUs=%d)",
+                 n_cpus, min(n_cpus, 8), os.cpu_count() or 0)
+    except Exception as e:
+        log.debug("torch threads config skipped: %s", e)
+
+def _get_doctr_model():
+    global _doctr_model_cache
+    if _doctr_model_cache is None:
+        _configure_torch_threads()
+        _doctr_model_cache = _doctr_ocr_predictor(
+            det_arch="db_resnet50", reco_arch="crnn_vgg16_bn", pretrained=True
+        )
+    return _doctr_model_cache
+
 try:
    from detectors.hospital_filter import HospitalFilter
    _HOSPITAL_FILTER_AVAILABLE = True
@@ -1093,16 +1144,6 @@ def _apply_admin_identifier_hits(full_raw: str, audit: List["PiiHit"], cfg: Dict

 # ----------------- Extraction -----------------

-_doctr_model_cache = None
-
-def _get_doctr_model():
-    global _doctr_model_cache
-    if _doctr_model_cache is None:
-        _doctr_model_cache = _doctr_ocr_predictor(
-            det_arch="db_resnet50", reco_arch="crnn_vgg16_bn", pretrained=True
-        )
-    return _doctr_model_cache
-
 def _extract_page_layout_aware(page) -> str:
    """Extrait le texte d'une page PyMuPDF en gérant les layouts multi-colonnes.

@@ -1763,7 +1804,7 @@ def _kv_value_only_mask(line: str, audit: List[PiiHit], page_idx: int, cfg: Dict
    parts = SPLITTER.split(line, maxsplit=1)
    # Une ligne narrative qui se termine par ` ;` ou ` :` produit un split
    # avec une "value" vide. La "key" contient alors tout le narratif —
-    # incluant d'éventuels force_term (`CHUXX`, `CONCERTATION`…) qui doivent
+    # incluant d'éventuels force_term (`CHUXX`, sigle local...) qui doivent
    # être masqués. Idem si la "key" fait plus de 5 mots : c'est très
    # probablement du narratif, pas un libellé `Label : valeur`.
    if len(parts) == 2 and parts[1].strip() and len(parts[0].split()) <= 5:
@@ -2977,6 +3018,11 @@ def _run_ner_on_original_text(
    Returns:
        Liste de NerDetection dédupliquée (par token+label+page+source).
    """
+    # H1 perf (D-19) : couvre le cas du PDF natif (texte riche, OCR sauté) où
+    # _get_doctr_model() n'est jamais appelé ; les NER torch (EDS-Pseudo, GLiNER)
+    # tourneraient alors mono-thread. Idempotent (no-op si déjà configuré par l'OCR).
+    _configure_torch_threads()
+
    detections: List[NerDetection] = []
    seen: set = set()  # (token_lower, label, page_idx, source) pour dédoublonnage

@@ -4274,7 +4320,9 @@ def redact_pdf_raster(original_pdf: Path, audit: List[PiiHit], out_pdf: Path, dp
            rects.extend(found)
        all_rects[pno] = rects

-    # Phase 2 : rasterisation parallèle (ProcessPoolExecutor)
+    # Phase 2 : rasterisation parallèle (ProcessPoolExecutor hors EXE,
+    # ThreadPoolExecutor en EXE PyInstaller pour éviter de relancer la GUI).
+    raster_t0 = time.perf_counter()
    n_pages = len(doc)
    rects_as_tuples = {
        pno: [(r.x0, r.y0, r.x1, r.y1) for r in rects]
@@ -4313,12 +4361,26 @@ def redact_pdf_raster(original_pdf: Path, audit: List[PiiHit], out_pdf: Path, dp
        for pno in range(n_pages)
    ]

-    # Mode frozen (PyInstaller --onefile) : ProcessPoolExecutor relance l'exe
-    # et ouvre des fenêtres GUI fantômes → séquentiel obligatoire
-    if getattr(sys, 'frozen', False) or n_pages <= 2:
+    frozen = bool(getattr(sys, 'frozen', False))
+    disable_threads = os.getenv("ANON_DISABLE_RASTER_THREADS", "").lower() in {"1", "true", "yes", "on"}
+    if n_pages <= 2:
+        log.info("Raster PDF: mode=sequential pages=%d dpi=%d reason=small_pdf", n_pages, dpi)
+        results = sorted([_rasterize_page(t) for t in tasks], key=lambda x: x[0])
+    elif frozen and not disable_threads:
+        n_workers = min(n_pages, os.cpu_count() or 4)
+        log.info("Raster PDF: mode=threads pages=%d workers=%d dpi=%d frozen=1", n_pages, n_workers, dpi)
+        try:
+            with ThreadPoolExecutor(max_workers=n_workers) as pool:
+                results = sorted(pool.map(_rasterize_page, tasks), key=lambda x: x[0])
+        except Exception as e:
+            log.warning("Raster PDF threaded mode failed, fallback sequential: %s", e)
+            results = sorted([_rasterize_page(t) for t in tasks], key=lambda x: x[0])
+    elif frozen and disable_threads:
+        log.info("Raster PDF: mode=sequential pages=%d dpi=%d frozen=1 reason=env_disabled", n_pages, dpi)
        results = sorted([_rasterize_page(t) for t in tasks], key=lambda x: x[0])
    else:
        n_workers = min(n_pages, os.cpu_count() or 4)
+        log.info("Raster PDF: mode=processes pages=%d workers=%d dpi=%d frozen=0", n_pages, n_workers, dpi)
        with ProcessPoolExecutor(max_workers=n_workers) as pool:
            results = sorted(pool.map(_rasterize_page, tasks), key=lambda x: x[0])

@@ -4331,6 +4393,7 @@ def redact_pdf_raster(original_pdf: Path, audit: List[PiiHit], out_pdf: Path, dp
    _apply_pseudo_xmp_metadata(out)
    out.save(str(out_pdf), deflate=True, garbage=4, clean=True)
    out.close()
+    log.info("Raster PDF done: pages=%d output=%s duration=%.2fs", n_pages, out_pdf.name, time.perf_counter() - raster_t0)

 # ----------------- VLM pour PDFs scannés -----------------

@@ -4424,15 +4487,31 @@ def process_pdf(
    camembert_manager=None,
    quarantine_mgr: Optional["QuarantineManager"] = None,
 ) -> Dict[str, str]:
+    perf_t0 = time.perf_counter()
+    last_mark = perf_t0
+
+    def _perf_mark(stage: str) -> None:
+        nonlocal last_mark
+        now = time.perf_counter()
+        log.info("PERF %s: stage=%s duration=%.2fs total=%.2fs",
+                 pdf_path.name, stage, now - last_mark, now - perf_t0)
+        last_mark = now
+
+    log.info("PERF %s: start frozen=%s vector=%s raster=%s",
+             pdf_path.name, bool(getattr(sys, "frozen", False)), make_vector_redaction, also_make_raster_burn)
    out_dir.mkdir(parents=True, exist_ok=True)
    cfg = load_dictionaries(config_path)
+    _perf_mark("load_config")
    pages_text, tables_lines, ocr_used, ocr_word_map = extract_text_with_fallback_ocr(pdf_path)
+    _perf_mark("extract_text_ocr")

    # Q-1 B-3 : pré-flight texte vide. Si moins de SEUIL_TEXTE_MINI caractères
    # extraits, c'est probablement un scan non-OCRisé ou un document corrompu.
    # On NE traite PAS — quarantaine totale, le doc original est copié pour
    # ré-essai manuel.
    extracted_chars = sum(len(p) for p in pages_text)
+    log.info("PERF %s: pages=%d extracted_chars=%d ocr_used=%s ocr_pages=%d",
+             pdf_path.name, len(pages_text), extracted_chars, bool(ocr_used), len(ocr_word_map or {}))
    if extracted_chars < SEUIL_TEXTE_MINI:
        log.warning("Preflight failed for %s: only %d chars extracted (seuil=%d)",
                    pdf_path.name, extracted_chars, SEUIL_TEXTE_MINI)
@@ -4479,12 +4558,14 @@ def process_pdf(
        gliner_mgr=gliner_manager,
        camembert_mgr=camembert_manager,
    )
+    _perf_mark("regex_rules")

    # 1b) VLM (optionnel) — sur les PDFs scannés uniquement
    if ocr_used and vlm_manager is not None and VlmManager is not None:
        try:
            if vlm_manager.is_loaded():
                _apply_vlm_on_scanned_pdf(pdf_path, anon, ocr_word_map, vlm_manager)
+                _perf_mark("vlm_scan")
        except Exception:
            pass  # dégradation gracieuse

@@ -4498,9 +4579,11 @@ def process_pdf(
        else:
            final_text, hf_hits = apply_hf_ner_on_narrative(final_text, cfg, ner_manager, ner_thresholds)
        anon.audit.extend(hf_hits)
+    _perf_mark("ner_optional")

    # 3) Rescan selectif
    final_text = selective_rescan(final_text, cfg=cfg)
+    _perf_mark("selective_rescan")

    # 3a-bis) Nettoyage post-masquage : continuation orpheline d'un nom composé
    #     coupé par saut de ligne. Cas Trackare en colonnes : "NOCENT-EJNAINI"
@@ -4820,6 +4903,7 @@ def process_pdf(
        r"DOSSIER|NDA|EPISODE|RPPS|DATE_NAISSANCE|AGE|NIR|IBAN|OGC)\])\]+"
    )
    final_text = _RE_BRACKET_CLEAN.sub(r"\1", final_text)
+    _perf_mark("post_cleaning")

    # 6) Whitelist absolue : filtrer les hits qui matchent un terme whitelist
    # de la GUI (clé YAML whitelist_phrases). Filet de sécurité après tous les
@@ -4960,6 +5044,7 @@ def process_pdf(
        for hit in audit_for_file:
            f.write(json.dumps(hit.__dict__, ensure_ascii=False) + "\n")
    outputs = {"text": str(txt_path), "audit": str(audit_path)}
+    _perf_mark("write_text_audit")

    # PDFs
    if make_vector_redaction and fitz is not None:
@@ -4967,6 +5052,7 @@ def process_pdf(
        try:
            redact_pdf_vector(pdf_path, anon.audit, vec_path, ocr_word_map=ocr_word_map)
            outputs["pdf_vector"] = str(vec_path)
+            _perf_mark("pdf_vector")
        except Exception as e:
            # Q-1 D2/D3 : ne plus avaler silencieusement. Le texte (.pseudonymise.txt)
            # est déjà sorti avant ce bloc.
@@ -5023,6 +5109,8 @@ def process_pdf(
            ras_path = out_dir / f"{base}.redacted_raster.pdf"
            redact_pdf_raster(pdf_path, anon.audit, ras_path, ogc_label=ogc_label, ocr_word_map=ocr_word_map)
            outputs["pdf_raster"] = str(ras_path)
+            _perf_mark("pdf_raster")
+    log.info("PERF %s: done total=%.2fs outputs=%s", pdf_path.name, time.perf_counter() - perf_t0, sorted(outputs.keys()))
    return outputs