source: the INT8 recipe decision is D818 (renumbered from D798)

source/build-camstack-models.py

@@ -45,7 +45,7 @@ Formats:
   int8      convert_lib.quantize_openvino_int8: NNCF post-training quantisation
             of the ONNX, calibrated on the first N images (sorted by path,
             recursively; an export folder's frames/ only) of --calib-dir
-            (default $CAMSTACK_CALIB_DIR or ./calibration-images). Recipe (D798):
+            (default $CAMSTACK_CALIB_DIR or ./calibration-images). Recipe (D818):
             preset MIXED, full bias correction, the Detect head in float —
             its own ops (DFL + box decode) for a YOLOv9 head, all of it for
             YOLO26's end-to-end head; every conv before it is INT8
@@ -62,18 +62,18 @@ equivalent but not bit-identical IR (NNCF picks ranges from the data).
 Which recipe each hosted INT8 IR carries. All were built from the HOSTED ONNX
 with OpenVINO 2026.2.1 + NNCF 3.2.0 on the first N COCO val2017 images
 (reproduce with `--reuse-onnx`, the mirror's .onnx placed under --out):
-  D798 recipe (this script, since 2026-10-09 — MIXED, full bias correction,
+  D818 recipe (this script, since 2026-10-09 — MIXED, full bias correction,
   the Detect head in float; docs/benchmarks/2026-10-09-int8-recipe.md):
     yolov9{t,s,m,c}-{320,640}, yolo26{n,s,l} @640, yolo26n-320,
     yolo26{n,s}-256
   NNCF defaults, centred calibration (convert_lib 2026-10-08 → 2026-10-09,
-  the mirror's source/archive/convert_lib-before-2026-10-09.py — kept: the D798
+  the mirror's source/archive/convert_lib-before-2026-10-09.py — kept: the D818
   build was more accurate but 11-22 % slower on the Intel iGPU):
     yolo26{m,x} @640, yolo26{m,l}-256
   NNCF defaults, top-left calibration (convert_lib before 2026-10-08,
   source/archive/convert_lib-before-2026-10-08.py — kept, same reason):
     yolo26{s,m,l}-320
-The D798 build reproduces every replaced IR's .bin bit for bit
+The D818 build reproduces every replaced IR's .bin bit for bit
 (docs/benchmarks/2026-10-09-int8-recipe.md).
 
 Toolchain: each recipe names the Ultralytics / torch / coremltools / OpenVINO

source/convert_lib.py

@@ -160,7 +160,7 @@ def int8_ignored_patterns(head: str, end_to_end: bool) -> "list[str]":
 def quantize_openvino_int8(
     onnx_path: str, out_xml: str, calib_dir: str, imgsz: int, max_samples: int = 300
 ) -> None:
-    """NNCF post-training INT8. For an Ultralytics detector (D798): preset MIXED
+    """NNCF post-training INT8. For an Ultralytics detector (D818): preset MIXED
     (asymmetric activations), full bias correction, and the Detect head in
     float — its own ops for a DFL head, all of it for an end-to-end one
     (`int8_ignored_patterns`). Any other model keeps NNCF's defaults. Measured in
@@ -188,7 +188,7 @@ def quantize_openvino_int8(
             "fast_bias_correction": False,
             "ignored_scope": nncf.IgnoredScope(patterns=int8_ignored_patterns(head, end_to_end), validate=False),
         }
-    # Not an Ultralytics detector: NNCF defaults, as before D798 (unmeasured otherwise).
+    # Not an Ultralytics detector: NNCF defaults, as before D818 (unmeasured otherwise).
     quantized = nncf.quantize(ov_model, dataset, subset_size=len(images), **options)
     Path(out_xml).parent.mkdir(parents=True, exist_ok=True)
     ov.save_model(quantized, out_xml)