Upload Quasar_axrvi_ranker.py

Quasar_axrvi_ranker.py

@@ -3870,6 +3870,7 @@ class QCSAMCrossAssetLayer(nn.Module):
         # Diagnostics
         self.last_align_loss: float = 0.0
         self.last_qmha_diagnostics: dict = {}
+        self._fwd_count: int = 0          # counts forward calls for periodic logging
 
     def forward(
         self,
@@ -3946,8 +3947,27 @@ class QCSAMCrossAssetLayer(nn.Module):
         h_out = h + gate * delta                             # (B, N, d_model) float32
 
         # ── Mark first forward complete ───────────────────────────────────
+        self._fwd_count += 1
         QCSAMCrossAssetLayer._first_forward_complete = True
 
+        # ── QCSAM diagnostic logging ──────────────────────────────────────
+        # Log on the very first forward pass, then every 100 calls.
+        # (10 MC-samples × ~5s rank cycle → every ~50s in steady state.)
+        if self._fwd_count == 1 or self._fwd_count % 100 == 0:
+            gate_val = torch.sigmoid(self.residual_gate).item()
+            heads_info = " | ".join(
+                f"h{i}: attn_norm={hd.get('attn_matrix_norm', 0.0):.4f} "
+                f"align_h={hd.get('align_loss_h', 0.0):.6f}"
+                for i, hd in enumerate(self.last_qmha_diagnostics.get("heads", []))
+            )
+            logger.info(
+                f"🔬 [QCSAM] fwd#{self._fwd_count} | "
+                f"B={B} N={N} hilbert_dim={self.hilbert_dim} | "
+                f"gate={gate_val:.4f} | "
+                f"align_loss={self.last_align_loss:.6f} | "
+                f"{heads_info}"
+            )
+
         return h_out, align_loss
 
 
@@ -4244,6 +4264,20 @@ class AXRVINet(nn.Module):
         # Average align loss over MC samples
         mean_align = sum(mc_align) / len(mc_align) if mc_align else torch.tensor(0.0)
 
+        # ── QCSAM inference summary (every 20 mc_forward calls) ───────────
+        if not hasattr(self, "_mc_fwd_count"):
+            self._mc_fwd_count = 0
+        self._mc_fwd_count += 1
+        if self._mc_fwd_count % 20 == 1:
+            align_val = mean_align.item() if hasattr(mean_align, "item") else float(mean_align)
+            gate_val  = torch.sigmoid(self.qcsam_layer.residual_gate).item()
+            logger.info(
+                f"🔬 [QCSAM/Inference] mc_fwd#{self._mc_fwd_count} | "
+                f"samples={mc_samples} | "
+                f"mean_align_loss={align_val:.6f} | "
+                f"residual_gate={gate_val:.4f}"
+            )
+
         return {
             "significance_weight": mean_sig,
             "epistemic_variance":  epistemic_var,