Update README.md

README.md

@@ -96,4 +96,15 @@ model = torch.compile(model, mode="max-autotune")
 inputs = tokenizer("Monarch matrices are efficiently sparse.", return_tensors="pt").to("cuda")
 with torch.no_grad():
     outputs = model(**inputs)
-    print(outputs.logits)
+    print(outputs.logits)
+```
+
+## 🧠 The "Memory Paradox" (Read this!)
+
+You might notice that while the parameter count is lower (82M vs 110M), the peak VRAM usage during inference can be slightly higher than the baseline.
+
+**Why?**
+This is a **software artifact**, not a hardware limitation.
+
+* **Math:** Monarch Matrices decompose a dense layer into: $\text{BlockDiag}_1 \to \text{Permutation} \to \text{BlockDiag}_2$.
+* **Solution:** A custom **Fused Triton Kernel** (planned) would fuse these steps, keeping intermediate activations in the GPU's SRAM (L1 Cache). This would drop dynamic VRAM usage significantly below the baseline, matching the FLOPs reduction.