Minor YAML updates

Author: tanner-andrulis

examples/arches/nvdla.yaml

@@ -33,11 +33,7 @@ arch:
   - !Memory
     name: Register
     size: weight.bits_per_value if weight else 1
-    area: 0
-    leak_power: 0
-    actions:
-    - {name: read, energy: 0, latency: 0}
-    - {name: write, energy: 0, latency: 0}
+    component_class: Dummy
     tensors: {keep: weight}
 
   - !Compute

examples/workloads/gpt3_175B_kv_cache.yaml

@@ -20,9 +20,11 @@ workload:
   einsums:
   - {einsum: "I[b, m, d] = I_in[b, m, d]", is_copy_operation: True}
 
-  # V and K containing the new tokens only. Assume N_TOKENS >> N_NEW_TOKENS, so we're
-  # just going to ignore these tensors and assume that the concatenation with the full K
-  # and V is cheap relative to the movement of K and V.
+  # V and K containing the new tokens only. Note that these two tensors aren't used
+  # later. We assume that N_TOKENS >> N_NEW_TOKENS, making the full K and V much larger
+  # than these. In real transformers, we'd concatenate these with the full K and V, but
+  # since K >> K_new and V >> V_new, we can ignore these tensors and assume that the
+  # concatenation is cheap relative to the movement of K and V.
   - "V_new[b, m, h, e] = I[b, m, d] * WV[h, e, d]"
   - "K_new[b, m, h, e] = I[b, m, d] * WK[h, e, d]"
 

examples/workloads/gpt3_6.7B.yaml

@@ -34,12 +34,15 @@ workload:
   einsums:
   - einsum: "I[b, m, d] = I_in[b, m, d]"
     is_copy_operation: True
+
   - "V[b, m, h, e] = I[b, m, d] * WV[h, e, d]"
   - "K[b, m, h, e] = I[b, m, d] * WK[h, e, d]"
   - "Q[b, m, h, e] = I[b, m, d] * WQ[h, e, d]"
+
   - einsum: "QK[b, m, p, h] = Q[b, m, h, e] * K[b, M: p, h, e]"
     renames: {input: Q}
   - "QK_softmax[b, m, p, h] = QK[b, m, p, h]"
+
   - einsum: "AV[b, m, h, f] = QK_softmax[b, m, p, h] * V[b, M: p, h, E: f]"
     renames: {input: QK_softmax}
   - "Z[b, m, g] = AV[b, m, h, f] * WZ[h, f, g]"

examples/workloads/gpt3_6.7B_kv_cache.yaml

@@ -20,9 +20,11 @@ workload:
   einsums:
   - {einsum: "I[b, m, d] = I_in[b, m, d]", is_copy_operation: True}
 
-  # V and K containing the new tokens only. Assume N_TOKENS >> N_NEW_TOKENS, so we're
-  # just going to ignore these tensors and assume that the concatenation with the full K
-  # and V is cheap relative to the movement of K and V.
+  # V and K containing the new tokens only. Note that these two tensors aren't used
+  # later. We assume that N_TOKENS >> N_NEW_TOKENS, making the full K and V much larger
+  # than these. In real transformers, we'd concatenate these with the full K and V, but
+  # since K >> K_new and V >> V_new, we can ignore these tensors and assume that the
+  # concatenation is cheap relative to the movement of K and V.
   - "V_new[b, m, h, e] = I[b, m, d] * WV[h, e, d]"
   - "K_new[b, m, h, e] = I[b, m, d] * WK[h, e, d]"