[ET-VK][ez] Don't copy zeros for cache tensors (#15601)

Currently, cache tensors for SDPA are prepacked even though the mutable buffer data just contains zeros.

For fused SDPA, this step can be skipped.

Differential Revision: [D86226137](https://our.internmc.facebook.com/intern/diff/D86226137/)

Author: pytorchbot

backends/vulkan/runtime/graph/ops/glsl/sdpa_kv_cache_update.glsl

@@ -14,6 +14,10 @@ ${define_required_extensions(DTYPE)}
 
 layout(std430) buffer;
 
+#define DEBUG_MODE
+
+#extension GL_EXT_debug_printf : enable
+
 #include "common.glslh"
 
 ${layout_declare_tensor(B, "w", "t_cache", DTYPE, OUTPUT_STORAGE, is_scalar_array=False)}
@@ -80,13 +84,17 @@ void main() {
   const int S = projected_sizes.z;
   const int H = projected_sizes.y;
 
-  if (d4 >= D4 || s >= S || h >= H) {
+  const int c = s + input_pos; // idx along max_context_len dim
+  const int C = cache_sizes.z;
+
+  if (d4 >= D4 || c >= C || h >= H) {
     return;
   }
 
-  const int c = s + input_pos; // idx along max_context_len dim
-  const int C = cache_sizes.y;
+  IN_VEC4_T in_texel = IN_VEC4_T(0.0);
+  if (s < S) {
+    in_texel = read_projected_d4(d4, h, s, D4, H, S);
+  }
 
-  IN_VEC4_T in_texel = read_projected_d4(d4, h, s, D4, H, S);
   write_cache_d4(in_texel, d4, c, h, D4, C, H);
 }

backends/vulkan/runtime/graph/ops/impl/SDPA.cpp

@@ -575,9 +575,9 @@ void compute_attn_weight_with_kv_cache_impl(
 
   utils::StorageType cache_storage = graph.storage_type_of(q_projected);
   const ValueRef k_cache =
-      prepack_standard(graph, k_cache_data, cache_storage, utils::kWidthPacked);
+      graph.add_tensor_like(k_cache_data, cache_storage, utils::kWidthPacked);
   const ValueRef v_cache =
-      prepack_standard(graph, v_cache_data, cache_storage, utils::kWidthPacked);
+      graph.add_tensor_like(v_cache_data, cache_storage, utils::kWidthPacked);
 
   update_cache_impl(graph, {k_projected, k_cache, input_pos_symint, -1});
   update_cache_impl(graph, {v_projected, v_cache, input_pos_symint, -1});