[ET-VK][LlaMa] Split SDPA + KV cache operator into SDPA operator and KV cache update operator

## Context

#7413 and #7412 split the `sdpa_with_kv_cache` operator into two separate operators, `update_cache` and `custom_sdpa` to decouple the cache update step from the actual SDPA computation.

As a result, SDPA is no longer being delegated on Vulkan because of this interface change. To rectify this, Vulkan must also split `sdpa_with_kv_cache` into two operators.

Note that during this diff the new operators are not partitioned yet because of complications caused by assertion ops in the graph. The next diff adds a pass to remove such assertion ops which allows the new operators to be partitioned.

Differential Revision: [D68916952](https://our.internmc.facebook.com/intern/diff/D68916952/)

[ghstack-poisoned]

Author: SS-JIA

backends/vulkan/_passes/insert_prepack_nodes.py

@@ -60,6 +60,12 @@ def prepack_not_required(node: torch.fx.Node) -> bool:
             )
             # This pass assumes that the SpecPropPass() has already been applied
             assert "spec" in node.meta
+            # Mutable buffers will not be marked as constant, but it might as well be
+            # for the purposes of memory planning. Mark it as a constant tensor so that
+            # it is handled correctly by the memory planning pass.
+            if not node.meta["spec"].const:
+                assert is_param_node(program, node)
+                node.meta["spec"].const = True
             # Validate that the original node is marked as a constant. Constant tensors
             # do not participate in memory planning.
             assert node.meta["spec"].const
@@ -68,7 +74,9 @@ def prepack_not_required(node: torch.fx.Node) -> bool:
             # Set the mem_obj_id to -1 to indicate that this node requires a dedicated
             # memory object.
             prepack_node.meta["spec"].mem_obj_id = -1
-            node.replace_all_uses_with(prepack_node, lambda x, y=prepack_node: x != y)
+            node.replace_all_uses_with(
+                prepack_node, lambda x, y=prepack_node: (x != y and x.op != "output")
+            )
 
     program.graph.eliminate_dead_code()
     return program

backends/vulkan/_passes/tag_memory_meta_pass.py

@@ -220,9 +220,7 @@ def should_delay_annotation(self, node: torch.fx.Node) -> bool:
 
     # noqa
     def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
-        sorted_nodes: NodeList = topo_sort(list(graph_module.graph.nodes))
-
-        for node in sorted_nodes:
+        for node in graph_module.graph.nodes:
             if not self.should_annotate(node) or self.should_delay_annotation(node):
                 continue
 

backends/vulkan/op_registry.py

@@ -478,7 +478,7 @@ def register_convolution_op(features: OpFeatures):
 
 
 @update_features("llama::sdpa_with_kv_cache")
-def register_sdpa_op(features: OpFeatures):
+def register_sdpa_with_kv_cache_op(features: OpFeatures):
     features.texture_impl = TextureImplFeatures(
         valid_packed_dims={PackedDim.WIDTH},
     )
@@ -489,6 +489,20 @@ def register_sdpa_op(features: OpFeatures):
     return features
 
 
+# TODO(ssjia) allow registration after remove assertions pass is implemented
+# @update_features(["llama::update_cache", exir_ops.edge.llama.custom_sdpa.default])
+def register_sdpa_ops(features: OpFeatures):
+    features.texture_impl = TextureImplFeatures(
+        valid_packed_dims={PackedDim.WIDTH},
+    )
+    features.resize_fn = False
+    features.buffer_impl = False
+    features.texture_impl = TextureImplFeatures(
+        valid_packed_dims={PackedDim.WIDTH},
+    )
+    return features
+
+
 @update_features(exir_ops.edge.et_vk.apply_rotary_emb.default)
 def register_rotary_emb_op(features: OpFeatures):
     features.texture_impl = TextureImplFeatures(

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

@@ -176,17 +176,32 @@ void resize_sdpa_out(
   graph->get_tensor(out)->virtual_resize(graph->sizes_of(q_projected));
 }
 
-void sdpa_with_kv_cache_impl(
-    ComputeGraph& graph,
-    const std::vector<ValueRef>& args) {
+void update_cache_impl(ComputeGraph& graph, const std::vector<ValueRef>& args) {
+  int arg_idx = 0;
+  const ValueRef value = args[arg_idx++];
+  const ValueRef cache = args[arg_idx++];
+  const ValueRef input_pos_symint = args[arg_idx++];
+  const ValueRef out = args[arg_idx++];
+
+  // Unused variables
+  (void)out;
+
+  VK_CHECK_COND(graph.size_at<int32_t>(-4, value) == 1);
+  VK_CHECK_COND(graph.size_at<int32_t>(-4, cache) == 1);
+  VK_CHECK_COND(
+      graph.size_at<int32_t>(-1, value) == graph.size_at<int32_t>(-1, cache));
+  VK_CHECK_COND(
+      graph.size_at<int32_t>(-2, value) == graph.size_at<int32_t>(-2, cache));
+
+  add_kv_cache_update_node(graph, input_pos_symint, value, cache);
+}
+
+void sdpa_impl(ComputeGraph& graph, const std::vector<ValueRef>& args) {
   int arg_idx = 0;
   const ValueRef q_projected = args[arg_idx++];
-  const ValueRef k_projected = args[arg_idx++];
-  const ValueRef v_projected = args[arg_idx++];
-  const ValueRef k_cache_data = args[arg_idx++];
-  const ValueRef v_cache_data = args[arg_idx++];
+  const ValueRef k_cache = args[arg_idx++];
+  const ValueRef v_cache = args[arg_idx++];
   const ValueRef input_pos_symint = args[arg_idx++];
-  const ValueRef sequence_len = args[arg_idx++];
   const ValueRef attn_mask = args[arg_idx++];
   const ValueRef dropout_p = args[arg_idx++];
   const ValueRef is_causal = args[arg_idx++];
@@ -195,23 +210,20 @@ void sdpa_with_kv_cache_impl(
   // Output tensors
   const ValueRef out = args[arg_idx++];
 
-  // Unused variables
-  (void)sequence_len;
-
   // Batches must be 1
   VK_CHECK_COND(graph.size_at<int32_t>(-4, q_projected) == 1);
-  VK_CHECK_COND(graph.size_at<int32_t>(-4, k_projected) == 1);
-  VK_CHECK_COND(graph.size_at<int32_t>(-4, v_projected) == 1);
+  VK_CHECK_COND(graph.size_at<int32_t>(-4, k_cache) == 1);
+  VK_CHECK_COND(graph.size_at<int32_t>(-4, v_cache) == 1);
   // k and v projected must have the same shape
-  VK_CHECK_COND(graph.sizes_of(k_projected) == graph.sizes_of(v_projected));
+  VK_CHECK_COND(graph.sizes_of(k_cache) == graph.sizes_of(v_cache));
   // head dim must match between tensors
   VK_CHECK_COND(
       graph.size_at<int32_t>(-1, q_projected) ==
-      graph.size_at<int32_t>(-1, k_projected));
+      graph.size_at<int32_t>(-1, k_cache));
   // All tensors must have the packed dim be the width (head) dimension
   VK_CHECK_COND(graph.packed_dim_of(q_projected) == WHCN::kWidthDim);
-  VK_CHECK_COND(graph.packed_dim_of(k_projected) == WHCN::kWidthDim);
-  VK_CHECK_COND(graph.packed_dim_of(v_projected) == WHCN::kWidthDim);
+  VK_CHECK_COND(graph.packed_dim_of(k_cache) == WHCN::kWidthDim);
+  VK_CHECK_COND(graph.packed_dim_of(v_cache) == WHCN::kWidthDim);
   // Some variables are not supported yet
   VK_CHECK_COND(
       graph.val_is_none(dropout_p) ||
@@ -222,16 +234,8 @@ void sdpa_with_kv_cache_impl(
       graph.val_is_none(is_causal) || graph.extract_scalar<bool>(is_causal));
   VK_CHECK_COND(graph.val_is_none(attn_mask));
 
-  const ValueRef k_cache =
-      prepack_standard_like(graph, k_cache_data, q_projected);
-  const ValueRef v_cache =
-      prepack_standard_like(graph, v_cache_data, q_projected);
-
   const int32_t max_seq_len = graph.size_at<int32_t>(1, k_cache);
 
-  add_kv_cache_update_node(graph, input_pos_symint, k_projected, k_cache);
-  add_kv_cache_update_node(graph, input_pos_symint, v_projected, v_cache);
-
   // Slice caches from 0 to input_pos + sequence_len
   const ValueRef k_cache_sliced = graph.add_tensor_view(k_cache);
   const ValueRef v_cache_sliced = graph.add_tensor_view(v_cache);
@@ -257,7 +261,7 @@ void sdpa_with_kv_cache_impl(
 
   // Repeat interleave
   const int64_t num_heads = graph.size_at<int64_t>(2, q_projected);
-  const int64_t num_kv_heads = graph.size_at<int64_t>(2, k_projected);
+  const int64_t num_kv_heads = graph.size_at<int64_t>(2, k_cache);
 
   const ValueRef num_repeats =
       graph.add_scalar<int64_t>(num_heads / num_kv_heads);
@@ -331,8 +335,52 @@ void sdpa_with_kv_cache_impl(
       new ExecuteNode(resize_sdpa_out, {q_projected, out}));
 }
 
+void sdpa_with_kv_cache_impl(
+    ComputeGraph& graph,
+    const std::vector<ValueRef>& args) {
+  int arg_idx = 0;
+  const ValueRef q_projected = args[arg_idx++];
+  const ValueRef k_projected = args[arg_idx++];
+  const ValueRef v_projected = args[arg_idx++];
+  const ValueRef k_cache_data = args[arg_idx++];
+  const ValueRef v_cache_data = args[arg_idx++];
+  const ValueRef input_pos_symint = args[arg_idx++];
+  const ValueRef sequence_len = args[arg_idx++];
+  const ValueRef attn_mask = args[arg_idx++];
+  const ValueRef dropout_p = args[arg_idx++];
+  const ValueRef is_causal = args[arg_idx++];
+  const ValueRef scale = args[arg_idx++];
+
+  // Output tensors
+  const ValueRef out = args[arg_idx++];
+
+  (void)sequence_len;
+
+  const ValueRef k_cache =
+      prepack_standard_like(graph, k_cache_data, q_projected);
+  const ValueRef v_cache =
+      prepack_standard_like(graph, v_cache_data, q_projected);
+
+  update_cache_impl(graph, {k_projected, k_cache, input_pos_symint, -1});
+  update_cache_impl(graph, {v_projected, v_cache, input_pos_symint, -1});
+
+  sdpa_impl(
+      graph,
+      {q_projected,
+       k_cache,
+       v_cache,
+       input_pos_symint,
+       attn_mask,
+       dropout_p,
+       is_causal,
+       scale,
+       out});
+}
+
 REGISTER_OPERATORS {
   VK_REGISTER_OP(sdpa_with_kv_cache.default, sdpa_with_kv_cache_impl);
+  VK_REGISTER_OP(update_cache.default, update_cache_impl);
+  VK_REGISTER_OP(llama.custom_sdpa.default, sdpa_impl);
 }
 
 } // namespace vkcompute