Skip to content

[ET-VK][ez] Don't copy zeros for cache tensors #15580

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Nov 5, 2025
Merged

[ET-VK][ez] Don't copy zeros for cache tensors #15580

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 12 additions & 4 deletions backends/vulkan/runtime/graph/ops/glsl/sdpa_kv_cache_update.glsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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)}
Expand Down Expand Up @@ -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);
}
4 changes: 2 additions & 2 deletions backends/vulkan/runtime/graph/ops/impl/SDPA.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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});
Expand Down
Loading