Skip to content

ggml-webgpu: tune subgroup split (d_split) in flash_attn_vec #25418

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
yomaytk merged 1 commit into from
Jul 8, 2026
+16 −23
Merged

ggml-webgpu: tune subgroup split (d_split) in flash_attn_vec #25418

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
25 changes: 9 additions & 16 deletions ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2821,23 +2821,16 @@ class ggml_webgpu_shader_lib {
variant.resize(variant.size() - (sizeof("_mask") - 1));
variant += "_mask_blk";
}
uint32_t vec_ne = 1u;
if (key.common.k_type == GGML_TYPE_F16 && key.common.v_type == GGML_TYPE_F16 &&
key.common.head_dim_qk == key.common.head_dim_v) {
switch (key.common.head_dim_qk) {
case 64:
case 192:
case 576:
vec_ne = 2u;
break;
case 96:
vec_ne = 4u;
break;
default:
break;
}

uint32_t d_split = context.min_subgroup_size;
if (key.common.k_type == GGML_TYPE_F16 && key.common.v_type == GGML_TYPE_F16) {
const uint32_t D = key.common.head_dim_qk | key.common.head_dim_v;
const uint32_t D_lsb = D & (~(D - 1u));
d_split = std::min(std::min(context.min_subgroup_size, 4u), std::max(D_lsb / 4u, 1u));

@yomaytk yomaytk Jul 7, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Vulkan sets this upper limit of D_SPLIT to 8u, but I can confirm that 4u works better on both M2 and V100.

}
defines.push_back(std::string("VEC_NE=") + std::to_string(vec_ne) + "u");

defines.push_back(std::string("D_SPLIT=") + std::to_string(d_split));
variant += "_dsplit" + std::to_string(d_split);

auto pipeline_decisions = std::make_shared<ggml_webgpu_flash_attn_vec_decisions>(decisions);
webgpu_pipeline pipeline =
Expand Down
14 changes: 7 additions & 7 deletions ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,9 +39,6 @@ enable subgroups;
#define KV_GRANULARITY 8
#define KV_TILE 16
#define WG_SIZE 64
#ifndef VEC_NE
#define VEC_NE 4u
#endif

#define KV_BLOCKS (KV_TILE / KV_GRANULARITY)

Expand Down Expand Up @@ -367,11 +364,11 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,

// accumulate q block * k block into registers across the entire KV tile
if (!skip_tile) {
let num_of_threads = subgroup_size / VEC_NE;
let num_of_threads:u32 = D_SPLIT;
let tx = sg_inv_id % num_of_threads;
let ty = sg_inv_id / num_of_threads;
if (subgroup_id == 0u && q_row_start < params.seq_len_q) {
for (var kv_base : u32 = 0u; kv_base < KV_TILE; kv_base += VEC_NE) {
for (var kv_base : u32 = 0u; kv_base < KV_TILE; kv_base += subgroup_size / D_SPLIT) {
let kv_idx = kv_base + ty;
var partial_sum: f32 = 0.0;
let kv_valid = kv_idx < KV_TILE && (kv_tile + kv_idx) < params.seq_len_kv;
Expand Down Expand Up @@ -486,15 +483,18 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
if (!skip_tile) {
// we have P (KV_TILE) in inter_shmem and V (KV_TILE x head_dim_v) in kv_shmem
// we want to compute O += P * V across the full KV tile
let ne_threads : u32 = VEC_NE;
let ne_threads : u32 = subgroup_size / D_SPLIT;
let nl_threads = max(1u, subgroup_size / ne_threads);
let tx_pv = sg_inv_id % nl_threads;
let ty_pv = sg_inv_id / nl_threads;
if (subgroup_id == 0u && q_row_start < params.seq_len_q) {
for (var vec_col = tx_pv; vec_col < (HEAD_DIM_V / 4u); vec_col += nl_threads) {
var lo = vec4<f32>(0.0, 0.0, 0.0, 0.0);
for (var cc = 0u; cc < KV_TILE / ne_threads; cc += 1u) {
for (var cc = 0u; cc * ne_threads < KV_TILE; cc += 1u) {
let kv_idx = cc * ne_threads + ty_pv;
if (kv_idx >= KV_TILE) {
continue;
}
let v_row = kv_tile + kv_idx;
if (v_row >= params.seq_len_kv) {
continue;
Expand Down
Loading