EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

vulkan: support arbitrary KV dimension in flash attention (#16160)

Browse Source 
The "Clamp" spec constant is already based on whether KV is a multiple of Bc,
so use that to control whether bounds checking is performed. Add bounds checking
to the scalar and coopmat1 paths. Coopmat2 didn't need any changes (the K/V
tensors are already optionally clamped, nothing else needed to be changed).
This commit is contained in:
Jeff Bolz
2025-09-27 16:43:39 -04:00
committed by GitHub
parent 8656f5de68
commit e6d65fb02d
3 changed files with 38 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
View File

@@ -117,6 +117,9 @@ void main() {
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
[[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
#if BLOCK_SIZE > 1
uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
@@ -155,7 +158,11 @@ void main() {
uint32_t c = (idx + tid) % Bc;
uint32_t r = (idx + tid) / Bc;
if (idx + tid < Bc * Br) {
masksh[c][r] = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
if (!KV_bounds_check || j * Bc + c < KV) {
masksh[c][r] = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
} else {
masksh[c][r] = float(0);
}
}
}
barrier();
@@ -172,8 +179,11 @@ void main() {
float rowmaxf[Br], Pf[Br][cols_per_thread], rowsumf[Br], eMf[Br], Moldf[Br];
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
rowmaxf[r] = Sf[r][0];
rowmaxf[r] = NEG_FLT_MAX_OVER_2;
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
rowmaxf[r] = max(rowmaxf[r], Sf[r][c]);
}
Moldf[r] = Mf[r];
@@ -190,6 +200,9 @@ void main() {
// Compute sum across row of P
rowsumf[r] = 0.0;
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
rowsumf[r] += Pf[r][c];
}
@@ -203,6 +216,9 @@ void main() {
}
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
#if BLOCK_SIZE > 1
uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);