mozempk/ggml-org-llama-cpp-mirror
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-11-11 22:32:07 +01:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Preprocess FA mask to detect all-neg-inf and all-zero. (#19281)

Browse Source 
Write out a 2-bit code per block and avoid loading the mask when it
matches these two common cases.

Apply this optimization when the mask is relatively large (i.e. prompt
processing).
This commit is contained in:
Jeff Bolz
2026-02-05 09:26:38 -06:00
committed by GitHub
parent 3795cc1e89
commit 449ec2ab07
8 changed files with 360 additions and 123 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
+20 -19
View File
@@ -94,6 +94,10 @@ void main() {
}
}
const uint32_t mo_stride = CEIL_DIV(KV, 16 * Bc);
// mo_offset will point to the tile starting at row i*Br and col 0
uint32_t mo_offset = mo_stride * i;
#if BLOCK_SIZE > 1
uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / BLOCK_BYTE_SIZE;
uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / BLOCK_BYTE_SIZE;
@@ -104,15 +108,28 @@ void main() {
uint32_t m_offset = gqa_iq1*KV;
if (p.nem2 != 1 || p.nem3 != 1) {
m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
mo_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * CEIL_DIV(p.nem1, Br) * mo_stride;
}
uint32_t mask_opt = 0;
uint32_t mask_opt_idx = ~0;
[[dont_unroll]]
for (uint32_t j = start_j; j < end_j; ++j) {
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
if (USE_MASK_OPT && mask_opt_idx != j / 16) {
mask_opt_idx = j / 16;
mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
}
uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
// skip this block
continue;
}
// Only load if the block is not all zeros
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0 && mask_opt_bits != MASK_OPT_ALL_ZERO) {
bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
float max_mask = NEG_FLT_MAX_OVER_2;
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
uint32_t c = (idx + tid) % Bc;
uint32_t r = (idx + tid) / Bc;
@@ -120,25 +137,12 @@ void main() {
if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
masksh[c][r] = m;
max_mask = max(max_mask, m);
} else {
masksh[c][r] = float(0);
}
}
}
// skip the block if the mask is entirely -inf
bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
barrier();
if (gl_SubgroupInvocationID == 0) {
tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
}
barrier();
[[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
max_mask = max(max_mask, tmpsh[s]);
}
if (max_mask <= NEG_FLT_MAX_OVER_2) {
continue;
}
}
float Sf[Br][cols_per_thread];
@@ -185,7 +189,7 @@ void main() {
}
}
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0 && mask_opt_bits != MASK_OPT_ALL_ZERO) {
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
float mvf = masksh[c * cols_per_iter + col_tid][r];
@@ -256,9 +260,6 @@ void main() {
barrier();
}
// prevent race on tmpsh
barrier();
// reduce across threads
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {