CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-11 10:36:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Support FA with any multiple of 8 head sizes (#15537)

Browse Source 
The scalar FA shader already handled multiples of 8. The coopmat1 FA
shader assumed 16x16x16 and the shared memory allocations need the HSK
dimensions padded to a multiple of 16. NVIDIA's coopmat2 implementation
requires multiples of 16 for N and K, and needs the matrix dimensions
padded and loads clamped.

Store the FA pipelines in a map, indexed by the pipeline state.
This commit is contained in:
Jeff Bolz
2025-08-24 04:24:25 -05:00
committed by GitHub
parent a9c6ffcbfa
commit c9a24fb932
5 changed files with 143 additions and 137 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
View File

@@ -9,6 +9,10 @@ layout (constant_id = 4) const uint32_t HSV = 32;
layout (constant_id = 5) const uint32_t Clamp = 0;
layout (constant_id = 6) const uint32_t D_split = 16;
// Round up head sizes to a multiple of 16, for coopmat1/coopmat2 paths
const uint32_t HSK_pad = (HSK + 15) & ~15;
const uint32_t HSV_pad = (HSV + 15) & ~15;
layout (push_constant) uniform parameter {
uint32_t N;
uint32_t KV;

23
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
View File

@@ -46,14 +46,14 @@ const uint32_t MatBc = 16;
shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
const uint32_t qstride = HSK / 4 + 2; // in units of f16vec4
const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
shared f16vec4 Qf[Br * qstride];
// Avoid padding for hsk==256 to make it fit in 48KB shmem.
const uint32_t sfshstride = (HSK <= 128) ? (Br + 8) : Br;
shared ACC_TYPE sfsh[Bc * sfshstride];
const uint32_t kshstride = HSK / 4 + 2; // in units of f16vec4
const uint32_t kshstride = HSK_pad / 4 + 2; // in units of f16vec4
shared f16vec4 ksh[Bc * kshstride];
shared float slope[Br];
@@ -74,6 +74,21 @@ void main() {
#define tile_row(r) (row_tid * rows_per_thread + (r))
// Zero-initialize shared memory for Q/K when HSK is not a multiple of 16 (HSK_pad > HSK).
if ((HSK % 16) != 0) {
[[unroll]] for (uint i = 0; i < Br * qstride; i += gl_WorkGroupSize.x) {
if (i + tid < Br * qstride) {
Qf[i + tid] = f16vec4(0);
}
}
[[unroll]] for (uint i = 0; i < Bc * kshstride; i += gl_WorkGroupSize.x) {
if (i + tid < Bc * kshstride) {
ksh[i + tid] = f16vec4(0);
}
}
barrier();
}
uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
[[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
@@ -151,14 +166,14 @@ void main() {
}
barrier();
// K * Q^T -> S^T: Bc x HSK * HSK x Br -> Bc x Br
// K * Q^T -> S^T: Bc x HSK_pad * HSK_pad x Br -> Bc x Br
// Bc split across workgroup (four subgroups), loop over HSK in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
// This is written transposed in order to allow for N being 8 if implementations need it
coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
for (uint32_t d = 0; d < HSK / 16; ++d) {
for (uint32_t d = 0; d < HSK_pad / 16; ++d) {
coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;

36
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
View File

@@ -104,16 +104,16 @@ void main() {
tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseAccumulator> Q;
coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA> Qf16;
coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseAccumulator> Q;
coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA> Qf16;
uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK));
coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK_pad));
Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA>(Q);
Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA>(Q);
Qf16 *= float16_t(p.scale);
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
@@ -140,10 +140,10 @@ void main() {
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
coopmat<float16_t, gl_ScopeWorkgroup, HSK, Bc, gl_MatrixUseB> K_T;
coopmat<float16_t, gl_ScopeWorkgroup, HSK_pad, Bc, gl_MatrixUseB> K_T;
uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK), tensorViewTranspose DECODEFUNC);
coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK_pad), tensorViewTranspose DECODEFUNC);
S = coopMatMulAdd(Qf16, K_T, S);
if (p.logit_softcap != 0.0f) {
@@ -208,31 +208,31 @@ void main() {
rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
rowsum = coopMatMulAdd(P_A, One, rowsum);
coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV, gl_MatrixUseB> V;
coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV_pad, gl_MatrixUseB> V;
uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
coopMatLoadTensorNV(V, data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV) DECODEFUNC);
coopMatLoadTensorNV(V, data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV_pad) DECODEFUNC);
L = eM*L + rowsum;
// This is the "diagonal" matrix in the paper, but since we do componentwise
// multiply rather than matrix multiply it has the diagonal element smeared
// across the row
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> eMdiag;
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> eMdiag;
// resize eM by using smear/reduce
coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
// multiply with fp16 accumulation, then add to O.
coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
PV = coopMatMulAdd(P_A, V, PV);
O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(PV);
O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(PV);
}
// If there is split_k, then the split_k resolve shader does the final
// division by L. Store the intermediate O value and per-row m and L values.
if (p.k_num > 1) {
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
@@ -243,16 +243,16 @@ void main() {
return;
}
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Ldiag;
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Ldiag;
// resize L by using smear/reduce
coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> S;
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> S;
coopMatPerElementNV(S, S, perElemOpGetSink, iq2);
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Mr;
coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Mr;
// resize M by using smear/reduce
coopMatReduceNV(Mr, M, gl_CooperativeMatrixReduceRowNV, smearReduce);
@@ -285,7 +285,7 @@ void main() {
uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
if (p.gqa_ratio > 1) {
coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
} else {
@@ -295,6 +295,6 @@ void main() {
// permute dimensions
tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV), tensorViewPermute);
coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV_pad), tensorViewPermute);
}
}