Vulkan: flash attention for DeepSeek models (#584)

* vulkan: support mixed/deepseekR1 FA head sizes (#14509)

* vulkan: better parameterize FA by head sizes

* vulkan: support mixed/deepseekR1 FA head sizes

* Fix the FA cherry-pick

---------

Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

4 files changed, 83 insertions, 80 deletions

diff --git a/ggml/src/vulkan-shaders/flash_attn.comp b/ggml/src/vulkan-shaders/flash_attn.comp
index ce230a8f..454b3411 100644
--- a/ggml/src/vulkan-shaders/flash_attn.comp
+++ b/ggml/src/vulkan-shaders/flash_attn.comp
@@ -11,7 +11,8 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t D_per_thread = D / D_split;
+const uint32_t HSK_per_thread = HSK / D_split;
+const uint32_t HSV_per_thread = HSV / D_split;
 
 const uint32_t cols_per_iter = WorkGroupSize / D_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
@@ -29,7 +30,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    uint32_t offset = (iq2 + r) * D + c;
+    uint32_t offset = (iq2 + r) * HSV + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -38,7 +39,7 @@ shared FLOAT_TYPE tmpsh[WorkGroupSize];
 shared vec4 tmpshv4[WorkGroupSize];
 
 shared float masksh[Bc][Br];
-shared vec4 Qf[Br][D / 4];
+shared vec4 Qf[Br][HSK / 4];
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
@@ -53,18 +54,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (D / 4);
-        uint32_t r = (idx + tid) / (D / 4);
-        if (r < Br && d < D / 4 &&
+    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (HSK / 4);
+        uint32_t r = (idx + tid) / (HSK / 4);
+        if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
             Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
         }
     }
     barrier();
 
-    vec4 Of[Br][D_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    vec4 Of[Br][HSV_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] = vec4(0.0);
         }
@@ -112,7 +113,7 @@ void main() {
 
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[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);
                 uint ib = coord / BLOCK_SIZE;
@@ -191,14 +192,14 @@ void main() {
             Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
         }
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
                 Of[r][d] = eMf[r] * Of[r][d];
             }
         }
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[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);
                 uint ib = coord / BLOCK_SIZE;
@@ -255,7 +256,7 @@ void main() {
         Lf[r] = tmpsh[d_tid];
         barrier();
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
             Of[r][d] = eMf * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -277,11 +278,11 @@ void main() {
     // 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) {
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -289,7 +290,7 @@ void main() {
             }
         }
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
                 perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -305,18 +306,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] *= Lfrcp[r];
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -326,9 +327,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (i * Br + r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + r) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }
diff --git a/ggml/src/vulkan-shaders/flash_attn_base.comp b/ggml/src/vulkan-shaders/flash_attn_base.comp
index 61d90e2d..1d3e6387 100644
--- a/ggml/src/vulkan-shaders/flash_attn_base.comp
+++ b/ggml/src/vulkan-shaders/flash_attn_base.comp
@@ -4,10 +4,10 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
 layout (constant_id = 1) const uint32_t Br = 1;
 layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-layout (constant_id = 4) const uint32_t Clamp = 0;
-layout (constant_id = 5) const uint32_t D_split = 16;
-
+layout (constant_id = 3) const uint32_t HSK = 32;
+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;
 
 layout (push_constant) uniform parameter {
     uint32_t N;
diff --git a/ggml/src/vulkan-shaders/flash_attn_cm1.comp b/ggml/src/vulkan-shaders/flash_attn_cm1.comp
index da478be2..ad7594fe 100644
--- a/ggml/src/vulkan-shaders/flash_attn_cm1.comp
+++ b/ggml/src/vulkan-shaders/flash_attn_cm1.comp
@@ -13,7 +13,9 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t D_per_thread = D / D_split;
+const uint32_t HSK_per_thread = HSK / D_split;
+const uint32_t HSV_per_thread = HSV / D_split;
+
 const uint32_t row_split = 4;
 const uint32_t rows_per_thread = Br / row_split;
 const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split;
@@ -32,7 +34,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    uint32_t offset = (iq2 + r) * D + c;
+    uint32_t offset = (iq2 + r) * HSV + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -44,14 +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 = D / 4 + 2; // in units of f16vec4
+const uint32_t qstride = HSK / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
-// Avoid padding for D==256 to make it fit in 48KB shmem.
-const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
+// 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 = D / 4 + 2; // in units of f16vec4
+const uint32_t kshstride = HSK / 4 + 2; // in units of f16vec4
 shared f16vec4 ksh[Bc * kshstride];
 
 shared float slope[Br];
@@ -74,18 +76,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (D / 4);
-        uint32_t r = (idx + tid) / (D / 4);
-        if (r < Br && d < D / 4 &&
+    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (HSK / 4);
+        uint32_t r = (idx + tid) / (HSK / 4);
+        if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
             Qf[r * qstride + d] = f16vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale);
         }
     }
     barrier();
 
-    ACC_TYPEV4 Of[rows_per_thread][D_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    ACC_TYPEV4 Of[rows_per_thread][HSV_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Of[r][d] = ACC_TYPEV4(0.0);
         }
@@ -127,10 +129,10 @@ void main() {
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
-        [[unroll]] for (uint32_t idx = 0; idx < Bc * D / 4; idx += gl_WorkGroupSize.x) {
-            uint32_t d = (idx + tid) % (D / 4);
-            uint32_t c = (idx + tid) / (D / 4);
-            if (c < Bc && d < D / 4) {
+        [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
+            uint32_t d = (idx + tid) % (HSK / 4);
+            uint32_t c = (idx + tid) / (HSK / 4);
+            if (c < Bc && d < HSK / 4) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
                 uint ib = coord / BLOCK_SIZE;
@@ -145,14 +147,14 @@ void main() {
         }
         barrier();
 
-        // K * Q^T -> S^T: Bc x D * D x Br -> Bc x Br
-        // Bc split across workgroup (four subgroups), loop over D in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
+        // K * Q^T -> S^T: Bc x HSK * HSK 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 < D / 16; ++d) {
+        for (uint32_t d = 0; d < HSK / 16; ++d) {
             coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
 
             uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
@@ -202,7 +204,7 @@ void main() {
             eMf[r] = exp(Moldf - Mf[r]);
         }
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                 Of[r][d] = float16_t(eMf[r]) * Of[r][d];
             }
@@ -217,7 +219,7 @@ void main() {
                 Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]);
                 Lf[r] += Pf[r];
             }
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[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);
                 uint ib = coord / BLOCK_SIZE;
@@ -280,7 +282,7 @@ void main() {
     }
 
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
             Of[r][d] = float16_t(eMf[r]) * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -300,11 +302,11 @@ void main() {
     // 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) {
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
                     }
@@ -312,7 +314,7 @@ void main() {
             }
         }
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
                 perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -328,18 +330,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Of[r][d] *= float16_t(Lfrcp[r]);
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
                     }
@@ -349,9 +351,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (i * Br + tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * D + (i * Br + tile_row(r)) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }
diff --git a/ggml/src/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/vulkan-shaders/flash_attn_cm2.comp
index 6acf67a0..91caa184 100644
--- a/ggml/src/vulkan-shaders/flash_attn_cm2.comp
+++ b/ggml/src/vulkan-shaders/flash_attn_cm2.comp
@@ -61,8 +61,8 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    if (r < N && c < D) {
-        uint32_t offset = (iq2 + r) * D + c;
+    if (r < N && c < HSV) {
+        uint32_t offset = (iq2 + r) * HSV + c;
         data_o[o_offset + offset] = D_TYPE(elem);
     }
     return elem;
@@ -86,9 +86,9 @@ void main() {
     tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE);
 #endif
 
-    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, D);
-    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
-    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
+    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, HSK);
+    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, HSK);
+    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, HSV);
 
     // hint to the compiler that strides are aligned for the aligned variant of the shader
     if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
@@ -104,16 +104,16 @@ void main() {
     tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
     tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
 
-    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Q;
-    coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Qf16;
+    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseAccumulator> Q;
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, 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, D));
+    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK));
 
-    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA>(Q);
+    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+    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, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -135,10 +135,10 @@ 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, D, Bc, gl_MatrixUseB> K_T;
+        coopmat<float16_t, gl_ScopeWorkgroup, HSK, 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, D), tensorViewTranspose DECODEFUNC);
+        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK), tensorViewTranspose DECODEFUNC);
         S = coopMatMulAdd(Qf16, K_T, S);
 
         if (p.logit_softcap != 0.0f) {
@@ -203,42 +203,42 @@ 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, D, gl_MatrixUseB> V;
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV, 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, D) DECODEFUNC);
+        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV) 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, D, gl_MatrixUseAccumulator> eMdiag;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, 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, D, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
         PV = coopMatMulAdd(P_A, V, PV);
 
-        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(PV);
+        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, 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, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
 
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
         coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
         return;
     }
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Ldiag;
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Ldiag;
 
     // resize L by using smear/reduce
     coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
@@ -250,18 +250,18 @@ void main() {
 
     O = Ldiag*O;
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
-    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
     if (p.gqa_ratio > 1) {
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
     } else {
         tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV);
-        tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D);
+        tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, HSV);
 
         // 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, D), tensorViewPermute);
+        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV), tensorViewPermute);
     }
 }