whispercpp 1.3.2 → 1.3.4

data/ext/sources/ggml/src/ggml-vulkan/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);
         }
@@ -99,6 +100,10 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1 || p.nem3 != 1) {
+        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+    }
 
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
@@ -112,7 +117,10 @@ 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) {
+            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);
                 uint ib = coord / BLOCK_SIZE;
@@ -144,13 +152,17 @@ void main() {
             }
         }
 
-        if (p.mask != 0) {
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
 
             [[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;
                 if (idx + tid < Bc * Br) {
-                    masksh[c][r] = float(data_m[(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();
@@ -167,8 +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];
@@ -185,20 +200,26 @@ 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];
             }
 
             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) {
+            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);
                 uint ib = coord / BLOCK_SIZE;
@@ -255,7 +276,7 @@ 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 +298,11 @@ 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 +310,7 @@ 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);
@@ -300,23 +321,47 @@ void main() {
         return;
     }
 
+    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
+
+            float ms = 1.0f;
+            float vs = 1.0f;
+
+            if (sink > Mf[r]) {
+                ms = exp(Mf[r] - sink);
+
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                    Of[r][d] *= ms;
+                }
+            } else {
+                vs = exp(sink - Mf[r]);
+            }
+
+            Lf[r] = Lf[r]*ms + vs;
+        }
+    }
+
     float Lfrcp[Br];
     [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
         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];
+#if defined(ACC_TYPE_MAX)
+            Of[r][d] = clamp(Of[r][d], -vec4(ACC_TYPE_MAX), vec4(ACC_TYPE_MAX));
+#endif
         }
     }
 
-    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 +371,9 @@ 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]);
                     }
                 }
             }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp

@@ -4,10 +4,16 @@ 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;
 
+// 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;
+
+const bool KV_bounds_check = Clamp != 0;
 
 layout (push_constant) uniform parameter {
     uint32_t N;
@@ -24,6 +30,8 @@ layout (push_constant) uniform parameter {
     uint32_t nev2;
     uint32_t nev3;
     uint32_t nem1;
+    uint32_t nem2;
+    uint32_t nem3;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -34,14 +42,12 @@ layout (push_constant) uniform parameter {
     uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
-    uint32_t nb31;
 
     float scale;
     float max_bias;
     float logit_softcap;
 
-    uint32_t mask;
-    uint32_t n_head_log2;
+    uint32_t mask_n_head_log2;
     float m0;
     float m1;
 
@@ -50,35 +56,59 @@ layout (push_constant) uniform parameter {
     uint32_t k_num;
 } p;
 
-layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
+#define SINK_ENABLE_BIT (1<<24)
+#define MASK_ENABLE_BIT (1<<16)
+#define N_LOG2_MASK 0xFFFF
+
+layout (binding = 4) readonly buffer S {float data_s[];};
+
+layout (binding = 5) writeonly buffer O {D_TYPE data_o[];};
 
 #if defined(A_TYPE_PACKED16)
 #define BINDING_IDX_K 0
 #define BINDING_IDX_V 1
-layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2];
+layout (binding = 1) readonly buffer K_PACKED16 {A_TYPE_PACKED16 k_data_packed16[];} k_packed;
+layout (binding = 2) readonly buffer V_PACKED16 {A_TYPE_PACKED16 v_data_packed16[];} v_packed;
 #endif
 
 #if defined(DATA_A_Q4_0)
 #define BLOCK_BYTE_SIZE 18
 
 vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
-    uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
-    uint shift = (iqs & 0x10) >> 2;
-    vui_lo >>= shift;
-    vui_hi >>= shift;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+    if (binding_idx == BINDING_IDX_K) {
+        uint vui_lo = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
+        uint vui_hi = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
+        uint shift = (iqs & 0x10) >> 2;
+        vui_lo >>= shift;
+        vui_hi >>= shift;
+
+        return float(k_packed.k_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+    } else {
+        uint vui_lo = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
+        uint vui_hi = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
+        uint shift = (iqs & 0x10) >> 2;
+        vui_lo >>= shift;
+        vui_hi >>= shift;
+
+        return float(v_packed.v_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+    }
 }
 #endif
 
 #if defined(DATA_A_Q8_0)
 #define BLOCK_BYTE_SIZE 34
 vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
-    const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
+    if (binding_idx == BINDING_IDX_K) {
+        const i8vec2 v0 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
+        const i8vec2 v1 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
+
+        return float(k_packed.k_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+    } else {
+        const i8vec2 v0 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
+        const i8vec2 v1 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
 
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+        return float(v_packed.v_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+    }
 }
 #endif
 
@@ -100,12 +130,22 @@ ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const i
 {
     const uint32_t h = iq2 + (r % p.gqa_ratio);
 
-    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
-    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
+    uint32_t n_head_log2 = p.mask_n_head_log2 & N_LOG2_MASK;
+
+    const ACC_TYPE base = ACC_TYPE(h < n_head_log2 ? p.m0 : p.m1);
+    const int      exph = int(h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1);
 
     return ACC_TYPE(pow(base, ACC_TYPE(exph)));
 }
 
+// Load the sink value, indexed by Q's dimension 2.
+ACC_TYPE perElemOpGetSink(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
+{
+    const uint32_t h = iq2 + (r % p.gqa_ratio);
+
+    return ACC_TYPE(data_s[h]);
+}
+
 uint32_t i, N, KV, split_k_index, Tr, start_j, end_j,
          iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
          q_stride, k_stride, v_stride, m_stride;

data/ext/sources/ggml/src/ggml-vulkan/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_pad / 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_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 ksh[Bc * kshstride];
 
 shared float slope[Br];
@@ -72,20 +74,35 @@ 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 * 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);
         }
@@ -123,36 +140,43 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1 || p.nem3 != 1) {
+        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+    }
 
     [[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) {
+                f16vec4 K_Tf = f16vec4(0);
+                if (!KV_bounds_check || j * Bc + c < KV) {
 #if BLOCK_SIZE > 1
-                uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
-                uint ib = coord / BLOCK_SIZE;
-                uint iqs = (coord % BLOCK_SIZE);
-                f16vec4 K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
+                    uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
+                    uint ib = coord / BLOCK_SIZE;
+                    uint iqs = (coord % BLOCK_SIZE);
+                    K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
 #else
-                f16vec4 K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
+                    K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
 #endif
+                }
 
                 ksh[c * kshstride + d] = K_Tf;
             }
         }
         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_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 < D / 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;
@@ -176,12 +200,14 @@ void main() {
             barrier();
         }
 
-        if (p.mask != 0) {
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
             [[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;
                 if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
-                    sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]));
+                    if (!KV_bounds_check || j * Bc + c < KV) {
+                        sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]));
+                    }
                 }
             }
             barrier();
@@ -189,8 +215,11 @@ void main() {
 
         float eMf[rows_per_thread];
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            float rowmaxf = sfsh[tile_row(r) + (0 * cols_per_iter + col_tid) * sfshstride];
+            float rowmaxf = 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 = max(rowmaxf, float(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride]));
             }
             float Moldf = Mf[r];
@@ -202,9 +231,9 @@ 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];
+                Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             }
         }
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
@@ -212,12 +241,15 @@ 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;
+            }
             float Pf[rows_per_thread];
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                 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;
@@ -227,7 +259,7 @@ void main() {
                 vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
 #endif
                 [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                    Of[r][d] += float16_t(Pf[r]) * ACC_TYPEV4(Vf);
+                    Of[r][d] += ACC_TYPE(Pf[r]) * ACC_TYPEV4(Vf);
                 }
             }
         }
@@ -280,9 +312,9 @@ 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];
+            Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             tmpshv4[tid] = Of[r][d];
 
             barrier();
@@ -300,11 +332,11 @@ 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 +344,7 @@ 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);
@@ -323,23 +355,47 @@ void main() {
         return;
     }
 
+    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
+
+            float ms = 1.0f;
+            float vs = 1.0f;
+
+            if (sink > Mf[r]) {
+                ms = exp(Mf[r] - sink);
+
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                    Of[r][d] *= ACC_TYPE(ms);
+                }
+            } else {
+                vs = exp(sink - Mf[r]);
+            }
+
+            Lf[r] = Lf[r]*ms + vs;
+        }
+    }
+
     float Lfrcp[rows_per_thread];
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         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]);
+            Of[r][d] *= ACC_TYPE(Lfrcp[r]);
+#if defined(ACC_TYPE_MAX)
+            Of[r][d] = clamp(Of[r][d], -ACC_TYPE_MAX, ACC_TYPE_MAX);
+#endif
         }
     }
 
-    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 +405,9 @@ 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]);
                     }
                 }
             }