whispercpp 1.3.2 → 1.3.4

data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -392,7 +392,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
     }
 }
 
-template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup, bool last_iter>
+template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles,
+    bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup, bool last_iter>
 static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const float2 * const __restrict__ Q_f2,
         const half2  * const __restrict__ K_h2,
@@ -408,7 +409,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const int stride_K,
         const int stride_V,
         const int stride_mask,
-        const int jt,
         half2        * const __restrict__ tile_Q,
         half2        * const __restrict__ tile_K,
         half2        * const __restrict__ tile_V,
@@ -418,7 +418,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         float        * const __restrict__ KQ_max,
         float        * const __restrict__ KQ_rowsum,
         const int kb0) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
     typedef fattn_mma_f16_config<DKQ, DV> c;
 
 #ifdef CP_ASYNC_AVAILABLE
@@ -455,7 +455,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         cp_async_wait_all();
         __syncthreads();
         flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, c::nbatch_fa, use_cp_async>
-            (V_h2 + k_VKQ_0*stride_V, tile_V, nbatch_V2, stride_V);
+            (V_h2 + int64_t(k_VKQ_0)*stride_V, tile_V, nbatch_V2, stride_V);
     } else {
         constexpr bool use_cp_async = nstages == 1;
         if (ncols2 > 1 || mask_h2) {
@@ -471,7 +471,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         if (nstages <= 1) {
             constexpr bool use_cp_async = nstages == 1;
             flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, c::nbatch_fa, use_cp_async>
-                (K_h2 + k_VKQ_0*stride_K + k0_start, tile_K, k0_diff, stride_K);
+                (K_h2 + int64_t(k_VKQ_0)*stride_K + k0_start, tile_K, k0_diff, stride_K);
             if (use_cp_async) {
                 cp_async_wait_all();
             }
@@ -652,9 +652,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         float KQ_max_scale[cols_per_thread];
 #pragma unroll
         for (int col = 0; col < cols_per_thread; ++col) {
-            KQ_max_scale[col] = expf(KQ_max[col] - KQ_max_new[col]);
+            const float KQ_max_diff = KQ_max[col] - KQ_max_new[col];
+            KQ_max_scale[col] = expf(KQ_max_diff);
             KQ_max[col] = KQ_max_new[col];
 
+            *((uint32_t *) &KQ_max_scale[col]) *= KQ_max_diff >= SOFTMAX_FTZ_THRESHOLD;
+
             // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
             KQ_rowsum[col] = KQ_max_scale[col]*KQ_rowsum[col] + KQ_rowsum_add[col];
         }
@@ -712,7 +715,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     (mask_h2 + (k_VKQ_0 + c::nbatch_fa)/2, tile_mask, stride_mask);
             }
             flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, c::nbatch_fa, use_cp_async>
-                (K_h2 + (k_VKQ_0 + c::nbatch_fa)*stride_K, tile_K, nbatch_K2, stride_K);
+                (K_h2 + int64_t(k_VKQ_0 + c::nbatch_fa)*stride_K, tile_K, nbatch_K2, stride_K);
         }
     }
 
@@ -729,7 +732,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         if (nstages <= 1 && i0_start < reusable_cutoff) {
             constexpr bool use_cp_async = nstages == 1;
             flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, c::nbatch_fa, use_cp_async>
-                (V_h2 + k_VKQ_0*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V);
+                (V_h2 + int64_t(k_VKQ_0)*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V);
             if (use_cp_async) {
                 cp_async_wait_all();
             }
@@ -764,17 +767,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         }
     }
 #else
-    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
-    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
-    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V);
-    GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K);
-    GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K);
-    GGML_UNUSED(tile_V); GGML_UNUSED(tile_mask); GGML_UNUSED(Q_B);
-    GGML_UNUSED(VKQ_C); GGML_UNUSED(KQ_max); GGML_UNUSED(KQ_rowsum);
-    GGML_UNUSED(kb0); GGML_UNUSED(tile_Q);
+    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup,
+        scale, slope, logit_softcap, ne01, ne02,
+        stride_K, stride_V, stride_mask,
+        tile_Q, tile_K, tile_V, tile_mask,
+        Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
 }
 
 template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup>
@@ -783,6 +782,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const half2  * const __restrict__ K_h2,
         const half2  * const __restrict__ V_h2,
         const half2  * const __restrict__ mask_h2,
+        const float  * const __restrict__ sinks_f,
         float2       * const __restrict__ dstk,
         float2       * const __restrict__ dstk_fixup,
         const float scale,
@@ -798,7 +798,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const int jt,
         const int kb0_start,
         const int kb0_stop) {
-#ifdef NEW_MMA_AVAILABLE
+#ifdef TURING_MMA_AVAILABLE
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
     typedef fattn_mma_f16_config<DKQ, DV> c;
@@ -917,21 +917,22 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                 (mask_h2 + kb0_start*c::nbatch_fa/2, tile_mask, stride_mask);
         }
         flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, c::nbatch_fa, use_cp_async>
-            (K_h2 + kb0_start*c::nbatch_fa*stride_K, tile_K, nbatch_K2, stride_K);
+            (K_h2 + int64_t(kb0_start)*c::nbatch_fa*stride_K, tile_K, nbatch_K2, stride_K);
     }
 
     // Iterate over ne11 == previous tokens:
-    for (int kb0 = kb0_start; kb0 < kb0_stop-1; ++kb0) {
+    int kb0 = kb0_start;
+    for (; kb0 < kb0_stop-1; ++kb0) {
         constexpr bool last_iter = false;
         flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter>
             (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap,
-             ne01, ne02, stride_K, stride_V, stride_mask, jt, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
+             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     }
     { // kb0_start is always < kb0_stop so the last iter can be executed unconditionally.
         constexpr bool last_iter = true;
         flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter>
             (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap,
-             ne01, ne02, stride_K, stride_V, stride_mask, jt, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0_stop-1);
+             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     }
 
     // With multi-stage loading there is no __syncthreads at the end of the iter,
@@ -954,6 +955,52 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         }
     }
 
+    // If attention sinks are used, potentially re-scale if KQ_max is small.
+    // Also add the sink as a value to KQ_rowsum, this is done after synchonization of KQ_rowsum
+    //     so it's being done unconditionally for every thread.
+    if (!is_fixup && (np == 1 || threadIdx.y % np == 0) && sinks_f) {
+        float KQ_max_scale[cols_per_thread];
+#pragma unroll
+        for (int col = 0; col < cols_per_thread; ++col) {
+            static_assert(ntiles == 1 || ntiles == 2, "ntiles > 2 not implemented");
+            const int jc = ntiles == 1 ? 2*tile_C_VKQ::get_j(col/2) + col % 2 : tile_C_VKQ_16::get_i(col);
+            const float sink = sinks_f[jc % ncols2];
+
+            const float KQ_max_new = fmaxf(KQ_max[col], sink);
+            const float KQ_max_diff = KQ_max[col] - KQ_max_new;
+            KQ_max_scale[col] = expf(KQ_max_diff);
+            KQ_max[col] = KQ_max_new;
+
+            *((uint32_t *) &KQ_max_scale[col]) *= KQ_max_diff >= SOFTMAX_FTZ_THRESHOLD;
+
+            const float KQ_max_add = expf(sink - KQ_max_new);
+            KQ_rowsum[col] = KQ_max_scale[col]*KQ_rowsum[col] + KQ_max_add;
+        }
+
+        if (ntiles == 1) {
+            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[1]);
+#pragma unroll
+            for (int i = 0; i < DV/tile_C_VKQ::I; ++i) {
+#pragma unroll
+                for (int l = 0; l < tile_C_VKQ::ne; ++l) {
+                    VKQ_C[i].x[l] *= KQ_max_scale_h2;
+                }
+            }
+        } else {
+#pragma unroll
+            for (int col = 0; col < cols_per_thread; ++col) {
+                const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[col], KQ_max_scale[col]);
+#pragma unroll
+                for (int i = 0; i < DV/tile_C_VKQ_16::J; ++i) {
+#pragma unroll
+                    for (int l0 = 0; l0 < tile_C_VKQ_16::ne; l0 += 2) {
+                        VKQ_C_16[i*ntiles/2 + col/2].x[l0 + col % 2] *= KQ_max_scale_h2;
+                    }
+                }
+            }
+        }
+    }
+
     // Combine VKQ accumulator values if np > 1.
     // It's also faster to do small writes to shared memory, then large write to VRAM than to do small writes to VRAM.
     // So also write VKQ accumulators to shared memory in column-major format if np == 1.
@@ -1186,14 +1233,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         }
     }
 #else
-    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
-    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
-    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_Q1);
-    GGML_UNUSED(stride_Q2); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V); GGML_UNUSED(stride_mask);
-    GGML_UNUSED(jt); GGML_UNUSED(kb0_start); GGML_UNUSED(kb0_stop);
+    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dstk_fixup,
+        scale, slope, logit_softcap, ne01, ne02,
+        stride_Q1, stride_Q2, stride_K, stride_V, stride_mask,
+        jt, kb0_start, kb0_stop);
     NO_DEVICE_CODE;
-#endif // NEW_MMA_AVAILABLE
+#endif // TURING_MMA_AVAILABLE
 }
 
 template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla>
@@ -1203,6 +1248,8 @@ static __global__ void flash_attn_ext_f16(
         const char * __restrict__ K,
         const char * __restrict__ V,
         const char * __restrict__ mask,
+        const char * __restrict__ sinks,
+        const int  * __restrict__ KV_max,
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
@@ -1211,30 +1258,14 @@ static __global__ void flash_attn_ext_f16(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int nb31,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int nb21,
-        const int nb22,
-        const int nb23,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
-#if defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE)
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
+#if defined(FLASH_ATTN_AVAILABLE) && defined(TURING_MMA_AVAILABLE)
 
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(DKQ == 128 || DKQ == 256)) {
@@ -1246,7 +1277,7 @@ static __global__ void flash_attn_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
-#endif __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
 
     static_assert(!mla || DKQ >= DV, "MLA needs DKQ >= DV");
 
@@ -1269,8 +1300,8 @@ static __global__ void flash_attn_ext_f16(
     constexpr int kb_niter = FATTN_KQ_STRIDE / c::nbatch_fa; // Number of kernel iterations per assigned KQ slice.
 
     // kbc == k block continuous, current index in continuous ijk space.
-    int       kbc      = (blockIdx.x + 0)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
-    const int kbc_stop = (blockIdx.x + 1)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
+    int       kbc      = (blockIdx.x + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    const int kbc_stop = (blockIdx.x + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
 
     // If the seams of 2 CUDA blocks fall within an output tile their results need to be combined.
     // For this we need to track both the block that starts the tile (needs_fixup) and the block that finishes the tile (is_fixup).
@@ -1279,32 +1310,42 @@ static __global__ void flash_attn_ext_f16(
     // kb0 == k start index when in the output tile.
     int kb0_start = kbc % iter_k;
     int kb0_stop  = min(iter_k, kb0_start + kbc_stop - kbc);
+
     while (kbc < kbc_stop && kb0_stop == iter_k) {
-        const int channel = kbc / (iter_k*iter_j);
-        const int jt      = (kbc - channel*iter_k*iter_j) / iter_k; // j index of current tile.
+        const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
+        const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2
+        const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile.
+
+        const int head0 = zt * ncols2;
 
-        const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
-        const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-        const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
-        float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
+        const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02* head0);
+        const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
+        const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
+            (const half2  *) (mask + nb33*(sequence % ne33) + nb31*jt*ncols1);
+        float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head0) * (DV/2);
 
-        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
+        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio));
+        const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr;
 
-        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f;
+        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f;
 
         const int kb0_start_kernel = kb0_start * kb_niter;
-        const int kb0_stop_kernel  = kb0_stop  * kb_niter;
+        int       kb0_stop_kernel  = kb0_stop  * kb_niter;
+
+        if (KV_max) {
+            kb0_stop_kernel = min(kb0_stop_kernel, KV_max[sequence*iter_j + jt] / c::nbatch_fa);
+        }
 
         constexpr bool is_fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         if (kb0_start == 0) {
             constexpr bool needs_fixup = false; // CUDA block is working on an entire tile.
             flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup>
-                (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
+                (Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
                  ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
         } else {
             constexpr bool needs_fixup = true; // CUDA block is working on the beginning of a tile.
             flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup>
-                (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
+                (Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
                  ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
         }
 
@@ -1319,39 +1360,47 @@ static __global__ void flash_attn_ext_f16(
         return;
     }
 
-    const int channel = kbc / (iter_k*iter_j);
-    const int jt      = (kbc - channel*iter_k*iter_j) / iter_k; // j index of current tile.
+    const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
+    const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2
+    const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile.
+
+    const int head0 = zt * ncols2;
 
-    const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
-    const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-    const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
-    float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
+    const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02* head0);
+    const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
+    const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
+        (const half2  *) (mask + nb33*(sequence % ne33) + nb31*jt*ncols1);
+    float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head0) * (DV/2);
 
-    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
+    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio));
+    const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr;
 
-    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f;
+    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f;
 
     const int kb0_start_kernel = kb0_start * kb_niter;
-    const int kb0_stop_kernel  = kb0_stop  * kb_niter;
+    int       kb0_stop_kernel  = kb0_stop  * kb_niter;
+
+    if (KV_max) {
+        kb0_stop_kernel = min(kb0_stop_kernel, KV_max[sequence*iter_j + jt] / c::nbatch_fa);
+    }
 
     constexpr bool is_fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     constexpr bool needs_fixup = false;
     flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup>
-        (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
+        (Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
          ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
-    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
-    GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
-#endif // defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE)
+#endif // defined(FLASH_ATTN_AVAILABLE) && defined(TURING_MMA_AVAILABLE)
 }
 
 template <int DKQ, int DV, int ncols1, int ncols2>
@@ -1401,24 +1450,24 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
         constexpr bool use_logit_softcap = false;
         fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla>;
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
         if (!shared_memory_limit_raised[id]) {
             CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
             shared_memory_limit_raised[id] = true;
         }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     } else {
         constexpr bool use_logit_softcap = true;
         fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla>;
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
         if (!shared_memory_limit_raised[id]) {
             CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
             shared_memory_limit_raised[id] = true;
         }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     }
 
     launch_fattn<DV, ncols1, ncols2>