vulkan: Rewrite synchronization to allow some overlap between nodes (#15489)

Track a list of nodes that need synchronization, and only sync if the new node
depends on them (or overwrites them). This allows some overlap which can
improve performance, and centralizes a big chunk of the synchronization logic.

The remaining synchronization logic involves writes to memory other than the
nodes, e.g. for dequantization or split_k. Each of these allocations has a bool
indicating whether they were in use and need to be synced. This should be
checked before they are written to, and set to true after they are done being
consumed.

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1231,6 +1231,14 @@ struct ggml_backend_vk_context {
     vk_pipeline_struct * prealloc_y_last_pipeline_used {};
     const ggml_tensor * prealloc_y_last_tensor_used {};
 
+    // Track which nodes have been used since the last sync, and whether they were written to
+    std::vector<const ggml_tensor *> unsynced_nodes_written;
+    std::vector<const ggml_tensor *> unsynced_nodes_read;
+    // Track which prealloc buffers have pending reads that need to be synchronized.
+    // These are checked before writing to the buffer (and call ggml_vk_sync_buffers if set),
+    // and set to true after the buffer contents are consumed.
+    bool prealloc_x_need_sync, prealloc_y_need_sync, prealloc_split_k_need_sync;
+
     vk_buffer buffer_pool[MAX_VK_BUFFERS];
 
     vk_context_ref compute_ctx;
@@ -1906,14 +1914,18 @@ static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) {
     return { buf, 0, VK_WHOLE_SIZE };
 }
 
-static void ggml_vk_sync_buffers(vk_context& ctx) {
+static void ggml_vk_sync_buffers(ggml_backend_vk_context* ctx, vk_context& subctx) {
     VK_LOG_DEBUG("ggml_vk_sync_buffers()");
 
-    const bool transfer_queue = ctx->p->q->transfer_only;
+    const bool transfer_queue = subctx->p->q->transfer_only;
 
-    ctx->s->buffer.pipelineBarrier(
-        ctx->p->q->stage_flags,
-        ctx->p->q->stage_flags,
+    if (ctx) {
+        ctx->prealloc_x_need_sync = ctx->prealloc_y_need_sync = ctx->prealloc_split_k_need_sync = false;
+    }
+
+    subctx->s->buffer.pipelineBarrier(
+        subctx->p->q->stage_flags,
+        subctx->p->q->stage_flags,
         {},
         { {
           { !transfer_queue ? (vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) : (vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) },
@@ -4898,7 +4910,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
             }
         }
 
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(ctx, subctx);
         subctx->s->buffer.copyBuffer(buf->buffer, dst->buffer, slices);
         return;
     }
@@ -4913,7 +4925,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
     ggml_vk_ensure_sync_staging_buffer(ctx->device, copy_size);
     VkBufferCopy buf_copy{ 0, offset, copy_size };
 
-    ggml_vk_sync_buffers(subctx);
+    ggml_vk_sync_buffers(ctx, subctx);
     vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
 
     for (uint64_t i3 = 0; i3 < ne3; i3++) {
@@ -4967,7 +4979,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
             }
         }
 
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(nullptr, subctx);
         subctx->s->buffer.copyBuffer(buf->buffer, dst->buffer, slices);
         return;
     }
@@ -4988,7 +5000,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
         offset,
         copy_size};
 
-    ggml_vk_sync_buffers(subctx);
+    ggml_vk_sync_buffers(nullptr, subctx);
     vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging_buffer->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
 
     if (width == spitch) {
@@ -5068,7 +5080,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size
 
     if (buf != nullptr) {
         // Memory is pinned, use as staging buffer
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(nullptr, subctx);
         subctx->s->buffer.copyBuffer(src->buffer, buf->buffer, slices);
 
         return;
@@ -5085,7 +5097,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size
 
     vk_buffer& staging_buffer = src->device->sync_staging;
 
-    ggml_vk_sync_buffers(subctx);
+    ggml_vk_sync_buffers(nullptr, subctx);
     subctx->s->buffer.copyBuffer(src->buffer, staging_buffer->buffer, slices);
 
     deferred_memcpy(dst, staging_buffer->ptr, copy_size, &subctx->out_memcpys);
@@ -5275,13 +5287,16 @@ static void ggml_vk_matmul(
         uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3,
         uint32_t padded_n) {
         VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", padded_n: " << padded_n << ")");
-    ggml_vk_sync_buffers(subctx);
     if (split_k == 1) {
         const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3, padded_n };
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, pc, { m, n, batch });
         return;
     }
 
+    if (ctx->prealloc_split_k_need_sync) {
+        ggml_vk_sync_buffers(ctx, subctx);
+    }
+
     GGML_ASSERT(batch_stride_d == m * n);
 
     // Round the split size up to a multiple of 256 (k-quant alignment)
@@ -5291,9 +5306,10 @@ static void ggml_vk_matmul(
     const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k_split, ne02, ne12, broadcast2, broadcast3, padded_n };
     // Make sure enough workgroups get assigned for split k to work
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
-    ggml_vk_sync_buffers(subctx);
+    ggml_vk_sync_buffers(ctx, subctx);
     const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2, { m * n * batch, 1, 1 });
+    ctx->prealloc_split_k_need_sync = true;
 }
 
 static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, uint32_t m, uint32_t n, bool aligned, ggml_type src0_type) {
@@ -5338,7 +5354,6 @@ static void ggml_vk_matmul_id(
         "m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " <<
         "batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " <<
         "n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")");
-    ggml_vk_sync_buffers(subctx);
     const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d,
                                               nei0, nei1, nbi1, ne11, padded_n };
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, pc, { m, nei1, n_as });
@@ -5469,8 +5484,8 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     };
     init_pushconst_fastdiv(pc);
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, pc, elements);
+    ggml_vk_sync_buffers(ctx, subctx);
 }
 
 static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
@@ -5488,8 +5503,8 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
 
     vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
 
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{ne}, { ne, 1, 1 });
+    ggml_vk_sync_buffers(ctx, subctx);
 }
 
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5684,12 +5699,23 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         GGML_ASSERT(qy_sz == y_sz);
     }
 
+    if (x_non_contig || qx_needs_dequant) {
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    if (y_non_contig || quantize_y) {
+        if (ctx->prealloc_y_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+
     if (x_non_contig) {
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_sync_buffers(ctx, subctx);
     }
     if (y_non_contig) {
         if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
@@ -5728,6 +5754,13 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21,
         split_k, ne12*ne13, ne02, ne12, r2, r3, padded_n
     );  // NOLINT
+
+    if (x_non_contig || qx_needs_dequant) {
+        ctx->prealloc_x_need_sync = true;
+    }
+    if (y_non_contig || quantize_y) {
+        ctx->prealloc_y_need_sync = true;
+    }
 }
 
 static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5874,6 +5907,17 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         GGML_ASSERT(qy_sz == y_sz);
     }
 
+    if (x_non_contig) {
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    if (y_non_contig) {
+        if (ctx->prealloc_y_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+
     if (x_non_contig) {
         GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
@@ -5917,10 +5961,16 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         stride_batch_x, stride_batch_y, stride_batch_d,
         (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
     };
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
                               { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
                               pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
+
+    if (x_non_contig) {
+        ctx->prealloc_x_need_sync = true;
+    }
+    if (y_non_contig) {
+        ctx->prealloc_y_need_sync = true;
+    }
 }
 
 static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -6007,7 +6057,6 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
         workgroups_z /= gqa_ratio;
     }
 
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { 1, (uint32_t)ne01, workgroups_z });
 }
 
@@ -6094,7 +6143,6 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 
     // compute
     const std::array<uint32_t, 12> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)), nb03, nb13, nb23 };
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
         { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { (uint32_t)ne03, (uint32_t)ne01, (uint32_t)ne12 });
 }
@@ -6306,13 +6354,24 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         GGML_ASSERT(qy_sz == y_sz);
     }
 
+    if (x_non_contig || qx_needs_dequant) {
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    if (y_non_contig) {
+        if (ctx->prealloc_y_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+
     if (x_non_contig) {
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
             { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_sync_buffers(ctx, subctx);
     }
     if (y_non_contig) {
         if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
@@ -6343,6 +6402,13 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         stride_batch_x, stride_batch_y, ne20*ne21,
         n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n
     );  // NOLINT
+
+    if (x_non_contig || qx_needs_dequant) {
+        ctx->prealloc_x_need_sync = true;
+    }
+    if (y_non_contig) {
+        ctx->prealloc_y_need_sync = true;
+    }
 }
 
 static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, bool dryrun = false) {
@@ -6502,6 +6568,17 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         GGML_ASSERT(qy_sz == y_sz);
     }
 
+    if (x_non_contig) {
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    if (y_non_contig) {
+        if (ctx->prealloc_y_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+
     if (x_non_contig) {
         GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
@@ -6538,11 +6615,17 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         (uint32_t)x_ne, stride_batch_y, (uint32_t)(ne20*ne21),
         (uint32_t)nei0, (uint32_t)ne11,
     };
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
         { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
         vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23}, vk_subbuffer{ d_ids, ids_buf_offset, ids_sz } },
         pc, { groups_x, (uint32_t)nei0, groups_z });
+
+    if (x_non_contig) {
+        ctx->prealloc_x_need_sync = true;
+    }
+    if (y_non_contig) {
+        ctx->prealloc_y_need_sync = true;
+    }
 }
 
 static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
@@ -6925,9 +7008,11 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                               mask_n_head_log2, m0, m1,
                                               gqa_ratio, split_kv, split_k };
 
-    ggml_vk_sync_buffers(subctx);
-
     if (split_k > 1) {
+        if (ctx->prealloc_split_k_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {
                                         vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE},
@@ -6943,7 +7028,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     // cancel out the divide by wg_denoms[0].
                                     pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
 
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(ctx, subctx);
         const std::array<uint32_t, 5> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, split_k, (sinks != nullptr) };
         ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
                                     {
@@ -6952,6 +7037,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                         vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                     },
                                     pc2, { (uint32_t)ne1, HSV, (uint32_t)ne3 });
+        ctx->prealloc_split_k_need_sync = true;
     } else {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {
@@ -7820,7 +7906,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             subbuf_y = { d_X, 0, x_sz };
         }
 
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_SOFT_MAX) {
         // Empty src1 and src2 is possible in soft_max, but the shader needs a buffer
@@ -7838,7 +7923,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             subbuf_z = { d_X, 0, x_sz };
         }
 
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
         // Empty src2 is possible in rope, but the shader needs a buffer
@@ -7849,30 +7933,23 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             subbuf_z = { d_X, 0, x_sz };
         }
 
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_IM2COL) {
         // im2col uses only src1 and dst buffers
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_COUNT_EQUAL) {
-        ggml_vk_sync_buffers(subctx);
         // count_equal assumes that destination buffer is initialized with zeroes
         ggml_vk_buffer_memset_async(subctx, d_D, d_buf_offset, 0, d_sz);
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(ctx, subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_OPT_STEP_SGD) {
         // OPT_STEP_SGD works on src0, it does not need dst
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz } }, pc, elements);
     } else if (use_src2) {
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (use_src1) {
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else {
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     }
 }
@@ -7999,7 +8076,6 @@ static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx,
         elements = { ne, 1, 1 };
     }
 
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
         {
             vk_subbuffer{ buf[0], offset[0], VK_WHOLE_SIZE },
@@ -8112,8 +8188,6 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
         src_buf_ctxs[i] = (ggml_backend_vk_buffer_context *)dst->src[i]->buffer->context;
     }
 
-    ggml_vk_sync_buffers(subctx);
-
     vk_buffer d_D = nullptr, d_srcs[7] = { nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr };
     size_t dst_offset = 0, src_offsets[7] = { 0, 0, 0, 0, 0, 0, 0 };
     bool dst_uma = false, srcs_uma[7] = { false, false, false, false, false, false, false };
@@ -8251,8 +8325,6 @@ static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_cont
     ggml_backend_vk_buffer_context * gv_buf_ctx = (ggml_backend_vk_buffer_context *)gv->buffer->context;
     ggml_backend_vk_buffer_context * p_buf_ctx = (ggml_backend_vk_buffer_context *)p->buffer->context;
 
-    ggml_vk_sync_buffers(subctx);
-
     vk_buffer d_X = nullptr, d_G = nullptr, d_GM = nullptr, d_GV = nullptr, d_P = nullptr;
     size_t x_offset = 0, g_offset = 0, gm_offset = 0, gv_offset = 0, p_offset = 0;
     bool X_uma = false, G_uma = false, GM_uma = false, GV_uma = false, P_uma = false;
@@ -9964,6 +10036,83 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         }
     }
 
+    if (!dryrun) {
+        // This logic detects dependencies between modes in the graph and calls ggml_vk_sync_buffers
+        // to synchronize them. This handles most "normal" synchronization when computing the graph, and when
+        // there is no auxiliary memory use, it shouldn't be necessary to call ggml_vk_sync_buffers
+        // outside of this logic. When a node uses one of the prealloc buffers for something like
+        // dequantization or split_k, additional synchronization is needed between those passes.
+        bool need_sync = false;
+
+        // Check whether "node" requires synchronization. The node requires synchronization if it
+        // overlaps in memory with another unsynchronized node and at least one of them is a write.
+        // Destination nodes are checked against both the written/read lists. Source nodes are only
+        // checked against the written list. Two nodes overlap in memory if they come from the same
+        // buffer and the tensor or view ranges overlap.
+        auto const &overlaps_unsynced = [&](const ggml_tensor *node, const std::vector<const ggml_tensor *> &unsynced_nodes) -> bool {
+            if (unsynced_nodes.size() == 0) {
+                return false;
+            }
+            auto n_base = vk_tensor_offset(node) + node->view_offs;
+            auto n_size = ggml_nbytes(node);
+            ggml_backend_vk_buffer_context * a_buf_ctx = (ggml_backend_vk_buffer_context *)node->buffer->context;
+            vk_buffer a_buf = a_buf_ctx->dev_buffer;
+            for (auto &other : unsynced_nodes) {
+                ggml_backend_vk_buffer_context * o_buf_ctx = (ggml_backend_vk_buffer_context *)other->buffer->context;
+                vk_buffer o_buf = o_buf_ctx->dev_buffer;
+                if (a_buf == o_buf) {
+                    auto o_base = vk_tensor_offset(other) + other->view_offs;
+                    auto o_size = ggml_nbytes(other);
+
+                    if ((o_base <= n_base && n_base < o_base + o_size) ||
+                        (n_base <= o_base && o_base < n_base + n_size)) {
+                        return true;
+                    }
+                }
+            }
+            return false;
+        };
+
+        // For all fused ops, check if the destination node or any of the source
+        // nodes require synchronization.
+        for (int32_t i = 0; i < ctx->num_additional_fused_ops + 1 && !need_sync; ++i) {
+            const ggml_tensor *cur_node = cgraph->nodes[node_idx + i];
+            if (overlaps_unsynced(cur_node, ctx->unsynced_nodes_read) || overlaps_unsynced(cur_node, ctx->unsynced_nodes_written)) {
+                need_sync = true;
+                break;
+            }
+            for (uint32_t j = 0; j < GGML_MAX_SRC; ++j) {
+                if (!cur_node->src[j]) {
+                    continue;
+                }
+                if (overlaps_unsynced(cur_node->src[j], ctx->unsynced_nodes_written)) {
+                    need_sync = true;
+                    break;
+                }
+            }
+        }
+        if (need_sync) {
+            VK_LOG_DEBUG("node_idx=" << i << " sync");
+            ctx->unsynced_nodes_written.clear();
+            ctx->unsynced_nodes_read.clear();
+            ggml_vk_sync_buffers(ctx, compute_ctx);
+        } else {
+            VK_LOG_DEBUG("node_idx=" << i << " unsynced");
+        }
+        // Add the last fused node and all fused source nodes to the unsynchronized list.
+        const ggml_tensor * last_node = cgraph->nodes[node_idx + ctx->num_additional_fused_ops];
+        ctx->unsynced_nodes_written.push_back(last_node);
+        for (int32_t i = 0; i < ctx->num_additional_fused_ops + 1; ++i) {
+            const ggml_tensor *cur_node = cgraph->nodes[node_idx + i];
+            for (uint32_t j = 0; j < GGML_MAX_SRC; ++j) {
+                if (!cur_node->src[j]) {
+                    continue;
+                }
+                ctx->unsynced_nodes_read.push_back(cur_node->src[j]);
+            }
+        }
+    }
+
     switch (node->op) {
     case GGML_OP_REPEAT:
         ggml_vk_repeat(ctx, compute_ctx, src0, node, dryrun);
@@ -10427,6 +10576,10 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
     ctx->gc.temp_buffers.clear();
     ctx->prealloc_y_last_pipeline_used = {};
 
+    ctx->unsynced_nodes_written.clear();
+    ctx->unsynced_nodes_read.clear();
+    ctx->prealloc_x_need_sync = ctx->prealloc_y_need_sync = ctx->prealloc_split_k_need_sync = false;
+
     ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
     ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);