Merge branch 'concedo_experimental' into crokeso

Author: Nexesenex

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

@@ -517,6 +517,8 @@ struct vk_device_struct {
 
     ggml_backend_buffer_type buffer_type;
 
+    bool disable_fusion;
+
 #ifdef GGML_VULKAN_MEMORY_DEBUG
     std::unique_ptr<vk_memory_logger> memory_logger;
 #endif
@@ -652,6 +654,7 @@ struct vk_flash_attn_push_constants {
     uint32_t nev3;
     uint32_t nem1;
     uint32_t nem2;
+    uint32_t nem3;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -667,8 +670,7 @@ struct vk_flash_attn_push_constants {
     float max_bias;
     float logit_softcap;
 
-    uint32_t mask;
-    uint32_t n_head_log2;
+    uint32_t mask_n_head_log2;
     float m0;
     float m1;
 
@@ -1107,8 +1109,8 @@ static size_t vk_skip_checks;
 static size_t vk_output_tensor;
 
 static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name);
-static void ggml_vk_check_results_0(ggml_tensor * tensor);
-static void ggml_vk_check_results_1(ggml_tensor * tensor);
+static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx);
+static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx);
 #endif
 
 typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
@@ -3531,6 +3533,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->idx = idx;
 
+        device->disable_fusion = getenv("GGML_VK_DISABLE_FUSION") != nullptr;
+
         return device;
     }
 
@@ -6135,6 +6139,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     const uint32_t nem1 = mask ? mask->ne[1] : 0;
     const uint32_t nem2 = mask ? mask->ne[2] : 0;
+    const uint32_t nem3 = mask ? mask->ne[3] : 0;
 
     const uint32_t HSK = nek0;
     const uint32_t HSV = nev0;
@@ -6202,7 +6207,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 
     if (N == 1 && qk_ratio > 1 && qk_ratio <= max_gqa &&
-        qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) {
+        qk_ratio * nek2 == neq2 && nek2 == nev2 && nem2 <= 1) {
         // grouped query attention - make the N dimension equal to gqa_ratio, reduce
         // workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1
         // and change addressing calculations to index Q's dimension 2.
@@ -6372,17 +6377,19 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         }
     }
 
+    uint32_t mask_n_head_log2 = ((mask != nullptr) << 16) | n_head_log2;
+
     const vk_flash_attn_push_constants pc = { N, KV,
                                               (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3,
                                               (uint32_t)neq2, (uint32_t)neq3,
                                               (uint32_t)nek2, (uint32_t)nek3,
                                               (uint32_t)nev2, (uint32_t)nev3,
-                                              nem1, nem2,
+                                              nem1, nem2, nem3,
                                               q_stride, (uint32_t)nbq2, (uint32_t)nbq3,
                                               k_stride, (uint32_t)nbk2, (uint32_t)nbk3,
                                               v_stride, (uint32_t)nbv2, (uint32_t)nbv3,
                                               scale, max_bias, logit_softcap,
-                                              mask != nullptr, n_head_log2, m0, m1,
+                                              mask_n_head_log2, m0, m1,
                                               gqa_ratio, split_kv, split_k };
 
     ggml_vk_sync_buffers(subctx);
@@ -7675,8 +7682,7 @@ static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun);
 }
 
-static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-    float * op_params = (float *)dst->op_params;
+static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, float * op_params, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
@@ -8906,7 +8912,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
     }
 }
 
-static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_tensor* tensor, int tensor_idx, bool use_fence, bool almost_ready);
+static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_cgraph * cgraph, ggml_tensor* tensor, int tensor_idx, bool use_fence, bool almost_ready);
 
 // Returns true if node has enqueued work into the queue, false otherwise
 // If submit is true the current all operations queued so far are being submitted to Vulkan to overlap cmdlist creation and GPU execution.
@@ -9167,9 +9173,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
             // fused rms_norm + mul
             ggml_tensor *mul = cgraph->nodes[node_idx + 1];
             ggml_tensor *other_src = mul->src[0] == node ? mul->src[1] : mul->src[0];
-            ggml_vk_rms_norm(ctx, compute_ctx, src0, other_src, mul, dryrun);
+            ggml_vk_rms_norm(ctx, compute_ctx, src0, other_src, mul, (float *)node->op_params, dryrun);
         } else {
-            ggml_vk_rms_norm(ctx, compute_ctx, src0, src0, node, dryrun);
+            ggml_vk_rms_norm(ctx, compute_ctx, src0, src0, node, (float *)node->op_params, dryrun);
         }
         break;
     case GGML_OP_RMS_NORM_BACK:
@@ -9329,7 +9335,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         ctx->compute_ctx.reset();
 
-        bool ok = ggml_vk_compute_forward(ctx, node_begin, node_idx_begin, false, almost_ready);
+        bool ok = ggml_vk_compute_forward(ctx, cgraph, node_begin, node_idx_begin, false, almost_ready);
         if (!ok) {
             if (node->op == GGML_OP_UNARY) {
                 std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast<ggml_unary_op>(node->op_params[0])) << ")" << std::endl;
@@ -9344,7 +9350,8 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     return true;
 }
 
-static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx, bool use_fence = true, bool almost_ready = false) {
+static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, ggml_tensor * tensor, int tensor_idx, bool use_fence = true, bool almost_ready = false) {
+    GGML_UNUSED(cgraph);
     ggml_backend_buffer * buf = nullptr;
 
     switch (tensor->op) {
@@ -9454,7 +9461,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     // Only run if ctx hasn't been submitted yet
     if (!subctx->seqs.empty()) {
 #ifdef GGML_VULKAN_CHECK_RESULTS
-        ggml_vk_check_results_0(tensor);
+        ggml_vk_check_results_0(ctx, cgraph, tensor_idx);
         use_fence = true;
 #endif
 
@@ -9474,7 +9481,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
             ggml_vk_wait_for_fence(ctx);
         }
 #ifdef GGML_VULKAN_CHECK_RESULTS
-        ggml_vk_check_results_1(tensor);
+        ggml_vk_check_results_1(ctx, cgraph, tensor_idx);
 #endif
     }
 
@@ -9921,6 +9928,37 @@ static bool ggml_vk_is_empty(ggml_tensor * node) {
     return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
 }
 
+static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
+    if (!ggml_can_fuse(cgraph, node_idx, ops)) {
+        return false;
+    }
+
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_RMS_NORM && ops.begin()[1] == GGML_OP_MUL) {
+        // additional constraints specific to this fusion
+        const ggml_tensor *rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor *mul = cgraph->nodes[node_idx + 1];
+
+        GGML_ASSERT(rms_norm->src[0]->type == GGML_TYPE_F32);
+        GGML_ASSERT(rms_norm->type == GGML_TYPE_F32);
+        // rms_norm only supports f32
+        if (mul->src[0]->type != GGML_TYPE_F32 ||
+            mul->src[1]->type != GGML_TYPE_F32 ||
+            mul->type != GGML_TYPE_F32) {
+            return false;
+        }
+        // if rms_norm is the B operand, then we don't handle broadcast
+        if (rms_norm == mul->src[1] &&
+            mul->src[0]->ne[1] != rms_norm->ne[1]) {
+            return false;
+        }
+        // rms_norm shader assumes contiguous rows
+        if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+            return false;
+        }
+    }
+    return true;
+}
+
 static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
@@ -9934,7 +9972,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
     uint64_t total_mat_mul_bytes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        if (ggml_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
             ctx->num_additional_fused_ops = 1;
         }
         ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
@@ -10004,7 +10042,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
             mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
         }
 
-        if (ggml_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
             ctx->num_additional_fused_ops = 1;
         }
 
@@ -10327,12 +10365,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (op->src[3] && op->src[3]->type != GGML_TYPE_F16) {
                     return false;
                 }
-                // TODO: support broadcast
-                // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14449, but
-                //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
-                if (op->src[0]->ne[3] != 1 || (op->src[3] && op->src[3]->ne[2] != 1)) {
-                    return false;
-                }
                 // It's straightforward to support different K/V dequant, but would
                 // significantly increase the number of pipelines
                 if (op->src[1]->type != op->src[2]->type) {
@@ -10787,11 +10819,21 @@ void * comp_result;
 size_t comp_size;
 size_t comp_nb[GGML_MAX_DIMS];
 size_t check_counter = 0;
-static void ggml_vk_check_results_0(ggml_tensor * tensor) {
+static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
+    ggml_tensor * tensor = cgraph->nodes[tensor_idx];
     if (tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
 
+    bool fused_rms_norm_mul = false;
+    int rms_norm_idx = -1;
+    if (ctx->num_additional_fused_ops == 1 &&
+        tensor->op == GGML_OP_RMS_NORM &&
+        cgraph->nodes[tensor_idx + 1]->op == GGML_OP_MUL) {
+        fused_rms_norm_mul = true;
+        tensor = cgraph->nodes[tensor_idx + 1];
+    }
+
     check_counter++;
     if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
         return;
@@ -10819,6 +10861,15 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
 
     for (int i = 0; i < 6; i++) {
         ggml_tensor * srci = tensor->src[i];
+        if (fused_rms_norm_mul) {
+            rms_norm_idx = tensor->src[0]->op == GGML_OP_RMS_NORM ? 0 : 1;
+            ggml_tensor *rms_norm = tensor->src[rms_norm_idx];
+            switch (i) {
+            case 0: srci = rms_norm->src[0]; break;
+            case 1: srci = tensor->src[1 - rms_norm_idx]; break;
+            default: continue;
+            }
+        }
         if (srci == nullptr) {
             continue;
         }
@@ -10876,7 +10927,12 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
     } else if (tensor->op == GGML_OP_SUB) {
         tensor_clone = ggml_sub(ggml_ctx, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_MUL) {
-        tensor_clone = ggml_mul(ggml_ctx, src_clone[0], src_clone[1]);
+        if (fused_rms_norm_mul) {
+            tensor_clone = ggml_rms_norm(ggml_ctx, src_clone[0], *(float *)tensor->src[rms_norm_idx]->op_params);
+            tensor_clone = ggml_mul(ggml_ctx, tensor_clone, src_clone[1 - rms_norm_idx]);
+        } else {
+            tensor_clone = ggml_mul(ggml_ctx, src_clone[0], src_clone[1]);
+        }
     } else if (tensor->op == GGML_OP_DIV) {
         tensor_clone = ggml_div(ggml_ctx, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_CONCAT) {
@@ -11067,10 +11123,10 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         GGML_ABORT("fatal error");
     }
 
-    ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
-    ggml_build_forward_expand(cgraph, tensor_clone);
+    ggml_cgraph * cgraph_cpu = ggml_new_graph(ggml_ctx);
+    ggml_build_forward_expand(cgraph_cpu, tensor_clone);
 
-    ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 8);
+    ggml_graph_compute_with_ctx(ggml_ctx, cgraph_cpu, 8);
 
     if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
         ggml_vk_print_tensor(tensor_clone, "tensor_clone");
@@ -11093,10 +11149,19 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
     VK_LOG_DEBUG("END ggml_vk_check_results_0(" << tensor->name << ")");
 }
 
-static void ggml_vk_check_results_1(ggml_tensor * tensor) {
+static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
+    ggml_tensor * tensor = cgraph->nodes[tensor_idx];
     if (tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
+    bool fused_rms_norm_mul = false;
+    if (ctx->num_additional_fused_ops == 1 &&
+        tensor->op == GGML_OP_RMS_NORM &&
+        cgraph->nodes[tensor_idx + 1]->op == GGML_OP_MUL) {
+        fused_rms_norm_mul = true;
+        tensor = cgraph->nodes[tensor_idx + 1];
+    }
+
     if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
         return;
     }

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp

@@ -101,8 +101,8 @@ void main() {
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
     uint32_t m_offset = 0;
-    if (p.nem2 != 1) {
-        m_offset = (iq3 % p.nem2) * p.nem1 * KV;
+    if (p.nem2 != 1 || p.nem3 != 1) {
+        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
     }
 
     [[dont_unroll]]
@@ -149,7 +149,7 @@ 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;

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp

@@ -25,6 +25,7 @@ layout (push_constant) uniform parameter {
     uint32_t nev3;
     uint32_t nem1;
     uint32_t nem2;
+    uint32_t nem3;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -40,8 +41,7 @@ layout (push_constant) uniform parameter {
     float max_bias;
     float logit_softcap;
 
-    uint32_t mask;
-    uint32_t n_head_log2;
+    uint32_t mask_n_head_log2;
     float m0;
     float m1;
 
@@ -50,6 +50,9 @@ layout (push_constant) uniform parameter {
     uint32_t k_num;
 } p;
 
+#define MASK_ENABLE_BIT (1<<16)
+#define N_LOG2_MASK 0xFFFF
+
 layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
 
 #if defined(A_TYPE_PACKED16)
@@ -100,8 +103,10 @@ 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)));
 }

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -126,8 +126,8 @@ void main() {
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
     uint32_t m_offset = 0;
-    if (p.nem2 != 1) {
-        m_offset = (iq3 % p.nem2) * p.nem1 * KV;
+    if (p.nem2 != 1 || p.nem3 != 1) {
+        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
     }
 
     [[dont_unroll]]
@@ -182,7 +182,7 @@ 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;