fix: chunk MiniMax-H3 QKV projections within backend buffer limits (#2286)

src/core/ggml_runner.cpp

@@ -211,6 +211,22 @@ ggml_backend_t GGMLRunner::backend_for_weight(const ggml_tensor* tensor) const {
     return nullptr;
 }
 
+static bool backend_can_allocate_op(ggml_backend_t backend, const ggml_tensor* node) {
+    auto buft        = ggml_backend_get_default_buffer_type(backend);
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+    size_t max_size  = ggml_backend_buft_get_max_size(buft);
+    max_size -= max_size % alignment;
+    if (ggml_backend_buft_get_alloc_size(buft, node) > max_size) {
+        return false;
+    }
+    for (auto src : node->src) {
+        if (src != nullptr && ggml_backend_buft_get_alloc_size(buft, src) > max_size) {
+            return false;
+        }
+    }
+    return true;
+}
+
 void GGMLRunner::pin_multi_device_nodes(ggml_backend_sched_t sched, ggml_cgraph* gf, ggml_cgraph* original_graph) {
     if (sched == nullptr || gf == nullptr) {
         return;
@@ -235,7 +251,14 @@ void GGMLRunner::pin_multi_device_nodes(ggml_backend_sched_t sched, ggml_cgraph*
             node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE) {
             continue;
         }
-        if (ggml_backend_supports_op(current, node)) {
+        // Operation support can allow tensors larger than the backend's allocation limit.
+        // Check inputs too, since cross-backend copies must also fit a single allocation.
+        if (!backend_can_allocate_op(current, node)) {
+            auto cpu = workspace_.cpu_backend();
+            if (cpu != nullptr && ggml_backend_supports_op(cpu, node)) {
+                ggml_backend_sched_set_tensor_backend(sched, node, cpu);
+            }
+        } else if (ggml_backend_supports_op(current, node)) {
             ggml_backend_sched_set_tensor_backend(sched, node, current);
         }
     }

src/core/vulkan/int8.cpp

@@ -715,10 +715,10 @@ namespace sd_vulkan_i8 {
         const Parameters pc = {n, rows, k, op.bias ? 1u : 0u};
         auto pipeline       = op.kind == SDW4Kind::W4A4 ? SD_VK_PIPELINE(mul_mat_w4a4_convrot, 6, Parameters) : SD_VK_PIPELINE(mul_mat_w4a8_convrot, 8, Parameters);
         uint32_t tile_cols = 64, tile_rows = 32;
-        const bool large_w4a4  = op.kind == SDW4Kind::W4A4 && rows >= 512 && d.compute.shared_memory >= w4a4_large_shared;
-        const bool large_w4a8  = op.kind == SDW4Kind::W4A8 && rows >= 512 && d.compute.shared_memory >= w4a8_large_shared;
-        const bool decode_w4a4 = op.kind == SDW4Kind::W4A4 && d.w4a4_cm2 && k % 256 == 0 && n % 64 == 0 && rows >= 128 &&
-                                 aligned(op.weight, 16) && sizeof(sd_vulkan_shaders::mul_mat_w4a4_convrot_decode_cm2) > 4;
+        const bool large_w4a4      = op.kind == SDW4Kind::W4A4 && rows >= 512 && d.compute.shared_memory >= w4a4_large_shared;
+        const bool large_w4a8      = op.kind == SDW4Kind::W4A8 && rows >= 512 && d.compute.shared_memory >= w4a8_large_shared;
+        const bool decode_w4a4     = op.kind == SDW4Kind::W4A4 && d.w4a4_cm2 && k % 256 == 0 && n % 64 == 0 && rows >= 128 &&
+                                     aligned(op.weight, 16) && sizeof(sd_vulkan_shaders::mul_mat_w4a4_convrot_decode_cm2) > 4;
         const uint32_t constants[] = {64, large_w4a4 || large_w4a8 ? 128u : 64u, 64};
         const char* force_dp4a     = std::getenv("GGML_VK_W4_FORCE_DP4A");
         if (!(force_dp4a && std::atoi(force_dp4a)) && aligned(op.input, 16) &&

src/model/common/ggml_block.hpp

@@ -974,7 +974,8 @@ public:
             w = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, w, prefix + "weight");
         }
         x = ggml_rms_norm(ctx->ggml_ctx, x, eps);
-        x = ggml_mul_inplace(ctx->ggml_ctx, x, w);
+        // The affine multiply may run on a different backend after normalization falls back.
+        x = ggml_mul(ctx->ggml_ctx, x, w);
         return x;
     }
 };

src/model/diffusion/minimax_h3.hpp

@@ -4,6 +4,7 @@
 #include <algorithm>
 #include <cinttypes>
 #include <cmath>
+#include <functional>
 #include <set>
 #include <string>
 #include <tuple>
@@ -243,8 +244,68 @@ namespace MiniMaxH3 {
             blocks["out_proj"] = std::make_shared<Linear>(inner, hidden_size, false);
         }
 
-        ggml_tensor* project_qkv(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            return std::dynamic_pointer_cast<Linear>(blocks["qkv_proj"])->forward(ctx, x);
+        std::vector<ggml_tensor*> split_qkv(GGMLRunnerContext* ctx, ggml_tensor* projected) {
+            std::vector<ggml_tensor*> result;
+            for (int i = 0; i < 3; ++i) {
+                result.push_back(ggml_view_4d(ctx->ggml_ctx, projected, head_dim, heads,
+                                              projected->ne[1], projected->ne[2] * projected->ne[3],
+                                              head_dim * projected->nb[0], projected->nb[1], projected->nb[2],
+                                              i * heads * head_dim * projected->nb[0]));
+            }
+            return result;
+        }
+
+        std::vector<ggml_tensor*> project_qkv(GGMLRunnerContext* ctx,
+                                              const std::vector<ggml_tensor*>& inputs,
+                                              const std::function<ggml_tensor*(ggml_tensor*, size_t)>& prepare_input = {}) {
+            auto projection  = std::dynamic_pointer_cast<Linear>(blocks["qkv_proj"]);
+            int64_t sequence = 0;
+            for (auto input : inputs) {
+                sequence += input->ne[1];
+            }
+            int64_t chunk_size = sequence;
+            if (ctx->backend != nullptr) {
+                auto buft        = ggml_backend_get_default_buffer_type(ctx->backend);
+                size_t alignment = ggml_backend_buft_get_alignment(buft);
+                size_t max_size  = ggml_backend_buft_get_max_size(buft);
+                max_size -= max_size % alignment;
+                size_t token_bytes = ggml_row_size(GGML_TYPE_F32, 3 * heads * head_dim) * inputs[0]->ne[2] * inputs[0]->ne[3];
+                chunk_size         = static_cast<int64_t>(std::min<size_t>(chunk_size, std::max<size_t>(1, max_size / token_bytes)));
+            }
+            if (chunk_size == sequence) {
+                ggml_tensor* projected = nullptr;
+                for (size_t i = 0; i < inputs.size(); ++i) {
+                    auto input = prepare_input ? prepare_input(inputs[i], i) : inputs[i];
+                    auto part  = projection->forward(ctx, input);
+                    projected  = projected == nullptr ? part : ggml_concat(ctx->ggml_ctx, projected, part, 1);
+                }
+                return split_qkv(ctx, projected);
+            }
+
+            // Split Q/K/V before concatenating tokens so the full packed QKV never exceeds the buffer limit.
+            std::vector<ggml_tensor*> chunks[3];
+            for (size_t i = 0; i < inputs.size(); ++i) {
+                auto input = inputs[i];
+                for (int64_t start = 0; start < input->ne[1]; start += chunk_size) {
+                    auto part = ggml_ext_slice(ctx->ggml_ctx, input, 1, start, std::min(start + chunk_size, input->ne[1]), false);
+                    if (!ggml_is_contiguous(part)) {
+                        part = ggml_cont(ctx->ggml_ctx, part);
+                    }
+                    // Keep normalization/modulation adjacent to the projection's input quantization.
+                    if (prepare_input) {
+                        part = prepare_input(part, i);
+                    }
+                    auto qkv = split_qkv(ctx, projection->forward(ctx, part));
+                    for (int j = 0; j < 3; ++j) {
+                        chunks[j].push_back(ggml_cont(ctx->ggml_ctx, qkv[j]));
+                    }
+                }
+            }
+            std::vector<ggml_tensor*> result;
+            for (auto& parts : chunks) {
+                result.push_back(ggml_ext_vec_concat(ctx->ggml_ctx, parts, 2));
+            }
+            return result;
         }
 
         ggml_tensor* project_output(GGMLRunnerContext* ctx, ggml_tensor* x) {
@@ -252,22 +313,15 @@ namespace MiniMaxH3 {
         }
 
         ggml_tensor* attend(GGMLRunnerContext* ctx,
-                            ggml_tensor* projected,
+                            const std::vector<ggml_tensor*>& qkv,
                             ggml_tensor* pe = nullptr) {
-            auto q_norm      = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
-            auto k_norm      = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
-            int64_t sequence = projected->ne[1];
-            int64_t batch    = projected->ne[2] * projected->ne[3];
-            auto view        = [&](int index) {
-                return ggml_view_4d(ctx->ggml_ctx, projected, head_dim, heads, sequence, batch,
-                                    head_dim * projected->nb[0], projected->nb[1], projected->nb[2],
-                                    index * heads * head_dim * projected->nb[0]);
-            };
-            auto q = view(0);
-            auto k = view(1);
-            auto v = view(2);
-            q      = q_norm->forward(ctx, q);
-            k      = k_norm->forward(ctx, k);
+            auto q_norm = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
+            auto k_norm = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
+            auto q      = qkv[0];
+            auto k      = qkv[1];
+            auto v      = qkv[2];
+            q           = q_norm->forward(ctx, q);
+            k           = k_norm->forward(ctx, k);
             if (pe != nullptr) {
                 q = apply_partial_rope(ctx, q, pe);
                 k = apply_partial_rope(ctx, k, pe);
@@ -290,7 +344,7 @@ namespace MiniMaxH3 {
         ggml_tensor* forward(GGMLRunnerContext* ctx,
                              ggml_tensor* x,
                              ggml_tensor* pe = nullptr) {
-            return project_output(ctx, attend(ctx, project_qkv(ctx, x), pe));
+            return project_output(ctx, attend(ctx, project_qkv(ctx, {x}), pe));
         }
     };
 
@@ -447,16 +501,9 @@ namespace MiniMaxH3 {
             auto modulate = [&](ggml_tensor* value, const std::vector<ggml_tensor*>& row, int shift, int scale) {
                 return ggml_ext_modulate(ctx, value, row[shift], row[scale]);
             };
-            ggml_tensor* qkv = nullptr;
-            for (size_t i = 0; i < segments.size(); ++i) {
-                auto h = modulate(norm1->forward(ctx, inputs[i]), rows[i], 0, 1);
-                h      = attn->project_qkv(ctx, h);
-                if (qkv == nullptr) {
-                    qkv = h;
-                    continue;
-                }
-                qkv = ggml_concat(ctx->ggml_ctx, qkv, h, 1);
-            }
+            auto qkv      = attn->project_qkv(ctx, inputs, [&](ggml_tensor* input, size_t i) {
+                return modulate(norm1->forward(ctx, input), rows[i], 0, 1);
+            });
             auto attended = attn->attend(ctx, qkv, pe);
             std::vector<ggml_tensor*> projections;
             for (const auto& segment : segments) {