Merge branch 'upstream' into concedo_experimental

# Conflicts: # .devops/openvino.Dockerfile # .github/workflows/build-cache.yml # .github/workflows/build-openvino.yml # .github/workflows/build-self-hosted.yml # .github/workflows/release.yml # docs/backend/OPENVINO.md # docs/backend/SYCL.md # docs/ops.md # docs/ops/SYCL.csv # ggml/src/ggml-opencl/ggml-opencl.cpp # ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl # ggml/src/ggml-openvino/.clang-format # ggml/src/ggml-openvino/CMakeLists.txt # ggml/src/ggml-openvino/ggml-decoder.cpp # ggml/src/ggml-openvino/ggml-decoder.h # ggml/src/ggml-openvino/ggml-openvino-extra.cpp # ggml/src/ggml-openvino/ggml-openvino-extra.h # ggml/src/ggml-openvino/ggml-openvino.cpp # ggml/src/ggml-openvino/ggml-quants.cpp # ggml/src/ggml-openvino/ggml-quants.h # ggml/src/ggml-openvino/openvino/decoder.h # ggml/src/ggml-openvino/openvino/frontend.h # ggml/src/ggml-openvino/openvino/input_model.h # ggml/src/ggml-openvino/openvino/node_context.h # ggml/src/ggml-openvino/openvino/op/cont.cpp # ggml/src/ggml-openvino/openvino/op/cpy.cpp # ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp # ggml/src/ggml-openvino/openvino/op/get_rows.cpp # ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp # ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp # ggml/src/ggml-openvino/openvino/op/mulmat.cpp # ggml/src/ggml-openvino/openvino/op/permute.cpp # ggml/src/ggml-openvino/openvino/op/reshape.cpp # ggml/src/ggml-openvino/openvino/op/rms_norm.cpp # ggml/src/ggml-openvino/openvino/op/rope.cpp # ggml/src/ggml-openvino/openvino/op/set_rows.cpp # ggml/src/ggml-openvino/openvino/op/softmax.cpp # ggml/src/ggml-openvino/openvino/op/transpose.cpp # ggml/src/ggml-openvino/openvino/op/unary_silu.cpp # ggml/src/ggml-openvino/openvino/op/view.cpp # ggml/src/ggml-openvino/openvino/op_table.cpp # ggml/src/ggml-openvino/openvino/op_table.h # ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h # ggml/src/ggml-openvino/openvino/translate_session.cpp # ggml/src/ggml-openvino/openvino/translate_session.h # ggml/src/ggml-openvino/openvino/utils.cpp # ggml/src/ggml-openvino/openvino/utils.h # ggml/src/ggml-openvino/utils.cpp # ggml/src/ggml-openvino/utils.h # ggml/src/ggml-sycl/common.hpp # ggml/src/ggml-sycl/ggml-sycl.cpp
2026-09-06 21:11:24 +02:00 · 2026-06-17 16:06:00 +08:00
parent 94653a9be4 d5376cf5d7
commit b8b7763c76
17 changed files with 424 additions and 151 deletions
@@ -11,8 +11,13 @@
 #include <sstream>
 #include <thread>
 #include <vector>
+#include <algorithm>

 #if defined(_WIN32)
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#       define NOMINMAX
+#    endif
 #    include <io.h>
 #    include <windows.h>
 #    define isatty _isatty
@@ -62,16 +67,15 @@ static const char* g_col[] = {
 };

 struct common_log_entry {
-    enum ggml_log_level level;
-
-    bool prefix;
-
-    int64_t timestamp;
+    enum ggml_log_level level {GGML_LOG_LEVEL_INFO};

    std::vector<char> msg;

-    // signals the worker thread to stop
-    bool is_end;
+    int64_t timestamp { 0 };
+    bool is_end       { false }; // signals the worker thread to stop
+    bool prefix       { false };
+
+    common_log_entry(size_t size = 256) : msg(size) { }

    void print(FILE * file = nullptr) const {
        FILE * fcur = file;
@@ -122,22 +126,15 @@ struct common_log_entry {
 };

 struct common_log {
-    // default capacity - will be expanded if needed
-    common_log() : common_log(256) {}
-
-    common_log(size_t capacity) {
-        file = nullptr;
-        prefix = false;
+    // default capacity
+    common_log(size_t capacity = 512) {
+        file       = nullptr;
+        prefix     = false;
        timestamps = false;
-        running = false;
-        t_start = t_us();
-
-        // initial message size - will be expanded if longer messages arrive
-        entries.resize(capacity);
-        for (auto & entry : entries) {
-            entry.msg.resize(256);
-        }
+        running    = false;
+        t_start    = t_us();

+        queue.resize(capacity, common_log_entry(256));
        head = 0;
        tail = 0;

@@ -152,9 +149,10 @@ struct common_log {
    }

 private:
-    std::mutex mtx;
-    std::thread thrd;
-    std::condition_variable cv;
+    std::mutex              mtx;
+    std::thread             thrd;
+    std::condition_variable cv_new;  // new entry
+    std::condition_variable cv_full; // wait on full

    FILE * file;

@@ -164,24 +162,53 @@ private:

    int64_t t_start;

-    // ring buffer of entries
-    std::vector<common_log_entry> entries;
+    // queue of entries
+    std::vector<common_log_entry> queue;
    size_t head;
    size_t tail;

-    // worker thread copies into this
-    common_log_entry cur;
+    bool print_entry(const common_log_entry & e) const {
+        if (e.is_end) return true;
+
+        e.print();
+        if (file) {
+            e.print(file);
+        }
+        return false;
+    }
+
+    bool flush_queue(size_t start_head, size_t end_tail, size_t & out_head) const {
+        bool stop = false;
+        size_t h = start_head;
+        while (h != end_tail && !stop) {
+            stop = print_entry(queue[h]);
+            h = (h + 1) % queue.size();
+        }
+        out_head = h;
+        return stop;
+    }

 public:
+    bool is_full() const {
+        return ((tail + 1) % queue.size()) == head;
+    }
+
+    bool is_empty() const {
+        return head == tail;
+    }
+
    void add(enum ggml_log_level level, const char * fmt, va_list args) {
-        std::lock_guard<std::mutex> lock(mtx);
+        std::unique_lock<std::mutex> lock(mtx);
+
+        // block if the queue is full
+        cv_full.wait(lock, [this]() { return !running || !is_full(); });

        if (!running) {
            // discard messages while the worker thread is paused
            return;
        }

-        auto & entry = entries[tail];
+        auto & entry = queue[tail];

        {
            // cannot use args twice, so make a copy in case we need to expand the buffer
@@ -216,38 +243,16 @@ public:
            va_end(args_copy);
        }

-        entry.level = level;
-        entry.prefix = prefix;
+        entry.is_end    = false;
+        entry.level     = level;
+        entry.prefix    = prefix;
        entry.timestamp = 0;
        if (timestamps) {
            entry.timestamp = t_us() - t_start;
        }
-        entry.is_end = false;

-        tail = (tail + 1) % entries.size();
-        if (tail == head) {
-            // expand the buffer
-            std::vector<common_log_entry> new_entries(2*entries.size());
-
-            size_t new_tail = 0;
-
-            do {
-                new_entries[new_tail] = std::move(entries[head]);
-
-                head     = (head     + 1) % entries.size();
-                new_tail = (new_tail + 1);
-            } while (head != tail);
-
-            head = 0;
-            tail = new_tail;
-
-            for (size_t i = tail; i < new_entries.size(); i++) {
-                new_entries[i].msg.resize(256);
-            }
-
-            entries = std::move(new_entries);
-        }
-        cv.notify_one();
+        tail = (tail + 1) % queue.size();
+        cv_new.notify_one();
    }

    void resume() {
@@ -261,23 +266,24 @@ public:

        thrd = std::thread([this]() {
            while (true) {
-                {
-                    std::unique_lock<std::mutex> lock(mtx);
-                    cv.wait(lock, [this]() { return head != tail; });
-                    cur = entries[head];
+                std::unique_lock<std::mutex> lock(mtx);
+                cv_new.wait(lock, [this]() { return !is_empty(); });

-                    head = (head + 1) % entries.size();
-                }
+                size_t cached_head = head;
+                size_t cached_tail = tail;

-                if (cur.is_end) {
+                lock.unlock(); // drop the lock during flush
+
+                size_t next_head;
+                bool stop = flush_queue(cached_head, cached_tail, next_head);
+
+                lock.lock();
+                head = next_head;
+                cv_full.notify_all();
+
+                if (stop) {
                    break;
                }
-
-                cur.print(); // stdout and stderr
-
-                if (file) {
-                    cur.print(file);
-                }
            }
        });
    }
@@ -293,13 +299,13 @@ public:
            running = false;

            // push an entry to signal the worker thread to stop
-            {
-                auto & entry = entries[tail];
-                entry.is_end = true;
+            auto & entry = queue[tail];
+            entry.is_end = true;
+            tail = (tail + 1) % queue.size();

-                tail = (tail + 1) % entries.size();
-            }
-            cv.notify_one();
+            // wakeup everyone
+            cv_new.notify_one();
+            cv_full.notify_all();
        }

        thrd.join();
@@ -422,7 +422,7 @@ std::map<std::string, std::string> merge_maps(const std::map<std::string, std::s
    return result;
 }

-static std::vector<std::future<void>> compiles;
+static std::deque<std::future<void>> compiles;
 void string_to_spv(std::string name, const std::string& source, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false, const std::string& suffix = "") {
    name = name + (f16acc ? "_f16acc" : "") + (coopmat ? "_cm1" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32")) + suffix;
    std::string out_path = join_paths(output_dir, name + ".spv");
@@ -452,6 +452,11 @@ void string_to_spv(std::string name, const std::string& source, const std::map<s
        string_to_spv_func, name, in_path, out_path, defines, coopmat, generate_dep_file, std::move(slot)));
    // Don't write the same dep file from multiple processes
    generate_dep_file = false;
+
+    // Clean up completed futures - don't accumulate virtual memory for completed threads' stacks.
+    while (!compiles.empty() && compiles.front().wait_for(std::chrono::seconds(0)) == std::future_status::ready) {
+        compiles.pop_front();
+    }
 }

 void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool coopmat2, bool f16acc, bool dot2 = false) {
@@ -1088,6 +1088,10 @@ ggml_tensor * llm_graph_context::build_lora_mm(
          ggml_tensor * w_s) const {
    ggml_tensor * res = ggml_mul_mat(ctx0, w, cur);

+    if (w_s) {
+        res = ggml_mul(ctx0, res, w_s);
+    }
+
    for (const auto & lora : *loras) {
        llama_adapter_lora_weight * lw = lora.first->get_weight(w);
        if (lw == nullptr) {
@@ -1106,18 +1110,24 @@ ggml_tensor * llm_graph_context::build_lora_mm(
        res = ggml_add(ctx0, res, ab_cur);
    }

-    if (w_s) {
-        res = ggml_mul(ctx0, res, w_s);
-    }
-
    return res;
 }

 ggml_tensor * llm_graph_context::build_lora_mm_id(
          ggml_tensor * w,   // ggml_tensor * as
          ggml_tensor * cur, // ggml_tensor * b
-          ggml_tensor * ids) const {
+          ggml_tensor * ids,
+          ggml_tensor * w_s) const {
    ggml_tensor * res = ggml_mul_mat_id(ctx0, w, cur, ids);
+
+    if (w_s) {
+        const int64_t n_expert = w_s->ne[0];
+        const int64_t n_tokens = cur->ne[2];
+        ggml_tensor * s = ggml_reshape_3d(ctx0, w_s, 1, n_expert, 1);
+        s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
+        s = ggml_get_rows(ctx0, s, ids);
+        res = ggml_mul(ctx0, res, s);
+    }
    for (const auto & lora : *loras) {
        llama_adapter_lora_weight * lw = lora.first->get_weight(w);
        if (lw == nullptr) {
@@ -1269,6 +1279,29 @@ ggml_tensor * llm_graph_context::build_ffn(
     llm_ffn_op_type   type_op,
   llm_ffn_gate_type   type_gate,
                 int   il) const {
+    // NVFP4 support is currently restricted to
+    // 1) LORA absence (*_s would be applied after LORA residual, which is incorrect)
+    // 2) bias absense (*_s would be applied after bias addition, which is incorrect)
+    // TODO: disambiguate LLM-architectural scales (which use *_s) from NVFP4 scale_2 (which also uses *_s currently)
+    auto has_lora = [this](ggml_tensor * w) {
+        if (!w) {
+            return false;
+        }
+        for (const auto & lora : *loras) {
+            if (lora.first->get_weight(w) != nullptr) {
+                return true;
+            }
+        }
+        return false;
+    };
+
+    GGML_ASSERT(!up_s   || !up_b   || !up   || up->type   != GGML_TYPE_NVFP4);
+    GGML_ASSERT(!gate_s || !gate_b || !gate || gate->type != GGML_TYPE_NVFP4);
+    GGML_ASSERT(!down_s || !down_b || !down || down->type != GGML_TYPE_NVFP4);
+    GGML_ASSERT(!up_s   || !up   || up->type   != GGML_TYPE_NVFP4 || !has_lora(up));
+    GGML_ASSERT(!gate_s || !gate || gate->type != GGML_TYPE_NVFP4 || !has_lora(gate));
+    GGML_ASSERT(!down_s || !down || down->type != GGML_TYPE_NVFP4 || !has_lora(down));
+
    ggml_tensor * tmp = up ? build_lora_mm(up, cur) : cur;
    cb(tmp, "ffn_up", il);

@@ -1627,23 +1660,18 @@ ggml_tensor * llm_graph_context::build_moe_ffn(

    if (gate_up_exps) {
        // merged gate_up path: one mul_mat_id, then split into gate and up views
-        ggml_tensor * gate_up = build_lora_mm_id(gate_up_exps, cur, selected_experts); // [n_ff*2, n_expert_used, n_tokens]
+        ggml_tensor * gate_up = build_lora_mm_id(gate_up_exps, cur, selected_experts, up_exps_s); // [n_ff*2, n_expert_used, n_tokens]
        cb(gate_up, "ffn_moe_gate_up", il);

+        if (up_exps_s) {
+            cb(gate_up, "ffn_moe_gate_up_scaled", il);
+        }
+
        if (gate_up_exps_b) {
            gate_up = ggml_add_id(ctx0, gate_up, gate_up_exps_b, selected_experts);
            cb(gate_up, "ffn_moe_gate_up_biased", il);
        }

-        // apply per-expert scale2 to merged gate_up (use up_exps_s since gate and up are fused)
-        if (up_exps_s) {
-            ggml_tensor * s = ggml_reshape_3d(ctx0, up_exps_s, 1, n_expert, 1);
-            s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
-            s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
-            gate_up = ggml_mul(ctx0, gate_up, s);
-            cb(gate_up, "ffn_moe_gate_up_scaled", il);
-        }
-
        const int64_t n_ff = gate_up->ne[0] / 2;
        cur = ggml_view_3d(ctx0, gate_up, n_ff, gate_up->ne[1], gate_up->ne[2], gate_up->nb[1], gate_up->nb[2], 0);
        cb(cur, "ffn_moe_gate", il);
@@ -1651,43 +1679,33 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
        cb(up, "ffn_moe_up", il);
    } else {
        // separate gate and up path
-        up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+        up = build_lora_mm_id(up_exps, cur, selected_experts, up_exps_s); // [n_ff, n_expert_used, n_tokens]
        cb(up, "ffn_moe_up", il);

+        if (up_exps_s) {
+            cb(up, "ffn_moe_up_scaled", il);
+        }
+
        if (up_exps_b) {
            up = ggml_add_id(ctx0, up, up_exps_b, selected_experts);
            cb(up, "ffn_moe_up_biased", il);
        }

-        // apply per-expert scale2 to up
-        if (up_exps_s) {
-            ggml_tensor * s = ggml_reshape_3d(ctx0, up_exps_s, 1, n_expert, 1);
-            s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
-            s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
-            up = ggml_mul(ctx0, up, s);
-            cb(up, "ffn_moe_up_scaled", il);
-        }
-
        if (gate_exps) {
-            cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+            cur = build_lora_mm_id(gate_exps, cur, selected_experts, gate_exps_s); // [n_ff, n_expert_used, n_tokens]
            cb(cur, "ffn_moe_gate", il);
        } else {
            cur = up;
        }

+        if (gate_exps_s) {
+            cb(cur, "ffn_moe_gate_scaled", il);
+        }
+
        if (gate_exps_b) {
            cur = ggml_add_id(ctx0, cur, gate_exps_b, selected_experts);
            cb(cur, "ffn_moe_gate_biased", il);
        }
-
-        // apply per-expert scale2 to gate
-        if (gate_exps_s) {
-            ggml_tensor * s = ggml_reshape_3d(ctx0, gate_exps_s, 1, n_expert, 1);
-            s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
-            s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
-            cur = ggml_mul(ctx0, cur, s);
-            cb(cur, "ffn_moe_gate_scaled", il);
-        }
    }

    const bool has_gate = gate_exps || gate_up_exps;
@@ -1759,23 +1777,18 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
            GGML_ABORT("fatal error");
    }

-    experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
+    experts = build_lora_mm_id(down_exps, cur, selected_experts, down_exps_s); // [n_embd, n_expert_used, n_tokens]
    cb(experts, "ffn_moe_down", il);

+    if (down_exps_s) {
+        cb(experts, "ffn_moe_down_scaled", il);
+    }
+
    if (down_exps_b) {
        experts = ggml_add_id(ctx0, experts, down_exps_b, selected_experts);
        cb(experts, "ffn_moe_down_biased", il);
    }

-    // apply per-expert scale2 to down
-    if (down_exps_s) {
-        ggml_tensor * s = ggml_reshape_3d(ctx0, down_exps_s, 1, n_expert, 1);
-        s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
-        s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
-        experts = ggml_mul(ctx0, experts, s);
-        cb(experts, "ffn_moe_down_scaled", il);
-    }
-
    if (!weight_before_ffn) {
        experts = ggml_mul(ctx0, experts, weights);
        cb(experts, "ffn_moe_weighted", il);
@@ -853,11 +853,12 @@ struct llm_graph_context {
              ggml_tensor * cur,
              ggml_tensor * w_s = nullptr) const;

-    // do mat_mul_id, while optionally apply lora
+    // do mat_mul_id, while optionally apply lora and per-expert scale
    ggml_tensor * build_lora_mm_id(
              ggml_tensor * w,   // ggml_tensor * as
              ggml_tensor * cur, // ggml_tensor * b
-              ggml_tensor * ids) const;
+              ggml_tensor * ids,
+              ggml_tensor * w_s = nullptr) const;

    ggml_tensor * build_norm(
             ggml_tensor * cur,
@@ -41,6 +41,7 @@
 		DATA_ERROR_HANDLED_ATTR,
 		BOOL_TRUE_STRING,
 		SETTINGS_KEYS,
+		CODE_BLOCK_HEADER_CLASS,
 		MERMAID_WRAPPER_CLASS,
 		MERMAID_BLOCK_CLASS,
 		MERMAID_LANGUAGE,
@@ -53,7 +54,11 @@
 		SVG_TAG_PREFIX,
 		SVG_SOURCE_ATTR,
 		SVG_RENDERED_ATTR,
-		SVG_INLINE_SHADOW_STYLE
+		SVG_INLINE_SHADOW_STYLE,
+		TOGGLE_SOURCE_BTN_CLASS,
+		DIAGRAM_VIEW_MODE_ATTR,
+		DIAGRAM_VIEW_RENDERED,
+		DIAGRAM_VIEW_SOURCE
 	} from '$lib/constants';
 	import { ColorMode, UrlProtocol } from '$lib/enums';
 	import { FileTypeText } from '$lib/enums/files.enums';
@@ -501,6 +506,23 @@
 	async function handleMermaidClick(event: MouseEvent) {
 		const target = event.target as HTMLElement;

+		// Toggle a diagram block between its rendered view and its source view.
+		// Shared by mermaid and svg, css drives the visibility from the wrapper mode.
+		const toggleBtn = target.closest(`.${TOGGLE_SOURCE_BTN_CLASS}`);
+		if (toggleBtn) {
+			event.preventDefault();
+			event.stopPropagation();
+
+			const wrapper = toggleBtn.closest(`.${MERMAID_WRAPPER_CLASS}, .${SVG_WRAPPER_CLASS}`);
+			if (!wrapper) return;
+
+			const isSource = wrapper.getAttribute(DIAGRAM_VIEW_MODE_ATTR) === DIAGRAM_VIEW_SOURCE;
+			const next = isSource ? DIAGRAM_VIEW_RENDERED : DIAGRAM_VIEW_SOURCE;
+			wrapper.setAttribute(DIAGRAM_VIEW_MODE_ATTR, next);
+			toggleBtn.setAttribute('aria-pressed', String(!isSource));
+			return;
+		}
+
 		// Check if clicking on copy or preview button in mermaid block
 		const copyBtn = target.closest(`.${MERMAID_WRAPPER_CLASS} .copy-code-btn`);
 		const previewBtn = target.closest(`.${MERMAID_WRAPPER_CLASS} .preview-code-btn`);
@@ -573,6 +595,11 @@
 			}
 		}

+		// A click on the header chrome targets the action buttons, never the
+		// diagram. Guard so a header click can not fall through to the click to
+		// zoom branches below, whatever the scroll position or stacking.
+		if (target.closest(`.${CODE_BLOCK_HEADER_CLASS}`)) return;
+
 		// Open preview when clicking the svg block itself. A final block carries its
 		// source, a streaming block does not and is mirrored live into the dialog.
 		const svgEl = target.closest(`.${SVG_BLOCK_CLASS}`);
@@ -300,7 +300,8 @@ div.markdown-user-content :global(.table-wrapper) {
 }

 .markdown-content :global(.copy-code-btn),
-.markdown-content :global(.preview-code-btn) {
+.markdown-content :global(.preview-code-btn),
+.markdown-content :global(.toggle-source-btn) {
 	display: flex;
 	align-items: center;
 	justify-content: center;
@@ -312,15 +313,22 @@ div.markdown-user-content :global(.table-wrapper) {
 }

 .markdown-content :global(.copy-code-btn:hover),
-.markdown-content :global(.preview-code-btn:hover) {
+.markdown-content :global(.preview-code-btn:hover),
+.markdown-content :global(.toggle-source-btn:hover) {
 	transform: scale(1.05);
 }

 .markdown-content :global(.copy-code-btn:active),
-.markdown-content :global(.preview-code-btn:active) {
+.markdown-content :global(.preview-code-btn:active),
+.markdown-content :global(.toggle-source-btn:active) {
 	transform: scale(0.95);
 }

+/* Pressed state marks the source view as active */
+.markdown-content :global(.toggle-source-btn[aria-pressed='true']) {
+	color: var(--primary);
+}
+
 .markdown-content :global(.code-block-wrapper pre) {
 	background: transparent;
 	margin: 0;
@@ -629,8 +637,8 @@ div.markdown-user-content :global(.table-wrapper) {
 	overflow-y: auto;
 	overflow-x: auto;
 	display: flex;
-	align-items: center;
-	justify-content: center;
+	align-items: safe center;
+	justify-content: safe center;
 	padding: 3rem 1rem 1rem;
 }

@@ -645,7 +653,9 @@ div.markdown-user-content :global(.table-wrapper) {
 	overflow-y: visible;
 }

-/* Diagram block uses same header styling as code blocks */
+/* Diagram block uses same header styling as code blocks. The header floats over
+   scrollable diagram content and stays transparent, so the overflow shows up to
+   the box edge. It keeps a z-index so it stays the click target above content. */
 .markdown-content :global(.mermaid-block-wrapper .code-block-header),
 .markdown-content :global(.svg-block-wrapper .code-block-header) {
 	display: flex;
@@ -657,6 +667,7 @@ div.markdown-user-content :global(.table-wrapper) {
 	top: 0;
 	left: 0;
 	right: 0;
+	z-index: 2;
 }

 .markdown-content :global(.mermaid-block-wrapper .code-block-actions),
@@ -683,6 +694,31 @@ div.markdown-user-content :global(.table-wrapper) {
 	padding: 3rem 1rem;
 }

+/* Source view stays hidden while the block renders, css swaps the two views
+   from the wrapper mode so the click handler only flips one attribute. The view
+   reuses the code block scroll container, so it matches the app code blocks. */
+.markdown-content :global(.diagram-source) {
+	display: none;
+	text-align: left;
+}
+
+.markdown-content :global(.diagram-source pre) {
+	background: transparent;
+	margin: 0;
+	border-radius: 0;
+	border: none;
+	font-size: 0.875rem;
+}
+
+.markdown-content :global([data-view-mode='source'] .mermaid-scroll-container),
+.markdown-content :global([data-view-mode='source'] .svg-scroll-container) {
+	display: none;
+}
+
+.markdown-content :global([data-view-mode='source'] .diagram-source) {
+	display: block;
+}
+
 /* Streaming mermaid block - empty preview box */
 .mermaid-streaming-block {
 	min-height: 300px;
@@ -7,12 +7,16 @@ import type { Element, ElementContent } from 'hast';
 import {
 	CODE_BLOCK_HEADER_CLASS,
 	CODE_BLOCK_ACTIONS_CLASS,
+	CODE_BLOCK_SCROLL_CONTAINER_CLASS,
 	CODE_LANGUAGE_CLASS,
 	COPY_CODE_BTN_CLASS,
 	PREVIEW_CODE_BTN_CLASS,
+	TOGGLE_SOURCE_BTN_CLASS,
+	DIAGRAM_SOURCE_CLASS,
 	RELATIVE_CLASS,
 	COPY_ICON_SVG,
-	PREVIEW_ICON_SVG
+	PREVIEW_ICON_SVG,
+	CODE_ICON_SVG
 } from '$lib/constants';

 export interface BlockIdGenerator {
@@ -32,14 +36,16 @@ export function createIconElement(svg: string): Element {
 }

 /**
- * Creates a button element with icon.
+ * Creates a button element with icon. Extra properties merge onto the button,
+ * which lets a stateful button carry attributes like aria-pressed.
 */
 export function createButton(
 	className: string,
 	title: string,
 	iconSvg: string,
 	id: string,
-	idAttribute: string
+	idAttribute: string,
+	extraProperties: Record<string, string> = {}
 ): Element {
 	return {
 		type: 'element',
@@ -48,7 +54,8 @@ export function createButton(
 			className: [className],
 			[idAttribute]: id,
 			title,
-			type: 'button'
+			type: 'button',
+			...extraProperties
 		},
 		children: [createIconElement(iconSvg)]
 	};
@@ -72,6 +79,52 @@ export function createPreviewButton(
 	return createButton(PREVIEW_CODE_BTN_CLASS, title, PREVIEW_ICON_SVG, id, idAttribute);
 }

+/**
+ * Creates a button that toggles a diagram block between its rendered view and
+ * its source view. aria-pressed starts false, the rendered view is the default.
+ */
+export function createToggleSourceButton(
+	id: string,
+	idAttribute: string,
+	title: string = 'Toggle source'
+): Element {
+	return createButton(TOGGLE_SOURCE_BTN_CLASS, title, CODE_ICON_SVG, id, idAttribute, {
+		'aria-pressed': 'false'
+	});
+}
+
+/**
+ * Creates a source view for a diagram block. It reuses the code block scroll
+ * container so it matches the app code blocks, and wraps the highlighted code
+ * element captured at transform time. A missing code element falls back to a
+ * plain code node built from the raw source.
+ */
+export function createSourceView(
+	codeElement: Element | undefined,
+	source: string,
+	language: string
+): Element {
+	const code: Element = codeElement ?? {
+		type: 'element',
+		tagName: 'code',
+		properties: { className: ['hljs', `language-${language}`] },
+		children: [{ type: 'text', value: source }]
+	};
+	return {
+		type: 'element',
+		tagName: 'div',
+		properties: { className: [DIAGRAM_SOURCE_CLASS, CODE_BLOCK_SCROLL_CONTAINER_CLASS] },
+		children: [
+			{
+				type: 'element',
+				tagName: 'pre',
+				properties: {},
+				children: [code]
+			}
+		]
+	};
+}
+
 /**
 * Creates a block header with language label and action buttons.
 */
@@ -116,14 +169,17 @@ export function createScrollContainer(preElement: Element, scrollContainerClass:
 }

 /**
- * Creates a wrapper element with header and scroll container.
+ * Creates a wrapper element with header and scroll container. Extra children
+ * append after the scroll container, which lets a block carry a source view
+ * alongside its rendered output.
 */
 export function createWrapper(
 	header: Element,
 	preElement: Element,
 	wrapperClass: string,
 	scrollContainerClass: string,
-	additionalAttributes?: Record<string, string>
+	additionalAttributes?: Record<string, string>,
+	extraChildren: Element[] = []
 ): Element {
 	return {
 		type: 'element',
@@ -132,7 +188,7 @@ export function createWrapper(
 			className: [wrapperClass, RELATIVE_CLASS],
 			...additionalAttributes
 		} as Element['properties'],
-		children: [header, createScrollContainer(preElement, scrollContainerClass)]
+		children: [header, createScrollContainer(preElement, scrollContainerClass), ...extraChildren]
 	};
 }

@@ -19,12 +19,17 @@ import {
 	MERMAID_BLOCK_CLASS,
 	MERMAID_LANGUAGE,
 	MERMAID_SYNTAX_ATTR,
-	MERMAID_ID_ATTR
+	MERMAID_ID_ATTR,
+	DIAGRAM_VIEW_MODE_ATTR,
+	DIAGRAM_VIEW_RENDERED
 } from '$lib/constants';
+import type { DiagramPreData } from './pre-transform';
 import {
 	createBlockHeader,
 	createCopyButton,
 	createPreviewButton,
+	createToggleSourceButton,
+	createSourceView,
 	createWrapper,
 	generateBlockId
 } from './code-block-utils';
@@ -75,16 +80,23 @@ export const rehypeEnhanceMermaidBlocks: Plugin<[], Root> = () => {

 			const actions = [
 				createCopyButton(mermaidId, MERMAID_ID_ATTR, 'Copy mermaid syntax'),
+				createToggleSourceButton(mermaidId, MERMAID_ID_ATTR, 'Toggle mermaid source'),
 				createPreviewButton(mermaidId, MERMAID_ID_ATTR, 'Preview diagram')
 			];

 			const header = createBlockHeader(MERMAID_LANGUAGE, mermaidId, MERMAID_ID_ATTR, actions);
+			const preservedCode = (node.data as DiagramPreData | undefined)?.sourceCode;
+			const sourceView = createSourceView(preservedCode, diagramText, MERMAID_LANGUAGE);
 			const wrapper = createWrapper(
 				header,
 				node,
 				MERMAID_WRAPPER_CLASS,
 				MERMAID_SCROLL_CONTAINER_CLASS,
-				{ [MERMAID_ID_ATTR]: mermaidId }
+				{
+					[MERMAID_ID_ATTR]: mermaidId,
+					[DIAGRAM_VIEW_MODE_ATTR]: DIAGRAM_VIEW_RENDERED
+				},
+				[sourceView]
 			);

 			// Replace pre with wrapper in parent
@@ -18,12 +18,17 @@ import {
 	SVG_BLOCK_CLASS,
 	SVG_LANGUAGE,
 	SVG_SOURCE_ATTR,
-	SVG_ID_ATTR
+	SVG_ID_ATTR,
+	DIAGRAM_VIEW_MODE_ATTR,
+	DIAGRAM_VIEW_RENDERED
 } from '$lib/constants';
+import type { DiagramPreData } from './pre-transform';
 import {
 	createBlockHeader,
 	createCopyButton,
 	createPreviewButton,
+	createToggleSourceButton,
+	createSourceView,
 	createWrapper,
 	generateBlockId
 } from './code-block-utils';
@@ -65,13 +70,24 @@ export const rehypeEnhanceSvgBlocks: Plugin<[], Root> = () => {

 			const actions = [
 				createCopyButton(svgId, SVG_ID_ATTR, 'Copy svg source'),
+				createToggleSourceButton(svgId, SVG_ID_ATTR, 'Toggle svg source'),
 				createPreviewButton(svgId, SVG_ID_ATTR, 'Preview svg')
 			];

 			const header = createBlockHeader(SVG_LANGUAGE, svgId, SVG_ID_ATTR, actions);
-			const wrapper = createWrapper(header, node, SVG_WRAPPER_CLASS, SVG_SCROLL_CONTAINER_CLASS, {
-				[SVG_ID_ATTR]: svgId
-			});
+			const preservedCode = (node.data as DiagramPreData | undefined)?.sourceCode;
+			const sourceView = createSourceView(preservedCode, svgSource, SVG_LANGUAGE);
+			const wrapper = createWrapper(
+				header,
+				node,
+				SVG_WRAPPER_CLASS,
+				SVG_SCROLL_CONTAINER_CLASS,
+				{
+					[SVG_ID_ATTR]: svgId,
+					[DIAGRAM_VIEW_MODE_ATTR]: DIAGRAM_VIEW_RENDERED
+				},
+				[sourceView]
+			);

 			// Replace pre with wrapper in parent
 			(parent.children as ElementContent[])[index] = wrapper;
@@ -2,6 +2,15 @@ import type { Plugin } from 'unified';
 import type { Root, Element, ElementContent, Text } from 'hast';
 import { visit } from 'unist-util-visit';

+/**
+ * Metadata a diagram pre carries on its unist data field. The source code holds
+ * the highlighted code element captured before the pre became a render target,
+ * which the enhancer reuses to build a matching source view.
+ */
+export interface DiagramPreData {
+	sourceCode: Element;
+}
+
 /**
 * Recursively extracts all text content from a HAST node.
 * Handles nested elements (e.g., span wrappers from syntax highlighting).
@@ -69,7 +78,10 @@ export function createPreTransform(
 					properties: {
 						className: [targetClass]
 					},
-					children: [{ type: 'text', value: text } as Text]
+					children: [{ type: 'text', value: text } as Text],
+					// Keep the highlighted code element so the block can offer a source
+					// view that matches the app code blocks without re highlighting.
+					data: { sourceCode: codeElement } satisfies DiagramPreData
 				};

 				(parent.children as ElementContent[])[index] = pre;
@@ -0,0 +1,9 @@
+// Shared constants for diagram blocks (mermaid and svg) that toggle between a
+// rendered view and a source view. The wrapper carries the active mode, css
+// drives the visibility, the click handler only flips the attribute.
+
+export const DIAGRAM_VIEW_MODE_ATTR = 'data-view-mode';
+export const DIAGRAM_VIEW_RENDERED = 'rendered';
+export const DIAGRAM_VIEW_SOURCE = 'source';
+export const DIAGRAM_SOURCE_CLASS = 'diagram-source';
+export const TOGGLE_SOURCE_BTN_CLASS = 'toggle-source-btn';
@@ -39,3 +39,5 @@ export const MODALITY_LABELS = {
 export const COPY_ICON_SVG = `<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-copy-icon lucide-copy"><rect width="14" height="14" x="8" y="8" rx="2" ry="2"/><path d="M4 16c-1.1 0-2-.9-2-2V4c0-1.1.9-2 2-2h10c1.1 0 2 .9 2 2"/></svg>`;

 export const PREVIEW_ICON_SVG = `<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-eye lucide-eye-icon"><path d="M2.062 12.345a1 1 0 0 1 0-.69C3.5 7.73 7.36 5 12 5s8.5 2.73 9.938 6.655a1 1 0 0 1 0 .69C20.5 16.27 16.64 19 12 19s-8.5-2.73-9.938-6.655"/><circle cx="12" cy="12" r="3"/></svg>`;
+
+export const CODE_ICON_SVG = `<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-code lucide-code-icon"><path d="m16 18 6-6-6-6"/><path d="m8 6-6 6 6 6"/></svg>`;
@@ -30,6 +30,7 @@ export * from './literal-html';
 export * from './markdown';
 export * from './mermaid-blocks';
 export * from './svg-blocks';
+export * from './diagram-blocks';
 export * from './max-bundle-size';
 export * from './mcp';
 export * from './mcp-form';
@@ -84,3 +84,8 @@ export const MCP_TRANSPORT_ICONS: Record<MCPTransportType, Component> = {
 	[MCPTransportType.STREAMABLE_HTTP]: Globe,
 	[MCPTransportType.SSE]: Radio
 };
+
+/** Standard SSE endpoint path indicators */
+export const MCP_SSE_ENDPOINT = '/sse';
+export const MCP_SSE_ENDPOINT_SLASH = '/sse/';
+export const MCP_SSE_ENDPOINT_QUERY = '/sse?';
@@ -16,7 +16,8 @@ import {
 	DEFAULT_MCP_CONFIG,
 	DEFAULT_CLIENT_VERSION,
 	DEFAULT_IMAGE_MIME_TYPE,
-	MCP_PARTIAL_REDACT_HEADERS
+	MCP_PARTIAL_REDACT_HEADERS,
+	CORS_PROXY_ENDPOINT
 } from '$lib/constants';
 import {
 	MCPConnectionPhase,
@@ -236,6 +237,36 @@ export class MCPService {

 		return {
 			fetch: async (input, init) => {
+				if (useProxy && typeof window !== 'undefined') {
+					let requestUrlStr = '';
+					if (typeof input === 'string') {
+						requestUrlStr = input;
+					} else if (input instanceof URL) {
+						requestUrlStr = input.href;
+					}
+
+					if (requestUrlStr) {
+						const parsedRequestUrl = new URL(requestUrlStr, window.location.origin);
+						if (
+							parsedRequestUrl.origin === window.location.origin &&
+							!parsedRequestUrl.pathname.includes(CORS_PROXY_ENDPOINT)
+						) {
+							const originalConfigUrl = new URL(config.url);
+							const realTargetUrl = new URL(
+								parsedRequestUrl.pathname + parsedRequestUrl.search,
+								originalConfigUrl.origin
+							);
+							const proxiedUrl = buildProxiedUrl(realTargetUrl.href);
+
+							if (typeof input === 'string') {
+								input = proxiedUrl.href;
+							} else if (input instanceof URL) {
+								input = proxiedUrl;
+							}
+						}
+					}
+				}
+
 				const startedAt = performance.now();
 				const requestHeaders = new Headers(baseInit.headers);

@@ -403,6 +434,32 @@ export class MCPService {
 			};
 		}

+		if (config.transport === MCPTransportType.SSE) {
+			const url = useProxy ? buildProxiedUrl(config.url) : new URL(config.url);
+			const { fetch: diagnosticFetch, disable: stopPhaseLogging } = this.createDiagnosticFetch(
+				serverName,
+				config,
+				requestInit,
+				url,
+				useProxy,
+				onLog
+			);
+
+			if (import.meta.env.DEV && import.meta.env.VITE_DEBUG) {
+				console.log(`[MCPService] Creating SSE transport for ${url.href}`);
+			}
+
+			return {
+				transport: new SSEClientTransport(url, {
+					requestInit,
+					fetch: diagnosticFetch,
+					eventSourceInit: { fetch: diagnosticFetch }
+				}),
+				type: MCPTransportType.SSE,
+				stopPhaseLogging
+			};
+		}
+
 		const url = useProxy ? buildProxiedUrl(config.url) : new URL(config.url);
 		const { fetch: diagnosticFetch, disable: stopPhaseLogging } = this.createDiagnosticFetch(
 			serverName,
@@ -19,7 +19,10 @@ import {
 	DISPLAY_NAME_SEPARATOR_REGEX,
 	PATH_SEPARATOR,
 	RESOURCE_TEXT_CONTENT_SEPARATOR,
-	DEFAULT_RESOURCE_FILENAME
+	DEFAULT_RESOURCE_FILENAME,
+	MCP_SSE_ENDPOINT,
+	MCP_SSE_ENDPOINT_SLASH,
+	MCP_SSE_ENDPOINT_QUERY
 } from '$lib/constants';
 import {
 	Database,
@@ -41,10 +44,22 @@ import type { MimeTypeUnion } from '$lib/types/common';
 export function detectMcpTransportFromUrl(url: string): MCPTransportType {
 	const normalized = url.trim().toLowerCase();

-	return normalized.startsWith(UrlProtocol.WEBSOCKET) ||
+	if (
+		normalized.startsWith(UrlProtocol.WEBSOCKET) ||
 		normalized.startsWith(UrlProtocol.WEBSOCKET_SECURE)
-		? MCPTransportType.WEBSOCKET
-		: MCPTransportType.STREAMABLE_HTTP;
+	) {
+		return MCPTransportType.WEBSOCKET;
+	}
+
+	if (
+		normalized.endsWith(MCP_SSE_ENDPOINT) ||
+		normalized.endsWith(MCP_SSE_ENDPOINT_SLASH) ||
+		normalized.includes(MCP_SSE_ENDPOINT_QUERY)
+	) {
+		return MCPTransportType.SSE;
+	}
+
+	return MCPTransportType.STREAMABLE_HTTP;
 }

 /**
@@ -41,7 +41,7 @@ if (LLAMA_BUILD_BORINGSSL)
    set(FIPS OFF CACHE BOOL "Enable FIPS (BoringSSL)")

    set(BORINGSSL_GIT "https://boringssl.googlesource.com/boringssl" CACHE STRING "BoringSSL git repository")
-    set(BORINGSSL_VERSION "0.20260526.0" CACHE STRING "BoringSSL version")
+    set(BORINGSSL_VERSION "0.20260616.0" CACHE STRING "BoringSSL version")

    message(STATUS "Fetching BoringSSL version ${BORINGSSL_VERSION}")