llama-cpp-bin 9743.0.0__tar.gz → 9745.0.0__tar.gz

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9743.0.0
+Version: 9745.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/common/peg-parser.cpp

@@ -6,13 +6,14 @@
 #include "unicode.h"
 
 #include <algorithm>
+#include <deque>
 #include <initializer_list>
 #include <map>
 #include <memory>
 #include <nlohmann/json.hpp>
 #include <regex>
+#include <set>
 #include <stdexcept>
-#include <unordered_set>
 
 // Trick to catch missing branches
 template <typename T>
@@ -88,40 +89,7 @@ struct trie {
         return match_result{match_result::NO_MATCH};
     }
 
-    struct prefix_and_next {
-        std::vector<uint32_t> prefix;
-        std::vector<uint32_t> next_chars;
-    };
-
-    std::vector<prefix_and_next> collect_prefix_and_next() {
-        std::vector<uint32_t>        prefix;
-        std::vector<prefix_and_next> result;
-        collect_prefix_and_next(0, prefix, result);
-        return result;
-    }
-
   private:
-    void collect_prefix_and_next(size_t index, std::vector<uint32_t> & prefix, std::vector<prefix_and_next> & out) {
-        if (!nodes[index].is_word) {
-            if (!nodes[index].children.empty()) {
-                std::vector<uint32_t> chars;
-                chars.reserve(nodes[index].children.size());
-                for (const auto & p : nodes[index].children) {
-                    chars.push_back(p.first);
-                }
-                out.emplace_back(prefix_and_next{prefix, chars});
-            }
-        }
-
-        for (const auto & p : nodes[index].children) {
-            uint32_t ch = p.first;
-            auto child = p.second;
-            prefix.push_back(ch);
-            collect_prefix_and_next(child, prefix, out);
-            prefix.pop_back();
-        }
-    }
-
     size_t create_node() {
         size_t index = nodes.size();
         nodes.emplace_back();
@@ -153,6 +121,65 @@ struct trie {
     }
 };
 
+// Aho-Corasick automaton
+struct aho_corasick {
+    trie                t;
+    std::vector<size_t> fail;      // failure links
+    std::vector<size_t> order;     // states in BFS order
+    std::vector<bool>   terminal;  // match states (directly or via a suffix link)
+    std::set<uint32_t>  alphabet;  // every character with a transition
+
+    aho_corasick(const std::vector<std::string> & strings) : t(strings) {
+        const auto & nodes = t.nodes;
+        const size_t n = nodes.size();
+
+        fail.assign(n, 0);
+        order.reserve(n);
+
+        std::deque<size_t> queue{ 0 };
+        while (!queue.empty()) {
+            size_t u = queue.front();
+            queue.pop_front();
+            order.push_back(u);
+            for (const auto & [ch, v] : nodes[u].children) {
+                if (u != 0) {
+                    size_t f = fail[u];
+                    while (f && nodes[f].children.find(ch) == nodes[f].children.end()) {
+                        f = fail[f];
+                    }
+                    auto it = nodes[f].children.find(ch);
+                    fail[v] = (it != nodes[f].children.end() && it->second != v) ? it->second : 0;
+                }
+                queue.push_back(v);
+            }
+        }
+
+        terminal.assign(n, false);
+        for (size_t u : order) {
+            terminal[u] = nodes[u].is_word || (u != 0 && terminal[fail[u]]);
+        }
+
+        for (const auto & node : nodes) {
+            for (const auto & [ch, v] : node.children) {
+                alphabet.insert(ch);
+            }
+        }
+    }
+
+    size_t num_states()          const { return t.nodes.size(); }
+    bool   is_terminal(size_t s) const { return terminal[s]; }
+
+    // follow failure links until a transition on `ch` exists.
+    size_t next(size_t state, uint32_t ch) const {
+        const auto & nodes = t.nodes;
+        while (state && nodes[state].children.find(ch) == nodes[state].children.end()) {
+            state = fail[state];
+        }
+        auto it = nodes[state].children.find(ch);
+        return it != nodes[state].children.end() ? it->second : 0;
+    }
+};
+
 static std::pair<uint32_t, size_t> parse_hex_escape(const std::string & str, size_t pos, int hex_count) {
     if (pos + hex_count > str.length()) {
         return {0, 0};
@@ -992,12 +1019,12 @@ void common_peg_arena::resolve_refs() {
 }
 
 std::string common_peg_arena::dump(common_peg_parser_id id) const {
-    std::unordered_set<common_peg_parser_id> visited;
+    std::set<common_peg_parser_id> visited;
     return dump_impl(id, visited);
 }
 
 std::string common_peg_arena::dump_impl(common_peg_parser_id                       id,
-                                        std::unordered_set<common_peg_parser_id> & visited) const {
+                                        std::set<common_peg_parser_id> & visited) const {
     // Check for cycles
     if (visited.count(id)) {
         return "[cycle]";
@@ -1502,61 +1529,74 @@ static std::string gbnf_escape_char_class(uint32_t c) {
     return std::string(buf);
 }
 
-static std::string gbnf_excluding_pattern(const std::vector<std::string> & strings) {
-    trie matcher(strings);
-    auto pieces = matcher.collect_prefix_and_next();
+// GBNF grammar matching strings that contain no string in `strings` as a
+// substring. Emits the complement of an Aho-Corasick automaton DFA and returns
+// the start state rule name.
+//
+// ref: https://github.com/ggml-org/llama.cpp/pull/24839
+static std::string gbnf_excluding_grammar(const common_grammar_builder & builder,
+                                          const std::string &            prefix,
+                                          const std::vector<std::string> & strings) {
+    aho_corasick ac(strings);
 
-    std::string pattern;
-    std::string trailing;  // optional proper-prefix of a delimiter, allowed only at the very end
-    for (size_t i = 0; i < pieces.size(); ++i) {
-        if (i > 0) {
-            pattern += " | ";
+    auto state_name = [&](size_t s) -> std::string {
+        if (s == 0) {
+            return prefix;
         }
+        std::string num = std::to_string(s);
+        num = num.size() == 1 ? ("0" + num) : num;
+        return prefix + "-" + num;
+    };
 
-        const auto & pre = pieces[i].prefix;
-        const auto & chars = pieces[i].next_chars;
-
-        std::string cls;
-        cls.reserve(chars.size());
+    auto char_class = [](const std::vector<uint32_t> & chars, bool negate) {
+        std::string s = negate ? "[^" : "[";
         for (uint32_t ch : chars) {
-            cls += gbnf_escape_char_class(ch);
-        }
-
-        if (!pre.empty()) {
-            std::string pre_literal = gbnf_format_literal(common_unicode_cpts_to_utf8(pre));
-            pattern += pre_literal + " [^" + cls + "]";
-            // Each interior alternative consumes a delimiter-prefix plus a disambiguating
-            // char, so the repetition alone cannot match a value that *ends* on a proper
-            // prefix of a delimiter (e.g. a trailing "\n" when the delimiter is
-            // "\n</parameter>\n"). The runtime until() (greedy first-match) accepts such
-            // values, so without this the grammar would reject input the parser accepts.
-            // Allow the value to terminate on any proper prefix as an optional tail.
-            // This makes the grammar a slight superset of the runtime language (a value
-            // may end on the longest prefix, which greedy first-match would not itself
-            // produce); harmless for constrained generation, which only needs to admit
-            // every runtime-valid string.
-            if (!trailing.empty()) {
-                trailing += " | ";
-            }
-            trailing += pre_literal;
-        } else {
-            pattern += "[^" + cls + "]";
+            s += gbnf_escape_char_class(ch);
         }
+        return s + "]";
+    };
+
+    for (size_t q = 0; q < ac.num_states(); q++) {
+        if (ac.is_terminal(q)) {
+            continue; // match states are dropped
+        }
+
+        std::map<size_t, std::vector<uint32_t>> buckets;
+        std::vector<uint32_t> excluded;
+        for (uint32_t c : ac.alphabet) {
+            size_t d = ac.next(q, c);
+            if (ac.is_terminal(d)) {
+                excluded.push_back(c); // completes a forbidden string -> omit
+            } else if (d != 0) {
+                buckets[d].push_back(c); // specific non-root destination
+                excluded.push_back(c);
+            }
+        }
+
+        std::string rhs = "|"; // every state is accepting
+        for (const auto & [d, chars] : buckets) {
+            rhs += " " + char_class(chars, false) + " " + state_name(d) + " |";
+        }
+        rhs += " " + char_class(excluded, true) + " " + state_name(0);
+
+        builder.add_rule(state_name(q), rhs);
     }
 
-    std::string result = "(" + pattern + ")*";
-    if (!trailing.empty()) {
-        result += " (" + trailing + ")?";
+    // An empty delimiter makes the start state terminal. Emit an entry rule
+    // that matches nothing so the returned reference stays valid.
+    if (ac.is_terminal(0)) {
+        builder.add_rule(prefix, "|");
     }
-    return result;
+
+    return state_name(0);
 }
 
-static std::unordered_set<std::string> collect_reachable_rules(
+static std::set<std::string> collect_reachable_rules(
     const common_peg_arena & arena,
     const common_peg_parser_id & rule
 ) {
-    std::unordered_set<std::string> reachable;
-    std::unordered_set<std::string> visited;
+    std::set<std::string> reachable;
+    std::set<std::string> visited;
 
     std::function<void(common_peg_parser_id)> visit = [&](common_peg_parser_id id) {
         const auto & parser = arena.get(id);
@@ -1765,7 +1805,7 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
                 if (p.delimiters.empty()) {
                     return ".*";
                 }
-                return gbnf_excluding_pattern(p.delimiters);
+                return gbnf_excluding_grammar(builder, "until-" + std::to_string(id), p.delimiters);
             } else if constexpr (std::is_same_v<T, common_peg_schema_parser>) {
                 if (schema_delegates(p)) {
                     return to_gbnf(p.child);
@@ -1789,7 +1829,7 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
     };
 
     // Collect reachable rules
-    std::unordered_set<std::string> reachable_rules;
+    std::set<std::string> reachable_rules;
 
     if (lazy) {
         // Collect rules reachable from trigger rules

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/common/peg-parser.h

@@ -3,8 +3,8 @@
 #include <nlohmann/json_fwd.hpp>
 
 #include <memory>
+#include <set>
 #include <unordered_map>
-#include <unordered_set>
 #include <string>
 #include <string_view>
 #include <functional>
@@ -335,7 +335,7 @@ class common_peg_arena {
     friend class common_peg_parser_builder;
 
   private:
-    std::string dump_impl(common_peg_parser_id id, std::unordered_set<common_peg_parser_id> & visited) const;
+    std::string dump_impl(common_peg_parser_id id, std::set<common_peg_parser_id> & visited) const;
 
     common_peg_parser_id add_parser(common_peg_parser_variant parser);
     void add_rule(const std::string & name, common_peg_parser_id id);

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/common/speculative.cpp

@@ -905,7 +905,13 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
     int32_t n_embd = 0;
 
-    bool is_mem_shared = false;
+    // One MTP draft driver, three modes (set once in the ctor):
+    //   is_mem_shared (gemma4): shares the target KV, runs all heads in one graph.
+    //   chain_heads (step35): n_mtp_layers trained heads, one per draft step.
+    //   neither (qwen35 / qwen35moe): a single trained MTP head.
+    int32_t n_mtp_layers  = 1;
+    bool    is_mem_shared = false;   // gemma4
+    bool    chain_heads   = false;   // derived in the ctor: n_mtp_layers > 1 && !is_mem_shared
 
     // Per-sequence cross-batch carryover: pair (h_p, x_{p+1}) at MTP pos p+1.
     // The last h-row of one process() call needs the first token of the NEXT
@@ -920,10 +926,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
     std::vector<std::vector<float>> verify_h;
     std::vector<int32_t> verify_h_rows;
 
-    // Per-seq draft length from the last draft() call, used in accept() to
-    // roll back ctx_dft's recurrent state past the AR draft's redundant
-    // pre-advancement before process() mirrored the verify batch.
-    std::vector<uint16_t> last_n_drafted;
+    std::vector<int>                i_last;
+    std::vector<std::vector<float>> chain_h;
 
     common_speculative_impl_draft_mtp(const common_params_speculative & params, uint32_t n_seq)
         : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq)
@@ -936,6 +940,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         n_embd = llama_model_n_embd_out(llama_get_model(ctx_dft));
         GGML_ASSERT(n_embd == llama_model_n_embd(llama_get_model(ctx_tgt)) &&
                 "MTP input row width must match the target h_nextn width");
+        n_mtp_layers = std::max(1, (int) llama_model_n_layer_nextn(llama_get_model(ctx_dft)));
 
         LOG_INF("%s: adding speculative implementation 'draft-mtp'\n", __func__);
         LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f, n_embd=%d, backend_sampling=%d\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min, n_embd, (int) this->params.backend_sampling);
@@ -982,16 +987,25 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);
 
         is_mem_shared = llama_get_ctx_other(ctx_dft) == ctx_tgt;
+        chain_heads   = n_mtp_layers > 1 && !is_mem_shared;
+
+        if (chain_heads) {
+            this->params.n_max = std::min(this->params.n_max, n_mtp_layers);
+            
+            chain_h.assign(n_seq, {});
+            for (auto & c : chain_h) {
+                c.reserve((size_t) (this->params.n_max + 1) * n_embd);
+            }
+        }
 
         pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));
 
+        i_last.assign(n_seq, -1);
         i_batch_beg.assign(n_seq, -1);
         i_batch_end.assign(n_seq, -1);
 
         verify_h.assign(n_seq, {});
         verify_h_rows.assign(n_seq, 0);
-
-        last_n_drafted.assign(n_seq, 0);
     }
 
     ~common_speculative_impl_draft_mtp() override {
@@ -1097,9 +1111,34 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                 set_h(i_batch_beg[seq_id], pending_h[seq_id].data());
             }
 
-            const int32_t rc = llama_decode(ctx_dft, batch);
-            if (rc != 0) {
-                LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (pos=%d)\n", __func__, (int) rc, (int) batch_in.pos[0]);
+            auto * mem_dft = llama_get_memory(ctx_dft);
+
+            bool ok = true;
+            for (int head = 0; head < n_mtp_layers; ++head) {
+                if (chain_heads) {
+                    // ref: https://github.com/ggml-org/llama.cpp/pull/24340/changes#r3413498544
+                    for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                        if (i_batch_beg[seq_id] < 0) {
+                            continue;
+                        }
+                        llama_memory_seq_rm(mem_dft, seq_id, batch_in.pos[i_batch_beg[seq_id]], -1);
+                    }
+                    llama_set_nextn_layer_offset(ctx_dft, head);
+                }
+
+                const int32_t rc = llama_decode(ctx_dft, batch);
+                if (rc != 0) {
+                    LOG_ERR("%s: llama_decode(ctx_dft) head=%d failed rc=%d (pos=%d)\n",
+                            __func__, head, (int) rc, (int) batch_in.pos[0]);
+                    ok = false;
+                    break;
+                }
+            }
+
+            if (chain_heads) {
+                llama_set_nextn_layer_offset(ctx_dft, 0); // restore default for non-draft decodes
+            }
+            if (!ok) {
                 return false;
             }
         }
@@ -1134,7 +1173,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         int n_drafting = 0;
         std::vector<bool> drafting(n_seq);
 
-        const float * h_row = nullptr;
         const size_t row_bytes = (size_t) n_embd * sizeof(float);
 
         for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
@@ -1149,22 +1187,43 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
             common_sampler_reset(smpls[seq_id].get());
 
             common_batch_add(batch, dp.id_last, dp.n_past, { seq_id }, true);
+            std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, pending_h[seq_id].data(), row_bytes);
 
-            h_row = pending_h[seq_id].data();
-            std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
-        }
+            i_last[seq_id] = batch.n_tokens - 1;
 
-        int ret = llama_decode(ctx_dft, batch);
-        if (ret != 0) {
-            LOG_WRN("%s: llama_decode returned %d\n", __func__, ret);
-            return;
+            if (chain_heads) {
+                chain_h[seq_id].assign(pending_h[seq_id].begin(), pending_h[seq_id].end());
+            }
         }
 
         int i = 0;
 
         while (n_drafting > 0) {
-            int i_batch = 0;
+            // each step decodes under a different head, i.e. a different decoder layer, and
+            // KV is per layer. process() filled this layer's KV only for positions < n_past
+            // (prompt + accepted prefix) — nothing in the draft region yet. so reset the
+            // draft region (the seq_rm lower bound is n_past, leaving the prompt KV intact)
+            // and select head i so it rebuilds its own layer's KV there; decoding just the
+            // latest token would leave its attention reading cells only another head wrote.
+            if (chain_heads) {
+                auto * mem_dft = llama_get_memory(ctx_dft);
+                for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                    if (drafting[seq_id]) {
+                        llama_memory_seq_rm(mem_dft, seq_id, dparams[seq_id].n_past, -1);
+                    }
+                }
+                llama_set_nextn_layer_offset(ctx_dft, i);
+            }
 
+            int ret = llama_decode(ctx_dft, batch);
+            if (ret != 0) {
+                LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret);
+                break;
+            }
+
+            // rebuild the batch for the next step: the growing-KV paths re-add only the
+            // new token (the KV already holds the prefix), while chained heads re-add the
+            // whole prefix at the next head. dropped sequences are simply not re-added.
             common_batch_clear(batch);
 
             for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
@@ -1174,9 +1233,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
                 auto * smpl = smpls[seq_id].get();
 
-                common_sampler_sample(smpl, ctx_dft, i_batch, true);
-                h_row = llama_get_embeddings_nextn_ith(ctx_dft, i_batch);
-                ++i_batch;
+                common_sampler_sample(smpl, ctx_dft, i_last[seq_id], true);
+                const float * h_row = llama_get_embeddings_nextn_ith(ctx_dft, i_last[seq_id]);
 
                 const auto * cur_p = common_sampler_get_candidates(smpl, true);
 
@@ -1210,30 +1268,41 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                     continue;
                 }
 
-                if (is_mem_shared) {
+                if (chain_heads) {
+                    // ref: https://github.com/ggml-org/llama.cpp/pull/24340#discussion_r3448031546
+                    chain_h[seq_id].insert(chain_h[seq_id].end(), h_row, h_row + n_embd);
+
+                    const int n_rows = (int) result.size() + 1; // id_last + tokens drafted so far
+                    for (int t = 0; t < n_rows; ++t) {
+                        const llama_token tok = (t == 0) ? dp.id_last : result[t - 1];
+                        common_batch_add(batch, tok, dp.n_past + t, { seq_id }, t == n_rows - 1);
+                        std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd,
+                                    chain_h[seq_id].data() + (size_t) t * n_embd, row_bytes);
+                    }
+                } else if (is_mem_shared) {
                     // note: with shared memory (e.g. Gemma4 assistants) we use the same position for all draft tokens
                     // ref: https://github.com/huggingface/transformers/blob/effde20942e3f82a1b97449f60b3a48c5ff96145/docs/source/en/model_doc/gemma4_assistant.md?plain=1#L36-L37
                     common_batch_add(batch, id, dp.n_past, { seq_id }, true);
+                    std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
                 } else {
                     common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
+                    std::memcpy(batch.embd + (size_t) (batch.n_tokens - 1) * n_embd, h_row, row_bytes);
                 }
-                std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
-            }
 
-            if (batch.n_tokens == 0) {
-                break;
+                i_last[seq_id] = batch.n_tokens - 1;
             }
 
-            // evaluate the drafted tokens on the draft model
-            ret = llama_decode(ctx_dft, batch);
-            if (ret != 0) {
-                LOG_WRN("%s: llama_decode[%d] returned %d\n", __func__, i, ret);
+            if (batch.n_tokens == 0) {
                 break;
             }
 
             ++i;
         }
 
+        if (chain_heads) {
+            llama_set_nextn_layer_offset(ctx_dft, 0); // restore default for non-draft decodes
+        }
+
         for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
             auto & dp = dparams[seq_id];
             if (!dp.drafting) {
@@ -1243,8 +1312,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
             if (dp.result->size() < (size_t) params.n_min) {
                 dp.result->clear();
             }
-
-            last_n_drafted[seq_id] = (uint16_t) dp.result->size();
         }
     }
 
@@ -1857,7 +1924,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
 
         bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE));
         bool has_draft_eagle3 = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3)) && params.draft.ctx_dft != nullptr;
-        bool has_mtp = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP)) && params.draft.ctx_dft != nullptr;
+        bool has_draft_mtp    = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP))    && params.draft.ctx_dft != nullptr;
 
 
 
@@ -1895,7 +1962,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
         if (has_draft_eagle3) {
             configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params));
         }
-        if (has_mtp) {
+        if (has_draft_mtp) {
             configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, params));
         }
     }

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/include/llama.h

@@ -558,14 +558,15 @@ extern "C" {
     LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
     LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);
 
-    LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_embd     (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_embd_out (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
-    LLAMA_API int32_t llama_model_n_swa      (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_ctx_train  (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd       (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd_inp   (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd_out   (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_layer      (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_layer_nextn(const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_head       (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_head_kv    (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_swa        (const struct llama_model * model);
 
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/src/llama-context.cpp

@@ -1156,6 +1156,10 @@ void llama_context::set_embeddings_layer_inp(uint32_t lid, bool enable) {
     sched_need_reserve = true;
 }
 
+void llama_context::set_nextn_layer_offset(int32_t offset) {
+    cparams.nextn_layer_offset = offset;
+}
+
 void llama_context::set_causal_attn(bool value) {
     LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value);
 
@@ -3699,6 +3703,10 @@ void llama_set_embeddings_layer_inp(llama_context * ctx, uint32_t lid, bool valu
     ctx->set_embeddings_layer_inp(lid, value);
 }
 
+void llama_set_nextn_layer_offset(llama_context * ctx, int32_t offset) {
+    ctx->set_nextn_layer_offset(offset);
+}
+
 llama_memory_t llama_get_memory(const struct llama_context * ctx) {
     if (!ctx) {
         return nullptr;

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/src/llama-context.h

@@ -115,6 +115,7 @@ struct llama_context {
     void set_embeddings (bool value);
     void set_embeddings_nextn(bool value, bool masked);
     void set_embeddings_layer_inp(uint32_t lid, bool enable);
+    void set_nextn_layer_offset(int32_t offset);
     void set_causal_attn(bool value);
     void set_warmup(bool value);
 

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/src/llama-cparams.h

@@ -18,6 +18,8 @@ struct llama_cparams {
     int32_t  n_threads;       // number of threads to use for generation
     int32_t  n_threads_batch; // number of threads to use for batch processing
 
+    int32_t  nextn_layer_offset = 0;
+
     float rope_freq_base;
     float rope_freq_scale;
 

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/src/llama-ext.h

@@ -95,6 +95,11 @@ LLAMA_API llama_memory_breakdown llama_get_memory_breakdown(const struct llama_c
 // If masked == false, output the embeddings for all tokens in the batch regardless of batch.logits
 LLAMA_API void llama_set_embeddings_nextn(struct llama_context * ctx, bool value, bool masked);
 
+// Select which appended NextN block the DECODER_MTP graph runs (offset past
+// the trunk: il = n_layer() + offset). Used by the speculative NextN driver to
+// chain multiple trained NextN heads. Default 0 (first head).
+LLAMA_API void llama_set_nextn_layer_offset(struct llama_context * ctx, int32_t offset);
+
 // mirrors:
 // LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 LLAMA_API float * llama_get_embeddings_nextn(struct llama_context * ctx);

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/src/llama-graph.h

@@ -682,9 +682,16 @@ struct llm_graph_params {
             }
         }
 
+        // TODO: https://github.com/ggml-org/llama.cpp/pull/24340#discussion_r3448035248
+        if (cparams.nextn_layer_offset != other.cparams.nextn_layer_offset) {
+            return false;
+        }
+
         return
-            cparams.embeddings  == other.cparams.embeddings  &&
-            cparams.causal_attn == other.cparams.causal_attn &&
+            cparams.embeddings              == other.cparams.embeddings              &&
+            cparams.embeddings_nextn        == other.cparams.embeddings_nextn        &&
+            cparams.embeddings_nextn_masked == other.cparams.embeddings_nextn_masked &&
+            cparams.causal_attn             == other.cparams.causal_attn             &&
             arch  == other.arch  &&
             gtype == other.gtype &&
             cvec  == other.cvec  &&

{llama_cpp_bin-9743.0.0 → llama_cpp_bin-9745.0.0}/llama.cpp/src/llama-model.cpp

@@ -2312,6 +2312,10 @@ int32_t llama_model_n_layer(const llama_model * model) {
     return model->hparams.n_layer();
 }
 
+int32_t llama_model_n_layer_nextn(const llama_model * model) {
+    return model->hparams.n_layer_nextn;
+}
+
 int32_t llama_model_n_head(const llama_model * model) {
     return model->hparams.n_head();
 }