ggml update to b7108 (#12992)

* Revert "vulkan: temporary cary of vulkan fixes (#12971)"

This reverts commit 3a9e8e9f.

* ggml update to b7087

* fix argsort on metal

* update to b7108

* fix bakllava regression

This model lacks the metadata for the projector type.

* update to b7209

* fix TopK perf

* only build arm code on arm

llama/llama.cpp/src/llama-grammar.cpp

@@ -6,8 +6,10 @@
 
 #include <cmath>
 #include <algorithm>
+#include <cstdint>
 #include <stdexcept>
 
+#define MAX_REPETITION_THRESHOLD 2000
 //
 // helpers
 //
@@ -345,8 +347,10 @@ const char * llama_grammar_parser::parse_sequence(
     size_t last_sym_start = rule.size();
     const char * pos = src;
 
-    auto handle_repetitions = [&](int min_times, int max_times) {
-
+    // use UINT64_MAX as the empty value because we aligned to the proper uint64_t type so -1 can't be used
+    // (though it's technically the same as -1 now)
+    auto handle_repetitions = [&](uint64_t min_times, uint64_t max_times) {
+        bool no_max = max_times == UINT64_MAX;
         if (last_sym_start == rule.size()) {
             throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
         }
@@ -373,20 +377,20 @@ const char * llama_grammar_parser::parse_sequence(
             rule.resize(last_sym_start);
         } else {
             // Repeat the previous elements (min_times - 1) times
-            for (int i = 1; i < min_times; i++) {
+            for (uint64_t i = 1; i < min_times; i++) {
                 rule.insert(rule.end(), prev_rule.begin(), prev_rule.end());
             }
         }
 
         uint32_t last_rec_rule_id = 0;
-        auto n_opt = max_times < 0 ? 1 : max_times - min_times;
+        auto n_opt = no_max ? 1 : max_times - min_times;
 
         llama_grammar_rule rec_rule(prev_rule);
-        for (int i = 0; i < n_opt; i++) {
+        for (uint64_t i = 0; i < n_opt; i++) {
             rec_rule.resize(prev_rule.size());
             uint32_t rec_rule_id = generate_symbol_id( rule_name);
-            if (i > 0 || max_times < 0) {
-                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
+            if (i > 0 || no_max) {
+                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, no_max ? rec_rule_id : last_rec_rule_id});
             }
             rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
             rec_rule.push_back({LLAMA_GRETYPE_END, 0});
@@ -478,10 +482,10 @@ const char * llama_grammar_parser::parse_sequence(
                 throw std::runtime_error(std::string("expecting an int at ") + pos);
             }
             const char * int_end = parse_int(pos);
-            int min_times = std::stoul(std::string(pos, int_end - pos));
+            uint64_t min_times = std::stoul(std::string(pos, int_end - pos));
             pos = parse_space(int_end, is_nested);
 
-            int max_times = -1;
+            uint64_t max_times = UINT64_MAX; // default: no max limit
 
             if (*pos == '}') {
                 max_times = min_times;
@@ -502,6 +506,10 @@ const char * llama_grammar_parser::parse_sequence(
             } else {
                 throw std::runtime_error(std::string("expecting ',' at ") + pos);
             }
+            bool has_max = max_times != UINT64_MAX;
+            if (min_times > MAX_REPETITION_THRESHOLD || (has_max && max_times > MAX_REPETITION_THRESHOLD)) {
+                throw std::runtime_error(std::string("number of repetitions exceeds sane defaults, please reduce the number of repetitions"));
+            }
             handle_repetitions(min_times, max_times);
         } else {
             break;

llama/llama.cpp/src/llama-graph.cpp

@@ -810,6 +810,9 @@ ggml_tensor * llm_graph_context::build_ffn(
             GGML_ABORT("fatal error");
     }
 
+    //expand here so that we can fuse ffn gate
+    ggml_build_forward_expand(gf, cur);
+
     if (gate && type_gate == LLM_FFN_PAR) {
         cur = ggml_mul(ctx0, cur, tmp);
         cb(cur, "ffn_gate_par", il);
@@ -958,14 +961,14 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         // organize experts into n_expert_groups
         ggml_tensor * selection_groups = ggml_reshape_3d(ctx0, selection_probs, n_exp_per_group, hparams.n_expert_groups, n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens]
 
-        ggml_tensor * group_scores = ggml_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
+        ggml_tensor * group_scores = ggml_argsort_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
         group_scores = ggml_get_rows(ctx0, ggml_reshape_4d(ctx0, selection_groups, 1, selection_groups->ne[0], selection_groups->ne[1], selection_groups->ne[2]), group_scores); // [1, 2, n_expert_groups, n_tokens]
 
         // get top n_group_used expert groups
         group_scores = ggml_sum_rows(ctx0, ggml_reshape_3d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2], group_scores->ne[3])); // [1, n_expert_groups, n_tokens]
         group_scores = ggml_reshape_2d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2]); // [n_expert_groups, n_tokens]
 
-        ggml_tensor * expert_groups = ggml_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
+        ggml_tensor * expert_groups = ggml_argsort_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
         cb(expert_groups, "ffn_moe_group_topk", il);
 
         // mask out the other groups
@@ -976,7 +979,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     }
 
     // select experts
-    ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
+    ggml_tensor * selected_experts = ggml_argsort_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
     cb(selected_experts->src[0], "ffn_moe_argsort", il);
     cb(selected_experts, "ffn_moe_topk", il);
 
@@ -1006,10 +1009,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens]
         cb(weights_sum, "ffn_moe_weights_sum", il);
 
-        if (arch == LLM_ARCH_BAILINGMOE2) {
-            weights_sum = ggml_scale_bias(ctx0, weights_sum, 1.0, 1e-20);
-            cb(weights_sum, "ffn_moe_weights_sum_biased", il);
-        }
+        // Avoid division by zero, clamp to smallest number representable by F16
+        weights_sum = ggml_clamp(ctx0, weights_sum, 6.103515625e-5, INFINITY);
+        cb(weights_sum, "ffn_moe_weights_sum_clamped", il);
 
         weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
         cb(weights, "ffn_moe_weights_norm", il);
@@ -1091,6 +1093,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             GGML_ABORT("fatal error");
     }
 
+    //expand here so that we can fuse ffn gate
+    ggml_build_forward_expand(gf, cur);
+
     experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
@@ -1137,7 +1142,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
 // input embeddings with optional lora
 ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
-    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd = hparams.n_embd_inp();
 
     auto inp = std::make_unique<llm_graph_input_embd>();
 
@@ -1274,7 +1279,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     //    return cur;
     //}
 
-    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
+    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
     const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
     cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
@@ -1587,9 +1592,10 @@ ggml_tensor * llm_graph_context::build_attn(
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
     // by doing so, the number of splits in the graph is reduced
+    // expand k later to enable rope fusion which directly writes into k-v cache
     ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
+    ggml_build_forward_expand(gf, k_cur);
 
     const auto * mctx_cur = inp->mctx;
 
@@ -2030,7 +2036,7 @@ int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buck
 
     if (bidirectional) {
         relative_bucket += (relative_position > 0) * n_buckets;
-        relative_position = abs(relative_position);
+        relative_position = std::abs(relative_position);
     } else {
         relative_position = -std::min<int32_t>(relative_position, 0);
     }

llama/llama.cpp/src/llama-hparams.cpp

@@ -60,6 +60,16 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
     return n_head/n_head_kv;
 }
 
+uint32_t llama_hparams::n_embd_inp() const {
+    uint32_t n_embd_inp = n_embd;
+
+    if (n_deepstack_layers > 0) {
+        n_embd_inp += n_embd * n_deepstack_layers;
+    }
+
+    return n_embd_inp;
+}
+
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
@@ -148,7 +158,7 @@ bool llama_hparams::is_recurrent(uint32_t il) const {
 }
 
 uint32_t llama_hparams::n_pos_per_embd() const {
-    return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
+    return rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1;
 }
 
 bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {

llama/llama.cpp/src/llama-hparams.h

@@ -6,7 +6,7 @@
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 384  // Kimi-K2
+#define LLAMA_MAX_EXPERTS 512 // Qwen3 Next
 
 enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_NONE           = 0,
@@ -185,6 +185,9 @@ struct llama_hparams {
     std::array<float, LLAMA_MAX_LAYERS> xielu_beta;
     std::array<float, LLAMA_MAX_LAYERS> xielu_eps;
 
+    // qwen3vl deepstack
+    uint32_t n_deepstack_layers = 0;
+
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
@@ -226,6 +229,9 @@ struct llama_hparams {
 
     uint32_t n_gqa(uint32_t il = 0) const;
 
+    // dimension of main + auxiliary input embeddings
+    uint32_t n_embd_inp() const;
+
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
 

llama/llama.cpp/src/llama-impl.cpp

@@ -23,7 +23,7 @@ time_meas::~time_meas() {
     if (t_start_us >= 0) {
         t_acc += ggml_time_us() - t_start_us;
     }
-    }
+}
 
 void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     ggml_log_set(log_callback, user_data);

llama/llama.cpp/src/llama-kv-cache-iswa.cpp

@@ -45,7 +45,9 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
 
     const uint32_t size_base = kv_size;
 
-    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch, n_pad));
+    // note: the SWA cache is always padded to 256 for performance
+    //       https://github.com/ggml-org/llama.cpp/issues/17037
+    uint32_t size_swa = GGML_PAD(std::min(size_base, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);
 
     // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
     if (swa_full) {

llama/llama.cpp/src/llama-kv-cache.cpp

@@ -8,6 +8,7 @@
 #include <algorithm>
 #include <cassert>
 #include <cmath>
+#include <cstring>
 #include <limits>
 #include <map>
 #include <stdexcept>
@@ -37,8 +38,15 @@ llama_kv_cache::llama_kv_cache(
 
     const uint32_t n_layer_kv = hparams.n_layer_kv();
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -53,13 +61,12 @@ llama_kv_cache::llama_kv_cache(
                 return nullptr;
             }
 
-            ctx_map[buft] = ctx;
-            ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
 
             return ctx;
         }
 
-        return it->second;
+        return it->second.get();
     };
 
     GGML_ASSERT(n_stream == 1 || n_stream == n_seq_max);
@@ -167,11 +174,8 @@ llama_kv_cache::llama_kv_cache(
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
-    for (auto it : ctx_map) {
-        auto * buft = it.first;
-        auto * ctx  = it.second;
-
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+    for (auto & [buft, ctx] : ctx_map) {
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for kv cache");
         }
@@ -179,7 +183,7 @@ llama_kv_cache::llama_kv_cache(
         LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
 
         ggml_backend_buffer_clear(buf, 0);
-        bufs.emplace_back(buf);
+        ctxs_bufs.emplace_back(std::move(ctx), buf);
     }
 
     {
@@ -203,7 +207,7 @@ void llama_kv_cache::clear(bool data) {
     }
 
     if (data) {
-        for (auto & buf : bufs) {
+        for (auto & [_, buf] : ctxs_bufs) {
             ggml_backend_buffer_clear(buf.get(), 0);
         }
     }
@@ -334,6 +338,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll
             llama_pos pos   = v_cells[s0].pos_get(i);
             llama_pos shift = v_cells[s0].get_shift(i);
 
+            llama_kv_cell_ext ext = v_cells[s0].ext_get(i);
+
             if (shift != 0) {
                 pos -= shift;
                 assert(pos >= 0);
@@ -345,6 +351,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll
             if (shift != 0) {
                 v_cells[s1].pos_add(i, shift);
             }
+
+            v_cells[s1].ext_set(i, ext);
         }
     }
 
@@ -379,6 +387,7 @@ void llama_kv_cache::seq_keep(llama_seq_id seq_id) {
 
 void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
+    GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_add() is only supported for n_pos_per_embd() == 1");
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
     auto & head  = v_heads[seq_to_stream[seq_id]];
@@ -423,6 +432,7 @@ void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, ll
 
 void llama_kv_cache::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
+    GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_div() is only supported for n_pos_per_embd() == 1");
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
 
@@ -472,8 +482,8 @@ llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const ggml_backend_buffer_ptr & buf_ptr : bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -896,6 +906,14 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch &
 
             cells.pos_set(idx, ubatch.pos[i]);
 
+            if (ubatch.is_pos_2d()) {
+                llama_kv_cell_ext ext {
+                    /*.x =*/ ubatch.pos[i + ubatch.n_tokens*2],
+                    /*.y =*/ ubatch.pos[i + ubatch.n_tokens],
+                };
+                cells.ext_set(idx, ext);
+            }
+
             for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) {
                 cells.seq_add(idx, ubatch.seq_id[i][s]);
             }
@@ -957,10 +975,14 @@ bool llama_kv_cache::get_has_shift() const {
 uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const {
     uint32_t result = 0;
 
+    // pad the n_kv value so that the graph remains constant across batches and can be reused
+    // note: this also helps some backends with performance (f.ex https://github.com/ggml-org/llama.cpp/pull/16812#issuecomment-3455112220)
+    const uint32_t n_pad_cur = std::max(n_pad, 256u);
+
     for (uint32_t s = 0; s < sinfo.n_stream(); ++s) {
         const auto & cells = v_cells[sinfo.strm[s]];
 
-        result = std::max(std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad))), result);
+        result = std::max(std::min(cells.size(), std::max(n_pad_cur, GGML_PAD(cells.used_max_p1(), n_pad_cur))), result);
     }
 
     return result;
@@ -1239,6 +1261,11 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
 
                 const llama_pos p1 = ubatch->pos[i];
 
+                // for M-RoPE
+                const bool is_2d = ubatch->is_pos_2d();
+                const llama_pos p1_x = is_2d ? ubatch->pos[i + ubatch->n_tokens*2] : 0;
+                const llama_pos p1_y = is_2d ? ubatch->pos[i + ubatch->n_tokens]   : 0;
+
                 const uint64_t idst = n_kv*(h*n_stream*n_tps_pad + s*n_tps_pad + ii);
 
                 for (uint32_t j = 0; j < n_kv; ++j) {
@@ -1258,6 +1285,14 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
                         continue;
                     }
 
+                    // M-RoPE causal mask
+                    if (causal_attn && is_2d && p0 == p1) {
+                        const auto & p0_ext = cells.ext_get(j);
+                        if (p0_ext.is_2d_gt(p1_x, p1_y)) {
+                            continue;
+                        }
+                    }
+
                     // apply SWA if any
                     if (is_masked_swa(p0, p1)) {
                         continue;
@@ -1298,7 +1333,7 @@ void llama_kv_cache::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch
 size_t llama_kv_cache::total_size() const {
     size_t size = 0;
 
-    for (const auto & buf : bufs) {
+    for (const auto & [_, buf] : ctxs_bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
     }
 
@@ -1340,7 +1375,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_beta_slow  = cparams.yarn_beta_slow;
 
     const auto & n_rot     = hparams.n_rot;
-    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE
+    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
                                 // @ngxson : this is a workaround
                                 // for M-RoPE, we want to rotate the whole vector when doing KV shift
                                 // a normal RoPE should work, we just need to use the correct ordering
@@ -1551,6 +1586,9 @@ void llama_kv_cache::state_write_meta(llama_io_write_i & io, const cell_ranges_t
             io.write(&pos,      sizeof(pos));
             io.write(&n_seq_id, sizeof(n_seq_id));
 
+            // TODO: we also need to save llama_kv_cell_ext when apply_ubatch() support loading it
+            //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
+
             for (const auto & seq_id : seq_ids) {
                 io.write(&seq_id, sizeof(seq_id));
             }
@@ -1696,6 +1734,8 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             return false;
         }
 
+        // TODO: we cannot yet restore llama_kv_cell_ext as the apply_ubatch() does not support it yet
+        //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
         apply_ubatch(sinfo, ubatch);
 
         const auto head_cur = sinfo.head();
@@ -2010,8 +2050,3 @@ void llama_kv_cache_context::set_input_kq_mask(ggml_tensor * dst, const llama_ub
 void llama_kv_cache_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     kv->set_input_pos_bucket(dst, ubatch);
 }
-
-uint32_t llama_kv_cache::get_padding(const llama_cparams & cparams) {
-    // the FA kernels require padding to avoid extra runtime boundary checks
-    return cparams.flash_attn ? 256u : 32u;
-}

llama/llama.cpp/src/llama-kv-cache.h

@@ -19,8 +19,6 @@ struct llama_context;
 
 class llama_kv_cache : public llama_memory_i {
 public:
-    static uint32_t get_padding(const llama_cparams & cparams);
-
     struct stream_copy_info {
         bool empty() const {
             assert(ssrc.size() == sdst.size());
@@ -217,8 +215,8 @@ private:
     // this is the SWA type of the cache - not to be confused with the model SWA type
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
-    std::vector<ggml_context_ptr>        ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
     // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method

llama/llama.cpp/src/llama-kv-cells.h

@@ -5,9 +5,27 @@
 
 #include <bitset>
 #include <cassert>
-#include <vector>
-#include <set>
+#include <cstring>
 #include <map>
+#include <set>
+#include <vector>
+
+struct llama_kv_cell_ext {
+    // 2D spatial positions, typically used for M-RoPE
+    llama_pos x = 0;
+    llama_pos y = 0;
+
+    // return true if the current 2D spatial position is greater than other
+    bool is_2d_gt(llama_pos ox, llama_pos oy) const {
+        return (y > oy) || (y == oy && x > ox);
+    }
+
+    void reset() {
+        static_assert(std::is_trivially_copyable_v<llama_kv_cell_ext>);
+
+        memset(this, 0, sizeof(*this));
+    }
+};
 
 // meta information about KV cells that can be part of multiple sequences at the same time
 // TODO: add unit tests
@@ -16,6 +34,7 @@ public:
     void reset() {
         for (uint32_t i = 0; i < pos.size(); ++i) {
             pos[i]   = -1;
+            ext[i].reset();
             shift[i] =  0;
             seq[i].reset();
         }
@@ -43,6 +62,7 @@ public:
 
     void resize(uint32_t n) {
         pos.resize(n);
+        ext.resize(n);
         shift.resize(n);
         seq.resize(n);
 
@@ -108,6 +128,7 @@ public:
             const auto idx = i + j;
 
             res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
             res.seq[j] = seq[idx];
 
             assert(shift[idx] == 0);
@@ -126,6 +147,7 @@ public:
             const auto idx = idxs[j];
 
             res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
             res.seq[j] = seq[idx];
 
             assert(shift[idx] == 0);
@@ -154,6 +176,7 @@ public:
             }
 
             pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
             seq[idx] = other.seq[j];
 
             if (pos[idx] != -1) {
@@ -184,6 +207,7 @@ public:
             }
 
             pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
             seq[idx] = other.seq[j];
 
             if (pos[idx] != -1) {
@@ -203,6 +227,7 @@ public:
         seq[i].reset();
 
         pos[i] = -1;
+        ext[i].reset();
         shift[i] = 0;
 
         used.erase(i);
@@ -221,6 +246,7 @@ public:
 
         if (seq[i].none()) {
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -250,6 +276,7 @@ public:
             seq[i].reset();
 
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -340,6 +367,13 @@ public:
         return pos[i];
     }
 
+    const llama_kv_cell_ext & ext_get(uint32_t i) const {
+        assert(i < pos.size());
+        assert(pos[i] != -1);
+
+        return ext[i];
+    }
+
     // note: call only if the cell is not empty
     llama_pos get_shift(uint32_t i) const {
         assert(i < pos.size());
@@ -368,6 +402,11 @@ public:
         used.insert(i);
     }
 
+    void ext_set(uint32_t i, llama_kv_cell_ext p) {
+        assert(i < ext.size());
+        ext[i] = p;
+    }
+
     // pos[i] = pos[i] + d
     // sets "has_shift" to true
     // note: call only if the cell is not empty
@@ -424,6 +463,9 @@ private:
 
     std::vector<llama_pos> pos;
 
+    // stores extra info per cell
+    std::vector<llama_kv_cell_ext> ext;
+
     // this array accumulates any applied shifts to the pos array since the last reset_shift() call
     // this is used to queue multiple updates to the pos array, which in the end can be applied in one go:
     //

llama/llama.cpp/src/llama-memory-recurrent.cpp

@@ -7,6 +7,7 @@
 
 #include <algorithm>
 #include <cassert>
+#include <cstring>
 #include <limits>
 #include <map>
 #include <stdexcept>
@@ -32,8 +33,15 @@ llama_memory_recurrent::llama_memory_recurrent(
     cells.clear();
     cells.resize(mem_size);
 
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     // create a context for each buffer type
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -48,13 +56,12 @@ llama_memory_recurrent::llama_memory_recurrent(
                 return nullptr;
             }
 
-            ctx_map[buft] = ctx;
-            ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
 
             return ctx;
         }
 
-        return it->second;
+        return it->second.get();
     };
 
     r_l.resize(n_layer);
@@ -93,17 +100,14 @@ llama_memory_recurrent::llama_memory_recurrent(
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
-    for (auto it : ctx_map) {
-        auto * buft = it.first;
-        auto * ctx  = it.second;
-
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+    for (auto & [buft, ctx] : ctx_map) {
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for rs cache");
         }
         ggml_backend_buffer_clear(buf, 0);
         LLAMA_LOG_INFO("%s: %10s RS buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
-        bufs.emplace_back(buf);
+        ctxs_bufs.emplace_back(std::move(ctx), buf);
     }
 
     {
@@ -129,7 +133,7 @@ void llama_memory_recurrent::clear(bool data) {
     used = 0;
 
     if (data) {
-        for (auto & buf : bufs) {
+        for (auto & [_, buf] : ctxs_bufs) {
             ggml_backend_buffer_clear(buf.get(), 0);
         }
     }
@@ -147,7 +151,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
-    // models like Mamba or RWKV can't have a state partially erased
+    // models like Mamba or RWKV can't have a state partially erased at the end
+    // of the sequence because their state isn't preserved for previous tokens
     if (seq_id >= (int64_t) size) {
         // could be fatal
         return false;
@@ -156,8 +161,8 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         int32_t & tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
             const auto & cell = cells[tail_id];
-            // partial intersection is invalid
-            if ((0 < p0 && p0 < cell.pos) || (0 < p1 && p1 <= cell.pos)) {
+            // partial intersection is invalid if it includes the final pos
+            if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
                 //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
                 return false;
             }
@@ -364,8 +369,8 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_memory_recurrent::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const ggml_backend_buffer_ptr & buf_ptr : bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -662,7 +667,7 @@ bool llama_memory_recurrent::get_can_shift() const {
 
 size_t llama_memory_recurrent::total_size() const {
     size_t size = 0;
-    for (const auto & buf : bufs) {
+    for (const auto & [_, buf] : ctxs_bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
     }
 

llama/llama.cpp/src/llama-memory-recurrent.h

@@ -109,8 +109,8 @@ private:
 
     const uint32_t n_seq_max = 1;
 
-    std::vector<ggml_context_ptr>        ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
 
     size_t total_size() const;
 

llama/llama.cpp/src/llama-model.h

@@ -77,6 +77,7 @@ enum llm_type {
     LLM_TYPE_16B,
     LLM_TYPE_20B,
     LLM_TYPE_22B,
+    LLM_TYPE_26B,
     LLM_TYPE_27B,
     LLM_TYPE_30B,
     LLM_TYPE_32B,
@@ -113,8 +114,10 @@ enum llm_type {
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air
+    LLM_TYPE_230B_A10B, // Minimax M2
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big
     LLM_TYPE_355B_A32B, // GLM-4.5
@@ -234,6 +237,7 @@ struct llama_layer {
     struct ggml_tensor * wk_enc    = nullptr;
     struct ggml_tensor * wv_enc    = nullptr;
     struct ggml_tensor * wo_enc    = nullptr;
+    struct ggml_tensor * wqkv_gate = nullptr;
 
     // attention bias
     struct ggml_tensor * bq   = nullptr;
@@ -307,6 +311,9 @@ struct llama_layer {
     struct ggml_tensor * ssm_conv1d_b = nullptr;
     struct ggml_tensor * ssm_dt_b     = nullptr;
 
+    // qwen3next
+    struct ggml_tensor * ssm_beta_alpha = nullptr;
+
     // rwkv
     struct ggml_tensor * time_mix_w1         = nullptr;
     struct ggml_tensor * time_mix_w2         = nullptr;
@@ -385,6 +392,13 @@ struct llama_layer {
     // openai-moe
     struct ggml_tensor * attn_sinks = nullptr;
 
+    // cogvlm
+    struct ggml_tensor * visexp_attn_wqkv = nullptr;
+    struct ggml_tensor * visexp_attn_wo   = nullptr;
+    struct ggml_tensor * visexp_ffn_gate  = nullptr;
+    struct ggml_tensor * visexp_ffn_down  = nullptr;
+    struct ggml_tensor * visexp_ffn_up    = nullptr;
+
     // xIELU activation parameters for Apertus
     struct ggml_tensor * ffn_act_alpha_n = nullptr;
     struct ggml_tensor * ffn_act_alpha_p = nullptr;
@@ -503,9 +517,8 @@ struct llama_model {
 
     ggml_tensor * get_rope_factors(const llama_cparams & cparams, int il) const;
 
-    // note: can mutate `cparams`
     // TODO: move this to new llm_arch_model_i interface
-    llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const;
+    llama_memory_i * create_memory(const llama_memory_params & params, const llama_cparams & cparams) const;
 
     // TODO: move this to new llm_arch_model_i interface
     ggml_cgraph * build_graph(const llm_graph_params & params) const;

llama/llama.cpp/src/llama-quant.cpp

@@ -653,7 +653,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                 gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
                 // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
-                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
+                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)std::abs(o.val_i64));
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
                 gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {
@@ -681,7 +681,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             }
             LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
             continue;
-        } else if (remapped_name != it.first) {
+        }
+
+        if (remapped_name != it.first) {
             ggml_set_name(it.second.tensor, remapped_name.c_str());
             LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
         }
@@ -726,13 +728,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     {
         const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
         // attention layers have a non-zero number of kv heads
-        int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
+        int32_t n_layer_attn = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
         if (llama_model_has_encoder(&model)) {
-            // now n_attn_layer is the number of attention layers in the encoder
+            // now n_layer_attn is the number of attention layers in the encoder
             // for each decoder block, there are 2 attention layers
-            n_attn_layer += 2 * model.hparams.dec_n_layer;
+            n_layer_attn += 2 * model.hparams.dec_n_layer;
         }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
+
+        // note: for linear-attention models (such as Qwen3 Next) this is the number of linear layers
+        const int32_t n_layer_recr = std::count(model.hparams.recurrent_layer_arr.begin(), model.hparams.recurrent_layer_arr.end(), true);
+
+        LLAMA_LOG_INFO("%s: n_layer_attn = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_attn, n_layer_recr, pruned_attention_w);
+
+        GGML_ASSERT((qs.n_attention_wv == n_layer_attn - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected");
     }
 
     size_t total_size_org = 0;

llama/llama.cpp/src/llama-sampling.cpp

@@ -4,6 +4,7 @@
 #include "llama-vocab.h"
 #include "llama-grammar.h"
 
+#include <array>
 #include <algorithm>
 #include <cassert>
 #include <cfloat>
@@ -471,9 +472,6 @@ static void llama_sampler_chain_reset(struct llama_sampler * smpl) {
     for (auto * smpl : chain->samplers) {
         llama_sampler_reset(smpl);
     }
-
-    chain->t_sample_us = 0;
-    chain->n_sample    = 0;
 }
 
 static struct llama_sampler * llama_sampler_chain_clone(const struct llama_sampler * smpl) {
@@ -1625,10 +1623,12 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
     auto * ctx = new llama_sampler_grammar;
 
     if (grammar_str != nullptr && grammar_str[0] != '\0') {
+        std::string trigger_pattern;
+        llama_grammar * grammar = nullptr;
         // TODO: remove trigger_words support.
         if (trigger_words != nullptr && num_trigger_words > 0) {
             GGML_ASSERT(trigger_patterns == nullptr && num_trigger_patterns == 0);
-            std::string trigger_pattern("[\\s\\S]*?(");
+            trigger_pattern = "[\\s\\S]*?(";
             for (size_t i = 0; i < num_trigger_words; ++i) {
                 static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
                 if (i > 0) {
@@ -1637,15 +1637,17 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                 trigger_pattern += std::regex_replace(trigger_words[i], special_chars, "\\$0");
             }
             trigger_pattern += ")[\\s\\S]*";
-            const auto * trigger_pattern_c = trigger_pattern.c_str();
-            trigger_patterns = &trigger_pattern_c;
-            num_trigger_patterns = 1;
+
+            std::array<const char *, 1> tmp_trigger_patterns = { trigger_pattern.c_str() };
+            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, tmp_trigger_patterns.data(), tmp_trigger_patterns.size(), trigger_tokens, num_trigger_tokens);
+        } else {
+            grammar = llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens);
         }
         *ctx = {
             /* .vocab        = */ vocab,
             /* .grammar_str  = */ grammar_str,
             /* .grammar_root = */ grammar_root,
-            /* .grammar      = */ llama_grammar_init_impl(vocab, nullptr, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens),
+            /* .grammar      = */ grammar,
         };
         if (!ctx->grammar) {
             delete ctx;
@@ -2665,8 +2667,7 @@ struct llama_perf_sampler_data llama_perf_sampler(const struct llama_sampler * c
 void llama_perf_sampler_print(const struct llama_sampler * chain) {
     const auto data = llama_perf_sampler(chain);
 
-    LLAMA_LOG_INFO("%s:    sampling time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, data.t_sample_ms, data.n_sample, data.t_sample_ms / data.n_sample, 1e3 / data.t_sample_ms * data.n_sample);
+    LLAMA_LOG_INFO("%s:    samplers time = %10.2f ms / %5d runs\n", __func__, data.t_sample_ms, data.n_sample);
 }
 
 void llama_perf_sampler_reset(struct llama_sampler * chain) {
@@ -2676,5 +2677,6 @@ void llama_perf_sampler_reset(struct llama_sampler * chain) {
 
     auto * ctx = (struct llama_sampler_chain *) chain->ctx;
 
-    ctx->t_sample_us = ctx->n_sample = 0;
+    ctx->t_sample_us = 0;
+    ctx->n_sample    = 0;
 }

llama/llama.cpp/src/llama-vocab.cpp

@@ -401,6 +401,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_GPT4O:
+            case LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2:
                 regex_exprs = {
                     // original regex from tokenizer.json
                     // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -442,6 +443,17 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_AFMOE:
+                regex_exprs = {
+                    // Digit handling - uses custom implementation in unicode.cpp
+                    // Groups digits with leading 1-2 based on total length modulo 3
+                    "\\p{AFMoE_digits}",
+                    // CJK and Asian scripts (using direct Unicode literals)
+                    "[一-鿿㐀-䶿豈-﫿぀-ゟ゠-ヿ･-ﾟ⼀-⿟เ-๿຀-໿ក-៿က-႟ꩠ-ꩿꧠ-꧿가-힯ᄀ-ᇿ]+",
+                    // Main BPE pattern
+                    "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\\r\\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1012,7 +1024,7 @@ private:
         }
     private:
         uint32_t get_node(size_t index) {
-            if (index > xcda_array_size) {
+            if (index >= xcda_array_size) {
                 throw std::runtime_error("Index out of array bounds in XCDA array!");
             }
             return xcda_array[index];
@@ -1269,6 +1281,7 @@ struct llm_tokenizer_plamo2 : llm_tokenizer {
 
         // Build suffix list in lexicographical order of reversed strings
         std::vector<std::string> suffixes;
+        suffixes.reserve(suffix_to_score.size() + 1);
         for (const auto & pair : suffix_to_score) {
             suffixes.push_back(pair.first);
         }
@@ -1981,6 +1994,14 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "grok-2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "afmoe") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_AFMOE;
+                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "minimax-m2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
+                clean_spaces = false;
             } else {
                 LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
                 pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;

llama/llama.cpp/src/llama-vocab.h

@@ -49,6 +49,8 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE   = 38,
     LLAMA_VOCAB_PRE_TYPE_GROK_2          = 39,
     LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
+    LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2      = 41,
+    LLAMA_VOCAB_PRE_TYPE_AFMOE           = 42,
 };
 
 struct LLM_KV;

llama/llama.cpp/src/llama.go

@@ -5,4 +5,8 @@ package llama
 // #cgo CPPFLAGS: -I${SRCDIR}/../../../ml/backend/ggml/ggml/include
 // #cgo windows CPPFLAGS: -D_WIN32_WINNT=0x0602
 import "C"
-import _ "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
+
+import (
+	_ "github.com/ollama/ollama/llama/llama.cpp/src/models"
+	_ "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
+)

llama/llama.cpp/src/models/afmoe.cpp

+#include "models.h"
+
+llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // MuP scaling: embeddings * sqrt(hidden_size)
+    // mup_enabled = true, hidden_size = 1024, scale = 32.0
+    inpL = ggml_scale(ctx0, inpL, sqrtf(float(n_embd)));
+    cb(inpL, "inp_embd_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+    auto * inp_attn = build_attn_inp_kv_iswa();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // dual attention normalization (pre)
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * attn_inp = cur;  // save input for gate computation
+
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            // compute gate from input
+            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
+            cb(gate, "attn_gate_proj", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+            // Q/K normalization
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            // RoPE only for sliding_attention layers
+            const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
+                                ((il + 1) % hparams.n_no_rope_layer_step) != 0;
+            if (use_rope) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur_rope", il);
+            }
+
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            cur = build_attn(inp_attn,
+                    NULL, NULL,  // wo will be applied after gating
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+
+            // attention gating: attn_out * sigmoid(gate) BEFORE o_proj
+            gate = ggml_sigmoid(ctx0, gate);
+            cb(gate, "attn_gate_sig", il);
+            cur = ggml_mul(ctx0, cur, gate);
+            cb(cur, "attn_gated", il);
+
+            // now apply output projection
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            cb(cur, "attn_o_proj", il);
+        }
+
+        // dual attention normalization (post)
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // dual ffn normalization (pre)
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // MoE or dense FFN
+        if ((uint32_t)il >= hparams.n_layer_dense_lead) {
+            // MoE layer with sigmoid routing, normalization, and scaling
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU,
+                    hparams.expert_weights_norm,           // norm_w (route_norm=True)
+                    hparams.expert_weights_scale,          // scale_w
+                    hparams.expert_weights_scale,          // w_scale (route_scale=2.826)
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // shared expert
+            if (hparams.n_expert_shared > 0) {
+                ggml_tensor * ffn_shexp = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            } else {
+                cur = moe_out;
+            }
+        } else {
+            // dense layer
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // dual ffn normalization (post)
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/apertus.cpp

+#include "models.h"
+
+
+
+llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos  = build_inp_pos();
+    auto *        inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+            cb(Vcur, "Vcur_pos", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network with xIELU activation
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // Up projection
+            ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);
+            cb(up, "ffn_up", il);
+
+            float alpha_n_val = hparams.xielu_alpha_n[il];
+            float alpha_p_val = hparams.xielu_alpha_p[il];
+            float beta_val    = hparams.xielu_beta[il];
+            float eps_val     = hparams.xielu_eps[il];
+
+            // Apply xIELU activation
+            ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);
+            cb(activated, "ffn_xielu", il);
+
+            // Down projection
+            cur = build_lora_mm(model.layers[il].ffn_down, activated);
+            cb(cur, "ffn_down", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}