init impl

src/llama-arch.cpp

@@ -108,6 +108,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
     { LLM_ARCH_BAILINGMOE2,      "bailingmoe2"      },
     { LLM_ARCH_DOTS1,            "dots1"            },
+    { LLM_ARCH_DOTS3NOTE,        "dots3note"        },
     { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_AFMOE,            "afmoe"            },
     { LLM_ARCH_LAGUNA,           "laguna"           },
@@ -263,6 +264,9 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
     { LLM_KV_ATTENTION_KEY_LENGTH_SWA,               "%s.attention.key_length_swa"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_SWA,             "%s.attention.value_length_swa"             },
+    { LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,           "%s.attention.key_length_mla_swa"           },
+    { LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA,         "%s.attention.value_length_mla_swa"         },
+    { LLM_KV_ATTENTION_KV_LORA_RANK_SWA,             "%s.attention.kv_lora_rank_swa"             },
     { LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,           "%s.attention.indexer.head_count"           },
     { LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,           "%s.attention.indexer.key_length"           },
     { LLM_KV_ATTENTION_INDEXER_TOP_K,                "%s.attention.indexer.top_k"                },
@@ -1023,6 +1027,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_DEEPSEEK32:
         case LLM_ARCH_DEEPSEEK4:
+        case LLM_ARCH_DOTS3NOTE:
         case LLM_ARCH_GLM_DSA:
         case LLM_ARCH_BITNET:
         case LLM_ARCH_T5:

src/llama-arch.h

@@ -113,6 +113,7 @@ enum llm_arch {
     LLM_ARCH_BAILINGMOE,
     LLM_ARCH_BAILINGMOE2,
     LLM_ARCH_DOTS1,
+    LLM_ARCH_DOTS3NOTE,
     LLM_ARCH_ARCEE,
     LLM_ARCH_AFMOE,
     LLM_ARCH_LAGUNA,
@@ -268,6 +269,9 @@ enum llm_kv {
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
     LLM_KV_ATTENTION_KEY_LENGTH_SWA,
     LLM_KV_ATTENTION_VALUE_LENGTH_SWA,
+    LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA,
+    LLM_KV_ATTENTION_KV_LORA_RANK_SWA,
     LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,
     LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
     LLM_KV_ATTENTION_INDEXER_TOP_K,

src/llama-graph.cpp

@@ -559,6 +559,12 @@ void llm_graph_input_attn_k_dsa::set_input(const llama_ubatch * ubatch) {
 
     mctx->get_mla()->set_input_kq_mask(self_kq_mask_mla, ubatch, cparams.causal_attn);
 
+    if (mctx->get_mla_swa()) {
+        mctx->get_mla_swa()->set_input_k_idxs(self_k_idxs_mla_swa, ubatch);
+
+        mctx->get_mla_swa()->set_input_kq_mask(self_kq_mask_mla_swa, ubatch, cparams.causal_attn);
+    }
+
     mctx->get_lid()->set_input_k_idxs(self_k_idxs_lid, ubatch);
 
     mctx->get_lid()->set_input_kq_mask(self_kq_mask_lid, ubatch, cparams.causal_attn);
@@ -579,6 +585,11 @@ bool llm_graph_input_attn_k_dsa::can_reuse(const llm_graph_params & params) {
     res &= can_reuse_kq_mask(self_kq_mask_mla, mctx->get_mla(), params.ubatch, params.cparams);
     res &= can_reuse_kq_mask(self_kq_mask_lid, mctx->get_lid(), params.ubatch, params.cparams);
 
+    if (mctx->get_mla_swa()) {
+        res &= self_k_idxs_mla_swa && self_k_idxs_mla_swa->ne[0] == params.ubatch.n_tokens;
+        res &= self_kq_mask_mla_swa && can_reuse_kq_mask(self_kq_mask_mla_swa, mctx->get_mla_swa(), params.ubatch, params.cparams);
+    }
+
     return res;
 }
 
@@ -2925,48 +2936,57 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_build_forward_expand(gf, v_cur);
     ggml_build_forward_expand(gf, k_cur);
 
-    const auto * mctx_cur = inp->mctx->get_mla();
+    const bool is_swa = hparams.is_swa(il);
+
+    // SWA layers use the dedicated window cache and have no indexer, so no top-k selection
+    GGML_ASSERT(!is_swa || top_k == nullptr);
+
+    const auto * mctx_cur = is_swa ? inp->mctx->get_mla_swa() : inp->mctx->get_mla();
 
     // store to KV cache
     {
-        const auto & k_idxs = inp->get_k_idxs_mla();
+        const auto & k_idxs = is_swa ? inp->get_k_idxs_mla_swa() : inp->get_k_idxs_mla();
 
         ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
     }
 
-    const auto & kq_mask = inp->get_kq_mask_mla();
+    const auto & kq_mask = is_swa ? inp->get_kq_mask_mla_swa() : inp->get_kq_mask_mla();
 
-    // prepare new kq mask - starts filled with -INFINITY
-    ggml_tensor * kq_mask_all = ggml_fill(ctx0, kq_mask, -INFINITY);
+    ggml_tensor * kq_mask_used = kq_mask;
 
-    // reshape KQ mask into tensor with rows of size 1:
-    // [n_kv, n_batch, 1, n_stream] -> [1, n_kv, n_batch, n_stream]
-    kq_mask_all = ggml_view_4d(ctx0, kq_mask_all, 1, kq_mask_all->ne[0], kq_mask_all->ne[1], kq_mask_all->ne[3], kq_mask_all->nb[0], kq_mask_all->nb[1], kq_mask_all->nb[2], 0);
+    if (top_k) {
+        // prepare new kq mask - starts filled with -INFINITY
+        ggml_tensor * kq_mask_all = ggml_fill(ctx0, kq_mask, -INFINITY);
 
-    // reshape top_k indices: [n_top_k, n_batch, 1, n_stream] -> [n_top_k, n_batch, n_stream, 1]
-    ggml_tensor * top_k_3d = ggml_view_4d(ctx0, top_k, top_k->ne[0], top_k->ne[1], top_k->ne[3], 1, top_k->nb[1], top_k->nb[2], top_k->ne[3]*top_k->nb[3], 0);
+        // reshape KQ mask into tensor with rows of size 1:
+        // [n_kv, n_batch, 1, n_stream] -> [1, n_kv, n_batch, n_stream]
+        kq_mask_all = ggml_view_4d(ctx0, kq_mask_all, 1, kq_mask_all->ne[0], kq_mask_all->ne[1], kq_mask_all->ne[3], kq_mask_all->nb[0], kq_mask_all->nb[1], kq_mask_all->nb[2], 0);
 
-    // prepare zero-filled tensor with rows of size 1: [1, n_top_k, n_batch, n_stream]
-    // this will be our source of zero values for unmasking top k mask elements
-    ggml_tensor * zeros = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, top_k_3d->ne[0], top_k_3d->ne[1], top_k_3d->ne[2]);
-    zeros = ggml_fill(ctx0, zeros, 0.0f);
+        // reshape top_k indices: [n_top_k, n_batch, 1, n_stream] -> [n_top_k, n_batch, n_stream, 1]
+        ggml_tensor * top_k_3d = ggml_view_4d(ctx0, top_k, top_k->ne[0], top_k->ne[1], top_k->ne[3], 1, top_k->nb[1], top_k->nb[2], top_k->ne[3]*top_k->nb[3], 0);
 
-    // modify KQ mask by unmasking elements that are in top_k indices
-    // ggml_set_rows([1, n_kv, n_batch, n_stream], [1, n_top_k, n_batch, n_stream], [n_top_k, n_batch, n_stream, 1])
-    ggml_tensor * kq_mask_top_k = ggml_set_rows(ctx0, kq_mask_all, zeros, top_k_3d);
+        // prepare zero-filled tensor with rows of size 1: [1, n_top_k, n_batch, n_stream]
+        // this will be our source of zero values for unmasking top k mask elements
+        ggml_tensor * zeros = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, top_k_3d->ne[0], top_k_3d->ne[1], top_k_3d->ne[2]);
+        zeros = ggml_fill(ctx0, zeros, 0.0f);
 
-    // reshape to restore the original shape of KQ mask:
-    // [1, n_kv, n_batch, n_stream] -> [n_kv, n_batch, 1, n_stream]
-    kq_mask_top_k = ggml_view_4d(ctx0, kq_mask_top_k, kq_mask_top_k->ne[1], kq_mask_top_k->ne[2], 1, kq_mask_top_k->ne[3], kq_mask_top_k->nb[2], kq_mask_top_k->nb[3], kq_mask_top_k->nb[3], 0);
+        // modify KQ mask by unmasking elements that are in top_k indices
+        // ggml_set_rows([1, n_kv, n_batch, n_stream], [1, n_top_k, n_batch, n_stream], [n_top_k, n_batch, n_stream, 1])
+        ggml_tensor * kq_mask_top_k = ggml_set_rows(ctx0, kq_mask_all, zeros, top_k_3d);
 
-    // combine with the original kq mask
-    kq_mask_top_k = ggml_add(ctx0, kq_mask_top_k, kq_mask);
+        // reshape to restore the original shape of KQ mask:
+        // [1, n_kv, n_batch, n_stream] -> [n_kv, n_batch, 1, n_stream]
+        kq_mask_top_k = ggml_view_4d(ctx0, kq_mask_top_k, kq_mask_top_k->ne[1], kq_mask_top_k->ne[2], 1, kq_mask_top_k->ne[3], kq_mask_top_k->nb[2], kq_mask_top_k->nb[3], kq_mask_top_k->nb[3], 0);
+
+        // combine with the original kq mask
+        kq_mask_used = ggml_add(ctx0, kq_mask_top_k, kq_mask);
+    }
 
     ggml_tensor * q = q_cur;
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask_top_k, sinks, v_mla, kq_scale, il);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask_used, sinks, v_mla, kq_scale, il);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -3207,6 +3227,13 @@ llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {
         inp->self_kq_mask_mla_cnv = inp->self_kq_mask_mla;
     }
 
+    if (mctx_cur->get_mla_swa()) {
+        inp->self_k_idxs_mla_swa = mctx_cur->get_mla_swa()->build_input_k_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask_mla_swa = build_attn_inp_kq_mask(ctx0, mctx_cur->get_mla_swa(), ubatch, cparams);
+        inp->self_kq_mask_mla_swa_cnv = inp->self_kq_mask_mla_swa;
+    }
+
     {
         inp->self_k_idxs_lid = mctx_cur->get_lid()->build_input_k_idxs(ctx0, ubatch);
 

src/llama-graph.h

@@ -404,19 +404,24 @@ public:
 
     bool can_reuse(const llm_graph_params & params) override;
 
-    ggml_tensor * get_k_idxs_mla() const { return self_k_idxs_mla; }
-    ggml_tensor * get_k_idxs_lid() const { return self_k_idxs_lid; }
+    ggml_tensor * get_k_idxs_mla()     const { return self_k_idxs_mla; }
+    ggml_tensor * get_k_idxs_mla_swa() const { return self_k_idxs_mla_swa; }
+    ggml_tensor * get_k_idxs_lid()     const { return self_k_idxs_lid; }
 
-    ggml_tensor * get_kq_mask_mla() const { return self_kq_mask_mla_cnv; }
-    ggml_tensor * get_kq_mask_lid() const { return self_kq_mask_lid; }
+    ggml_tensor * get_kq_mask_mla()     const { return self_kq_mask_mla_cnv; }
+    ggml_tensor * get_kq_mask_mla_swa() const { return self_kq_mask_mla_swa_cnv; }
+    ggml_tensor * get_kq_mask_lid()     const { return self_kq_mask_lid; }
 
-    ggml_tensor * self_k_idxs_mla = nullptr; // I64 [n_batch]
-    ggml_tensor * self_k_idxs_lid = nullptr; // I64 [n_batch]
+    ggml_tensor * self_k_idxs_mla     = nullptr; // I64 [n_batch]
+    ggml_tensor * self_k_idxs_mla_swa = nullptr; // I64 [n_batch]
+    ggml_tensor * self_k_idxs_lid     = nullptr; // I64 [n_batch]
 
-    ggml_tensor * self_kq_mask_mla     = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
-    ggml_tensor * self_kq_mask_mla_cnv = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]
-    ggml_tensor * self_kq_mask_lid     = nullptr; // F32     [n_kv, n_batch/n_stream, 1, n_stream]
-    ggml_tensor * self_kq_mask_lid_cnv = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_mla         = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_mla_cnv     = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_mla_swa     = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_mla_swa_cnv = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_lid         = nullptr; // F32     [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_lid_cnv     = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]
 
     ggml_tensor * self_k_rot_lid = nullptr;
 

src/llama-hparams.h

@@ -100,6 +100,11 @@ struct llama_hparams {
     uint32_t n_group_used       = 0;
     uint32_t n_group_experts    = 0;
 
+    // MLA + SWA (i.e. dots3note)
+    uint32_t n_lora_kv_swa           = 0;
+    uint32_t n_embd_head_k_mla_swa   = 0;
+    uint32_t n_embd_head_v_mla_swa   = 0;
+
     float    expert_group_scale   = 0.05f;
     float    expert_weights_scale = 0.0f;
     bool     expert_weights_norm  = false;

src/llama-kv-cache-dsa.cpp

@@ -20,6 +20,7 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
                      bool   unified,
                  uint32_t   kv_size,
                  uint32_t   n_seq_max,
+                 uint32_t   n_ubatch,
                  uint32_t   n_pad,
                  uint32_t   n_swa,
            llama_swa_type   swa_type,
@@ -28,12 +29,41 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
     const  layer_reuse_cb & reuse) :
     hparams_lid(model.hparams), n_stream(unified ? 1 : n_seq_max) {
 
+    const bool has_swa = swa_type != LLAMA_SWA_TYPE_NONE;
+
+    const layer_filter_cb filter_mla_full = [&](int32_t il) {
+        if (filter_mla && !filter_mla(il)) {
+            return false;
+        }
+
+        return !(has_swa && model.hparams.is_swa(il));
+    };
+
+    const layer_filter_cb filter_mla_swa = [&](int32_t il) {
+        if (filter_mla && !filter_mla(il)) {
+            return false;
+        }
+
+        return model.hparams.is_swa(il);
+    };
+
     LLAMA_LOG_INFO("%s: creating main KV cache, size = %u cells\n", __func__, kv_size);
 
     kv_mla = std::make_unique<llama_kv_cache>(
             model, model.hparams, type_k, type_v,
             v_trans, offload, unified, kv_size, n_seq_max, n_pad,
-            n_swa, swa_type, nullptr, filter_mla, reuse, nullptr);
+            0, LLAMA_SWA_TYPE_NONE, nullptr, filter_mla_full, reuse, nullptr);
+
+    if (has_swa) {
+        const uint32_t size_swa = GGML_PAD(std::min(kv_size, n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);
+
+        LLAMA_LOG_INFO("%s: creating SWA KV cache, size = %u cells\n", __func__, size_swa);
+
+        kv_mla_swa = std::make_unique<llama_kv_cache>(
+                model, model.hparams, type_k, type_v,
+                v_trans, offload, unified, size_swa, n_seq_max, n_pad,
+                n_swa, swa_type, nullptr, filter_mla_swa, reuse, nullptr);
+    }
 
     // we use llama_kv_cache for caching indexer keys
     // by hand-tweaking some hparams we fool it to create
@@ -47,14 +77,18 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
 
     LLAMA_LOG_INFO("%s: creating indexer KV cache, size = %u cells\n", __func__, kv_size);
 
+    // the lightning indexer only exists on full-attention layers, so no SWA handling here
     kv_lid = std::make_unique<llama_kv_cache>(
             model, hparams_lid, type_k, type_v,
             v_trans, offload, unified, kv_size, n_seq_max, n_pad,
-            n_swa, swa_type, nullptr, filter_lid, reuse, nullptr);
+            0, LLAMA_SWA_TYPE_NONE, nullptr, filter_lid, reuse, nullptr);
 }
 
 void llama_kv_cache_dsa::clear(bool data) {
     kv_mla->clear(data);
+    if (kv_mla_swa) {
+        kv_mla_swa->clear(data);
+    }
     kv_lid->clear(data);
 }
 
@@ -62,6 +96,9 @@ bool llama_kv_cache_dsa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1)
     bool res = true;
 
     res = res & kv_mla->seq_rm(seq_id, p0, p1);
+    if (kv_mla_swa) {
+        res = res & kv_mla_swa->seq_rm(seq_id, p0, p1);
+    }
     res = res & kv_lid->seq_rm(seq_id, p0, p1);
 
     return res;
@@ -69,21 +106,33 @@ bool llama_kv_cache_dsa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1)
 
 void llama_kv_cache_dsa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
     kv_mla->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    if (kv_mla_swa) {
+        kv_mla_swa->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    }
     kv_lid->seq_cp(seq_id_src, seq_id_dst, p0, p1);
 }
 
 void llama_kv_cache_dsa::seq_keep(llama_seq_id seq_id) {
     kv_mla->seq_keep(seq_id);
+    if (kv_mla_swa) {
+        kv_mla_swa->seq_keep(seq_id);
+    }
     kv_lid->seq_keep(seq_id);
 }
 
 void llama_kv_cache_dsa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     kv_mla->seq_add(seq_id, p0, p1, shift);
+    if (kv_mla_swa) {
+        kv_mla_swa->seq_add(seq_id, p0, p1, shift);
+    }
     kv_lid->seq_add(seq_id, p0, p1, shift);
 }
 
 void llama_kv_cache_dsa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     kv_mla->seq_div(seq_id, p0, p1, d);
+    if (kv_mla_swa) {
+        kv_mla_swa->seq_div(seq_id, p0, p1, d);
+    }
     kv_lid->seq_div(seq_id, p0, p1, d);
 }
 
@@ -97,6 +146,11 @@ llama_pos llama_kv_cache_dsa::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_dsa::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> mb = kv_mla->memory_breakdown();
+    if (kv_mla_swa) {
+        for (const auto & buft_size : kv_mla_swa->memory_breakdown()) {
+            mb[buft_size.first] += buft_size.second;
+        }
+    }
     for (const auto & buft_size : kv_lid->memory_breakdown()) {
         mb[buft_size.first] += buft_size.second;
     }
@@ -133,6 +187,14 @@ llama_memory_context_ptr llama_kv_cache_dsa::init_batch(
             break;
         }
 
+        llama_kv_cache::slot_info_vec_t sinfos_mla_swa;
+        if (kv_mla_swa) {
+            sinfos_mla_swa = kv_mla_swa->prepare(ubatches);
+            if (sinfos_mla_swa.empty()) {
+                break;
+            }
+        }
+
         auto sinfos_lid = kv_lid->prepare(ubatches);
         if (sinfos_lid.empty()) {
             break;
@@ -141,7 +203,7 @@ llama_memory_context_ptr llama_kv_cache_dsa::init_batch(
         assert(sinfos_mla.size() == sinfos_lid.size());
 
         return std::make_unique<llama_kv_cache_dsa_context>(
-                this, std::move(sinfos_mla), std::move(sinfos_lid), std::move(ubatches));
+                this, std::move(sinfos_mla), std::move(sinfos_mla_swa), std::move(sinfos_lid), std::move(ubatches));
     } while (false);
 
     return std::make_unique<llama_kv_cache_dsa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
@@ -157,17 +219,24 @@ llama_memory_context_ptr llama_kv_cache_dsa::init_update(llama_context * lctx, b
 
 bool llama_kv_cache_dsa::get_can_shift() const {
     return kv_mla->get_can_shift() &&
+           (!kv_mla_swa || kv_mla_swa->get_can_shift()) &&
            kv_lid->get_can_shift() &&
            kv_mla->get_size() == kv_lid->get_size();
 }
 
 void llama_kv_cache_dsa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
     kv_mla->state_write(io, seq_id, flags);
+    if (kv_mla_swa) {
+        kv_mla_swa->state_write(io, seq_id, flags);
+    }
     kv_lid->state_write(io, seq_id, flags);
 }
 
 void llama_kv_cache_dsa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
     kv_mla->state_read(io, seq_id, flags);
+    if (kv_mla_swa) {
+        kv_mla_swa->state_read(io, seq_id, flags);
+    }
     kv_lid->state_read(io, seq_id, flags);
 }
 
@@ -175,6 +244,10 @@ llama_kv_cache * llama_kv_cache_dsa::get_mla() const {
     return kv_mla.get();
 }
 
+llama_kv_cache * llama_kv_cache_dsa::get_mla_swa() const {
+    return kv_mla_swa.get();
+}
+
 llama_kv_cache * llama_kv_cache_dsa::get_lid() const {
     return kv_lid.get();
 }
@@ -188,8 +261,11 @@ llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(llama_memory_status statu
 llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
         llama_kv_cache_dsa * kv) :
     ctx_mla(kv->get_mla()->init_full()),
+    ctx_mla_swa(kv->get_mla_swa() ? kv->get_mla_swa()->init_full() : nullptr),
     ctx_lid(kv->get_lid()->init_full()),
-    status(llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status())) {
+    status(llama_memory_status_combine(
+                llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status()),
+                ctx_mla_swa ? ctx_mla_swa->get_status() : LLAMA_MEMORY_STATUS_NO_UPDATE)) {
 }
 
 llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
@@ -197,20 +273,27 @@ llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
         llama_context * lctx,
         bool optimize) :
     ctx_mla(kv->get_mla()->init_update(lctx, optimize)),
+    ctx_mla_swa(kv->get_mla_swa() ? kv->get_mla_swa()->init_update(lctx, optimize) : nullptr),
     ctx_lid(kv->get_lid()->init_update(lctx, optimize)),
-    status(llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status())) {
+    status(llama_memory_status_combine(
+                llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status()),
+                ctx_mla_swa ? ctx_mla_swa->get_status() : LLAMA_MEMORY_STATUS_NO_UPDATE)) {
 }
 
 llama_kv_cache_dsa_context::llama_kv_cache_dsa_context(
         llama_kv_cache_dsa * kv,
         slot_info_vec_t sinfos_mla,
+        slot_info_vec_t sinfos_mla_swa,
         slot_info_vec_t sinfos_lid,
         std::vector<llama_ubatch> ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
     ctx_mla(new llama_kv_cache_context(kv->get_mla(), std::move(sinfos_mla), this->ubatches)),
+    ctx_mla_swa(kv->get_mla_swa() ? new llama_kv_cache_context(kv->get_mla_swa(), std::move(sinfos_mla_swa), this->ubatches) : nullptr),
     ctx_lid(new llama_kv_cache_context(kv->get_lid(), std::move(sinfos_lid), this->ubatches)),
-    status(llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status())) {
+    status(llama_memory_status_combine(
+                llama_memory_status_combine(ctx_mla->get_status(), ctx_lid->get_status()),
+                ctx_mla_swa ? ctx_mla_swa->get_status() : LLAMA_MEMORY_STATUS_NO_UPDATE)) {
 }
 
 llama_kv_cache_dsa_context:: ~llama_kv_cache_dsa_context() = default;
@@ -219,6 +302,9 @@ bool llama_kv_cache_dsa_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     ctx_mla->next();
+    if (ctx_mla_swa) {
+        ctx_mla_swa->next();
+    }
     ctx_lid->next();
 
     if (++i_next >= ubatches.size()) {
@@ -234,6 +320,9 @@ bool llama_kv_cache_dsa_context::apply() {
     bool res = true;
 
     res = res & ctx_mla->apply();
+    if (ctx_mla_swa) {
+        res = res & ctx_mla_swa->apply();
+    }
     res = res & ctx_lid->apply();
 
     return res;
@@ -255,6 +344,12 @@ const llama_kv_cache_context * llama_kv_cache_dsa_context::get_mla() const {
     return static_cast<const llama_kv_cache_context *>(ctx_mla.get());
 }
 
+const llama_kv_cache_context * llama_kv_cache_dsa_context::get_mla_swa() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return static_cast<const llama_kv_cache_context *>(ctx_mla_swa.get());
+}
+
 const llama_kv_cache_context * llama_kv_cache_dsa_context::get_lid()  const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 

src/llama-kv-cache-dsa.h

@@ -11,6 +11,7 @@
 // utilizes two instances of llama_kv_cache:
 // - the first instance is for caching key tensors of the model,
 // - the second instance is for caching lightning indexer key tensors
+// when swa_type != NONE, a third instance holds the SWA layers (iswa-style, window-sized)
 
 class llama_kv_cache_dsa : public llama_memory_i {
 public:
@@ -23,6 +24,7 @@ public:
                          bool   unified,
                      uint32_t   kv_size,
                      uint32_t   n_seq_max,
+                     uint32_t   n_ubatch,
                      uint32_t   n_pad,
                      uint32_t   n_swa,
                llama_swa_type   swa_type,
@@ -69,8 +71,9 @@ public:
     // llama_kv_cache_dsa specific API
     //
 
-    llama_kv_cache * get_mla() const;
-    llama_kv_cache * get_lid() const;
+    llama_kv_cache * get_mla()     const;
+    llama_kv_cache * get_mla_swa() const; // null when the model has no SWA layers
+    llama_kv_cache * get_lid()     const;
 
 private:
     // we keep indexer KV cache hparams instance here as llama_kv_cache stores only reference to it
@@ -78,6 +81,7 @@ private:
     const uint32_t n_stream  = 1;
 
     std::unique_ptr<llama_kv_cache> kv_mla;
+    std::unique_ptr<llama_kv_cache> kv_mla_swa;
     std::unique_ptr<llama_kv_cache> kv_lid;
 };
 
@@ -102,6 +106,7 @@ public:
     llama_kv_cache_dsa_context(
             llama_kv_cache_dsa * kv,
             slot_info_vec_t sinfos_base,
+            slot_info_vec_t sinfos_swa,
             slot_info_vec_t sinfos_ik,
             std::vector<llama_ubatch> ubatches);
 
@@ -121,8 +126,9 @@ public:
     // llama_kv_cache_dsa_context specific API
     //
 
-    const llama_kv_cache_context * get_mla() const;
-    const llama_kv_cache_context * get_lid()  const;
+    const llama_kv_cache_context * get_mla()     const;
+    const llama_kv_cache_context * get_mla_swa() const; // null when the model has no SWA layers
+    const llama_kv_cache_context * get_lid()     const;
 
 private:
     //llama_kv_cache_dsa * kv;
@@ -133,6 +139,7 @@ private:
     std::vector<llama_ubatch> ubatches;
 
     const llama_memory_context_ptr ctx_mla;
+    const llama_memory_context_ptr ctx_mla_swa;
     const llama_memory_context_ptr ctx_lid;
 
     const llama_memory_status status;

src/llama-kv-cache.cpp

@@ -323,7 +323,8 @@ llama_kv_cache::llama_kv_cache(
             hparams.n_embd_head_k() % 64 == 0;
 
         // always create Hadamard rotation tensors for DeepSeek lightning indexers
-        if ((model.arch == LLM_ARCH_DEEPSEEK32 || model.arch == LLM_ARCH_DEEPSEEK4 || model.arch == LLM_ARCH_GLM_DSA) &&
+        if ((model.arch == LLM_ARCH_DEEPSEEK32 || model.arch == LLM_ARCH_DEEPSEEK4 ||
+                model.arch == LLM_ARCH_GLM_DSA || model.arch == LLM_ARCH_DOTS3NOTE) &&
                 hparams.n_embd_head_k_full == hparams.indexer_head_size) {
             attn_rot_k = true;
         }

src/llama-model-saver.cpp

@@ -30,6 +30,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
         case LLM_ARCH_MUSE_GLIMMER:
         case LLM_ARCH_MELLUM:
         case LLM_ARCH_LAGUNA:
+        case LLM_ARCH_DOTS3NOTE: // TODO: need to handle SWA pattern and MLA+SWA config
             return false;
         default:
             return true;

src/llama-model.cpp

@@ -194,6 +194,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_deepseek2ocr(params);
         case LLM_ARCH_DEEPSEEK32:
             return new llama_model_deepseek32(params);
+        case LLM_ARCH_DOTS3NOTE:
+            return new llama_model_dots3note(params);
         case LLM_ARCH_DEEPSEEK4:
             return new llama_model_deepseek4(params);
         case LLM_ARCH_GLM_DSA:
@@ -1907,7 +1909,9 @@ void llama_model::print_info() const {
             LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
         }
 
-        if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR || arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MISTRAL4) {
+        if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR ||
+                arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA ||
+                arch == LLM_ARCH_DOTS3NOTE || arch == LLM_ARCH_MISTRAL4) {
             LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
             LLAMA_LOG_INFO("%s: n_lora_q              = %d\n",     __func__, hparams.n_lora_q);
             LLAMA_LOG_INFO("%s: n_lora_kv             = %d\n",     __func__, hparams.n_lora_kv);
@@ -2125,6 +2129,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
             } break;
         case LLM_ARCH_GLM_DSA:
         case LLM_ARCH_DEEPSEEK32:
+        case LLM_ARCH_DOTS3NOTE:
             {
                 if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP && hparams.n_layer_nextn > 0) {
                     // The NextN/MTP draft head runs dense MLA (no DSA indexer), so the
@@ -2157,7 +2162,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     if (hparams.n_layer_nextn > 0) {
                         filter_mla = [&](uint32_t il) { return il < hparams.n_layer(); };
                     }
-                    llama_kv_cache::layer_filter_cb filter_lid = [&](uint32_t il) { return il < hparams.n_layer() && (arch != LLM_ARCH_GLM_DSA || hparams.is_indexer_full(il)); };
+                    const bool lid_by_types = arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_DOTS3NOTE;
+                    llama_kv_cache::layer_filter_cb filter_lid = [&](uint32_t il) { return il < hparams.n_layer() && (!lid_by_types || hparams.is_indexer_full(il)); };
 
                     res = new llama_kv_cache_dsa(
                             *this,
@@ -2168,6 +2174,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                             cparams.kv_unified,
                             cparams.n_ctx_seq,
                             cparams.n_seq_max,
+                            cparams.n_ubatch,
                             1,
                             hparams.n_swa,
                             hparams.swa_type,
@@ -2641,6 +2648,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_LLAMA_EMBED:
         case LLM_ARCH_MAINCODER:
         case LLM_ARCH_GLM_DSA:
+        case LLM_ARCH_DOTS3NOTE:
         case LLM_ARCH_NANBEIGE:
         case LLM_ARCH_POCKETTTS:
             return LLAMA_ROPE_TYPE_NORM;

src/models/dots3note.cpp

@@ -0,0 +1,510 @@
+#include "models.h"
+
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-dsa.h"
+
+// note: code adapted from deepseek32.cpp (DSA indexer + absorbed MLA) and step35.cpp (head-wise output gate)
+
+void llama_model_dots3note::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    hparams.f_norm_eps = 1e-6;  // eps for the indexer k_norm layer norm
+
+    // TODO: use MTP layer
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
+    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_all");
+
+    // MoE parameters
+    ml.get_key(LLM_KV_EXPERT_COUNT,               hparams.n_expert);
+    ml.get_key(LLM_KV_EXPERT_USED_COUNT,          hparams.n_expert_used);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,  hparams.n_layer_dense_lead);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,        hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,         hparams.expert_gating_func);
+
+    // MLA parameters of the full-attention layers
+    ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,      hparams.n_lora_q);
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl);
+
+    // MLA parameters of the sliding-window layers
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK_SWA,     hparams.n_lora_kv_swa);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,   hparams.n_embd_head_k_mla_swa);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA, hparams.n_embd_head_v_mla_swa);
+
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,       hparams.rope_freq_base_train_swa);
+    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer());
+
+    // DSA parameters - by default the indexer exists exactly on the full-attention layers
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head);
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size);
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K,      hparams.indexer_top_k);
+    for (uint32_t il = 0; il < hparams.n_layer(); ++il) {
+        hparams.is_indexer_full_impl[il] = hparams.is_swa(il) ? 0 : 1;
+    }
+    ml.get_key_or_arr(LLM_KV_ATTENTION_INDEXER_TYPES, hparams.is_indexer_full_impl, hparams.n_layer(), false);
+
+    switch (hparams.n_layer()) {
+        case 46: type = LLM_TYPE_UNKNOWN; break; // 288B-A19B
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_dots3note::load_arch_tensors(llama_model_loader & ml) {
+    LLAMA_LOAD_LOCALS;
+    GGML_UNUSED(ml);
+
+    if (!hparams.is_mla()) {
+        throw std::runtime_error("DOTS3NOTE architecture requires MLA");
+    }
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+
+    const int64_t q_lora_rank     = hparams.n_lora_q;
+    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (!output) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer_all; ++i) {
+        auto & layer = layers[i];
+
+        const bool is_mtp = i >= n_layer;
+        // the NextN/MTP block uses the sliding-attention geometry
+        const bool is_swa = is_mtp || hparams.is_swa(i);
+
+        // MTP tensors are preserved in the GGUF but there is no MTP graph yet
+        const int flags = is_mtp ? TENSOR_SKIP | TENSOR_NOT_REQUIRED : 0;
+
+        const int64_t n_head_l = hparams.n_head(i);
+
+        const int64_t kv_lora_rank       = is_swa ? hparams.n_lora_kv_swa         : hparams.n_lora_kv;
+        const int64_t n_embd_head_k_mla  = is_swa ? hparams.n_embd_head_k_mla_swa : hparams.n_embd_head_k_mla();
+        const int64_t n_embd_head_v_mla  = is_swa ? hparams.n_embd_head_v_mla_swa : hparams.n_embd_head_v_mla();
+        const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+
+        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, flags);
+        layer.attn_q_a_norm  = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM,  "weight", i), {q_lora_rank}, flags);
+        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags);
+        // norm applied on the shared rope key before rope
+        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_qk_rope}, flags);
+
+        layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags);
+        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head_l * n_embd_head_k_mla}, flags);
+
+        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags);
+
+        layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head_l}, flags);
+        layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head_l}, flags);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head_l * n_embd_head_v_mla, n_embd}, flags);
+
+        // head-wise sigmoid output gate
+        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, flags);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
+
+        // DSA indexer
+        if (!is_mtp && hparams.is_indexer_full(i)) {
+            layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {hparams.indexer_head_size}, flags);
+            layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {hparams.indexer_head_size}, flags);
+            layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, hparams.indexer_n_head}, flags);
+            layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, hparams.indexer_head_size}, flags);
+            layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
+        }
+
+        if (is_mtp || i < (int) hparams.n_layer_dense_lead) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
+        } else {
+            if (n_expert == 0 || n_expert_used == 0) {
+                throw std::runtime_error("n_expert and n_expert_used must be > 0");
+            }
+
+            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, flags);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, flags);
+
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, flags);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+        }
+
+        if (is_mtp) {
+            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), { 2 * n_embd, n_embd }, flags);
+            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), { n_embd }, flags);
+            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), { n_embd }, flags);
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), { n_embd, n_vocab }, flags);
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_dots3note::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_dots3note::graph::graph(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    GGML_ASSERT(hparams.is_mla());
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+
+    const int64_t n_indexer_head           = hparams.indexer_n_head;
+    const int64_t n_embd_indexer_head      = hparams.indexer_head_size;
+    const int64_t n_embd_indexer_head_rope = hparams.n_rot();
+    const int64_t n_embd_indexer_head_nope = n_embd_indexer_head - n_embd_indexer_head_rope;
+    const uint32_t n_indexer_top_k = hparams.indexer_top_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    llm_graph_input_attn_k_dsa * inp_attn_dsa = build_attn_inp_k_dsa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        const bool is_swa = hparams.is_swa(il);
+
+        const int64_t n_head_l = hparams.n_head(il);
+
+        const int64_t kv_lora_rank        = is_swa ? hparams.n_lora_kv_swa         : hparams.n_lora_kv;
+        const int64_t n_embd_head_k_mla   = is_swa ? hparams.n_embd_head_k_mla_swa : hparams.n_embd_head_k_mla();
+        const int64_t n_embd_head_v_mla   = is_swa ? hparams.n_embd_head_v_mla_swa : hparams.n_embd_head_v_mla();
+        const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+
+        const float kq_scale    = 1.0f/sqrtf(float(n_embd_head_k_mla));
+        const float freq_base_l = model.get_rope_freq_base(cparams, il);
+
+        // norm
+        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;
+
+            ggml_tensor * qr = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+            cb(qr, "qr", il);
+
+            qr = build_norm(qr, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+            cb(qr, "qr", il);
+
+            ggml_tensor * top_k = nullptr;
+
+            // lightning indexer (full-attention layers only)
+            if (!is_swa) {
+                ggml_tensor * indexer_q = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_q_b, qr);
+                cb(indexer_q, "indexer_q", il);
+
+                // split into {n_embd_indexer_head_rope, n_indexer_head, n_tokens}
+                ggml_tensor * indexer_q_pe =
+                    ggml_view_3d(ctx0, indexer_q, n_embd_indexer_head_rope, n_indexer_head, n_tokens,
+                                 ggml_row_size(indexer_q->type, n_embd_indexer_head),
+                                 ggml_row_size(indexer_q->type, n_embd_indexer_head) * n_indexer_head, 0);
+                cb(indexer_q_pe, "indexer_q_pe", il);
+
+                // and {n_embd_indexer_head_nope, n_indexer_head, n_tokens}
+                ggml_tensor * indexer_q_nope =
+                    ggml_view_3d(ctx0, indexer_q, n_embd_indexer_head_nope, n_indexer_head, n_tokens,
+                                 ggml_row_size(indexer_q->type, n_embd_indexer_head),
+                                 ggml_row_size(indexer_q->type, n_embd_indexer_head) * n_indexer_head,
+                                 ggml_row_size(indexer_q->type, n_embd_indexer_head_nope));
+                cb(indexer_q_nope, "indexer_q_nope", il);
+
+                indexer_q_pe = ggml_rope_ext(ctx0, indexer_q_pe, inp_pos, nullptr, n_rot,
+                                     LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(indexer_q_pe, "indexer_q_pe", il);
+
+                indexer_q = ggml_concat(ctx0, indexer_q_pe, indexer_q_nope, 0);
+                cb(indexer_q, "indexer_q", il);
+
+                ggml_tensor * indexer_k = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_k, cur);
+                cb(indexer_k, "indexer_k", il);
+
+                indexer_k = build_norm(indexer_k, model.layers[il].indexer_k_norm, model.layers[il].indexer_k_norm_b, LLM_NORM, il);
+                cb(indexer_k, "indexer_k", il);
+
+                // split into {n_embd_indexer_head_rope, 1, n_tokens}
+                ggml_tensor * indexer_k_pe =
+                    ggml_view_3d(ctx0, indexer_k, n_embd_indexer_head_rope, 1, n_tokens,
+                                 ggml_row_size(indexer_k->type, n_embd_indexer_head),
+                                 ggml_row_size(indexer_k->type, n_embd_indexer_head) * 1, 0);
+                cb(indexer_k_pe, "indexer_k_pe", il);
+
+                // and {n_embd_indexer_head_nope, 1, n_tokens}
+                ggml_tensor * indexer_k_nope =
+                    ggml_view_3d(ctx0, indexer_k, n_embd_indexer_head_nope, 1, n_tokens,
+                                 ggml_row_size(indexer_k->type, n_embd_indexer_head),
+                                 ggml_row_size(indexer_k->type, n_embd_indexer_head) * 1,
+                                 ggml_row_size(indexer_k->type, n_embd_indexer_head_nope));
+                cb(indexer_k_nope, "indexer_k_nope", il);
+
+                indexer_k_pe = ggml_rope_ext(ctx0, indexer_k_pe, inp_pos, nullptr, n_rot,
+                                     LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(indexer_k_pe, "indexer_k_pe", il);
+
+                indexer_k = ggml_concat(ctx0, indexer_k_pe, indexer_k_nope, 0);
+                cb(indexer_k, "indexer_k", il);
+
+                // perform Hadamard transform on indexer q and k
+                indexer_q = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_q);
+                cb(indexer_q, "indexer_q", il);
+                indexer_k = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_k);
+                cb(indexer_k, "indexer_k", il);
+
+                // store indexer keys to KV cache
+                const auto * mctx_lid = inp_attn_dsa->mctx->get_lid();
+                const auto & k_idxs_lid = inp_attn_dsa->get_k_idxs_lid();
+                ggml_build_forward_expand(gf, mctx_lid->cpy_k(ctx0, indexer_k, k_idxs_lid, il));
+
+                ggml_tensor * indexer_weights = ggml_mul_mat(ctx0, model.layers[il].indexer_proj, cur);
+                cb(indexer_weights, "indexer_weights", il);
+
+                indexer_k = mctx_lid->get_k(ctx0, il);
+
+                // split the batch into streams if needed
+                const auto n_stream = indexer_k->ne[3];
+                indexer_q = ggml_view_4d(ctx0, indexer_q, indexer_q->ne[0], indexer_q->ne[1], indexer_q->ne[2]/n_stream, n_stream, indexer_q->nb[1], indexer_q->nb[2], indexer_q->nb[3]/n_stream, 0);
+                indexer_weights = ggml_view_4d(ctx0, indexer_weights, indexer_weights->ne[0], indexer_weights->ne[1]/n_stream, indexer_weights->ne[2], n_stream, indexer_weights->nb[1], indexer_weights->nb[2]/n_stream, indexer_weights->nb[3]/n_stream, 0);
+
+                // pre-scale weights to avoid scaling operations on huge indexer_score tensor
+                indexer_weights = ggml_scale(ctx0, indexer_weights, 1.0f / sqrtf(float(n_embd_indexer_head * n_indexer_head)));
+                cb(indexer_weights, "indexer_weights", il);
+
+                ggml_tensor * indexer_score = nullptr;
+                if (cparams.fused_lid) {
+                    indexer_score = ggml_lightning_indexer(ctx0, indexer_q, indexer_k, indexer_weights, inp_attn_dsa->get_kq_mask_lid());
+                    cb(indexer_score, "indexer_score", il);
+                    res->add_fused_node({LLM_FUSED_OP_LIGHTNING_INDEXER, indexer_score, il});
+                } else {
+                    indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3);
+                    cb(indexer_q, "indexer_q", il);
+                    indexer_k = ggml_permute(ctx0, indexer_k, 0, 2, 1, 3);
+                    cb(indexer_k, "indexer_k", il);
+
+                    ggml_tensor * indexer_kq = ggml_mul_mat(ctx0, indexer_k, indexer_q);
+                    cb(indexer_kq, "indexer_kq", il);
+
+                    // ReLU requires contiguous tensors
+                    indexer_kq = ggml_cont(ctx0, ggml_permute(ctx0, indexer_kq, 2, 1, 0, 3));
+                    cb(indexer_kq, "indexer_kq", il);
+
+                    indexer_score = ggml_relu(ctx0, indexer_kq);
+                    cb(indexer_score, "indexer_score", il);
+
+                    indexer_score = ggml_mul(ctx0, indexer_score, indexer_weights);
+                    cb(indexer_score, "indexer_score", il);
+
+                    // sum by q n_indexer_head dimension
+                    indexer_score = ggml_sum_rows(ctx0, indexer_score);
+                    cb(indexer_score, "indexer_score", il);
+
+                    // permute result to match KQ mask
+                    indexer_score = ggml_cont(ctx0, ggml_permute(ctx0, indexer_score, 2, 1, 0, 3));
+                    cb(indexer_score, "indexer_score", il);
+
+                    ggml_tensor * indexer_kq_mask = inp_attn_dsa->get_kq_mask_lid();
+                    indexer_score = ggml_add(ctx0, indexer_score, indexer_kq_mask);
+                    cb(indexer_score, "indexer_score", il);
+                }
+
+                // get indices of top k indexer scores
+                uint32_t n_top_k = indexer_score->ne[0] < n_indexer_top_k ? indexer_score->ne[0] : n_indexer_top_k;
+                top_k = ggml_cont(ctx0, ggml_top_k(ctx0, indexer_score, n_top_k));
+                cb(top_k, "top_k", il);
+            }
+
+            ggml_tensor * q = ggml_mul_mat(ctx0, model.layers[il].wq_b, qr);
+            cb(q, "q", il);
+
+            // split into {n_embd_head_qk_nope, n_head_l, n_tokens}
+            ggml_tensor * q_nope =
+                ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head_l, n_tokens, ggml_row_size(q->type, n_embd_head_k_mla),
+                             ggml_row_size(q->type, n_embd_head_k_mla) * n_head_l, 0);
+            cb(q_nope, "q_nope", il);
+
+            // and {n_embd_head_qk_rope, n_head_l, n_tokens}
+            ggml_tensor * q_pe = ggml_view_3d(
+                ctx0, q, n_embd_head_qk_rope, n_head_l, n_tokens, ggml_row_size(q->type, n_embd_head_k_mla),
+                ggml_row_size(q->type, n_embd_head_k_mla) * n_head_l, ggml_row_size(q->type, n_embd_head_qk_nope));
+            cb(q_pe, "q_pe", il);
+
+            ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+            cb(kv_cmpr_pe, "kv_cmpr_pe", il);
+
+            // split into {kv_lora_rank, n_tokens}
+            ggml_tensor * kv_cmpr =
+                ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+                             ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            // and {n_embd_head_qk_rope, 1, n_tokens}
+            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+            cb(k_pe, "k_pe", il);
+
+            // norm on the shared rope key, applied before rope
+            k_pe = build_norm(k_pe, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
+            cb(k_pe, "k_pe", il);
+
+            q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(q_pe, "q_pe", il);
+
+            k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(k_pe, "k_pe", il);
+
+            kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            // MLA attention with the absorption optimization
+            {
+                // {n_embd_head_qk_nope, n_tokens, n_head_l}
+                q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+                cb(q_nope, "q_nope_perm", il);
+
+                // {n_embd_head_qk_nope, kv_lora_rank, n_head_l} x {n_embd_head_qk_nope, n_tokens, n_head_l}
+                ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                cb(q_nope_absorbed, "q_nope_absorbed", il);
+
+                // {kv_lora_rank, n_head_l, n_tokens}
+                q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+                cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, n_head_l, n_tokens}
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+                cb(Qcur, "Qcur", il);
+
+                kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+                cb(kv_cmpr, "kv_cmpr_reshape", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+                cb(Kcur, "Kcur", il);
+
+                // {kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Vcur = kv_cmpr;
+                cb(Vcur, "Vcur", il);
+
+                // apply the head-wise output gate before o_proj, so wo stays out of build_attn
+                cur = build_attn(inp_attn_dsa,
+                        nullptr, nullptr, nullptr,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, top_k, kq_scale, il);
+                cb(cur, "attn_out", il);
+
+                ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
+                cb(gate, "attn_gate", il);
+
+                gate = ggml_sigmoid(ctx0, gate);
+                cb(gate, "attn_gate_sigmoid", il);
+
+                // broadcast the per-head gate over the head dimension
+                ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, cur, n_embd_head_v_mla, n_head_l, n_tokens);
+                ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate,                1, n_head_l, n_tokens);
+                attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);
+                cb(attn_3d, "attn_gated", il);
+
+                cur = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v_mla * n_head_l, n_tokens);
+
+                cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
+                cb(cur, "attn_output", 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);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
+                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
+                model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            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,
+                hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il,
+                nullptr,
+                model.layers[il].ffn_gate_up_exps,
+                model.layers[il].ffn_up_exps_s,
+                model.layers[il].ffn_gate_exps_s,
+                model.layers[il].ffn_down_exps_s);
+            cb(moe_out, "ffn_moe_out", il);
+
+            ggml_tensor * ffn_shexp =
+                build_ffn(cur,
+                    model.layers[il].ffn_up_shexp, NULL, model.layers[il].ffn_up_shexp_s,
+                    model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s,
+                    model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,
+                    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);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        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;
+
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/models.h

@@ -1156,6 +1156,18 @@ struct llama_model_deepseek32 : public llama_model_base {
 };
 
 
+struct llama_model_dots3note : public llama_model_base {
+    llama_model_dots3note(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    struct graph : public llm_graph_context {
+        graph(const llama_model & model, const llm_graph_params & params);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
 struct llama_model_deepseek4 : public llama_model_base {
     llama_model_deepseek4(const struct llama_model_params & params) : llama_model_base(params) {}
     void load_arch_hparams(llama_model_loader & ml) override;

tests/test-llama-archs.cpp

@@ -104,6 +104,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
     } else if (arch == LLM_ARCH_DEEPSEEK2
             || arch == LLM_ARCH_DEEPSEEK32
             || arch == LLM_ARCH_GLM_DSA
+            || arch == LLM_ARCH_DOTS3NOTE
             || arch == LLM_ARCH_KIMI_LINEAR
             || arch == LLM_ARCH_MISTRAL4) {
         n_embd = 128;
@@ -163,6 +164,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
     if (arch == LLM_ARCH_DEEPSEEK2
             || arch == LLM_ARCH_DEEPSEEK32
             || arch == LLM_ARCH_GLM_DSA
+            || arch == LLM_ARCH_DOTS3NOTE
             || arch == LLM_ARCH_KIMI_LINEAR
             || arch == LLM_ARCH_MISTRAL4) {
         ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH,       uint32_t(576));
@@ -170,6 +172,15 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
         ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT,       uint32_t(64));
         ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   uint32_t(192));
         ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, uint32_t(128));
+        if (arch == LLM_ARCH_DOTS3NOTE) {
+            // SWA layers reuse the same MLA geometry as the full layers in this fixture
+            ms.add_kv(LLM_KV_ATTENTION_KV_LORA_RANK_SWA,     uint32_t(512));
+            ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_SWA,       uint32_t(576));
+            ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_SWA,     uint32_t(512));
+            ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,   uint32_t(192));
+            ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA, uint32_t(128));
+            ms.add_kv(LLM_KV_ROPE_FREQ_BASE_SWA,             10000.0f);
+        }
     } else if (arch == LLM_ARCH_MINIMAX_M3) {
         // partial rotary: n_rot must not exceed the indexer key length (64)
         ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT,       uint32_t(64));
@@ -192,7 +203,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
         ms.add_kv(LLM_KV_ROPE_FREQ_BASE_SWA,              10000.0f);
         // SWA pattern: every 5th layer is full attention (matches E2B layer_types)
         ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, uint32_t(5));
-    } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_MUSE_GLIMMER) {
+    } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_DOTS3NOTE) {
         std::vector<uint32_t> pattern;
         pattern.reserve(n_layer);
         for (uint32_t il = 0; il < n_layer; il++) {
@@ -348,6 +359,7 @@ static bool moe_mandatory(const llm_arch arch) {
         case LLM_ARCH_DEEPSEEK:
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_DEEPSEEK32:
+        case LLM_ARCH_DOTS3NOTE:
         case LLM_ARCH_GLM4_MOE:
         case LLM_ARCH_GLM_DSA:
         case LLM_ARCH_EXAONE_MOE:
@@ -436,7 +448,7 @@ static bool arch_supported(const llm_arch arch) {
 
     // FIXME: these hit scheduler/view-backed-output issues with WebGPU on CI.
 #ifdef GGML_USE_WEBGPU
-    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA) {
+    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_DOTS3NOTE) {
         return false;
     }
 #endif // GGML_USE_WEBGPU