model: add Mellum architecture (#23966)

* model: support for Mellum architecture

* model: improve mellum.py formatting

* model: improve mellum.py formatting once again

* deps: downgrade transformers to 4.57.6 (to fix CI)

* deps: remove huggingface_hub dependency

* deps: remove huggingface_hub from test requirements

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

src/llama-arch.cpp

@@ -135,6 +135,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_MAINCODER,        "maincoder"        },
     { LLM_ARCH_KIMI_LINEAR,      "kimi-linear"      },
     { LLM_ARCH_TALKIE,           "talkie"           },
+    { LLM_ARCH_MELLUM,           "mellum"           },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 

src/llama-arch.h

@@ -139,6 +139,7 @@ enum llm_arch {
     LLM_ARCH_MAINCODER,
     LLM_ARCH_KIMI_LINEAR,
     LLM_ARCH_TALKIE,
+    LLM_ARCH_MELLUM,
     LLM_ARCH_UNKNOWN,
 };
 

src/llama-model-saver.cpp

@@ -29,6 +29,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
         case LLM_ARCH_APERTUS:
         case LLM_ARCH_MIMO2:
         case LLM_ARCH_STEP35:
+        case LLM_ARCH_MELLUM:
             return false;
         default:
             return true;

src/llama-model.cpp

@@ -81,6 +81,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_mpt(params);
         case LLM_ARCH_STABLELM:
             return new llama_model_stablelm(params);
+        case LLM_ARCH_MELLUM:
+            return new llama_model_mellum(params);
         case LLM_ARCH_QWEN:
             return new llama_model_qwen(params);
         case LLM_ARCH_QWEN2:
@@ -764,6 +766,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_A13B:          return "A13B";
         case LLM_TYPE_7B_A1B:        return "7B.A1B";
         case LLM_TYPE_8B_A1B:        return "8B.A1B";
+        case LLM_TYPE_12B_A2_5B:     return "12B.A2.5B";
         case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_24B_A2B:       return "24B.A2B";
@@ -1816,7 +1819,11 @@ void llama_model::print_info() const {
             LLAMA_LOG_INFO("%s: n_ff_shexp            = %d\n",     __func__, hparams.n_ff_shexp);
         }
 
-        if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE || arch == LLM_ARCH_QWEN3VLMOE || arch == LLM_ARCH_RND1) {
+        if (arch == LLM_ARCH_MELLUM ||
+                arch == LLM_ARCH_QWEN3MOE ||
+                arch == LLM_ARCH_OPENAI_MOE ||
+                arch == LLM_ARCH_QWEN3VLMOE ||
+                arch == LLM_ARCH_RND1) {
             LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
         }
 
@@ -2404,6 +2411,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_MIMO2:
         case LLM_ARCH_STEP35:
         case LLM_ARCH_TALKIE:
+        case LLM_ARCH_MELLUM:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:

src/llama-model.h

@@ -116,6 +116,7 @@ enum llm_type {
     LLM_TYPE_A13B,
     LLM_TYPE_7B_A1B,
     LLM_TYPE_8B_A1B, // lfm2moe
+    LLM_TYPE_12B_A2_5B,
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_24B_A2B, // lfm2moe

src/llama-vocab.cpp

@@ -353,6 +353,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
             case LLAMA_VOCAB_PRE_TYPE_CODESHELL:
             case LLAMA_VOCAB_PRE_TYPE_EXAONE:
             case LLAMA_VOCAB_PRE_TYPE_MINERVA:
+            case LLAMA_VOCAB_PRE_TYPE_MELLUM2:
                 regex_exprs = {
                     "\\p{N}",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
@@ -2325,6 +2326,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "solar-open") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "mellum2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_MELLUM2;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }

src/llama-vocab.h

@@ -63,6 +63,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_MINICPM5          = 52,
     LLAMA_VOCAB_PRE_TYPE_WHITESPACE        = 53,
     LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI = 54,
+    LLAMA_VOCAB_PRE_TYPE_MELLUM2           = 55,
 };
 
 struct LLM_KV;

src/models/mellum.cpp

@@ -0,0 +1,225 @@
+#include "models.h"
+
+void llama_model_mellum::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
+
+    if (hparams.n_swa > 0) {
+        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+
+        uint32_t swa_period = 4;
+        const auto res = ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+        if (res) {
+            hparams.set_swa_pattern(swa_period);
+        } else {
+            ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
+        }
+
+        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+
+        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+    } else {
+        hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+    }
+
+    switch (hparams.n_layer) {
+        case 28: type = LLM_TYPE_12B_A2_5B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_mellum::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    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}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+
+        if (n_expert == 0) {
+            throw std::runtime_error("n_expert must be > 0 for Mellum");
+        }
+        if (n_expert_used == 0) {
+            throw std::runtime_error("n_expert_used must be > 0 for Mellum");
+        }
+
+        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_mellum::build_arch_graph(const llm_graph_params & params) const {
+    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+        return std::make_unique<graph<true>>(*this, params);
+    }
+    return std::make_unique<graph<false>>(*this, params);
+}
+
+template <bool iswa>
+llama_model_mellum::graph<iswa>::graph(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 == n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, nullptr,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
+                    n_embd_head, n_head, n_head_kv, il);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            const bool is_swa = hparams.is_swa(il);
+
+            if (is_swa) {
+                // For sliding window layers, use regular rope with no yarn rope scaling.
+                // This is achieved here by setting freq_scale and attn_factor to 1.
+                // We also set ext_factor to 0 to avoid a few unnecessary computations.
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
+                    0.0, 1.0, beta_fast, beta_slow
+                    );
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
+                    0.0, 1.0, beta_fast, beta_slow
+                    );
+            } else {
+                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
+                    );
+
+                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(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), 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);
+
+        // MoE
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, nullptr,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        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,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il,
+                    nullptr, nullptr,
+                    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);
+        cur = moe_out;
+
+        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, model.output_s);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+template struct llama_model_mellum::graph<false>;
+template struct llama_model_mellum::graph<true>;

src/models/models.h

@@ -411,6 +411,18 @@ struct llama_model_stablelm : public llama_model_base {
     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };
 
+struct llama_model_mellum : public llama_model_base {
+    llama_model_mellum(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;
+
+    template <bool iswa>
+    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_qwen : public llama_model_base {
     llama_model_qwen(const struct llama_model_params & params) : llama_model_base(params) {}