0006-add-mllama-support.patch

From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: jmorganca <jmorganca@gmail.com>
Date: Sun, 20 Apr 2025 16:12:36 -0700
Subject: [PATCH] add mllama support

adds support for the llama 3.2 vision architecture
---
 ggml/src/ggml-backend-reg.cpp |   6 +-
 include/llama.h               |   6 +
 src/llama-arch.cpp            |  44 +++++
 src/llama-arch.h              |  10 ++
 src/llama-batch.cpp           |   3 +
 src/llama-context.cpp         |  23 ++-
 src/llama-context.h           |   1 +
 src/llama-cparams.h           |   1 +
 src/llama-graph.cpp           |  25 +++
 src/llama-graph.h             |  12 ++
 src/llama-hparams.cpp         |   4 +
 src/llama-hparams.h           |   7 +
 src/llama-kv-cache.cpp        |  14 +-
 src/llama-model-loader.cpp    |   2 +
 src/llama-model.cpp           | 311 +++++++++++++++++++++++++++++++++-
 src/llama-model.h             |  12 ++
 src/llama-quant.cpp           |   4 +-
 tools/mtmd/llava.cpp          |   5 +-
 tools/mtmd/mtmd-helper.cpp    |   7 +-
 19 files changed, 475 insertions(+), 22 deletions(-)

diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 405d8e31..82ae1b5b 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -178,9 +178,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_CANN
         register_backend(ggml_backend_cann_reg());
 #endif
-#ifdef GGML_USE_BLAS
-        register_backend(ggml_backend_blas_reg());
-#endif
+// #ifdef GGML_USE_BLAS
+//         register_backend(ggml_backend_blas_reg());
+// #endif
 #ifdef GGML_USE_RPC
         register_backend(ggml_backend_rpc_reg());
 #endif
diff --git a/include/llama.h b/include/llama.h
index abedebdb..41beef21 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -258,6 +258,7 @@ extern "C" {
 
         llama_token  *  token;
         float        *  embd;
+        int32_t         n_embd;
         llama_pos    *  pos;
         int32_t      *  n_seq_id;
         llama_seq_id ** seq_id;
@@ -365,6 +366,7 @@ extern "C" {
         bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
         bool no_perf;     // whether to measure performance timings
         bool op_offload;  // whether to offload host tensor operations to device
+        bool cross_attn;  // whether to use cross attention
     };
 
     // model quantization parameters
@@ -464,6 +466,10 @@ extern "C" {
             struct llama_context_params   params),
             "use llama_init_from_model instead");
 
+    // TODO (jmorganca): this should most likely be passed in as part of a batch
+    // and not set on the context for all batches.
+    LLAMA_API void llama_set_cross_attention(struct llama_context * ctx, bool cross_attn_state);
+
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 5ab3f572..eb7b5325 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -6,6 +6,7 @@
 
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_LLAMA,            "llama"            },
+    { LLM_ARCH_MLLAMA,           "mllama"           },
     { LLM_ARCH_LLAMA4,           "llama4"           },
     { LLM_ARCH_DECI,             "deci"             },
     { LLM_ARCH_FALCON,           "falcon"           },
@@ -144,6 +145,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
+    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,       "%s.attention.cross_attention_layers"       },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
 
@@ -273,6 +275,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_MLLAMA,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_CROSS_ATTN_K_NORM,    "blk.%d.cross_attn_k_norm" },
+            { LLM_TENSOR_CROSS_ATTN_K_PROJ,    "blk.%d.cross_attn_k_proj" },
+            { LLM_TENSOR_CROSS_ATTN_O_PROJ,    "blk.%d.cross_attn_o_proj" },
+            { LLM_TENSOR_CROSS_ATTN_Q_NORM,    "blk.%d.cross_attn_q_norm" },
+            { LLM_TENSOR_CROSS_ATTN_Q_PROJ,    "blk.%d.cross_attn_q_proj" },
+            { LLM_TENSOR_CROSS_ATTN_V_PROJ,    "blk.%d.cross_attn_v_proj" },
+            { LLM_TENSOR_CROSS_ATTN_ATTN_GATE, "blk.%d.cross_attn_attn_gate" },
+            { LLM_TENSOR_CROSS_ATTN_MLP_GATE,  "blk.%d.cross_attn_mlp_gate" },
+        },
+    },
     {
         LLM_ARCH_DECI,
         {
@@ -1701,6 +1737,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     // this tensor is loaded for T5, but never used
     {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
     {LLM_TENSOR_BSKCN_TV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_CROSS_ATTN_K_NORM,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_CROSS_ATTN_K_PROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_CROSS_ATTN_O_PROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_CROSS_ATTN_Q_NORM,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_CROSS_ATTN_Q_PROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_CROSS_ATTN_V_PROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_CROSS_ATTN_ATTN_GATE,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_CROSS_ATTN_MLP_GATE,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CONV1D,                     {LLM_TENSOR_LAYER_INPUT,     GGML_OP_IM2COL}},
     {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
diff --git a/src/llama-arch.h b/src/llama-arch.h
index 525c1b7d..bc8a4f0b 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -11,6 +11,7 @@
 enum llm_arch {
     LLM_ARCH_LLAMA,
     LLM_ARCH_LLAMA4,
+    LLM_ARCH_MLLAMA,
     LLM_ARCH_DECI,
     LLM_ARCH_FALCON,
     LLM_ARCH_BAICHUAN,
@@ -148,6 +149,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
+    LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
@@ -349,6 +351,14 @@ enum llm_tensor {
     LLM_TENSOR_CLS,
     LLM_TENSOR_CLS_OUT,
     LLM_TENSOR_BSKCN_TV,
+    LLM_TENSOR_CROSS_ATTN_K_NORM,
+    LLM_TENSOR_CROSS_ATTN_K_PROJ,
+    LLM_TENSOR_CROSS_ATTN_O_PROJ,
+    LLM_TENSOR_CROSS_ATTN_Q_NORM,
+    LLM_TENSOR_CROSS_ATTN_Q_PROJ,
+    LLM_TENSOR_CROSS_ATTN_V_PROJ,
+    LLM_TENSOR_CROSS_ATTN_ATTN_GATE,
+    LLM_TENSOR_CROSS_ATTN_MLP_GATE,
     LLM_TENSOR_CONV1D,
     LLM_TENSOR_CONVNEXT_DW,
     LLM_TENSOR_CONVNEXT_NORM,
diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp
index a88b2fe3..241b316e 100644
--- a/src/llama-batch.cpp
+++ b/src/llama-batch.cpp
@@ -320,6 +320,7 @@ struct llama_batch llama_batch_get_one(
         /*n_tokens       =*/ n_tokens,
         /*tokens         =*/ tokens,
         /*embd           =*/ nullptr,
+        /*n_embd         =*/ 0,
         /*pos            =*/ nullptr,
         /*n_seq_id       =*/ nullptr,
         /*seq_id         =*/ nullptr,
@@ -332,6 +333,7 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_
         /*n_tokens       =*/ 0,
         /*tokens         =*/ nullptr,
         /*embd           =*/ nullptr,
+        /*n_embd         =*/ 0,
         /*pos            =*/ nullptr,
         /*n_seq_id       =*/ nullptr,
         /*seq_id         =*/ nullptr,
@@ -340,6 +342,7 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_
 
     if (embd) {
         batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
+        batch.n_embd = embd;
     } else {
         batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
     }
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index dca22d8b..c22687e4 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -514,7 +514,7 @@ float * llama_context::get_logits_ith(int32_t i) {
             throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
         }
 
-        return logits + j*model.vocab.n_tokens();
+        return logits + j*model.hparams.n_vocab;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
 #ifndef NDEBUG
@@ -632,6 +632,10 @@ void llama_context::set_warmup(bool value) {
     cparams.warmup = value;
 }
 
+void llama_context::set_cross_attn(bool value) {
+    cparams.cross_attn = value;
+}
+
 void llama_context::set_adapter_lora(
             llama_adapter_lora * adapter,
             float scale) {
@@ -709,7 +713,7 @@ int llama_context::encode(llama_batch & inp_batch) {
 
     const int64_t n_embd = hparams.n_embd;
 
-    llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);
+    llama_sbatch sbatch = llama_sbatch(batch, batch.n_embd, /* simple_split */ true, /* logits_all */ true);
 
     const llama_ubatch ubatch = sbatch.split_simple(n_tokens);
 
@@ -863,10 +867,9 @@ int llama_context::decode(llama_batch & inp_batch) {
 
     const llama_batch & batch = batch_allocr.batch;
 
-    const auto & vocab   = model.vocab;
     const auto & hparams = model.hparams;
 
-    const int32_t n_vocab = vocab.n_tokens();
+    const int32_t n_vocab = hparams.n_vocab;
 
     const int64_t n_tokens_all = batch.n_tokens;
     const int64_t n_embd       = hparams.n_embd;
@@ -1087,7 +1090,7 @@ int llama_context::decode(llama_batch & inp_batch) {
         // make the outputs have the same order they had in the user-provided batch
         // note: this is mostly relevant for recurrent models atm
         if (!sorted_output) {
-            const uint32_t n_vocab = model.vocab.n_tokens();
+            const uint32_t n_vocab = model.hparams.n_vocab;
             const uint32_t n_embd  = model.hparams.n_embd;
 
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
@@ -1142,12 +1145,11 @@ int llama_context::decode(llama_batch & inp_batch) {
 
 int32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto & hparams = model.hparams;
-    const auto & vocab   = model.vocab;
 
     const int64_t n_outputs_max = std::max<int64_t>(n_outputs, n_seq_max());
 
     const auto n_batch = cparams.n_batch;
-    const auto n_vocab = vocab.n_tokens();
+    const auto n_vocab = hparams.n_vocab;
     const auto n_embd  = hparams.n_embd;
 
     // TODO: use a per-batch flag for logits presence instead
@@ -1682,7 +1684,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
     {
         LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);
 
-        const uint64_t logits_size = std::min((uint64_t) this->logits_size, (uint64_t) n_outputs * model.vocab.n_tokens());
+        const uint64_t logits_size = std::min((uint64_t) this->logits_size, (uint64_t) n_outputs * model.hparams.n_vocab);
 
         io.write(&logits_size, sizeof(logits_size));
 
@@ -2091,6 +2093,7 @@ llama_context_params llama_context_default_params() {
         /*.flash_attn                  =*/ false,
         /*.no_perf                     =*/ true,
         /*.op_offload                  =*/ true,
+        /*.cross_attn                  =*/ false,
     };
 
     return result;
@@ -2216,6 +2219,10 @@ void llama_set_warmup(llama_context * ctx, bool warmup) {
     ctx->set_warmup(warmup);
 }
 
+void llama_set_cross_attention(struct llama_context * ctx, bool cross_attention) {
+    ctx->set_cross_attn(cross_attention);
+}
+
 void llama_synchronize(llama_context * ctx) {
     ctx->synchronize();
 }
diff --git a/src/llama-context.h b/src/llama-context.h
index c0ceacb1..c4ab242a 100644
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -71,6 +71,7 @@ struct llama_context {
     void set_embeddings (bool value);
     void set_causal_attn(bool value);
     void set_warmup(bool value);
+    void set_cross_attn(bool value);
 
     void set_adapter_lora(
             llama_adapter_lora * adapter,
diff --git a/src/llama-cparams.h b/src/llama-cparams.h
index 246fa577..7a6156ce 100644
--- a/src/llama-cparams.h
+++ b/src/llama-cparams.h
@@ -31,6 +31,7 @@ struct llama_cparams {
     bool no_perf;
     bool warmup;
     bool op_offload;
+    bool cross_attn;
 
     enum llama_pooling_type pooling_type;
 
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index b0e3f635..f14869cf 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -532,6 +532,12 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+void llm_graph_input_cross_attn_state::set_input(const llama_ubatch * ubatch) {
+    if (ubatch->embd) {
+        ggml_backend_tensor_set(cross_attn_state, ubatch->embd, 0, ggml_nbytes(cross_attn_state));
+    }
+}
+
 //
 // llm_graph_context
 //
@@ -1514,6 +1520,25 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
     return (llm_graph_input_attn_cross *) res->add_input(std::move(inp));
 }
 
+ggml_tensor * llm_graph_context::build_inp_cross_attn_state() const {
+    const int64_t n_embd = hparams.n_embd;
+
+    auto inp = std::make_unique<llm_graph_input_cross_attn_state>();
+
+    ggml_tensor * cur = nullptr;
+
+    inp->cross_attn_state = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd, 1601, 4);
+    ggml_set_input(inp->cross_attn_state);
+
+    cur = inp->cross_attn_state;
+
+    cb(cur, "inp_cross_attn_state", -1);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_cross * inp,
         ggml_cgraph * gf,
diff --git a/src/llama-graph.h b/src/llama-graph.h
index 832a8c09..5a322785 100644
--- a/src/llama-graph.h
+++ b/src/llama-graph.h
@@ -87,6 +87,7 @@ public:
 
     ggml_tensor * tokens = nullptr; // I32 [n_batch]
     ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
+    ggml_tensor * cross_attn_state; // F32 [4, n_embd, 1061]
 };
 
 class llm_graph_input_pos : public llm_graph_input_i {
@@ -284,6 +285,16 @@ public:
     const llama_cross * cross = nullptr;
 };
 
+class llm_graph_input_cross_attn_state : public llm_graph_input_i {
+public:
+    llm_graph_input_cross_attn_state()          = default;
+    virtual ~llm_graph_input_cross_attn_state() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * cross_attn_state; // F32 [4, n_embd, 1061]
+};
+
 //
 // llm_graph_result
 //
@@ -495,6 +506,7 @@ struct llm_graph_context {
     ggml_tensor * build_inp_cls() const;
     ggml_tensor * build_inp_s_copy() const;
     ggml_tensor * build_inp_s_mask() const;
+    ggml_tensor * build_inp_cross_attn_state() const;
 
     ggml_tensor * build_inp_cross_embd() const;
     ggml_tensor * build_inp_pos_bucket_enc() const;
diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
index 8a667960..6a02de03 100644
--- a/src/llama-hparams.cpp
+++ b/src/llama-hparams.cpp
@@ -85,3 +85,7 @@ bool llama_hparams::is_swa(uint32_t il) const {
 
     GGML_ABORT("fatal error");
 }
+
+bool llama_hparams::cross_attention_layers(uint32_t il) const {
+    return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
+}
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index 48dce407..b6fc7e6d 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -2,6 +2,8 @@
 
 #include "llama.h"
 
+#include <algorithm>
+
 #include <array>
 
 // bump if necessary
@@ -42,6 +44,7 @@ struct llama_hparams {
     uint32_t n_expert = 0;
     uint32_t n_expert_used = 0;
     uint32_t n_rel_attn_bkts = 0;
+    uint32_t n_vocab = 0;
 
     // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
     uint32_t n_embd_head_k_mla = 0;
@@ -56,6 +59,7 @@ struct llama_hparams {
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
 
     std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr = {};
+    std::array<uint32_t, LLAMA_MAX_LAYERS> cross_attn_layers;
 
     uint32_t n_layer_dense_lead = 0;
     uint32_t n_lora_q           = 0;
@@ -159,6 +163,9 @@ struct llama_hparams {
     // Block skip connection
     bool n_bskcn(uint32_t n, uint32_t il) const;
 
+    // cross attention layers
+    bool cross_attention_layers(uint32_t il) const;
+
     bool is_swa(uint32_t il) const;
 };
 
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
index 3dcad65b..a7b0a7eb 100644
--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
@@ -100,8 +100,16 @@ llama_kv_cache_unified::llama_kv_cache_unified(
             throw std::runtime_error("failed to create ggml context for kv cache");
         }
 
-        ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
-        ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size);
+        ggml_tensor * k, *v;
+
+        // for cross attention layers
+        if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layers(i)) {
+            k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
+            v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
+        } else {
+            k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
+            v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size);
+        }
         ggml_format_name(k, "cache_k_l%d", i);
         ggml_format_name(v, "cache_v_l%d", i);
         k_l.push_back(k);
@@ -446,7 +454,7 @@ void llama_kv_cache_unified::set_full() {
 llama_sbatch llama_kv_cache_unified::sbatch_init(
         const llama_batch & batch,
         bool logits_all) {
-    return llama_sbatch(batch, hparams.n_embd, true, logits_all);
+    return llama_sbatch(batch, batch.n_embd, true, logits_all);
 }
 
 llama_ubatch llama_kv_cache_unified::ubatch_next(
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index 7f6617fa..2acfd4a8 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -315,6 +315,8 @@ namespace GGUFMeta {
         return true;
     }
 
+    template bool llama_model_loader::get_arr<std::array<unsigned int, 512>>(enum llm_kv kid, std::array<unsigned int, 512>& result, bool required);
+
     template<typename T, size_t N_MAX>
     bool llama_model_loader::get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required) {
         const int kid = gguf_find_key(meta.get(), key.c_str());
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 831b68c0..e8298f56 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -433,6 +433,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
     // get general kv
     ml.get_key(LLM_KV_GENERAL_NAME, name, false);
+    ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab, false);
 
     // everything past this point is not vocab-related
     if (hparams.vocab_only) {
@@ -444,6 +445,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_BLOCK_COUNT,       hparams.n_layer);
     ml.get_key(LLM_KV_EXPERT_COUNT,      hparams.n_expert,      false);
     ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false);
+    ml.get_key(LLM_KV_VOCAB_SIZE,        hparams.n_vocab,       false);
 
     if (arch == LLM_ARCH_WAVTOKENIZER_DEC) {
         ml.get_key(LLM_KV_FEATURES_LENGTH, hparams.n_embd_features);
@@ -467,9 +469,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
     std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
     std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
+    std::fill(hparams.cross_attn_layers.begin(), hparams.cross_attn_layers.end(), -1);
 
     ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff_arr,   hparams.n_layer, false);
     ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer, false);
+    ml.get_arr(LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, hparams.cross_attn_layers, false);
 
     // n_head_kv is optional, default to n_head
     hparams.n_head_kv_arr = hparams.n_head_arr;
@@ -522,7 +526,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
         ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
 
-        if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON) {
+        if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_MLLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON) {
             if (hparams.n_rot != hparams.n_embd_head_k) {
                 throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
             }
@@ -585,6 +589,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.use_kq_norm = false;
                 }
             } break;
+        case LLM_ARCH_MLLAMA:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 40: type = LLM_TYPE_11B; break;
+                    case 100: type = LLM_TYPE_90B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_DECI:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1581,7 +1595,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         const int64_t n_embd_head_v = hparams.n_embd_head_v;
         const int64_t n_ff          = hparams.n_ff();
         const int64_t n_embd_gqa    = n_embd_v_gqa;
-        const int64_t n_vocab       = vocab.n_tokens();
+        const int64_t n_vocab       = hparams.n_vocab;
         const int64_t n_token_types = vocab.n_token_types();
         const int64_t n_rot         = hparams.n_rot;
         const int64_t n_expert      = hparams.n_expert;
@@ -1840,6 +1854,52 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
                     }
                 } break;
+            case LLM_ARCH_MLLAMA:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab+8}, 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}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (output == NULL) {
+                            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        if (hparams.cross_attention_layers(i)) {
+                            layer.cross_attn_k_norm = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_K_NORM,   "weight", i), {128}, 0);
+                            layer.cross_attn_k_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_K_PROJ,   "weight", i), {n_embd, 1024}, 0);
+                            layer.cross_attn_o_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_O_PROJ,   "weight", i), {n_embd, n_embd}, 0);
+                            layer.cross_attn_q_norm = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_Q_NORM, "weight", i), {128}, 0);
+                            layer.cross_attn_q_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_Q_PROJ, "weight", i), {n_embd, n_embd}, 0);
+                            layer.cross_attn_v_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_V_PROJ, "weight", i), {n_embd, 1024}, 0);
+                            layer.cross_attn_attn_gate = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_ATTN_GATE, i), {1}, 0);
+                            layer.cross_attn_mlp_gate = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_MLP_GATE, i), {1}, 0);
+                            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        } else {
+                            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+                            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        }
+                    }
+                } break;
             case LLM_ARCH_DECI:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -4756,6 +4816,246 @@ struct llm_build_llama : public llm_graph_context {
     }
 };
 
+struct llm_build_mllama: public llm_graph_context {
+    llm_build_mllama(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        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;
+        ggml_tensor * inpCAS;
+
+        inpL = build_inp_embd(model.tok_embd);
+        inpCAS = build_inp_cross_attn_state();
+
+          // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            if (hparams.cross_attention_layers(il)) {
+                if (!ubatch.embd && !cparams.cross_attn) {
+                    continue;
+                }
+
+                // cross attention layer
+                ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_q_proj, cur);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 0, 2, 1, 3));
+                cb(Qcur, "Qcur", il);
+
+                Qcur = build_norm(Qcur, model.layers[il].cross_attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur, * Vcur;
+                if (ubatch.embd) {
+                    Kcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_k_proj, inpCAS);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 6404);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = build_norm(Kcur, model.layers[il].cross_attn_k_norm, NULL, LLM_NORM_RMS, il);
+                    cb(Kcur, "Kcur", il);
+
+                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, kv_self->k_l[il]));
+
+                    Vcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_v_proj, inpCAS);
+                    cb(Vcur, "Vcur", il);
+
+                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, 6404);
+                    cb(Vcur, "Vcur", il);
+
+                    Vcur = ggml_permute(ctx0, Vcur, 0, 2, 1, 3);
+                    cb(Vcur, "Vcur", il);
+
+                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, kv_self->v_l[il]));
+                } else {
+                    Kcur = ggml_view_tensor(ctx0, kv_self->k_l[il]);
+                    cb(Kcur, "Kcur (view)", il);
+
+                    Vcur = ggml_view_tensor(ctx0, kv_self->v_l[il]);
+                    cb(Vcur, "Vcur (view)", il);
+                }
+
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, Kcur, Qcur);
+                cb(kq, "kq", il);
+
+                // TODO: apply causal masks
+                struct ggml_tensor * kq_soft_max = ggml_soft_max_ext(ctx0, kq, nullptr, 1.f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
+                cb(kq_soft_max, "kq_soft_max", il);
+
+                Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, Vcur));
+                cb(Vcur, "Vcur", il);
+
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, Vcur, kq_soft_max);
+                cb(kqv, "kqv", il);
+
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
+
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
+
+                cur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_o_proj, cur);
+                cb(cur, "cur", il);
+
+                // TODO: do this in place once?
+                cur = ggml_mul(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_attn_gate));
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+
+                // TODO: do this inplace once?
+                cur = ggml_add_inplace(ctx0, ggml_mul_inplace(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_mlp_gate)), ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                cur = build_cvec(cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            } else {
+                // self attention layer
+
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", 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);
+                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", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+
+                if (il == n_layer - 1) {
+                    // skip computing output for unused tokens
+                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                    n_tokens = n_outputs;
+                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+                }
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", 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, 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);
+    }
+};
+
 struct llm_build_deci : public llm_graph_context {
     llm_build_deci(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -12496,7 +12796,7 @@ struct llm_build_solar : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -13128,6 +13428,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_llama>(*this, params, gf);
             } break;
+        case LLM_ARCH_MLLAMA:
+            {
+                llm = std::make_unique<llm_build_mllama>(*this, params, gf);
+            } break;
         case LLM_ARCH_DECI:
             {
                 llm = std::make_unique<llm_build_deci>(*this, params, gf);
@@ -13489,6 +13793,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         // use what we call a normal RoPE, operating on pairs of consecutive head values
         case LLM_ARCH_LLAMA:
         case LLM_ARCH_LLAMA4:
+        case LLM_ARCH_MLLAMA:
         case LLM_ARCH_DECI:
         case LLM_ARCH_BAICHUAN:
         case LLM_ARCH_STARCODER:
diff --git a/src/llama-model.h b/src/llama-model.h
index 43746c7d..9281e629 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -11,6 +11,7 @@
 #include <string>
 #include <unordered_map>
 #include <vector>
+#include <stdexcept>
 
 struct llama_cparams;
 struct llama_ubatch;
@@ -74,6 +75,7 @@ enum llm_type {
     LLM_TYPE_40B,
     LLM_TYPE_65B,
     LLM_TYPE_70B,
+    LLM_TYPE_90B,
     LLM_TYPE_236B,
     LLM_TYPE_290B,
     LLM_TYPE_314B,
@@ -318,6 +320,16 @@ struct llama_layer {
 
     struct ggml_tensor * bskcn_tv = nullptr;
 
+    // cross attention
+    struct ggml_tensor * cross_attn_k_norm = nullptr;
+    struct ggml_tensor * cross_attn_k_proj = nullptr;
+    struct ggml_tensor * cross_attn_o_proj = nullptr;
+    struct ggml_tensor * cross_attn_q_norm = nullptr;
+    struct ggml_tensor * cross_attn_q_proj = nullptr;
+    struct ggml_tensor * cross_attn_v_proj = nullptr;
+    struct ggml_tensor * cross_attn_attn_gate = nullptr;
+    struct ggml_tensor * cross_attn_mlp_gate = nullptr;
+
     struct llama_layer_posnet posnet;
 
     struct llama_layer_convnext convnext;
diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp
index 820d5128..56531980 100644
--- a/src/llama-quant.cpp
+++ b/src/llama-quant.cpp
@@ -639,7 +639,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         if (llama_model_has_encoder(&model)) {
             n_attn_layer *= 3;
         }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
+        if (qs.n_attention_wv != n_attn_layer) {
+            LLAMA_LOG_WARN("%s: n_attention_wv is unexpected, expected: %d, found: %d\n", __func__, n_attn_layer, qs.n_attention_wv);
+        }
     }
 
     size_t total_size_org = 0;
diff --git a/tools/mtmd/llava.cpp b/tools/mtmd/llava.cpp
index ebef8b3c..b0eb79bb 100644
--- a/tools/mtmd/llava.cpp
+++ b/tools/mtmd/llava.cpp
@@ -462,7 +462,7 @@ struct llava_embd_batch {
     std::vector<llama_seq_id *> seq_ids;
     std::vector<int8_t>         logits;
     llama_batch batch;
-    llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+    llava_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
         pos     .resize(n_tokens);
         n_seq_id.resize(n_tokens);
         seq_ids .resize(n_tokens + 1);
@@ -474,6 +474,7 @@ struct llava_embd_batch {
             /*n_tokens       =*/ n_tokens,
             /*tokens         =*/ nullptr,
             /*embd           =*/ embd,
+            /*n_embd         =*/ n_embd,
             /*pos            =*/ pos.data(),
             /*n_seq_id       =*/ n_seq_id.data(),
             /*seq_id         =*/ seq_ids.data(),
@@ -497,7 +498,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
             n_eval = n_batch;
         }
         float * embd = image_embed->embed+i*n_embd;
-        llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
+        llava_embd_batch llava_batch = llava_embd_batch(embd, n_embd, n_eval, *n_past, 0);
         if (llama_decode(ctx_llama, llava_batch.batch)) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return false;
diff --git a/tools/mtmd/mtmd-helper.cpp b/tools/mtmd/mtmd-helper.cpp
index 7a328867..61ebdd43 100644
--- a/tools/mtmd/mtmd-helper.cpp
+++ b/tools/mtmd/mtmd-helper.cpp
@@ -58,7 +58,7 @@ struct decode_embd_batch {
     std::vector<llama_seq_id *> seq_ids;
     std::vector<int8_t>         logits;
     llama_batch batch;
-    decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+    decode_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
         pos     .resize(n_tokens * n_pos_per_embd);
         n_seq_id.resize(n_tokens);
         seq_ids .resize(n_tokens + 1);
@@ -69,6 +69,7 @@ struct decode_embd_batch {
             /*n_tokens       =*/ n_tokens,
             /*tokens         =*/ nullptr,
             /*embd           =*/ embd,
+            /*n_embd         =*/ n_embd,
             /*pos            =*/ pos.data(),
             /*n_seq_id       =*/ n_seq_id.data(),
             /*seq_id         =*/ seq_ids.data(),
@@ -131,6 +132,7 @@ struct decode_embd_batch {
             /*n_tokens       =*/ n_tokens,
             /*tokens         =*/ nullptr,
             /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
+            /*n_embd         =*/ batch.n_embd,
             /*pos            =*/ pos_ptr,
             /*n_seq_id       =*/ batch.n_seq_id + offset,
             /*seq_id         =*/ batch.seq_id   + offset,
@@ -166,7 +168,8 @@ int32_t mtmd_helper_decode_image_chunk(
     int32_t n_tokens = mtmd_image_tokens_get_n_tokens(image_tokens);
     int32_t i_batch = 0;
     int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
-    decode_embd_batch batch_embd(encoded_embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
+    int n_embd  = llama_model_n_embd(llama_get_model(lctx));
+    decode_embd_batch batch_embd(encoded_embd, n_embd, n_tokens, n_past, seq_id);
 
     const int nx = mtmd_image_tokens_get_nx(image_tokens);
     const int ny = mtmd_image_tokens_get_ny(image_tokens);