ollama/llama/patches/0010-add-mllama-support.patch

From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: jmorganca <jmorganca@gmail.com>
Date: Thu, 17 Oct 2024 15:18:22 -0700
Subject: [PATCH] add mllama support

mllama adds cross-attention layers to the standard llama architecture
it also requires a way to input a new tensor: cross_attention_state
once per generation

cross-attention layers don't change and so they are cached in the
kv cache once per run

remaining is to implement the cross attention mask
---
 include/llama.h |   4 +
 src/llama.cpp   | 456 ++++++++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 447 insertions(+), 13 deletions(-)

diff --git a/include/llama.h b/include/llama.h
index 7cae1bbe..122e3cf1 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -423,6 +423,10 @@ extern "C" {
                      struct llama_model * model,
             struct llama_context_params   params);
 
+    // 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_attn_state(struct llama_context * ctx, float * cross_attn_state);
+
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
diff --git a/src/llama.cpp b/src/llama.cpp
index 83b80b59..b189a19a 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -169,6 +169,7 @@ static std::string format(const char * fmt, ...) {
 
 enum llm_arch {
     LLM_ARCH_LLAMA,
+    LLM_ARCH_MLLAMA,
     LLM_ARCH_FALCON,
     LLM_ARCH_BAICHUAN,
     LLM_ARCH_GROK,
@@ -223,6 +224,7 @@ enum llm_arch {
 
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_LLAMA,           "llama"        },
+    { LLM_ARCH_MLLAMA,          "mllama"       },
     { LLM_ARCH_FALCON,          "falcon"       },
     { LLM_ARCH_GROK,            "grok"         },
     { LLM_ARCH_GPT2,            "gpt2"         },
@@ -330,6 +332,7 @@ 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_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_FREQ_BASE,
@@ -439,6 +442,7 @@ 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.%d" },
+    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers"   },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
@@ -613,6 +617,14 @@ 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,
 };
 
 static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@@ -642,6 +654,40 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        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_BAICHUAN,
         {
@@ -2390,6 +2436,7 @@ enum e_model {
     MODEL_40B,
     MODEL_65B,
     MODEL_70B,
+    MODEL_90B,
     MODEL_236B,
     MODEL_314B,
     MODEL_SMALL,
@@ -2434,6 +2481,7 @@ 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;
@@ -2502,10 +2550,11 @@ struct llama_hparams {
         if (this->n_expert      != other.n_expert)      return true;
         if (this->n_expert_used != other.n_expert_used) return true;
 
-        if (this->n_head_arr    != other.n_head_arr)    return true;
-        if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
-        if (this->n_ff_arr      != other.n_ff_arr)      return true;
-        if (this->n_bskcn_arr   != other.n_bskcn_arr)   return true;
+        if (this->n_head_arr        != other.n_head_arr)        return true;
+        if (this->n_head_kv_arr     != other.n_head_kv_arr)     return true;
+        if (this->n_ff_arr          != other.n_ff_arr)          return true;
+        if (this->n_bskcn_arr       != other.n_bskcn_arr)       return true;
+        if (this->cross_attn_layers != other.cross_attn_layers) return true;
 
         if (this->n_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
         if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
@@ -2623,6 +2672,10 @@ struct llama_hparams {
 
         GGML_ABORT("fatal error");
     }
+
+    bool cross_attention_layer(uint32_t il) const {
+        return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
+    }
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
@@ -2806,6 +2859,16 @@ struct llama_layer {
     struct ggml_tensor * ffn_down_scale;
 
     struct ggml_tensor * bskcn_tv;
+
+    // cross attention
+    struct ggml_tensor * cross_attn_k_norm;
+    struct ggml_tensor * cross_attn_k_proj;
+    struct ggml_tensor * cross_attn_o_proj;
+    struct ggml_tensor * cross_attn_q_norm;
+    struct ggml_tensor * cross_attn_q_proj;
+    struct ggml_tensor * cross_attn_v_proj;
+    struct ggml_tensor * cross_attn_attn_gate;
+    struct ggml_tensor * cross_attn_mlp_gate;
 };
 
 // very similar to llama_batch,
@@ -3452,6 +3515,12 @@ struct llama_context {
     struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
     struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
     struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
+
+    // TODO (jmorganca): this should most likely be passed in as part of a batch
+    // and not set on the context for all batches.
+    float * cross_attn_state = nullptr;
+    bool cross_attn_state_first_pass = true;
+    struct ggml_tensor * inp_cross_attn_state; // F32 [4, n_embd, 1061]
 };
 
 struct llama_lora_weight {
@@ -3686,6 +3755,18 @@ static bool llama_kv_cache_init(
     cache.v_l.reserve(n_layer);
 
     for (int i = 0; i < (int) n_layer; i++) {
+        // for cross attention layers
+        if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layer(i)) {
+            struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front();
+            ggml_tensor * k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
+            ggml_tensor * v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
+            ggml_format_name(k, "cache_k_l%d", i);
+            ggml_format_name(v, "cache_v_l%d", i);
+            cache.k_l.push_back(k);
+            cache.v_l.push_back(v);
+            continue;
+        }
+
         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
 
@@ -5460,12 +5541,14 @@ static void llm_load_hparams(
     }
 
     // zero-out the per-layer hparams
-    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.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);
-    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,       hparams.n_ff_arr,          hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT,      hparams.n_head_arr,        hparams.n_layer);
+    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;
@@ -5514,7 +5597,7 @@ static void llm_load_hparams(
 
         ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
 
-        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
+        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_MLLAMA || model.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));
             }
@@ -5554,6 +5637,16 @@ static void llm_load_hparams(
                     }
                 }
             } 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: model.type = e_model::MODEL_11B; break;
+                    case 100: model.type = e_model::MODEL_90B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_MINICPM:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -7249,6 +7342,55 @@ static bool llm_load_tensors(
                         layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head_qk_rope/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
                     }
                 } break;
+            case LLM_ARCH_MLLAMA:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab+8});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, 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 (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        if (hparams.cross_attention_layer(i)) {
+                            layer.cross_attn_k_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_NORM,   "weight", i), {128});
+                            layer.cross_attn_k_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_PROJ,   "weight", i), {n_embd, 1024});
+                            layer.cross_attn_o_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_O_PROJ,   "weight", i), {n_embd, n_embd});
+                            layer.cross_attn_q_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_NORM, "weight", i), {128});
+                            layer.cross_attn_q_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_PROJ, "weight", i), {n_embd, n_embd});
+                            layer.cross_attn_v_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_V_PROJ, "weight", i), {n_embd, 1024});
+                            layer.cross_attn_attn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_ATTN_GATE, i), {1});
+                            layer.cross_attn_mlp_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_MLP_GATE, i), {1});
+                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        } else {
+                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
+                            layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                            layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                            layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                            layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        }
+                    }
+                } break;
             case LLM_ARCH_GROK:
                 {
                     if (n_expert == 0) {
@@ -9093,7 +9235,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
 
         if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
             model.hparams.n_vocab != model.vocab.id_to_token.size()) {
-            throw std::runtime_error("vocab size mismatch");
+            LLAMA_LOG_WARN("%s: vocab mismatch %u !- %zu ...\n", __func__, model.hparams.n_vocab, model.vocab.id_to_token.size());
         }
 
         if (params.vocab_only) {
@@ -9178,7 +9320,7 @@ static struct ggml_tensor * llm_build_inp_embd(
 
         inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
     } else {
-       lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
+        lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
         inpL = lctx.inp_embd;
         ggml_set_input(lctx.inp_embd);
     }
@@ -9193,6 +9335,22 @@ static struct ggml_tensor * llm_build_inp_embd(
     return inpL;
 }
 
+static struct ggml_tensor * llm_build_inp_cross_attn_state(
+        struct ggml_context * ctx,
+       struct llama_context & lctx,
+        const llama_hparams & hparams,
+         const llm_build_cb & cb) {
+    const int64_t n_embd = hparams.n_embd;
+
+    struct ggml_tensor * inpCAS;
+    lctx.inp_cross_attn_state = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, 1601, 4);
+    cb(lctx.inp_cross_attn_state, "inp_cross_attn_state", -1);
+    ggml_set_input(lctx.inp_cross_attn_state);
+    inpCAS = lctx.inp_cross_attn_state;
+
+    return inpCAS;
+}
+
 static void llm_build_kv_store(
         struct ggml_context * ctx,
         const llama_hparams & hparams,
@@ -10167,6 +10325,7 @@ struct llm_build_context {
         lctx.inp_pos_bucket    = nullptr;
         lctx.inp_embd_enc      = nullptr;
         lctx.inp_KQ_mask_cross = nullptr;
+        lctx.inp_cross_attn_state = nullptr;
     }
 
     void free() {
@@ -10754,6 +10913,253 @@ struct llm_build_context {
                 LLM_NORM_RMS, cb, -1);
         cb(cur, "result_norm", -1);
 
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_mllama() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // 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);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+        struct ggml_tensor * inpCAS;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpCAS = llm_build_inp_cross_attn_state(ctx0, lctx, hparams, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            if (hparams.cross_attention_layer(il)) {
+                if (!lctx.cross_attn_state) {
+                    continue;
+                }
+
+                // cross attention layer
+                struct 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_permute(ctx0, Qcur, 0, 2, 1, 3);
+                cb(Qcur, "Qcur", il);
+
+                // TODO: is this required?
+                Qcur = ggml_cont(ctx0, Qcur);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].cross_attn_q_norm, NULL, LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur;
+                if (lctx.cross_attn_state_first_pass) {
+                    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_permute(ctx0, Kcur, 0, 2, 1, 3);
+                    cb(Kcur, "Kcur", il);
+
+                    // TODO: is this required?
+                    Kcur = ggml_cont(ctx0, Kcur);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].cross_attn_k_norm, NULL, LLM_NORM_RMS, cb, il);
+                    cb(Kcur, "Kcur", il);
+
+                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, kv_self.k_l[il]));
+                } else {
+                    Kcur = ggml_view_tensor(ctx0, kv_self.k_l[il]);
+                    cb(Kcur, "Kcur (view)", il);
+                }
+
+                struct ggml_tensor * Vcur;
+                if (lctx.cross_attn_state_first_pass) {
+                    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 {
+                    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);
+
+                kq = ggml_scale_inplace(ctx0, kq, 1.0f/sqrtf(float(n_embd_head)));
+                cb(kq, "kq_scaled", il);
+
+                // TODO: apply causal masks
+                struct ggml_tensor * kq_soft_max = ggml_soft_max_inplace(ctx0, kq);
+                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 = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, 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, cb, 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 = lctx.cvec.apply_to(ctx0, 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
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, 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);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, 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);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, 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_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, 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 = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, 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, cb, il);
+                cb(cur, "ffn_out", il);
+
+                cur = ggml_add(ctx0, cur, ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                cur = lctx.cvec.apply_to(ctx0, cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            }
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
         // lm_head
         cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
         cb(cur, "result_output", -1);
@@ -16501,6 +16907,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_llama();
             } break;
+        case LLM_ARCH_MLLAMA:
+            {
+                result = llm.build_mllama();
+            } break;
         case LLM_ARCH_BAICHUAN:
             {
                 result = llm.build_baichuan();
@@ -16773,6 +17183,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
         ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
     }
 
+    // TODO (jmorganca): this might copy a lot of data on every request of a
+    // single generation even though it doesn't change, so we should
+    // find a way to not set this more than one time per image
+    if (lctx.inp_cross_attn_state &&
+        lctx.inp_cross_attn_state->buffer) {
+        ggml_backend_tensor_set(lctx.inp_cross_attn_state, lctx.cross_attn_state, 0, hparams.n_embd * 1601 * 4 * ggml_element_size(lctx.inp_cross_attn_state));
+    }
+
     if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
         GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs");
         const int64_t n_tokens = batch.n_tokens;
@@ -17455,6 +17873,10 @@ static int llama_decode_internal(
 
         llama_set_inputs(lctx, ubatch);
 
+        // TODO: replace with something better to find out if its
+        // our first actual pass
+        lctx.cross_attn_state_first_pass = false;
+
         llama_graph_compute(lctx, gf, n_threads, threadpool);
 
         // update the kv ring buffer
@@ -18648,7 +19070,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         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;
@@ -19744,6 +20168,11 @@ struct llama_context * llama_new_context_with_model(
     return ctx;
 }
 
+void llama_set_cross_attn_state(struct llama_context * ctx, float * cross_attn_state) {
+    ctx->cross_attn_state_first_pass = true;
+    ctx->cross_attn_state = cross_attn_state;
+}
+
 void llama_free(struct llama_context * ctx) {
     delete ctx;
 }
@@ -19814,6 +20243,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values
         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_MLLAMA:
         case LLM_ARCH_BAICHUAN:
         case LLM_ARCH_STARCODER:
         case LLM_ARCH_PLAMO: