0005-solar-pro.patch

From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Mon, 16 Sep 2024 15:53:16 -0700
Subject: [PATCH] solar-pro

solar-pro introduces block skip connections where blocks are connected
to other, non-sequential blocks with a scale multiple

this change adds 4 new keys to store the skip connections and one new
tensor to store the scalar. the scalar is implemented a 1-dimensional
tensor with 2 elements dervied from the model's bskcn_tv configuration.
in general, the values are (bskcn_tv, 1 - bskcn_tv)
---
 src/llama-arch.cpp         |  21 +++++
 src/llama-arch.h           |   3 +
 src/llama-hparams.cpp      |   8 ++
 src/llama-hparams.h        |   5 ++
 src/llama-model-loader.cpp |   1 +
 src/llama-model.cpp        |  44 +++++++++++
 src/llama-model.h          |   3 +
 src/llama.cpp              | 152 ++++++++++++++++++++++++++++++++++++-
 8 files changed, 236 insertions(+), 1 deletion(-)

diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 97a1e7e5..a1e0ebcc 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -61,6 +61,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_GRANITE,          "granite"          },
     { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
     { LLM_ARCH_CHAMELEON,        "chameleon"        },
+    { LLM_ARCH_SOLAR,            "solar"            },
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
@@ -125,6 +126,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
     { 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_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -1271,6 +1273,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
         },
     },
+    {
+        LLM_ARCH_SOLAR,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { 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_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_BSKCN_TV,        "bskcn_tv" },
+        },
+    },
     {
         LLM_ARCH_WAVTOKENIZER_DEC,
         {
@@ -1429,6 +1449,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     // 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_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 122fdceb..77919578 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -65,6 +65,7 @@ enum llm_arch {
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
     LLM_ARCH_CHAMELEON,
+    LLM_ARCH_SOLAR,
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_UNKNOWN,
 };
@@ -129,6 +130,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -311,6 +313,7 @@ enum llm_tensor {
     LLM_TENSOR_ENC_OUTPUT_NORM,
     LLM_TENSOR_CLS,
     LLM_TENSOR_CLS_OUT,
+    LLM_TENSOR_BSKCN_TV,
     LLM_TENSOR_CONV1D,
     LLM_TENSOR_CONVNEXT_DW,
     LLM_TENSOR_CONVNEXT_NORM,
diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
index ea87b295..f3955de9 100644
--- a/src/llama-hparams.cpp
+++ b/src/llama-hparams.cpp
@@ -69,3 +69,11 @@ uint32_t llama_hparams::n_embd_v_s() const {
     // corresponds to Mamba's ssm_states size
     return ssm_d_state * ssm_d_inner;
 }
+
+bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {
+    if (il < n_layer) {
+        return n_bskcn_arr[n][il] > 0;
+    }
+
+    GGML_ABORT("fatal error");
+}
\ No newline at end of file
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index 1fe45410..1bdcdfd5 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -50,6 +50,8 @@ struct llama_hparams {
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
 
+    std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr = {};
+
     uint32_t n_layer_dense_lead = 0;
     uint32_t n_lora_q           = 0;
     uint32_t n_lora_kv          = 0;
@@ -133,6 +135,9 @@ struct llama_hparams {
 
     // dimension of the recurrent state embeddings
     uint32_t n_embd_v_s() const;
+
+    // Block skip connection
+    bool n_bskcn(uint32_t n, uint32_t il) const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index 05d58ad9..1252aca1 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -439,6 +439,7 @@ namespace GGUFMeta {
     // TODO: this is not very clever - figure out something better
     template bool llama_model_loader::get_key_or_arr<std::array<int, 4>>(enum llm_kv kid, std::array<int, 4> & result, uint32_t n, bool required);
     template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
+    template bool llama_model_loader::get_key_or_arr<uint32_t>(const std::string & key, std::array<uint32_t, 512> & result, uint32_t n, bool required);
 
 llama_model_loader::llama_model_loader(
         const std::string & fname,
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 36a0a009..ad1315c6 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -1238,6 +1238,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                }
             } break;
+        case LLM_ARCH_SOLAR:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                for (size_t i = 0; i < hparams.n_bskcn_arr.max_size(); ++i) {
+                    auto & bskcn = hparams.n_bskcn_arr[i];
+                    bskcn.fill(0);
+                    auto kv = LLM_KV(arch);
+                    ml.get_key_or_arr(format((kv(LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION) + ".%d").c_str(), i), bskcn, hparams.n_layer, false);
+                }
+
+                switch (hparams.n_layer) {
+                    case 64: type = LLM_TYPE_22B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_WAVTOKENIZER_DEC:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
@@ -3316,6 +3331,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 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_SOLAR:
+                {
+                    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}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                    }
+
+                    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);
+
+                        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.bskcn_tv = create_tensor(tn(LLM_TENSOR_BSKCN_TV, "weight", i), {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);
@@ -3900,6 +3943,7 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) {
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_CHAMELEON:
+        case LLM_ARCH_SOLAR:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
diff --git a/src/llama-model.h b/src/llama-model.h
index a7c30444..1afb0024 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -55,6 +55,7 @@ enum llm_type {
     LLM_TYPE_15B,
     LLM_TYPE_16B,
     LLM_TYPE_20B,
+    LLM_TYPE_22B,
     LLM_TYPE_30B,
     LLM_TYPE_32B,
     LLM_TYPE_34B,
@@ -281,6 +282,8 @@ struct llama_layer {
     struct ggml_tensor * ffn_up_scale   = nullptr;
     struct ggml_tensor * ffn_down_scale = nullptr;
 
+    struct ggml_tensor * bskcn_tv = nullptr;
+
     struct llama_layer_posnet posnet;
 
     struct llama_layer_convnext convnext;
diff --git a/src/llama.cpp b/src/llama.cpp
index ac85bfed..6d320ea4 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -7953,9 +7953,155 @@ struct llm_build_context {
         cb(img_logits, "img_logits", -1);
         cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
         cb(cur, "result_output", -1);
-
         ggml_build_forward_expand(gf, cur);
+        return gf;
+   }
+
+   ggml_cgraph * build_solar() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), 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;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, 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();
+
+        struct ggml_tensor * bskcn_1;
+        struct ggml_tensor * bskcn_2;
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            if (hparams.n_bskcn(0, il)) {
+                bskcn_1 = inpSA;
+            }
+
+            if (hparams.n_bskcn(1, il)) {
+                bskcn_2 = inpSA;
+            }
+
+            if (hparams.n_bskcn(2, il)) {
+                inpSA = ggml_add(
+                   ctx0,
+                   ggml_mul(ctx0, bskcn_1, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
+                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
+            }
+
+            if (hparams.n_bskcn(3, il)) {
+                inpSA = ggml_add(
+                   ctx0,
+                   ggml_mul(ctx0, bskcn_2, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
+                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
+            }
 
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // 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);
+        ggml_build_forward_expand(gf, cur);
         return gf;
     }
 
@@ -8398,6 +8544,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_chameleon();
             } break;
+        case LLM_ARCH_SOLAR:
+            {
+                result = llm.build_solar();
+            } break;
         case LLM_ARCH_WAVTOKENIZER_DEC:
             {
                 result = llm.build_wavtokenizer_dec();