ollama/llama/patches/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.cpp | 267 +++++++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 253 insertions(+), 14 deletions(-)

diff --git a/src/llama.cpp b/src/llama.cpp
index 9e292c4f..26be6254 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -196,6 +196,7 @@ enum llm_arch {
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
     LLM_ARCH_CHAMELEON,
+    LLM_ARCH_SOLAR,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -251,6 +252,7 @@ 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_UNKNOWN,         "(unknown)"    },
 };
 
@@ -308,6 +310,7 @@ 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,
@@ -411,20 +414,21 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_RESIDUAL_SCALE,                    "%s.residual_scale"                    },
     { LLM_KV_EMBEDDING_SCALE,                   "%s.embedding_scale"                   },
 
-    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
-    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
-    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"         },
-    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"              },
-    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"             },
-    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"           },
-    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"     },
-    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon" },
-    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                 },
-    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"            },
-    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
-    { 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_HEAD_COUNT,             "%s.attention.head_count"               },
+    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"            },
+    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"           },
+    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"                },
+    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"               },
+    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"             },
+    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"       },
+    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon"   },
+    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                   },
+    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"              },
+    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"             },
+    { 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.%d" },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,             "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,          "%s.rope.dimension_sections"              },
@@ -607,6 +611,7 @@ enum llm_tensor {
     LLM_TENSOR_ENC_OUTPUT_NORM,
     LLM_TENSOR_CLS,
     LLM_TENSOR_CLS_OUT,
+    LLM_TENSOR_BSKCN_TV,
 };
 
 static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_NAMES = {
@@ -1564,6 +1569,24 @@ 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_UNKNOWN,
         {
@@ -2425,6 +2448,7 @@ enum e_model {
     MODEL_15B,
     MODEL_16B,
     MODEL_20B,
+    MODEL_22B,
     MODEL_30B,
     MODEL_32B,
     MODEL_34B,
@@ -2475,6 +2499,8 @@ 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;
@@ -2546,6 +2572,7 @@ struct llama_hparams {
         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_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
         if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
@@ -2658,6 +2685,14 @@ struct llama_hparams {
             return ssm_d_state * ssm_d_inner;
         }
     }
+
+    bool n_bskcn(uint32_t n, uint32_t il = 0) const {
+        if (il < n_layer) {
+            return n_bskcn_arr[n][il] > 0;
+        }
+
+        GGML_ABORT("fatal error");
+    }
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
@@ -2844,6 +2879,8 @@ struct llama_layer {
     struct ggml_tensor * ffn_gate_scale;
     struct ggml_tensor * ffn_up_scale;
     struct ggml_tensor * ffn_down_scale;
+
+    struct ggml_tensor * bskcn_tv;
 };
 
 // very similar to llama_batch,
@@ -6247,6 +6284,21 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                }
             } break;
+        case LLM_ARCH_SOLAR:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                for (int i = 0; i < hparams.n_bskcn_arr.max_size(); ++i) {
+                    auto & bskcn = hparams.n_bskcn_arr.at(i);
+                    bskcn.fill(0);
+                    ml.get_key_or_arr(::format(LLM_KV_NAMES.at(LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION), LLM_ARCH_NAMES.at(ml.llm_kv.arch), i), bskcn, hparams.n_layer, false);
+                }
+
+                switch (hparams.n_layer) {
+                    case 64: model.type = e_model::MODEL_22B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            }
         default: (void)0;
     }
 
@@ -7239,6 +7291,7 @@ static const std::map<llm_tensor, llm_tensor_info> llm_tensor_info_mapping = {
     {LLM_TENSOR_FFN_UP_EXPS,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     // 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}}
 };
 
 // checks if the weight tensor can be used with the specified buffer type and device
@@ -9253,6 +9306,35 @@ static bool llm_load_tensors(
 
                         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:
+                {
+                    model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    {
+                        model.output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                        model.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 = model.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);
@@ -16671,6 +16753,158 @@ struct llm_build_context {
 
         return gf;
     }
+
+    ggml_cgraph * build_solar() {
+        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;
+
+        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;
+    }
 };
 
 static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@@ -16942,6 +17176,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_chameleon();
             } break;
+        case LLM_ARCH_SOLAR:
+            {
+                result = llm.build_solar();
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -20137,6 +20375,7 @@ enum llama_rope_type llama_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