Add support for Cohere2 (#341)

* Add support for Cohere2

* Fixe IQ4_NL on AVX2

* Command-A needs fp32 precision for K*Q

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

1 files changed, 199 insertions, 4 deletions

diff --git a/src/llama.cpp b/src/llama.cpp
index c870b09e..26ddeb2e 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -229,6 +229,7 @@ enum llm_arch {
     LLM_ARCH_JAIS,
     LLM_ARCH_GRANITE = 46,
     LLM_ARCH_GRANITE_MOE,
+    LLM_ARCH_COHERE2,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -279,6 +280,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_JAIS,            "jais"         },
     { LLM_ARCH_GRANITE,         "granite"      },
     { LLM_ARCH_GRANITE_MOE,     "granitemoe"   },
+    { LLM_ARCH_COHERE2,         "cohere2"      },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -1456,7 +1458,21 @@ 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_COHERE2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { 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_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2539,6 +2555,7 @@ struct llama_hparams {
         if (this->n_layer       != other.n_layer)       return true;
         if (this->n_rot         != other.n_rot)         return true;
         if (this->n_swa         != other.n_swa)         return true;
+        if (this->n_swa_pattern != other.n_swa_pattern) return false;
         if (this->n_embd_head_k != other.n_embd_head_k) return true;
         if (this->n_embd_head_v != other.n_embd_head_v) return true;
         if (this->n_expert      != other.n_expert)      return true;
@@ -5797,6 +5814,17 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_COHERE2:
+            {
+                hparams.n_swa_pattern = 4;
+                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
+                ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_8B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -6406,6 +6434,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
         LLAMA_LOG_INFO("%s: n_head_kv        = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_rot            = %u\n",     __func__, hparams.n_rot);
         LLAMA_LOG_INFO("%s: n_swa            = %u\n",     __func__, hparams.n_swa);
+        LLAMA_LOG_INFO("%s: n_swa_pattern    = %u\n",     __func__, hparams.n_swa_pattern);
         LLAMA_LOG_INFO("%s: n_embd_head_k    = %u\n",     __func__, hparams.n_embd_head_k);
         LLAMA_LOG_INFO("%s: n_embd_head_v    = %u\n",     __func__, hparams.n_embd_head_v);
         LLAMA_LOG_INFO("%s: n_gqa            = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il);        }, hparams.n_layer).c_str());
@@ -8397,6 +8426,34 @@ static bool llm_load_tensors(
                         layer.ffn_down   = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
                     }
                 } break;
+            case LLM_ARCH_COHERE2:
+                {
+                    model.tok_embd = create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+                    // output
+                    model.output_norm = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+                    // init output from the input tok embed
+                    model.output      = create_tensor(ctx_output_split, 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 = model.layers[i];
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        layer.attn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
+
+                        layer.wq = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd }, 0);
+                        layer.wk = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_gqa }, 0);
+                        layer.wv = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_gqa }, 0);
+                        layer.wo = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }, 0);
+
+                        layer.ffn_gate = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, 0);
+                        layer.ffn_down = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
+                        layer.ffn_up   = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, 0);
+                    }
+                }
+                break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -9340,7 +9397,7 @@ static struct ggml_tensor * llm_build_kqv(
         // For DeepSeek-2, it is perfectly fine with fp16 for PP, but I get gibberish when uding fp16 for TG.
         // Not sure if it is really a matter of insufficient precision, or I have made a mistake in the fattn-vec-f16 kernel.
         if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX ||
-            (model.arch == LLM_ARCH_DEEPSEEK2 && q->ne[1] <= 8)) {
+            (model.arch == LLM_ARCH_DEEPSEEK2 && q->ne[1] <= 8) || model.arch == LLM_ARCH_COHERE2) {
             ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
         }
         //ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
@@ -9364,7 +9421,8 @@ static struct ggml_tensor * llm_build_kqv(
 
             //ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
 
-            if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
+            if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2 ||
+                model.arch == LLM_ARCH_COHERE2) {
                 // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
                 // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
                 ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@@ -9423,7 +9481,8 @@ static struct ggml_tensor * llm_build_kqv(
                 auto k_i = ggml_view_3d(ctx, k, k->ne[0], k->ne[1], this_ne12, k->nb[1], k->nb[2], k->nb[2]*i02);
                 auto q_i = ggml_view_3d(ctx, q, q->ne[0], q->ne[1], this_ne12, q->nb[1], q->nb[2], q->nb[2]*i12);
                 auto kq_i = ggml_mul_mat(ctx, k_i, q_i);
-                if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
+                if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2 ||
+                    model.arch == LLM_ARCH_COHERE2) {
                     ggml_mul_mat_set_prec(kq_i, GGML_PREC_F32);
                 }
                 if (model.arch == LLM_ARCH_GROK) {
@@ -15013,6 +15072,137 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_cohere2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        const float f_logit_scale = hparams.f_logit_scale;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, 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)
+        // cohere2 requires different mask for layers using sliding window (SWA)
+        struct ggml_tensor * KQ_mask     = build_inp_KQ_mask();
+        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa();
+
+        // sliding window switch pattern
+        const int32_t sliding_window_pattern = 4;
+
+        for (int il = 0; il < n_layer; ++il) {
+            // three layers sliding window attention (window size 4096) and ROPE
+            // fourth layer uses global attention without positional embeddings
+            const bool           is_sliding = il % sliding_window_pattern < (sliding_window_pattern - 1);
+            struct ggml_tensor * KQ_mask_l = is_sliding ? KQ_mask_swa : KQ_mask;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+            struct ggml_tensor * ffn_inp = cur;
+
+            // self-attention
+            {
+                // rope freq factors for 128k context
+                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);
+                }
+
+                if (is_sliding) {
+                    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);
+                } else {
+                    // For non-sliding layers, just reshape without applying RoPE
+                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                    cb(Qcur, "Qcur", il);
+
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                    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_l, 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();
+                cur                              = ggml_get_rows(ctx0, cur, inp_out_ids);
+                inpL                             = ggml_get_rows(ctx0, inpL, inp_out_ids);
+                ffn_inp                          = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+            }
+
+            struct ggml_tensor * attn_out = cur;
+
+            // feed-forward network
+            {
+                cur = llm_build_ffn(ctx0, lctx, ffn_inp, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate,
+                                    NULL, NULL, model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR,
+                                    cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            // add together residual + FFN + self-attention
+            cur = ggml_add(ctx0, cur, inpL);
+            cur = ggml_add(ctx0, cur, attn_out);
+            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, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        if (f_logit_scale) {
+            cur = ggml_scale(ctx0, cur, f_logit_scale);
+        }
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     struct ggml_cgraph * build_t5_encoder() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
@@ -15813,6 +16003,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_bitnet_25();
             } break;
+        case LLM_ARCH_COHERE2:
+            {
+                result = llm.build_cohere2();
+            } break;
         case LLM_ARCH_T5:
             {
                 if (lctx.is_encoding) {
@@ -19486,6 +19680,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_CHATGLM:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
+        case LLM_ARCH_COHERE2:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2