@fugood/llama.node 1.2.0-rc.0 → 1.2.1

package/src/llama.cpp/ggml/src/ggml-cpu/ops.cpp

@@ -776,6 +776,24 @@ static void ggml_compute_forward_dup_f32(
                         id += ne00 * (ne01 - ir1);
                     }
                 }
+            } else if (dst->type == GGML_TYPE_I32) {
+                size_t id = 0;
+                int32_t * dst_ptr = (int32_t *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += ne00 * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = *src0_ptr;
+                                id++;
+                            }
+                        }
+                        id += ne00 * (ne01 - ir1);
+                    }
+                }
             } else {
                 GGML_ABORT("fatal error"); // TODO: implement
             }
@@ -947,6 +965,144 @@ static void ggml_compute_forward_dup_f32(
                 }
             }
         }
+    } else if (dst->type == GGML_TYPE_I32) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                i10 += ne00 * ir0;
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+
+                        *(int32_t *) dst_ptr = *(const float *) src0_ptr;
+
+                        if (++i10 == ne0) {
+                            i10 = 0;
+                            if (++i11 == ne1) {
+                                i11 = 0;
+                                if (++i12 == ne2) {
+                                    i12 = 0;
+                                    if (++i13 == ne3) {
+                                        i13 = 0;
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+                i10 += ne00 * (ne01 - ir1);
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    } else {
+        GGML_ABORT("fatal error"); // TODO: implement
+    }
+}
+
+static void ggml_compute_forward_dup_i32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    const int ith = params->ith; // thread index
+    const int nth = params->nth; // number of threads
+
+    // parallelize by rows
+    const int nr = ne01;
+    // number of rows per thread
+    const int dr = (nr + nth - 1) / nth;
+    // row range for this thread
+    const int ir0 = dr * ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    // dst counters
+
+    int64_t i10 = 0;
+    int64_t i11 = 0;
+    int64_t i12 = 0;
+    int64_t i13 = 0;
+
+    // TODO: not optimal, but works
+    if (dst->type == GGML_TYPE_F32) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                i10 += ne00 * ir0;
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+
+                        *(float *) dst_ptr = *(const int32_t *) src0_ptr;
+
+                        if (++i10 == ne0) {
+                            i10 = 0;
+                            if (++i11 == ne1) {
+                                i11 = 0;
+                                if (++i12 == ne2) {
+                                    i12 = 0;
+                                    if (++i13 == ne3) {
+                                        i13 = 0;
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+                i10 += ne00 * (ne01 - ir1);
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+            }
+        }
     } else {
         GGML_ABORT("fatal error"); // TODO: implement
     }
@@ -1177,6 +1333,10 @@ void ggml_compute_forward_dup(
             {
                 ggml_compute_forward_dup_f32(params, dst);
             } break;
+        case GGML_TYPE_I32:
+            {
+                ggml_compute_forward_dup_i32(params, dst);
+            } break;
         default:
             {
                 if (ggml_is_quantized(src0->type) && dst->type == GGML_TYPE_F32) {
@@ -8438,7 +8598,7 @@ static void ggml_compute_forward_timestep_embedding_f32(
             embed_data[j + half] = sinf(arg);
         }
         if (dim % 2 != 0 && ith == 0) {
-            embed_data[dim] = 0.f;
+            embed_data[2 * half] = 0.f;
         }
     }
 }

package/src/llama.cpp/src/llama-arch.cpp

@@ -96,6 +96,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DREAM,            "dream"            },
     { LLM_ARCH_SMALLTHINKER,     "smallthinker"     },
     { LLM_ARCH_LLADA,            "llada"            },
+    { LLM_ARCH_LLADA_MOE,        "llada-moe"        },
     { LLM_ARCH_SEED_OSS,         "seed_oss"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
@@ -137,7 +138,9 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
+    { LLM_KV_DECODER_BLOCK_COUNT,               "%s.decoder_block_count"               },
     { LLM_KV_ATTN_LOGIT_SOFTCAPPING,            "%s.attn_logit_softcapping"            },
+    { LLM_KV_ROUTER_LOGIT_SOFTCAPPING,          "%s.router_logit_softcapping"          },
     { LLM_KV_FINAL_LOGIT_SOFTCAPPING,           "%s.final_logit_softcapping"           },
     { LLM_KV_SWIN_NORM,                         "%s.swin_norm"                         },
     { LLM_KV_RESCALE_EVERY_N_LAYERS,            "%s.rescale_every_n_layers"            },
@@ -168,19 +171,25 @@ 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_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
+    { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
 
-    { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
-    { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
-    { LLM_KV_ROPE_FREQ_BASE,            "%s.rope.freq_base"                       },
-    { LLM_KV_ROPE_SCALE_LINEAR,         "%s.rope.scale_linear"                    },
-    { LLM_KV_ROPE_SCALING_TYPE,         "%s.rope.scaling.type"                    },
-    { LLM_KV_ROPE_SCALING_FACTOR,       "%s.rope.scaling.factor"                  },
-    { LLM_KV_ROPE_SCALING_ATTN_FACTOR,  "%s.rope.scaling.attn_factor"             },
-    { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
-    { LLM_KV_ROPE_SCALING_FINETUNED,    "%s.rope.scaling.finetuned"               },
-    { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier"     },
+    { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
+    { LLM_KV_ROPE_DIMENSION_SECTIONS,       "%s.rope.dimension_sections"              },
+    { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
+    { LLM_KV_ROPE_SCALE_LINEAR,             "%s.rope.scale_linear"                    },
+    { LLM_KV_ROPE_SCALING_TYPE,             "%s.rope.scaling.type"                    },
+    { LLM_KV_ROPE_SCALING_FACTOR,           "%s.rope.scaling.factor"                  },
+    { LLM_KV_ROPE_SCALING_ATTN_FACTOR,      "%s.rope.scaling.attn_factor"             },
+    { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,     "%s.rope.scaling.original_context_length" },
+    { LLM_KV_ROPE_SCALING_FINETUNED,        "%s.rope.scaling.finetuned"               },
+    { LLM_KV_ROPE_SCALING_YARN_LOG_MUL,     "%s.rope.scaling.yarn_log_multiplier"     },
+    { LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,  "%s.rope.scaling.yarn_ext_factor"         },
+    { LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor"        },
+    { LLM_KV_ROPE_SCALING_YARN_BETA_FAST,   "%s.rope.scaling.yarn_beta_fast"          },
+    { LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,   "%s.rope.scaling.yarn_beta_slow"          },
 
     { LLM_KV_SPLIT_NO,            "split.no"            },
     { LLM_KV_SPLIT_COUNT,         "split.count"         },
@@ -397,12 +406,16 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { 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_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
             { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
             { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
         },
@@ -2135,6 +2148,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_LLADA_MOE,
+        {
+            { 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_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { 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_INP,       "blk.%d.ffn_gate_inp" },
+            { 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_ARCH_SEED_OSS,
         {
@@ -2415,6 +2448,7 @@ bool llm_arch_is_diffusion(const llm_arch & arch) {
     switch (arch) {
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
+        case LLM_ARCH_LLADA_MOE:
             return true;
         default:
             return false;

package/src/llama.cpp/src/llama-arch.h

@@ -100,6 +100,7 @@ enum llm_arch {
     LLM_ARCH_DREAM,
     LLM_ARCH_SMALLTHINKER,
     LLM_ARCH_LLADA,
+    LLM_ARCH_LLADA_MOE,
     LLM_ARCH_SEED_OSS,
     LLM_ARCH_UNKNOWN,
 };
@@ -141,7 +142,9 @@ enum llm_kv {
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
+    LLM_KV_DECODER_BLOCK_COUNT,
     LLM_KV_ATTN_LOGIT_SOFTCAPPING,
+    LLM_KV_ROUTER_LOGIT_SOFTCAPPING,
     LLM_KV_FINAL_LOGIT_SOFTCAPPING,
     LLM_KV_SWIN_NORM,
     LLM_KV_RESCALE_EVERY_N_LAYERS,
@@ -172,6 +175,8 @@ enum llm_kv {
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_OUTPUT_SCALE,
+    LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
@@ -185,6 +190,10 @@ enum llm_kv {
     LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
     LLM_KV_ROPE_SCALING_FINETUNED,
     LLM_KV_ROPE_SCALING_YARN_LOG_MUL,
+    LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,
+    LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR,
+    LLM_KV_ROPE_SCALING_YARN_BETA_FAST,
+    LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,
 
     LLM_KV_SPLIT_NO,
     LLM_KV_SPLIT_COUNT,

package/src/llama.cpp/src/llama-chat.cpp

@@ -70,6 +70,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "hunyuan-dense",     LLM_CHAT_TEMPLATE_HUNYUAN_DENSE     },
     { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
     { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
+    { "grok-2",            LLM_CHAT_TEMPLATE_GROK_2            },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -204,6 +205,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_KIMI_K2;
     } else if (tmpl_contains("<seed:bos>")) {
         return LLM_CHAT_TEMPLATE_SEED_OSS;
+    } else if (tmpl_contains("'Assistant: '  + message['content'] + '<|separator|>")) {
+        return LLM_CHAT_TEMPLATE_GROK_2;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -763,6 +766,20 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<seed:bos>assistant\n";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GROK_2) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "System: " << trim(message->content) << "<|separator|>\n\n";
+            } else if (role == "user") {
+                ss << "Human: " << trim(message->content) << "<|separator|>\n\n";
+            } else if (role == "assistant") {
+                ss << "Assistant: " << message->content << "<|separator|>\n\n";
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
     } else {
         // template not supported
         return -1;

package/src/llama.cpp/src/llama-chat.h

@@ -50,6 +50,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_HUNYUAN_DENSE,
     LLM_CHAT_TEMPLATE_KIMI_K2,
     LLM_CHAT_TEMPLATE_SEED_OSS,
+    LLM_CHAT_TEMPLATE_GROK_2,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

package/src/llama.cpp/src/llama-context.cpp

@@ -35,10 +35,10 @@ llama_context::llama_context(
 
     cparams.n_threads        = params.n_threads;
     cparams.n_threads_batch  = params.n_threads_batch;
-    cparams.yarn_ext_factor  = params.yarn_ext_factor;
-    cparams.yarn_attn_factor = params.yarn_attn_factor;
-    cparams.yarn_beta_fast   = params.yarn_beta_fast;
-    cparams.yarn_beta_slow   = params.yarn_beta_slow;
+    cparams.yarn_ext_factor  = params.yarn_ext_factor  >= 0.0f ? params.yarn_ext_factor  : hparams.yarn_ext_factor;
+    cparams.yarn_attn_factor = params.yarn_attn_factor >= 0.0f ? params.yarn_attn_factor : hparams.yarn_attn_factor;
+    cparams.yarn_beta_fast   = params.yarn_beta_fast   >= 0.0f ? params.yarn_beta_fast   : hparams.yarn_beta_fast;
+    cparams.yarn_beta_slow   = params.yarn_beta_slow   >= 0.0f ? params.yarn_beta_slow   : hparams.yarn_beta_slow;
     cparams.embeddings       = params.embeddings;
     cparams.offload_kqv      = params.offload_kqv;
     cparams.no_perf          = params.no_perf;
@@ -181,7 +181,7 @@ llama_context::llama_context(
         // graph outputs buffer
         {
             // resized during inference when a batch uses more outputs
-            if ((uint32_t) output_reserve(params.n_seq_max) < params.n_seq_max) {
+            if (output_reserve(params.n_seq_max) < params.n_seq_max) {
                 throw std::runtime_error("failed to reserve initial output buffer");
             }
 
@@ -285,8 +285,8 @@ llama_context::llama_context(
         const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
-        // avoid reserving graphs with zero outputs
-        n_outputs = 1;
+        // avoid reserving graphs with zero outputs - assume one output per sequence
+        n_outputs = n_seqs;
 
         LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
 
@@ -1447,7 +1447,9 @@ ggml_status llama_context::graph_compute(
     if (backend_cpu != nullptr) {
         auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_cpu));
         auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool");
-        set_threadpool_fn(backend_cpu, tp);
+        if (set_threadpool_fn) {
+            set_threadpool_fn(backend_cpu, tp);
+        }
     }
 
     // set the number of threads for all the backends
@@ -2261,9 +2263,9 @@ llama_context_params llama_context_default_params() {
         /*.rope_freq_base              =*/ 0.0f,
         /*.rope_freq_scale             =*/ 0.0f,
         /*.yarn_ext_factor             =*/ -1.0f,
-        /*.yarn_attn_factor            =*/ 1.0f,
-        /*.yarn_beta_fast              =*/ 32.0f,
-        /*.yarn_beta_slow              =*/ 1.0f,
+        /*.yarn_attn_factor            =*/ -1.0f,
+        /*.yarn_beta_fast              =*/ -1.0f,
+        /*.yarn_beta_slow              =*/ -1.0f,
         /*.yarn_orig_ctx               =*/ 0,
         /*.defrag_thold                =*/ -1.0f,
         /*.cb_eval                     =*/ nullptr,

package/src/llama.cpp/src/llama-graph.cpp

@@ -1273,7 +1273,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
     // split the batch into streams if needed
     const auto n_stream = k->ne[3];
 
-    q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream);
+    q = ggml_view_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream, q->nb[1], q->nb[2], q->nb[3]/n_stream, 0);
 
     q = ggml_permute(ctx0, q, 0, 2, 1, 3);
     k = ggml_permute(ctx0, k, 0, 2, 1, 3);
@@ -1335,14 +1335,14 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         if (arch == LLM_ARCH_GROK) {
             // need to do the following:
-            // multiply by attn_output_multiplyer of 0.08838834764831845
+            // multiply by attn_output_multiplier
             // and then :
             // kq = 30 * tanh(kq / 30)
             // before the softmax below
 
-            kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, 0.08838834764831845f/30.0f));
+            kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, hparams.f_attn_out_scale / hparams.f_attn_logit_softcapping));
             cb(kq, "kq_tanh", il);
-            kq = ggml_scale(ctx0, kq, 30);
+            kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping);
             cb(kq, "kq_scaled", il);
         }
 
@@ -1431,7 +1431,8 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // [TAG_NO_CACHE_PAD]
     // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams
-    assert(!ubatch.equal_seqs() || (k_cur->ne[3] == 1 && k_cur->ne[3] == ubatch.n_seqs_unq));
+    //       but it might not be worth it: https://github.com/ggml-org/llama.cpp/pull/15636
+    //assert(!ubatch.equal_seqs() || (k_cur->ne[3] == 1 && k_cur->ne[3] == ubatch.n_seqs_unq));
 
     ggml_tensor * q = q_cur;
     ggml_tensor * k = k_cur;

package/src/llama.cpp/src/llama-hparams.h

@@ -82,8 +82,9 @@ struct llama_hparams {
     float f_norm_rms_eps;
     float f_norm_group_eps;
 
-    float f_attn_logit_softcapping  = 50.0f;
-    float f_final_logit_softcapping = 30.0f;
+    float f_attn_logit_softcapping   = 50.0f;
+    float f_router_logit_softcapping = 30.0f;
+    float f_final_logit_softcapping  = 30.0f;
 
     // for RWKV
     uint32_t rescale_every_n_layers = 0;
@@ -104,6 +105,11 @@ struct llama_hparams {
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul = 0.0f;
 
+    float    yarn_ext_factor  = -1.0f;
+    float    yarn_attn_factor =  1.0f;
+    float    yarn_beta_fast   = 32.0f;
+    float    yarn_beta_slow   =  1.0f;
+
     std::array<int, 4> rope_sections;
 
     // Sliding Window Attention (SWA)
@@ -136,10 +142,14 @@ struct llama_hparams {
     float f_embedding_scale = 0.0f;
     float f_attention_scale = 0.0f;
 
+    // grok-2
+    float    f_attn_out_scale = 0.0f;
+    uint32_t attn_temp_length = 0;
+
     bool causal_attn   = true;
     bool use_alibi     = false;
     bool attn_soft_cap = false;
-    bool use_kq_norm   = true;
+    bool use_kq_norm   = false;
 
     // for Classifiers
     uint32_t n_cls_out = 1;
@@ -159,6 +169,7 @@ struct llama_hparams {
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
+    uint32_t    dec_n_layer        = 0;
 
     enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;

package/src/llama.cpp/src/llama-kv-cache.cpp

@@ -1018,16 +1018,33 @@ ggml_tensor * llama_kv_cache::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggm
 
     const int32_t ikv = map_layer_ids.at(il);
 
-    auto * k = layers[ikv].k;
+    ggml_tensor * k = layers[ikv].k;
+
+    const int64_t n_embd_head = k_cur->ne[0];
+    const int64_t n_head      = k_cur->ne[1];
+    const int64_t n_tokens    = k_cur->ne[2];
+
+    const int64_t n_embd_gqa = n_embd_head*n_head;
 
-    const int64_t n_tokens = k_cur->ne[2];
+    // we can merge dims 0 and 1
+    // TODO: add ggml helper function for this?
+    GGML_ASSERT(ggml_row_size(k_cur->type, n_embd_head) == k_cur->nb[1]);
 
-    k_cur = ggml_reshape_2d(ctx, k_cur, k->ne[0], n_tokens);
+    k_cur = ggml_view_2d(ctx, k_cur, n_embd_gqa, n_tokens, k_cur->nb[2], 0);
 
-    if (k->ne[2] > 1) {
-        k = ggml_reshape_2d(ctx, k, k->ne[0], k->ne[1]*k->ne[2]);
+    const int64_t n_stream = k->ne[2];
+
+    if (n_stream > 1) {
+        const int64_t kv_size = get_size();
+
+        assert(n_embd_gqa == k->ne[0]);
+        assert(kv_size    == k->ne[1]);
+
+        // merge the buffer across all streams because the idxs are global
+        k = ggml_reshape_2d(ctx, k, n_embd_gqa, kv_size*n_stream);
     }
 
+    // store the current K values into the cache
     return ggml_set_rows(ctx, k, k_cur, k_idxs);
 }
 
@@ -1038,28 +1055,51 @@ ggml_tensor * llama_kv_cache::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggm
 
     auto * v = layers[ikv].v;
 
-    const int64_t n_embd_v_gqa = v_cur->ne[0]*v_cur->ne[1];
-    const int64_t n_tokens     = v_cur->ne[2];
+    const int64_t n_embd_head = v_cur->ne[0];
+    const int64_t n_head      = v_cur->ne[1];
+    const int64_t n_tokens    = v_cur->ne[2];
+
+    const int64_t n_embd_gqa = n_embd_head*n_head;
 
-    v_cur = ggml_reshape_2d(ctx, v_cur, n_embd_v_gqa, n_tokens);
+    // we can merge dims 0 and 1
+    GGML_ASSERT(ggml_row_size(v_cur->type, n_embd_head) == v_cur->nb[1]);
 
+    const int64_t n_stream = v->ne[2];
+
+    // take this branch when FA is enabled (the V cache is not transposed)
     if (!v_trans) {
-        if (v->ne[2] > 1) {
-            v = ggml_reshape_2d(ctx, v, v->ne[0], v->ne[1]*v->ne[2]);
+        v_cur = ggml_view_2d(ctx, v_cur, n_embd_gqa, n_tokens, v_cur->nb[2], 0);
+
+        if (n_stream > 1) {
+            const int64_t kv_size = get_size();
+
+            assert(n_embd_gqa == v->ne[0]);
+            assert(kv_size    == v->ne[1]);
+
+            // merge the buffer across all streams because the idxs are global
+            v = ggml_reshape_2d(ctx, v, n_embd_gqa, kv_size*n_stream);
         }
 
         return ggml_set_rows(ctx, v, v_cur, v_idxs);
     }
 
+    if (ggml_row_size(v_cur->type, n_embd_gqa) == v_cur->nb[2]) {
+        // we can merge dims 0, 1 and 2
+        v_cur = ggml_reshape_2d(ctx, v_cur, n_embd_gqa, n_tokens);
+    } else {
+        // otherwise -> make a copy to get contiguous data
+        v_cur = ggml_cont_2d   (ctx, v_cur, n_embd_gqa, n_tokens);
+    }
+
     // [TAG_V_CACHE_VARIABLE]
-    if (n_embd_v_gqa < v->ne[0]) {
-        v_cur = ggml_pad(ctx, v_cur, v->ne[0] - n_embd_v_gqa, 0, 0, 0);
+    if (n_embd_gqa < v->ne[0]) {
+        v_cur = ggml_pad(ctx, v_cur, v->ne[0] - n_embd_gqa, 0, 0, 0);
     }
 
-    // the row becomes a single element
-    ggml_tensor * v_view = ggml_reshape_2d(ctx, v, 1, v->ne[0]*v->ne[1]*v->ne[2]);
+    // in this branch the v_idxs are constructed in such a way that each row is a single head element
+    ggml_tensor * v_view = ggml_reshape_2d(ctx, v, 1, ggml_nelements(v));
 
-    v_cur = ggml_reshape_2d(ctx, v_cur, 1, v_cur->ne[0]*v_cur->ne[1]);
+    v_cur = ggml_reshape_2d(ctx, v_cur, 1, ggml_nelements(v_cur));
 
     return ggml_set_rows(ctx, v_view, v_cur, v_idxs);
 }

package/src/llama.cpp/src/llama-kv-cache.h

@@ -317,9 +317,17 @@ public:
     ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
+    // note: the heads in k_cur and v_cur should be layed out contiguously in memory
+    //   - k_cur  [n_embd_head_k, n_head_k, n_tokens]
+    //   - k_idxs [n_tokens]
+    //   - v_cur  [n_embd_head_v, n_head_v, n_tokens]
+    //   - v_idxs [n_tokens] or [n_tokens*n_embd_v_gqa] depending if V cache is transposed
     ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const;
     ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const;
 
+    // create destination indices for each head of the current batch for where it would be written in the KV cache
+    // the indices address the global KV cache (not per stream) - this is not relevant for the user of this API, but
+    //   helps understand the implementation logic of cpy_k and cpy_v
     ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
     ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;