talk-llama : sync llama.cpp

diff --git a/examples/talk-llama/llama.cpp b/examples/talk-llama/llama.cpp
index 3bb056dba2e6d6ab21fcd21a3ab70776ad01557d..d39ff94c7fae696ccba29c25cc5e16b09588fd77 100644 (file)
--- a/examples/talk-llama/llama.cpp
+++ b/examples/talk-llama/llama.cpp
@@ -1903,6 +1903,28 @@ static void llama_kv_cache_seq_shift(
     cache.head = new_head != cache.size ? new_head : 0;
 }
 
+static void llama_kv_cache_seq_div(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1,
+                          int   d) {
+    if (p0 < 0) p0 = 0;
+    if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            cache.has_shift = true;
+
+            {
+                llama_pos p_old = cache.cells[i].pos;
+                cache.cells[i].pos   /= d;
+                cache.cells[i].delta += cache.cells[i].pos - p_old;
+            }
+        }
+    }
+}
+
 //
 // model loading and saving
 //
@@ -2180,7 +2202,11 @@ struct llama_model_loader {
                     type_max   = type;
                 }
 
-                // LLAMA_LOG_INFO("%s: - tensor %4d: %32s %-8s [ %s ]\n", __func__, i, name, ggml_type_name(meta->type), llama_format_tensor_shape(meta).c_str());
+                // TODO: make runtime configurable
+#if 0
+                struct ggml_tensor * meta = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i));
+                LLAMA_LOG_INFO("%s: - tensor %4d: %32s %-8s [ %s ]\n", __func__, i, ggml_get_name(meta), ggml_type_name(type), llama_format_tensor_shape(meta).c_str());
+#endif
             }
 
             switch (type_max) {
@@ -2196,6 +2222,8 @@ struct llama_model_loader {
                 case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break;
                 case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break;
                 case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K;   break;
+                case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break;
+                case GGML_TYPE_IQ2_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS;  break;
                 default:
                     {
                         LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -2558,7 +2586,8 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0";
 
         // K-quants
-        case LLAMA_FTYPE_MOSTLY_Q2_K:   return "Q2_K";
+        case LLAMA_FTYPE_MOSTLY_Q2_K:   return "Q2_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q2_K_S: return "Q2_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large";
@@ -2567,6 +2596,8 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "Q5_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q6_K:   return "Q6_K";
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XSS - 2.0625 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw";
 
         default: return "unknown, may not work";
     }
@@ -2801,6 +2832,7 @@ static void llm_load_hparams(
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
                 switch (hparams.n_layer) {
+                    case 24: model.type = e_model::MODEL_1B; break;
                     case 32: model.type = e_model::MODEL_3B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
@@ -3117,7 +3149,15 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
     LLAMA_LOG_INFO("%s: model type       = %s\n",     __func__, llama_model_type_name(model.type));
     LLAMA_LOG_INFO("%s: model ftype      = %s\n",     __func__, llama_model_ftype_name(model.ftype).c_str());
-    LLAMA_LOG_INFO("%s: model params     = %.2f B\n", __func__, ml.n_elements*1e-9);
+    if (ml.n_elements >= 1e12) {
+        LLAMA_LOG_INFO("%s: model params     = %.2f T\n", __func__, ml.n_elements*1e-12);
+    } else if (ml.n_elements >= 1e9) {
+        LLAMA_LOG_INFO("%s: model params     = %.2f B\n", __func__, ml.n_elements*1e-9);
+    } else if (ml.n_elements >= 1e6) {
+        LLAMA_LOG_INFO("%s: model params     = %.2f M\n", __func__, ml.n_elements*1e-6);
+    } else {
+        LLAMA_LOG_INFO("%s: model params     = %.2f K\n", __func__, ml.n_elements*1e-3);
+    }
     if (ml.n_bytes < GiB) {
         LLAMA_LOG_INFO("%s: model size       = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0,        ml.n_bytes*8.0/ml.n_elements);
     } else {
@@ -4772,7 +4812,6 @@ struct llm_build_context {
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_gqa  == n_embd);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -4896,7 +4935,6 @@ struct llm_build_context {
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_gqa  == n_embd);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * pos;
@@ -4995,9 +5033,7 @@ struct llm_build_context {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_gqa  == n_embd);
 
         const int64_t n_rot = n_embd_head_k / 2;
 
@@ -5209,9 +5245,7 @@ struct llm_build_context {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_gqa  == n_embd);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -5304,7 +5338,6 @@ struct llm_build_context {
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_gqa  == n_embd);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -5400,7 +5433,6 @@ struct llm_build_context {
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_gqa  == n_embd);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -5727,7 +5759,6 @@ struct llm_build_context {
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_gqa  == n_embd);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * attn_norm_output;
@@ -5951,7 +5982,6 @@ struct llm_build_context {
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_gqa  == n_embd);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * pos;
@@ -8926,10 +8956,13 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             // TODO: explore better strategies
             new_type = GGML_TYPE_Q8_0;
         }
-    } else if (name.find("ffn_down.weight") != std::string::npos) {
+    } else if (name.find("ffn_down") != std::string::npos) {
         if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
+            if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q4_K;
+        }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
-            new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K
+            new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q5_K
                      : arch != LLM_ARCH_FALCON || use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q4_K
                      : GGML_TYPE_Q3_K;
         }
@@ -8938,14 +8971,14 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
             if (arch == LLM_ARCH_FALCON) {
-                new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K :
+                new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q6_K :
                            use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
             } else {
                 if (use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
             }
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
-        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < 4) {
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) {
             new_type = GGML_TYPE_Q5_K;
         }
         ++qs.i_feed_forward_w2;
@@ -8963,9 +8996,10 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
     }
-    else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) {
-        if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-    }
+    // IK: let's remove this, else Q2_K is almost the same as Q3_K_S
+    //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+    //}
     // This can be used to reduce the size of the Q5_K_S model.
     // The associated PPL increase is fully in line with the size reduction
     //else {
@@ -9014,6 +9048,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
         // K-quants
         case LLAMA_FTYPE_MOSTLY_Q2_K:   quantized_type = GGML_TYPE_Q2_K; break;
+        case LLAMA_FTYPE_MOSTLY_Q2_K_S: quantized_type = GGML_TYPE_Q2_K; break;
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:
         case LLAMA_FTYPE_MOSTLY_Q3_K_M:
         case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break;
@@ -9022,6 +9057,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_K_S:
         case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break;
         case LLAMA_FTYPE_MOSTLY_Q6_K:   quantized_type = GGML_TYPE_Q6_K; break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS:quantized_type = GGML_TYPE_IQ2_XXS; break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_XS :quantized_type = GGML_TYPE_IQ2_XS;  break;
 
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
@@ -9070,7 +9107,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) {
             ++qs.n_attention_wv;
         }
-        else if (name.find("ffn_down.weight") != std::string::npos) {
+        else if (name.find("ffn_down") != std::string::npos) {
             ++qs.n_feed_forward_w2;
         }
     }
@@ -10146,9 +10183,21 @@ void llama_kv_cache_seq_keep(struct llama_context * ctx, llama_seq_id seq_id) {
 }
 
 void llama_kv_cache_seq_shift(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
+    if (delta == 0) {
+        return;
+    }
+
     llama_kv_cache_seq_shift(ctx->kv_self, seq_id, p0, p1, delta);
 }
 
+void llama_kv_cache_seq_div(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    if (d == 1) {
+        return;
+    }
+
+    llama_kv_cache_seq_div(ctx->kv_self, seq_id, p0, p1, d);
+}
+
 // Returns the *maximum* size of the state
 size_t llama_get_state_size(const struct llama_context * ctx) {
     // we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
@@ -10881,7 +10930,7 @@ void llama_print_timings(struct llama_context * ctx) {
             __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval);
     LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval);
-    LLAMA_LOG_INFO("%s:       total time = %10.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms));
+    LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (timings.t_end_ms - timings.t_start_ms), (timings.n_p_eval + timings.n_eval));
 }
 
 void llama_reset_timings(struct llama_context * ctx) {

diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h
index 461d4604a1b5455c160078f2517ec30549d8d078..43d41b8f642b57f42e9fcb63b2d962654ea374be 100644 (file)
--- a/examples/talk-llama/llama.h
+++ b/examples/talk-llama/llama.h
@@ -103,6 +103,9 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XXS       = 19, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -484,6 +487,17 @@ extern "C" {
                        llama_pos   p1,
                        llama_pos   delta);
 
+    // Integer division of the positions by factor of `d > 1`
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_div(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                             int   d);
+
     //
     // State / sessions
     //