Merge branch 'ggerganov:master' into master

Makefile

@@ -507,7 +507,7 @@ ifdef LLAMA_CUDA
 		CUDA_PATH ?= /usr/local/cuda
 	endif
 	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
 	OBJS         += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
 	OBJS         += $(OBJS_CUDA_TEMP_INST)

README.md

@@ -209,6 +209,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
 - [AIKit](https://github.com/sozercan/aikit) (MIT)
+- [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
@@ -387,6 +388,30 @@ brew install llama.cpp
 ```
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
 
+### Nix
+
+On Mac and Linux, the Nix package manager can be used via
+```
+nix profile install nixpkgs#llama-cpp
+```
+For flake enabled installs.
+
+Or
+```
+nix-env --file '<nixpkgs>' --install --attr llama-cpp
+```
+For non-flake enabled installs.
+
+This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
+
+#### Flox
+
+On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
+```
+flox install llama-cpp
+```
+Flox follows the nixpkgs build of llama.cpp.
+
 ### Metal Build
 
 On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.

common/common.h

@@ -73,7 +73,6 @@ struct gpt_params {
     int32_t n_gpu_layers_draft    =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
     int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
     float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
-    int32_t n_beams               =     0; // if non-zero then use beam search of given width.
     int32_t grp_attn_n            =     1; // group-attention factor
     int32_t grp_attn_w            =   512; // group-attention width
     int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)

convert-hf-to-gguf.py

@@ -1632,6 +1632,12 @@ def set_gguf_parameters(self):
         super().set_gguf_parameters()
         if (n_experts := self.hparams.get("num_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+            logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
+        if (shared_expert_intermediate_size := self.hparams.get('shared_expert_intermediate_size')) is not None:
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size)
+            logger.info(f"gguf: expert shared feed forward length = {shared_expert_intermediate_size}")
 
     _experts: list[dict[str, Tensor]] | None = None
 

examples/infill/infill.cpp

@@ -223,7 +223,11 @@ int main(int argc, char ** argv) {
     inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
     embd_inp = inp_pfx;
     embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-    embd_inp.push_back(llama_token_middle(model));
+
+    const llama_token middle_token = llama_token_middle(model);
+    if (middle_token >= 0) {
+        embd_inp.push_back(middle_token);
+    }
 
     LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
     LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
@@ -528,7 +532,12 @@ int main(int argc, char ** argv) {
                 inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
                 embd_inp = inp_pfx;
                 embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-                embd_inp.push_back(llama_token_middle(model));
+
+                const llama_token middle_token = llama_token_middle(model);
+                if (middle_token >= 0) {
+                    embd_inp.push_back(middle_token);
+                }
+
                 embd.clear();
                 n_remain = params.n_predict;
                 n_past = 0;

examples/server/server.cpp

@@ -2038,7 +2038,12 @@ struct server_context {
                             prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
                             prefix_tokens.insert(prefix_tokens.end(),   llama_token_suffix(model));
                             prefix_tokens.insert(prefix_tokens.end(),   suffix_tokens.begin(), suffix_tokens.end());
-                            prefix_tokens.push_back(llama_token_middle(model));
+
+                            const llama_token middle_token = llama_token_middle(model);
+                            if (middle_token >= 0) {
+                                prefix_tokens.push_back(middle_token);
+                            }
+
                             prompt_tokens = prefix_tokens;
                         } else {
                             prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt

ggml-backend.c

@@ -1706,14 +1706,16 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     bool backend_ids_changed = false;
     for (int i = 0; i < sched->graph->n_nodes; i++) {
-        if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i]) {
+        if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i] &&
+            sched->bufts[sched->node_backend_ids[i]] != sched->bufts[sched->prev_node_backend_ids[i]]) {
             backend_ids_changed = true;
             break;
         }
     }
     if (!backend_ids_changed) {
         for (int i = 0; i < sched->graph->n_leafs; i++) {
-            if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i]) {
+            if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i] &&
+                sched->bufts[sched->leaf_backend_ids[i]] != sched->bufts[sched->prev_leaf_backend_ids[i]]) {
                 backend_ids_changed = true;
                 break;
             }
@@ -1977,6 +1979,15 @@ int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
     return sched->n_copies;
 }
 
+int ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched) {
+    return sched->n_backends;
+}
+
+ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i) {
+    GGML_ASSERT(i >= 0 && i < sched->n_backends);
+    return sched->backends[i];
+}
+
 size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
     int backend_index = ggml_backend_sched_backend_id(sched, backend);
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);

ggml-backend.h

@@ -182,6 +182,9 @@ extern "C" {
     // Initialize backend buffers from a measure graph
     GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
 
+    GGML_API int                  ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i);
+
     // Get the number of splits of the last graph
     GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
     GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);

ggml-impl.h

@@ -17,7 +17,7 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
-#if defined(_WIN32)
+#if defined(_MSC_VER)
 
 #define m512bh(p) p
 #define m512i(p) p

gguf-py/gguf/constants.py

@@ -33,21 +33,22 @@ class General:
         FILE_TYPE            = "general.file_type"
 
     class LLM:
-        VOCAB_SIZE                 = "{arch}.vocab_size"
-        CONTEXT_LENGTH             = "{arch}.context_length"
-        EMBEDDING_LENGTH           = "{arch}.embedding_length"
-        BLOCK_COUNT                = "{arch}.block_count"
-        LEADING_DENSE_BLOCK_COUNT  = "{arch}.leading_dense_block_count"
-        FEED_FORWARD_LENGTH        = "{arch}.feed_forward_length"
-        EXPERT_FEED_FORWARD_LENGTH = "{arch}.expert_feed_forward_length"
-        USE_PARALLEL_RESIDUAL      = "{arch}.use_parallel_residual"
-        TENSOR_DATA_LAYOUT         = "{arch}.tensor_data_layout"
-        EXPERT_COUNT               = "{arch}.expert_count"
-        EXPERT_USED_COUNT          = "{arch}.expert_used_count"
-        EXPERT_SHARED_COUNT        = "{arch}.expert_shared_count"
-        EXPERT_WEIGHTS_SCALE       = "{arch}.expert_weights_scale"
-        POOLING_TYPE               = "{arch}.pooling_type"
-        LOGIT_SCALE                = "{arch}.logit_scale"
+        VOCAB_SIZE                        = "{arch}.vocab_size"
+        CONTEXT_LENGTH                    = "{arch}.context_length"
+        EMBEDDING_LENGTH                  = "{arch}.embedding_length"
+        BLOCK_COUNT                       = "{arch}.block_count"
+        LEADING_DENSE_BLOCK_COUNT         = "{arch}.leading_dense_block_count"
+        FEED_FORWARD_LENGTH               = "{arch}.feed_forward_length"
+        EXPERT_FEED_FORWARD_LENGTH        = "{arch}.expert_feed_forward_length"
+        EXPERT_SHARED_FEED_FORWARD_LENGTH = "{arch}.expert_shared_feed_forward_length"
+        USE_PARALLEL_RESIDUAL             = "{arch}.use_parallel_residual"
+        TENSOR_DATA_LAYOUT                = "{arch}.tensor_data_layout"
+        EXPERT_COUNT                      = "{arch}.expert_count"
+        EXPERT_USED_COUNT                 = "{arch}.expert_used_count"
+        EXPERT_SHARED_COUNT               = "{arch}.expert_shared_count"
+        EXPERT_WEIGHTS_SCALE              = "{arch}.expert_weights_scale"
+        POOLING_TYPE                      = "{arch}.pooling_type"
+        LOGIT_SCALE                       = "{arch}.logit_scale"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"

gguf-py/gguf/gguf_writer.py

@@ -394,6 +394,9 @@ def add_feed_forward_length(self, length: int) -> None:
     def add_expert_feed_forward_length(self, length: int) -> None:
         self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
 
+    def add_expert_shared_feed_forward_length(self, length: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_SHARED_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+
     def add_parallel_residual(self, use: bool) -> None:
         self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
 

llama.cpp

@@ -286,6 +286,7 @@ enum llm_kv {
     LLM_KV_LEADING_DENSE_BLOCK_COUNT,
     LLM_KV_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
     LLM_KV_USE_PARALLEL_RESIDUAL,
     LLM_KV_TENSOR_DATA_LAYOUT,
     LLM_KV_EXPERT_COUNT,
@@ -364,21 +365,22 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_GENERAL_SOURCE_URL,            "general.source.url"                    },
     { LLM_KV_GENERAL_SOURCE_HF_REPO,        "general.source.huggingface.repository" },
 
-    { LLM_KV_VOCAB_SIZE,                    "%s.vocab_size"                 },
-    { LLM_KV_CONTEXT_LENGTH,                "%s.context_length"             },
-    { LLM_KV_EMBEDDING_LENGTH,              "%s.embedding_length"           },
-    { LLM_KV_BLOCK_COUNT,                   "%s.block_count"                },
-    { LLM_KV_LEADING_DENSE_BLOCK_COUNT,     "%s.leading_dense_block_count"  },
-    { LLM_KV_FEED_FORWARD_LENGTH,           "%s.feed_forward_length"        },
-    { LLM_KV_EXPERT_FEED_FORWARD_LENGTH,    "%s.expert_feed_forward_length" },
-    { LLM_KV_USE_PARALLEL_RESIDUAL,         "%s.use_parallel_residual"      },
-    { LLM_KV_TENSOR_DATA_LAYOUT,            "%s.tensor_data_layout"         },
-    { LLM_KV_EXPERT_COUNT,                  "%s.expert_count"               },
-    { LLM_KV_EXPERT_USED_COUNT,             "%s.expert_used_count"          },
-    { LLM_KV_EXPERT_SHARED_COUNT,           "%s.expert_shared_count"        },
-    { LLM_KV_EXPERT_WEIGHTS_SCALE,          "%s.expert_weights_scale"       },
-    { LLM_KV_POOLING_TYPE ,                 "%s.pooling_type"               },
-    { LLM_KV_LOGIT_SCALE,                   "%s.logit_scale"                },
+    { LLM_KV_VOCAB_SIZE,                        "%s.vocab_size"                        },
+    { LLM_KV_CONTEXT_LENGTH,                    "%s.context_length"                    },
+    { LLM_KV_EMBEDDING_LENGTH,                  "%s.embedding_length"                  },
+    { LLM_KV_BLOCK_COUNT,                       "%s.block_count"                       },
+    { LLM_KV_LEADING_DENSE_BLOCK_COUNT,         "%s.leading_dense_block_count"         },
+    { LLM_KV_FEED_FORWARD_LENGTH,               "%s.feed_forward_length"               },
+    { LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        "%s.expert_feed_forward_length"        },
+    { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
+    { LLM_KV_USE_PARALLEL_RESIDUAL,             "%s.use_parallel_residual"             },
+    { LLM_KV_TENSOR_DATA_LAYOUT,                "%s.tensor_data_layout"                },
+    { LLM_KV_EXPERT_COUNT,                      "%s.expert_count"                      },
+    { LLM_KV_EXPERT_USED_COUNT,                 "%s.expert_used_count"                 },
+    { LLM_KV_EXPERT_SHARED_COUNT,               "%s.expert_shared_count"               },
+    { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
+    { LLM_KV_POOLING_TYPE ,                     "%s.pooling_type"                      },
+    { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
 
     { LLM_KV_ATTENTION_HEAD_COUNT,          "%s.attention.head_count"             },
     { LLM_KV_ATTENTION_HEAD_COUNT_KV,       "%s.attention.head_count_kv"          },
@@ -1970,6 +1972,7 @@ struct llama_hparams {
     uint32_t n_lora_q = 0;
     uint32_t n_lora_kv = 0;
     uint32_t n_ff_exp = 0;
+    uint32_t n_ff_shexp = 0;
     uint32_t n_expert_shared = 0;
     float    expert_weights_scale = 0.0;
 
@@ -2018,6 +2021,7 @@ struct llama_hparams {
         if (this->n_lora_q           != other.n_lora_q)           return true;
         if (this->n_lora_kv          != other.n_lora_kv)          return true;
         if (this->n_ff_exp           != other.n_ff_exp)           return true;
+        if (this->n_ff_shexp         != other.n_ff_shexp)         return true;
         if (this->n_expert_shared    != other.n_expert_shared)    return true;
 
         if (this->rope_finetuned  != other.rope_finetuned)  return true;
@@ -4455,6 +4459,9 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_QWEN2MOE:
             {
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
+
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 24: model.type = e_model::MODEL_A2_7B; break;
@@ -5240,6 +5247,11 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
         LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
         LLAMA_LOG_INFO("%s: rope_yarn_log_mul    = %.4f\n",   __func__, hparams.rope_yarn_log_mul);
     }
+
+    if (model.arch == LLM_ARCH_QWEN2MOE) {
+        LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
+    }
 }
 
 // Returns false if cancelled by progress_callback
@@ -6026,16 +6038,17 @@ static bool llm_load_tensors(
                         GGML_ASSERT(hparams.n_expert_used > 0);
 
                         // MoE branch
-                        auto n_ff_exp = n_ff / hparams.n_expert_used;
+                        auto n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / hparams.n_expert_used;
                         layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert});
                         layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert});
                         layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert});
 
                         // Shared expert branch
+                        auto n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
                         layer.ffn_gate_inp_shexp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), {n_embd});
-                        layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd,   n_ff});
-                        layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {  n_ff, n_embd});
-                        layer.ffn_up_shexp   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd,   n_ff});
+                        layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp});
+                        layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd});
+                        layer.ffn_up_shexp   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp});
                     }
                 } break;
             case LLM_ARCH_PHI2:
@@ -16260,6 +16273,11 @@ struct llama_context * llama_new_context_with_model(
         params.flash_attn = false;
     }
 
+    if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) {
+        LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
     if (params.type_v != GGML_TYPE_F16 && !params.flash_attn) {
         LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
         return nullptr;

scripts/sync-ggml.last

@@ -1 +1 @@
-2aae01fd9b8f9399f343cf18f46f38996ef52e2c
+5653a195935ea3ac54652644c9daf154dbc1571b

sgemm.cpp

@@ -43,8 +43,10 @@
 // [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
 //     Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
 
+#if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Wpedantic"
 #pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
 
 #include "sgemm.h"
 #include "ggml-impl.h"