sync : ggml

examples/talk-llama/llama-vocab.h

@@ -8,6 +8,8 @@
 #include <map>
 #include <set>
 
+struct llm_tokenizer;
+
 struct llama_vocab {
     using id    = llama_token;
     using token = std::string;
@@ -65,7 +67,14 @@ struct llama_vocab {
 
     std::vector<char> precompiled_charsmap;
 
+    llm_tokenizer * tokenizer = nullptr;
+
+    llama_vocab() = default;
+    ~llama_vocab();
+
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const;
+
+    void init_tokenizer();
 };
 
 //

examples/talk-llama/llama.h

@@ -102,6 +102,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
         LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
         LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
+        LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
     };
 
     enum llama_rope_type {
@@ -192,6 +193,7 @@ extern "C" {
         LLAMA_POOLING_TYPE_MEAN = 1,
         LLAMA_POOLING_TYPE_CLS  = 2,
         LLAMA_POOLING_TYPE_LAST = 3,
+        LLAMA_POOLING_TYPE_RANK = 4, // used by reranking models to attach the classification head to the graph
     };
 
     enum llama_attention_type {
@@ -201,9 +203,9 @@ extern "C" {
     };
 
     enum llama_split_mode {
-        LLAMA_SPLIT_MODE_NONE    = 0, // single GPU
-        LLAMA_SPLIT_MODE_LAYER   = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_MODE_ROW     = 2, // split rows across GPUs
+        LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
+        LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
+        LLAMA_SPLIT_MODE_ROW   = 2, // split rows across GPUs
     };
 
     // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
@@ -871,7 +873,8 @@ extern "C" {
 
     // Get the embeddings for a sequence id
     // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
-    // shape: [n_embd] (1-dimensional)
+    // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[1] with the rank of the sequence
+    // otherwise: float[n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
 
     //
@@ -910,6 +913,8 @@ extern "C" {
     //
     // Tokenization
     //
+    // The API is thread-safe.
+    //
 
     /// @details Convert the provided text into tokens.
     /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.

ggml/include/ggml-metal.h

@@ -25,9 +25,6 @@
 #include <stddef.h>
 #include <stdbool.h>
 
-// max memory buffers that can be mapped to the device
-#define GGML_METAL_MAX_BUFFERS 64
-
 struct ggml_tensor;
 struct ggml_cgraph;
 
@@ -48,8 +45,6 @@ GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
 GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
 
-GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-
 GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);

ggml/include/ggml.h

@@ -229,14 +229,16 @@
 #define GGML_MAX_PARAMS         2048
 #define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
-#ifndef GGML_MAX_NAME
-#define GGML_MAX_NAME           64
 #define GGML_MAX_N_THREADS      512
+#define GGML_MAX_OP_PARAMS      64
 
+#ifndef GGML_MAX_NAME
+#   define GGML_MAX_NAME        64
 #endif
-#define GGML_MAX_OP_PARAMS      64
+
 #define GGML_DEFAULT_N_THREADS  4
 #define GGML_DEFAULT_GRAPH_SIZE 2048
+
 #if UINTPTR_MAX == 0xFFFFFFFF
     #define GGML_MEM_ALIGN 4
 #else
@@ -259,21 +261,21 @@
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
 #ifndef NDEBUG
-#define GGML_UNREACHABLE() do { fprintf(stderr, "statement should be unreachable\n"); abort(); } while(0)
+#   define GGML_UNREACHABLE() do { fprintf(stderr, "statement should be unreachable\n"); abort(); } while(0)
 #elif defined(__GNUC__)
-#define GGML_UNREACHABLE() __builtin_unreachable()
+#   define GGML_UNREACHABLE() __builtin_unreachable()
 #elif defined(_MSC_VER)
-#define GGML_UNREACHABLE() __assume(0)
+#   define GGML_UNREACHABLE() __assume(0)
 #else
-#define GGML_UNREACHABLE() ((void) 0)
+#   define GGML_UNREACHABLE() ((void) 0)
 #endif
 
 #ifdef __cplusplus
-#define GGML_NORETURN [[noreturn]]
+#   define GGML_NORETURN [[noreturn]]
 #elif defined(_MSC_VER)
-#define GGML_NORETURN __declspec(noreturn)
+#   define GGML_NORETURN __declspec(noreturn)
 #else
-#define GGML_NORETURN _Noreturn
+#   define GGML_NORETURN _Noreturn
 #endif
 
 #define GGML_ABORT(...) ggml_abort(__FILE__, __LINE__, __VA_ARGS__)
@@ -575,10 +577,10 @@ extern "C" {
 
     // this tensor...
     enum ggml_tensor_flag {
-        GGML_TENSOR_FLAG_INPUT    = 1, // ...is an input for the GGML compute graph
-        GGML_TENSOR_FLAG_OUTPUT   = 2, // ...is an output for the GGML compute graph
-        GGML_TENSOR_FLAG_PARAM    = 4, // ...contains trainable parameters
-        GGML_TENSOR_FLAG_LOSS     = 8, // ...defines loss for numerical optimization (multiple loss tensors add up)
+        GGML_TENSOR_FLAG_INPUT  =  1, // ...is an input for the GGML compute graph
+        GGML_TENSOR_FLAG_OUTPUT =  2, // ...is an output for the GGML compute graph
+        GGML_TENSOR_FLAG_PARAM  =  4, // ...contains trainable parameters
+        GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
     };
 
     // n-dimensional tensor
@@ -1408,14 +1410,14 @@ extern "C" {
     // supports 3D: a->ne[2] == b->ne[1]
     GGML_API struct ggml_tensor * ggml_get_rows(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * a,  // data
+            struct ggml_tensor  * b); // row indices
 
     GGML_API struct ggml_tensor * ggml_get_rows_back(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            struct ggml_tensor  * c);
+            struct ggml_tensor  * a,  // gradients of ggml_get_rows result
+            struct ggml_tensor  * b,  // row indices
+            struct ggml_tensor  * c); // data for ggml_get_rows, only used for its shape
 
     GGML_API struct ggml_tensor * ggml_diag(
         struct ggml_context     * ctx,
@@ -1566,9 +1568,9 @@ extern "C" {
     // a - dy
     GGML_API struct ggml_tensor * ggml_rope_back(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            struct ggml_tensor  * c,
+            struct ggml_tensor  * a, // gradients of ggml_rope result
+            struct ggml_tensor  * b, // positions
+            struct ggml_tensor  * c, // freq factors
             int                   n_dims,
             int                   mode,
             int                   n_ctx_orig,
@@ -2034,22 +2036,23 @@ extern "C" {
     // loss function
 
     GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
-            struct ggml_context         * ctx,
-            struct ggml_tensor          * a,
-            struct ggml_tensor          * b);
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // logits
+            struct ggml_tensor  * b); // labels
 
     GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
-            struct ggml_context         * ctx,
-            struct ggml_tensor          * a,
-            struct ggml_tensor          * b,
-            struct ggml_tensor          * c);
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // logits
+            struct ggml_tensor  * b,  // labels
+            struct ggml_tensor  * c); // gradients of cross_entropy_loss result
 
     // AdamW optimizer step
     // Paper: https://arxiv.org/pdf/1711.05101v3.pdf
     // PyTorch: https://pytorch.org/docs/stable/generated/torch.optim.AdamW.html
     GGML_API struct ggml_tensor * ggml_opt_step_adamw(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
+            struct ggml_tensor  * grad,
             float                 alpha,
             float                 beta1,
             float                 beta2,
@@ -2064,7 +2067,7 @@ extern "C" {
     GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
 
     GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
-    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate, bool keep);
+    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
 
     GGML_API void ggml_build_opt_adamw(
             struct ggml_context * ctx,
@@ -2507,6 +2510,9 @@ extern "C" {
     GGML_API int ggml_cpu_has_cann       (void);
     GGML_API int ggml_cpu_has_llamafile  (void);
 
+    // get the sve vector length in bytes
+    GGML_API int ggml_cpu_get_sve_cnt(void);
+
     //
     // Internal types and functions exposed for tests and benchmarks
     //