npm - @fugood/llama.node - Versions diffs - 0.3.2 → 0.3.3 - Mend

@fugood/llama.node 0.3.2 → 0.3.3

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (190) hide show

package/CMakeLists.txt +2 -0
package/bin/darwin/arm64/llama-node.node +0 -0
package/bin/darwin/x64/llama-node.node +0 -0
package/bin/linux/arm64/llama-node.node +0 -0
package/bin/linux/x64/llama-node.node +0 -0
package/bin/linux-vulkan/arm64/llama-node.node +0 -0
package/bin/linux-vulkan/x64/llama-node.node +0 -0
package/bin/win32/arm64/llama-node.node +0 -0
package/bin/win32/arm64/node.lib +0 -0
package/bin/win32/x64/llama-node.node +0 -0
package/bin/win32/x64/node.lib +0 -0
package/bin/win32-vulkan/arm64/llama-node.node +0 -0
package/bin/win32-vulkan/arm64/node.lib +0 -0
package/bin/win32-vulkan/x64/llama-node.node +0 -0
package/bin/win32-vulkan/x64/node.lib +0 -0
package/package.json +1 -1
package/src/DetokenizeWorker.cpp +1 -1
package/src/EmbeddingWorker.cpp +2 -2
package/src/LlamaCompletionWorker.cpp +8 -8
package/src/LlamaCompletionWorker.h +2 -2
package/src/LlamaContext.cpp +8 -9
package/src/TokenizeWorker.cpp +1 -1
package/src/common.hpp +4 -4
package/src/llama.cpp/.github/workflows/build.yml +43 -9
package/src/llama.cpp/.github/workflows/docker.yml +3 -0
package/src/llama.cpp/CMakeLists.txt +7 -4
package/src/llama.cpp/cmake/arm64-apple-clang.cmake +16 -0
package/src/llama.cpp/common/CMakeLists.txt +0 -2
package/src/llama.cpp/common/arg.cpp +642 -607
package/src/llama.cpp/common/arg.h +22 -22
package/src/llama.cpp/common/common.cpp +79 -281
package/src/llama.cpp/common/common.h +130 -100
package/src/llama.cpp/common/json-schema-to-grammar.cpp +1 -1
package/src/llama.cpp/common/log.cpp +50 -50
package/src/llama.cpp/common/log.h +18 -18
package/src/llama.cpp/common/ngram-cache.cpp +36 -36
package/src/llama.cpp/common/ngram-cache.h +19 -19
package/src/llama.cpp/common/sampling.cpp +116 -108
package/src/llama.cpp/common/sampling.h +20 -20
package/src/llama.cpp/docs/build.md +37 -17
package/src/llama.cpp/examples/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/batched/batched.cpp +14 -14
package/src/llama.cpp/examples/batched-bench/batched-bench.cpp +10 -11
package/src/llama.cpp/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +1 -1
package/src/llama.cpp/examples/cvector-generator/cvector-generator.cpp +9 -9
package/src/llama.cpp/examples/embedding/embedding.cpp +12 -12
package/src/llama.cpp/examples/eval-callback/eval-callback.cpp +8 -8
package/src/llama.cpp/examples/export-lora/export-lora.cpp +5 -5
package/src/llama.cpp/examples/gen-docs/gen-docs.cpp +7 -7
package/src/llama.cpp/examples/gritlm/gritlm.cpp +18 -18
package/src/llama.cpp/examples/imatrix/imatrix.cpp +20 -11
package/src/llama.cpp/examples/infill/infill.cpp +40 -86
package/src/llama.cpp/examples/llama-bench/llama-bench.cpp +42 -151
package/src/llama.cpp/examples/llama.android/llama/build.gradle.kts +1 -0
package/src/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp +11 -14
package/src/llama.cpp/examples/llava/clip.cpp +1 -0
package/src/llama.cpp/examples/llava/llava-cli.cpp +23 -23
package/src/llama.cpp/examples/llava/llava.cpp +37 -3
package/src/llama.cpp/examples/llava/minicpmv-cli.cpp +21 -21
package/src/llama.cpp/examples/lookahead/lookahead.cpp +26 -26
package/src/llama.cpp/examples/lookup/lookup-create.cpp +7 -7
package/src/llama.cpp/examples/lookup/lookup-merge.cpp +4 -4
package/src/llama.cpp/examples/lookup/lookup-stats.cpp +14 -14
package/src/llama.cpp/examples/lookup/lookup.cpp +29 -29
package/src/llama.cpp/examples/main/main.cpp +64 -109
package/src/llama.cpp/examples/parallel/parallel.cpp +18 -19
package/src/llama.cpp/examples/passkey/passkey.cpp +14 -14
package/src/llama.cpp/examples/perplexity/perplexity.cpp +99 -120
package/src/llama.cpp/examples/quantize-stats/quantize-stats.cpp +10 -9
package/src/llama.cpp/examples/retrieval/retrieval.cpp +13 -13
package/src/llama.cpp/examples/rpc/rpc-server.cpp +3 -1
package/src/llama.cpp/examples/save-load-state/save-load-state.cpp +34 -17
package/src/llama.cpp/examples/server/CMakeLists.txt +4 -13
package/src/llama.cpp/examples/server/server.cpp +553 -691
package/src/llama.cpp/examples/server/utils.hpp +312 -25
package/src/llama.cpp/examples/simple/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/simple/simple.cpp +128 -96
package/src/llama.cpp/examples/simple-chat/CMakeLists.txt +5 -0
package/src/llama.cpp/examples/simple-chat/simple-chat.cpp +197 -0
package/src/llama.cpp/examples/speculative/speculative.cpp +54 -51
package/src/llama.cpp/examples/tokenize/tokenize.cpp +2 -2
package/src/llama.cpp/ggml/CMakeLists.txt +15 -9
package/src/llama.cpp/ggml/include/ggml-amx.h +25 -0
package/src/llama.cpp/ggml/include/ggml-backend.h +46 -33
package/src/llama.cpp/ggml/include/ggml-blas.h +5 -3
package/src/llama.cpp/ggml/include/ggml-cann.h +9 -7
package/src/llama.cpp/ggml/include/ggml-cpp.h +38 -0
package/src/llama.cpp/ggml/include/ggml-cpu.h +177 -0
package/src/llama.cpp/ggml/include/ggml-cuda.h +12 -12
package/src/llama.cpp/ggml/include/ggml-kompute.h +7 -3
package/src/llama.cpp/ggml/include/ggml-metal.h +11 -7
package/src/llama.cpp/ggml/include/ggml-opt.h +216 -0
package/src/llama.cpp/ggml/include/ggml-rpc.h +9 -5
package/src/llama.cpp/ggml/include/ggml-sycl.h +18 -11
package/src/llama.cpp/ggml/include/ggml-vulkan.h +10 -8
package/src/llama.cpp/ggml/include/ggml.h +53 -393
package/src/llama.cpp/ggml/src/CMakeLists.txt +66 -1149
package/src/llama.cpp/ggml/src/ggml-aarch64.c +46 -3126
package/src/llama.cpp/ggml/src/ggml-aarch64.h +0 -20
package/src/llama.cpp/ggml/src/ggml-alloc.c +23 -27
package/src/llama.cpp/ggml/src/ggml-amx/CMakeLists.txt +107 -0
package/src/llama.cpp/ggml/src/ggml-amx/common.h +94 -0
package/src/llama.cpp/ggml/src/ggml-amx/ggml-amx.cpp +446 -0
package/src/llama.cpp/ggml/src/ggml-amx/mmq.cpp +2510 -0
package/src/llama.cpp/ggml/src/ggml-amx/mmq.h +17 -0
package/src/llama.cpp/ggml/src/ggml-backend-impl.h +6 -25
package/src/llama.cpp/ggml/src/ggml-backend-reg.cpp +195 -0
package/src/llama.cpp/ggml/src/ggml-backend.cpp +303 -864
package/src/llama.cpp/ggml/src/ggml-blas/CMakeLists.txt +91 -0
package/src/llama.cpp/ggml/src/{ggml-blas.cpp → ggml-blas/ggml-blas.cpp} +213 -65
package/src/llama.cpp/ggml/src/ggml-cann/CMakeLists.txt +46 -0
package/src/llama.cpp/ggml/src/{ggml-cann.cpp → ggml-cann/ggml-cann.cpp} +255 -149
package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +261 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.c +3560 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.h +30 -0
package/src/llama.cpp/ggml/src/{ggml-cpu-impl.h → ggml-cpu/ggml-cpu-impl.h} +0 -243
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c +10822 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.h +63 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +13970 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +663 -0
package/src/llama.cpp/ggml/src/{llamafile → ggml-cpu/llamafile}/sgemm.cpp +667 -1
package/src/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +155 -0
package/src/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +106 -0
package/src/llama.cpp/ggml/src/ggml-impl.h +366 -16
package/src/llama.cpp/ggml/src/ggml-kompute/CMakeLists.txt +162 -0
package/src/llama.cpp/ggml/src/{ggml-kompute.cpp → ggml-kompute/ggml-kompute.cpp} +238 -72
package/src/llama.cpp/ggml/src/ggml-metal/CMakeLists.txt +108 -0
package/src/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +249 -0
package/src/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt +100 -0
package/src/llama.cpp/ggml/src/ggml-opt.cpp +867 -0
package/src/llama.cpp/ggml/src/ggml-quants.c +187 -10692
package/src/llama.cpp/ggml/src/ggml-quants.h +78 -125
package/src/llama.cpp/ggml/src/ggml-rpc/CMakeLists.txt +11 -0
package/src/llama.cpp/ggml/src/{ggml-rpc.cpp → ggml-rpc/ggml-rpc.cpp} +475 -300
package/src/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt +81 -0
package/src/llama.cpp/ggml/src/ggml-sycl/backend.hpp +3 -0
package/src/llama.cpp/ggml/src/ggml-sycl/common.cpp +40 -0
package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp +258 -0
package/src/llama.cpp/ggml/src/ggml-sycl/concat.cpp +1 -0
package/src/llama.cpp/ggml/src/ggml-sycl/dpct/helper.hpp +2 -22
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.cpp +1011 -0
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.hpp +76 -0
package/src/llama.cpp/ggml/src/{ggml-sycl.cpp → ggml-sycl/ggml-sycl.cpp} +3584 -4142
package/src/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp +69 -67
package/src/llama.cpp/ggml/src/ggml-sycl/norm.cpp +3 -3
package/src/llama.cpp/ggml/src/ggml-sycl/outprod.cpp +56 -0
package/src/llama.cpp/ggml/src/ggml-sycl/outprod.hpp +11 -0
package/src/llama.cpp/ggml/src/ggml-sycl/presets.hpp +6 -0
package/src/llama.cpp/ggml/src/ggml-sycl/vecdotq.hpp +4 -4
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp +138 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.hpp +10 -0
package/src/llama.cpp/ggml/src/ggml-threading.cpp +12 -0
package/src/llama.cpp/ggml/src/ggml-threading.h +12 -0
package/src/llama.cpp/ggml/src/ggml-vulkan/CMakeLists.txt +78 -0
package/src/llama.cpp/ggml/src/{ggml-vulkan.cpp → ggml-vulkan/ggml-vulkan.cpp} +555 -623
package/src/llama.cpp/ggml/src/{vulkan-shaders → ggml-vulkan/vulkan-shaders}/vulkan-shaders-gen.cpp +125 -206
package/src/llama.cpp/ggml/src/ggml.c +4032 -19890
package/src/llama.cpp/include/llama.h +67 -33
package/src/llama.cpp/pocs/vdot/q8dot.cpp +4 -3
package/src/llama.cpp/pocs/vdot/vdot.cpp +8 -7
package/src/llama.cpp/src/CMakeLists.txt +2 -1
package/src/llama.cpp/src/llama-sampling.cpp +745 -105
package/src/llama.cpp/src/llama-sampling.h +21 -2
package/src/llama.cpp/src/llama-vocab.cpp +49 -9
package/src/llama.cpp/src/llama-vocab.h +35 -11
package/src/llama.cpp/src/llama.cpp +2636 -2406
package/src/llama.cpp/src/unicode-data.cpp +2 -2
package/src/llama.cpp/tests/CMakeLists.txt +1 -2
package/src/llama.cpp/tests/test-arg-parser.cpp +14 -14
package/src/llama.cpp/tests/test-backend-ops.cpp +185 -60
package/src/llama.cpp/tests/test-barrier.cpp +1 -0
package/src/llama.cpp/tests/test-chat-template.cpp +9 -5
package/src/llama.cpp/tests/test-json-schema-to-grammar.cpp +17 -4
package/src/llama.cpp/tests/test-log.cpp +2 -2
package/src/llama.cpp/tests/test-opt.cpp +853 -142
package/src/llama.cpp/tests/test-quantize-fns.cpp +22 -19
package/src/llama.cpp/tests/test-quantize-perf.cpp +16 -14
package/src/llama.cpp/tests/test-rope.cpp +1 -0
package/src/llama.cpp/tests/test-sampling.cpp +162 -137
package/src/llama.cpp/tests/test-tokenizer-0.cpp +7 -7
package/src/llama.cpp/tests/test-tokenizer-1-bpe.cpp +5 -5
package/src/llama.cpp/tests/test-tokenizer-1-spm.cpp +5 -5
package/src/llama.cpp/common/train.cpp +0 -1515
package/src/llama.cpp/common/train.h +0 -233
package/src/llama.cpp/examples/baby-llama/CMakeLists.txt +0 -5
package/src/llama.cpp/examples/baby-llama/baby-llama.cpp +0 -1639
package/src/llama.cpp/tests/test-grad0.cpp +0 -1683
/package/src/llama.cpp/ggml/{cmake → src/ggml-cpu/cmake}/FindSIMD.cmake +0 -0
/package/src/llama.cpp/ggml/src/{llamafile → ggml-cpu/llamafile}/sgemm.h +0 -0
/package/src/llama.cpp/ggml/src/{vulkan-shaders → ggml-vulkan/vulkan-shaders}/CMakeLists.txt +0 -0

package/src/llama.cpp/include/llama.h CHANGED Viewed

@@ -2,6 +2,7 @@
 #define LLAMA_H
 #include "ggml.h"
+#include "ggml-cpu.h"
 #include "ggml-backend.h"
 #include <stddef.h>
@@ -205,7 +206,7 @@ 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_ROW   = 2, // split layers and KV across GPUs, use tensor parallelism if supported
     };
     // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
@@ -217,6 +218,7 @@ extern "C" {
     typedef struct llama_token_data_array {
         // TODO: consider SoA
+        // NOTE: this pointer can be modified by the samplers
         llama_token_data * data;
         size_t size;
         int64_t selected; // this is the index in the data array (i.e. not the token id)
@@ -232,8 +234,11 @@ extern "C" {
     // - token  : the token ids of the input (used when embd is NULL)
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
+    //            (if set to NULL, the token position will be tracked automatically by llama_decode)
     // - seq_id : the sequence to which the respective token belongs
+    //            (if set to NULL, the sequence ID will be assumed to be 0)
     // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
+    //            (if set to NULL, only the logits for last token will be returned)
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -244,15 +249,6 @@ extern "C" {
         int32_t      *  n_seq_id;
         llama_seq_id ** seq_id;
         int8_t       *  logits; // TODO: rename this to "output"
-        // NOTE: helpers for smooth API transition - can be deprecated in the future
-        //       for future-proof code, use the above fields instead and ignore everything below
-        //
-        // pos[i] = all_pos_0 + i*all_pos_1
-        //
-        llama_pos    all_pos_0;  // used if pos == NULL
-        llama_pos    all_pos_1;  // used if pos == NULL
-        llama_seq_id all_seq_id; // used if seq_id == NULL
     } llama_batch;
     enum llama_model_kv_override_type {
@@ -279,10 +275,7 @@ extern "C" {
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
-        // main_gpu interpretation depends on split_mode:
-        // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model
-        // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results
-        // LLAMA_SPLIT_MODE_LAYER: ignored
+        // the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE
         int32_t main_gpu;
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
@@ -433,6 +426,7 @@ extern "C" {
     LLAMA_API bool llama_supports_mmap       (void);
     LLAMA_API bool llama_supports_mlock      (void);
     LLAMA_API bool llama_supports_gpu_offload(void);
+    LLAMA_API bool llama_supports_rpc        (void);
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
@@ -775,15 +769,15 @@ extern "C" {
     // Decoding
     //
-    // Return batch for single sequence of tokens starting at pos_0
+    // Return batch for single sequence of tokens
+    // The sequence ID will be fixed to 0
+    // The position of the tokens will be tracked automatically by llama_decode
     //
     // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
     //
     LLAMA_API struct llama_batch llama_batch_get_one(
                   llama_token * tokens,
-                      int32_t   n_tokens,
-                    llama_pos   pos_0,
-                 llama_seq_id   seq_id);
+                      int32_t   n_tokens);
     // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
     // Each token can be assigned up to n_seq_max sequence ids
@@ -803,7 +797,7 @@ extern "C" {
     // Processes a batch of tokens with the ecoder part of the encoder-decoder model.
     // Stores the encoder output internally for later use by the decoder cross-attention layers.
     //   0 - success
-    // < 0 - error
+    // < 0 - error. the KV cache state is restored to the state before this call
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -811,7 +805,7 @@ extern "C" {
     // Positive return values does not mean a fatal error, but rather a warning.
     //   0 - success
     //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    // < 0 - error
+    // < 0 - error. the KV cache state is restored to the state before this call
     LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -896,6 +890,7 @@ extern "C" {
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
+    LLAMA_API llama_token llama_token_eot(const struct llama_model * model); // end-of-turn
     LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
     LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
     LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
@@ -904,11 +899,17 @@ extern "C" {
     LLAMA_API bool llama_add_bos_token(const struct llama_model * model);
     LLAMA_API bool llama_add_eos_token(const struct llama_model * model);
-    // Codellama infill tokens
-    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
-    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
-    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
-    LLAMA_API llama_token llama_token_eot   (const struct llama_model * model); // End of infill middle
+    // infill tokens
+    DEPRECATED(LLAMA_API llama_token llama_token_prefix(const struct llama_model * model), "use llama_token_fim_pre instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_middle(const struct llama_model * model), "use llama_token_fim_mid instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_suffix(const struct llama_model * model), "use llama_token_fim_suf instead");
+    LLAMA_API llama_token llama_token_fim_pre(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_suf(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_mid(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_pad(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_rep(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_sep(const struct llama_model * model);
     //
     // Tokenization
@@ -1067,12 +1068,13 @@ extern "C" {
     // available samplers:
-    LLAMA_API struct llama_sampler * llama_sampler_init_greedy     (void);
-    LLAMA_API struct llama_sampler * llama_sampler_init_dist       (uint32_t seed);
+    LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
+    LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
-    LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void);
+    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
+        "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
@@ -1083,16 +1085,18 @@ extern "C" {
     /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
     LLAMA_API struct llama_sampler * llama_sampler_init_min_p      (float   p, size_t min_keep);
-    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API struct llama_sampler * llama_sampler_init_tail_free  (float   z, size_t min_keep);
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
     LLAMA_API struct llama_sampler * llama_sampler_init_typical    (float   p, size_t min_keep);
+    /// #details Updates the logits l_i` = l_i/t. When t <= 0.0f, the maximum logit is kept at it's original value, the rest are set to -inf
     LLAMA_API struct llama_sampler * llama_sampler_init_temp       (float   t);
     /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.
     LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext   (float   t, float   delta, float exponent);
+    /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335
+    LLAMA_API struct llama_sampler * llama_sampler_init_xtc        (float   p, float   t,     size_t min_keep, uint32_t seed);
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -1132,11 +1136,43 @@ extern "C" {
                                 bool   penalize_nl,     // consider newlines as a repeatable token
                                 bool   ignore_eos);     // ignore the end-of-sequence token
+    ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
+    LLAMA_API struct llama_sampler *    llama_sampler_init_dry(
+            const struct llama_model *  model,
+                               float    dry_multiplier,
+                               float    dry_base,
+                             int32_t    dry_allowed_length,
+                             int32_t    dry_penalty_last_n,
+                          const char ** seq_breakers,
+                              size_t    num_breakers);
     LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
                              int32_t   n_vocab,
                              int32_t   n_logit_bias,
               const llama_logit_bias * logit_bias);
+    // this sampler is meant to be used for fill-in-the-middle infilling
+    // it's supposed to be used after top_k + top_p sampling
+    //
+    // 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG
+    // 2. combine probs of tokens that have the same prefix
+    //
+    // example:
+    //
+    // - before:
+    //   "hel":   0.5
+    //   "hell":  0.2
+    //   "hello": 0.1
+    //   "dummy": 0.1
+    //
+    // - after:
+    //   "hel":   0.8
+    //   "dummy": 0.1
+    //
+    // 3. discard non-EOG tokens with low prob
+    // 4. if no tokens are left -> pick EOT
+    //
+    LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model);
     // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
     LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);
@@ -1208,8 +1244,6 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
-    LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx);
 #ifdef __cplusplus
 }
 #endif

package/src/llama.cpp/pocs/vdot/q8dot.cpp CHANGED Viewed

@@ -11,6 +11,7 @@
 #include <type_traits>
 #include <ggml.h>
+#include <ggml-cpu.h>
 constexpr int kVecSize = 1 << 16;
@@ -136,7 +137,7 @@ int main(int argc, char** argv) {
     auto ggml_type = type == 0 ? GGML_TYPE_Q4_0 : GGML_TYPE_Q4_1;
-    auto funcs = ggml_internal_get_type_traits(ggml_type);
+    const auto * funcs = ggml_get_type_traits_cpu(ggml_type);
     Stat simple, ggml;
@@ -156,8 +157,8 @@ int main(int argc, char** argv) {
         t1 = std::chrono::high_resolution_clock::now();
         float fs;
-        if (type == 0) funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x40.data(), 0, y.data(), 0, 1);
-        else funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x41.data(), 0, y.data(), 0, 1);
+        if (type == 0) funcs->vec_dot(kVecSize * QK4_1, &fs, 0, x40.data(), 0, y.data(), 0, 1);
+        else funcs->vec_dot(kVecSize * QK4_1, &fs, 0, x41.data(), 0, y.data(), 0, 1);
         t2 = std::chrono::high_resolution_clock::now();
         t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
         if (iloop > 3) ggml.addResult(fs, t);

package/src/llama.cpp/pocs/vdot/vdot.cpp CHANGED Viewed

@@ -9,6 +9,7 @@
 #include <array>
 #include <ggml.h>
+#include <ggml-cpu.h>
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -236,7 +237,7 @@ int main(int argc, char** argv) {
     int n4 = useQ4_1 ? kVecSize / QK4_1 : kVecSize / QK4_0; n4 = 64*((n4 + 63)/64);
     int n8 = kVecSize / QK8_0; n8 = 64*((n8 + 63)/64);
-    auto funcs = useQ4_1 ? ggml_internal_get_type_traits(GGML_TYPE_Q4_1) : ggml_internal_get_type_traits(GGML_TYPE_Q4_0);
+    const auto * funcs_cpu = ggml_get_type_traits_cpu(useQ4_1 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q4_0);
     std::vector<block_q4_0> q40;
     std::vector<block_q4_1> q41;
@@ -261,9 +262,9 @@ int main(int argc, char** argv) {
         // Note, we do not include this in the timing as in practical application
         // we already have the quantized model weights.
         if (useQ4_1) {
-            funcs.from_float(x1.data(), q41.data(), kVecSize);
+            funcs_cpu->from_float(x1.data(), q41.data(), kVecSize);
         } else {
-            funcs.from_float(x1.data(), q40.data(), kVecSize);
+            funcs_cpu->from_float(x1.data(), q40.data(), kVecSize);
         }
         // Now measure time the dot product needs using the "scalar" version above
@@ -282,10 +283,10 @@ int main(int argc, char** argv) {
             dot_q4_q8(kVecSize, &result, q40.data(), q8.data());
         }
         else {
-            auto vdot = ggml_internal_get_type_traits(funcs.vec_dot_type);
-            vdot.from_float(y1.data(), q8.data(), kVecSize);
-            if (useQ4_1) funcs.vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
-            else funcs.vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
+            const auto * vdot = ggml_get_type_traits_cpu(funcs_cpu->vec_dot_type);
+            vdot->from_float(y1.data(), q8.data(), kVecSize);
+            if (useQ4_1) funcs_cpu->vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
+            else funcs_cpu->vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
         }
         sumq += result;
         t2 = std::chrono::high_resolution_clock::now();

package/src/llama.cpp/src/CMakeLists.txt CHANGED Viewed

@@ -29,5 +29,6 @@ target_link_libraries(llama PUBLIC ggml)
 if (BUILD_SHARED_LIBS)
     set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
+    target_compile_definitions(llama PRIVATE LLAMA_BUILD)
+    target_compile_definitions(llama PUBLIC  LLAMA_SHARED)
 endif()