Merge branch 'master' of github.com:ggerganov/llama.cpp into phillip-kravtsov/support-adept-persimmon-8b

README.md

@@ -11,7 +11,8 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-- Parallel decoding + continuous batching support incoming: [#3228](https://github.com/ggerganov/llama.cpp/pull/3228) \
+- ‼️ Breaking change: `rope_freq_base` and `rope_freq_scale` must be set to zero to use the model default values: [#3401](https://github.com/ggerganov/llama.cpp/pull/3401)
+- Parallel decoding + continuous batching support added: [#3228](https://github.com/ggerganov/llama.cpp/pull/3228) \
   **Devs should become familiar with the new API**
 - Local Falcon 180B inference on Mac Studio
 
@@ -92,7 +93,8 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
 - [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) and its derivations (such as [baichuan-7b-sft](https://huggingface.co/hiyouga/baichuan-7b-sft))
 - [X] [Aquila-7B](https://huggingface.co/BAAI/Aquila-7B) / [AquilaChat-7B](https://huggingface.co/BAAI/AquilaChat-7B)
-- [X] Mistral AI v0.1
+- [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
+- [X] [Mistral AI v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 
 **Bindings:**
 
@@ -662,6 +664,8 @@ PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
 
 The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
 
+For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
+
 ### Instruction mode with Alpaca
 
 1. First, download the `ggml` Alpaca model into the `./models` folder

convert.py

@@ -439,7 +439,7 @@ Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
 def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
     #print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
     if n_head_kv is not None and n_head != n_head_kv:
-        n_head //= n_head_kv
+        n_head = n_head_kv
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
                 .swapaxes(1, 2)
                 .reshape(weights.shape))

examples/finetune/finetune.cpp

@@ -626,7 +626,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
     // KQ_pos - contains the positions
     struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
-    {
+    ggml_allocr_alloc(alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(alloc)) {
         int * data = (int *) KQ_pos->data;
         for (int i = 0; i < N; ++i) {
             data[i] = n_past + i;
@@ -786,6 +787,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
     GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
     ggml_allocr_alloc(alloc, t36->grad);
+    // KQ_pos
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
 
     // make sure base model tensors data cannot be used in viewable operations
     ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -334,7 +334,8 @@ static struct ggml_tensor * llama_build_train_graphs(
 
     // KQ_pos - contains the positions
     struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
-    {
+    ggml_allocr_alloc(alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(alloc)) {
         int * data = (int *) KQ_pos->data;
         for (int i = 0; i < N; ++i) {
             data[i] = n_past + i;

ggml.c

@@ -11583,6 +11583,7 @@ static void ggml_compute_forward_mul_mat(
               struct ggml_tensor * dst) {
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
+
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;

llama.cpp

@@ -6491,7 +6491,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         nthread = std::thread::hardware_concurrency();
     }
 
-    llama_model_loader ml(fname_inp, /*use_mmap*/ false);
+    // mmap consistently increases speed Linux, and also increases speed on Windows with
+    // hot cache. It may cause a slowdown on macOS, possibly related to free memory.
+#if defined(__linux__) || defined(_WIN32)
+    constexpr bool use_mmap = true;
+#else
+    constexpr bool use_mmap = false;
+#endif
+
+    llama_model_loader ml(fname_inp, use_mmap);
+    if (ml.use_mmap) {
+        ml.mapping.reset(new llama_mmap(&ml.file, /* prefetch */ 0, ggml_is_numa()));
+    }
 
     llama_model model;
     llm_load_arch(ml, model);
@@ -6569,10 +6580,12 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
         const std::string name = ggml_get_name(tensor);
 
-        if (read_data.size() < ggml_nbytes(tensor)) {
-            read_data.resize(ggml_nbytes(tensor));
+        if (!ml.use_mmap) {
+            if (read_data.size() < ggml_nbytes(tensor)) {
+                read_data.resize(ggml_nbytes(tensor));
+            }
+            tensor->data = read_data.data();
         }
-        tensor->data = read_data.data();
         ml.load_data_for(tensor);
 
         LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ",

llama.h

@@ -167,18 +167,18 @@ extern "C" {
 
     struct llama_context_params {
         uint32_t seed;            // RNG seed, -1 for random
-        uint32_t n_ctx;           // text context
-        uint32_t n_batch;         // prompt processing batch size
+        uint32_t n_ctx;           // text context, 0 = from model
+        uint32_t n_batch;         // prompt processing maximum batch size
         uint32_t n_threads;       // number of threads to use for generation
         uint32_t n_threads_batch; // number of threads to use for batch processing
 
         // ref: https://github.com/ggerganov/llama.cpp/pull/2054
-        float rope_freq_base;  // RoPE base frequency
-        float rope_freq_scale; // RoPE frequency scaling factor
+        float rope_freq_base;  // RoPE base frequency, 0 = from model
+        float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
-        bool f16_kv;     // use fp16 for KV cache
+        bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool embedding;  // embedding mode only
     };