Merge branch 'ggerganov:master' into fix

README.md

@@ -89,6 +89,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
 - [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
+- [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 

common/common.cpp

@@ -941,11 +941,37 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
 
 #ifdef LLAMA_USE_CURL
 
+#define CURL_MAX_RETRY 3
+#define CURL_RETRY_DELAY_SECONDS 2
+
+
 static bool starts_with(const std::string & str, const std::string & prefix) {
     // While we wait for C++20's std::string::starts_with...
     return str.rfind(prefix, 0) == 0;
 }
 
+static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_attempts, int retry_delay_seconds) {
+    int remaining_attempts = max_attempts;
+
+    while (remaining_attempts > 0) {
+        fprintf(stderr, "%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+
+        CURLcode res = curl_easy_perform(curl);
+        if (res == CURLE_OK) {
+            return true;
+        }
+
+        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
+        fprintf(stderr, "%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
+
+        remaining_attempts--;
+        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
+    }
+
+    fprintf(stderr, "%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
+    return false;
+}
+
 static bool llama_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
 
     // Initialize libcurl
@@ -1049,9 +1075,8 @@ static bool llama_download_file(const std::string & url, const std::string & pat
         curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
         curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
 
-        CURLcode res = curl_easy_perform(curl.get());
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (res != CURLE_OK) {
+        if (!was_perform_successful) {
-            fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
             return false;
         }
 
@@ -1126,11 +1151,10 @@ static bool llama_download_file(const std::string & url, const std::string & pat
         };
 
         // start the download
-        fprintf(stderr, "%s: downloading from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
+        fprintf(stderr, "%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
-                llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
+            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        auto res = curl_easy_perform(curl.get());
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (res != CURLE_OK) {
+        if (!was_perform_successful) {
-            fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
             return false;
         }
 

convert_hf_to_gguf_update.py

@@ -31,6 +31,7 @@ import re
 import requests
 import sys
 import json
+import shutil
 
 from hashlib import sha256
 from enum import IntEnum, auto
@@ -125,12 +126,27 @@ def download_model(model):
     if tokt == TOKENIZER_TYPE.UGM:
         files.append("spiece.model")
 
-    for file in files:
+    if os.path.isdir(repo):
-        save_path = f"models/tokenizers/{name}/{file}"
+        # If repo is a path on the file system, copy the directory
-        if os.path.isfile(save_path):
+        for file in files:
-            logger.info(f"{name}: File {save_path} already exists - skipping")
+            src_path = os.path.join(repo, file)
-            continue
+            dst_path = f"models/tokenizers/{name}/{file}"
-        download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
+            if os.path.isfile(dst_path):
+                logger.info(f"{name}: File {dst_path} already exists - skipping")
+                continue
+            if os.path.isfile(src_path):
+                shutil.copy2(src_path, dst_path)
+                logger.info(f"{name}: Copied {src_path} to {dst_path}")
+            else:
+                logger.warning(f"{name}: Source file {src_path} does not exist")
+    else:
+        # If repo is a URL, download the files
+        for file in files:
+            save_path = f"models/tokenizers/{name}/{file}"
+            if os.path.isfile(save_path):
+                logger.info(f"{name}: File {save_path} already exists - skipping")
+                continue
+            download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
 
 
 for model in models:

examples/batched-bench/batched-bench.cpp

@@ -3,32 +3,10 @@
 #include "llama.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-// mutates the input string
-static std::vector<int> parse_list(char * p) {
-    std::vector<int> ret;
-
-    char * q = p;
-
-    while (*p) {
-        if (*p == ',') {
-            *p = '\0';
-            ret.push_back(std::atoi(q));
-            q = p + 1;
-        }
-
-        ++p;
-    }
-
-    ret.push_back(std::atoi(q));
-
-    return ret;
-}
-
 static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);

examples/llava/minicpmv-cli.cpp

@@ -18,8 +18,8 @@ struct llava_context {
 };
 
 static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+    LOG_TEE("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
 static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
@@ -255,7 +255,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, show_additional_info)) {
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
         return 1;
     }
 

ggml/src/ggml-cann.cpp

@@ -2009,7 +2009,7 @@ GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) {
         GGML_CANN_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
         return nullptr;
     }
-
+    ggml_cann_set_device(ctx->device);
     ggml_backend_t cann_backend =
         new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
                          /* .interface = */ ggml_backend_cann_interface,

ggml/src/ggml-cuda/mmq.cu

@@ -26,7 +26,11 @@ void ggml_cuda_op_mul_mat_q(
     // nrows_dst == nrows of the matrix that the kernel writes into
     const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
 
-    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst};
+    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
+    // Also its fixup needs to allocate a temporary buffer in the memory pool.
+    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
+    const bool use_stream_k = compute_capability >= CC_VOLTA && compute_capability < CC_OFFSET_AMD && src1_ncols == ne11;
+    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
 
     switch (src0->type) {
         case GGML_TYPE_Q4_0:

ggml/src/ggml-cuda/mmq.cuh

@@ -2742,6 +2742,7 @@ struct mmq_args {
     int64_t ne00; int64_t ne01; int64_t stride01;
     int64_t ne10; int64_t ne11; int64_t stride11;
     int64_t ne0;
+    bool use_stream_k;
 };
 
 template<ggml_type type>
@@ -2777,8 +2778,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
     const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
     const dim3 block_nums_xy_tiling(nty, ntx, 1);
 
-    const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
+    if (!args.use_stream_k) {
-    if (!use_stream_k) {
         if (args.ne01 % mmq_y == 0) {
             constexpr bool need_check = false;
             mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>

src/llama.cpp

@@ -16080,19 +16080,21 @@ static int llama_decode_internal(
         return -1;
     }
 
-    for (uint32_t i = 0; i < n_tokens_all; ++i) {
-        if (batch_all.token[i] < 0 || (uint32_t)batch_all.token[i] >= lctx.model.vocab.n_vocab) {
-            LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch_all.token[i]);
-            return -1;
-        }
-    }
-
     const auto & model   = lctx.model;
     const auto & hparams = model.hparams;
     const auto & cparams = lctx.cparams;
 
     GGML_ASSERT((!batch_all.token && batch_all.embd) || (batch_all.token && !batch_all.embd)); // NOLINT
 
+    if (batch_all.token) {
+        for (uint32_t i = 0; i < n_tokens_all; ++i) {
+            if (batch_all.token[i] < 0 || (uint32_t)batch_all.token[i] >= model.vocab.n_vocab) {
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch_all.token[i]);
+                return -1;
+            }
+        }
+    }
+
     GGML_ASSERT(n_tokens_all <= cparams.n_batch);
 
     GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens");
@@ -16379,19 +16381,21 @@ static int llama_encode_internal(
         return -1;
     }
 
-    for (uint32_t i = 0; i < n_tokens; ++i) {
-        if (batch.token[i] < 0 || (uint32_t)batch.token[i] >= lctx.model.vocab.n_vocab) {
-            LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch.token[i]);
-            return -1;
-        }
-    }
-
     const auto & model   = lctx.model;
     const auto & hparams = model.hparams;
     const auto & cparams = lctx.cparams;
 
     GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
+    if (batch.token) {
+        for (uint32_t i = 0; i < n_tokens; ++i) {
+            if (batch.token[i] < 0 || (uint32_t)batch.token[i] >= model.vocab.n_vocab) {
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch.token[i]);
+                return -1;
+            }
+        }
+    }
+
     // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
     GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens");