Merge branch 'master' of github.com:ggerganov/llama.cpp

* 'master' of github.com:ggerganov/llama.cpp:
  fix embeddings when using CUDA (ggerganov#3657)
  llama : avoid fprintf in favor of LLAMA_LOG (ggerganov#3538)
  readme : update hot-topics & models, detail windows release in usage (ggerganov#3615)
  CLBlast: Fix temporary buffer size for f16 conversion (wsize)
  train-text-from-scratch : fix assert failure in ggml-alloc (ggerganov#3618)
  editorconfig : remove trailing spaces
  server : documentation of JSON return value of /completion endpoint (ggerganov#3632)
  save-load-state : fix example + add ci test (ggerganov#3655)
  readme : add Aquila2 links (ggerganov#3610)
  tokenizer : special token handling (ggerganov#3538)
  k-quants : fix quantization ranges (ggerganov#3646)
  llava : fix tokenization to not add bos between image embeddings and user prompt (ggerganov#3645)
  MPT : support GQA for replit-code-v1.5 (ggerganov#3627)
  Honor -ngl option for Cuda offloading in llava (ggerganov#3621)

README.md

@@ -10,7 +10,7 @@
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
-
+- ‼️ BPE tokenizer update: existing Falcon and Starcoder `.gguf` models will need to be reconverted: [#3252](https://github.com/ggerganov/llama.cpp/pull/3252)
 - ‼️ 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**
@@ -89,15 +89,17 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
 - [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
-- [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b)
+- [X] [Pygmalion/Metharme](#using-pygmalion-7b--metharme-7b)
 - [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] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
+- [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
 - [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
 - [X] [Mistral AI v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
-- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
+- [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
 - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
+- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
+
 
 **Bindings:**
 
@@ -206,7 +208,7 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
 
 ## Usage
 
-Here are the steps for the LLaMA-7B model.
+Here are the end-to-end binary build and model conversion steps for the LLaMA-7B model.
 
 ### Get the Code
 
@@ -573,6 +575,18 @@ python3 convert.py models/7B/
 
 When running the larger models, make sure you have enough disk space to store all the intermediate files.
 
+### Running on Windows with prebuilt binaries
+
+You will find prebuilt Windows binaries on the release page.
+
+Simply download and extract the latest zip package of choice: (e.g. `llama-b1380-bin-win-avx2-x64.zip`)
+
+From the unzipped folder, open a terminal/cmd window here and place a pre-converted `.gguf` model file. Test out the main example like so:
+
+```
+.\main -m llama-2-7b.Q4_0.gguf -n 128
+```
+
 ### Memory/Disk Requirements
 
 As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.

ci/run.sh

@@ -208,6 +208,8 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
+    (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
     function check_ppl {
         qnt="$1"
         ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
@@ -296,6 +298,7 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
     gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
     gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
     gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
@@ -382,6 +385,8 @@ function gg_run_open_llama_7b_v2 {
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
+    (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
     function check_ppl {
         qnt="$1"
         ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
@@ -470,6 +475,7 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
     gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
     gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
     #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"

common/common.cpp

@@ -879,21 +879,23 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 std::vector<llama_token> llama_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
-                        bool   add_bos) {
-    return llama_tokenize(llama_get_model(ctx), text, add_bos);
+                        bool   add_bos,
+                        bool   special) {
+    return llama_tokenize(llama_get_model(ctx), text, add_bos, special);
 }
 
 std::vector<llama_token> llama_tokenize(
     const struct llama_model * model,
            const std::string & text,
-                        bool   add_bos) {
+                        bool   add_bos,
+                        bool   special) {
     // upper limit for the number of tokens
     int n_tokens = text.length() + add_bos;
     std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);

common/common.h

@@ -137,12 +137,14 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 std::vector<llama_token> llama_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
-                        bool   add_bos);
+                        bool   add_bos,
+                        bool   special = false);
 
 std::vector<llama_token> llama_tokenize(
     const struct llama_model * model,
            const std::string & text,
-                        bool   add_bos);
+                        bool   add_bos,
+                        bool   special = false);
 
 // tokenizes a token into a piece
 // should work similar to Python's `tokenizer.id_to_piece`

common/train.cpp

@@ -863,7 +863,7 @@ size_t tokenize_file(
             (int) buf.size(),
             out_tokens.data(),
             (int) out_tokens.size(),
-            false);
+            false, false);
         if (n_tokens < 0) {
             out_tokens.resize(-n_tokens);
             n_tokens = llama_tokenize(
@@ -872,7 +872,7 @@ size_t tokenize_file(
                 (int) buf.size(),
                 out_tokens.data(),
                 (int) out_tokens.size(),
-                false);
+                false, false);
         }
         if (n_tokens >= 0) {
             out_tokens.resize(n_tokens);
@@ -966,15 +966,15 @@ size_t tokenize_file(
                     (int) buf_sample.size(),
                     tok_sample.data(),
                     (int) tok_sample.size(),
-                    false);
+                    false, false);
                 if (n_tokens < 0) {
                     tok_sample.resize(-n_tokens);
                     n_tokens = llama_tokenize(llama_get_model(lctx),
                         buf_sample.data(),
                         (int) buf_sample.size(),
                         tok_sample.data(),
                         (int) tok_sample.size(),
-                        false);
+                        false, false);
                     GGML_ASSERT(n_tokens >= 0);
                 }
                 GGML_ASSERT(n_tokens <= (int) tok_sample.size());

convert-mpt-hf-to-gguf.py

@@ -98,6 +98,8 @@ def parse_args() -> argparse.Namespace:
 gguf_writer.add_block_count(block_count)
 gguf_writer.add_feed_forward_length(4 * hparams["d_model"])
 gguf_writer.add_head_count(hparams["n_heads"])
+if kv_n_heads := hparams["attn_config"].get("kv_n_heads"):
+    gguf_writer.add_head_count_kv(kv_n_heads)
 gguf_writer.add_layer_norm_eps(1e-05)
 if hparams["attn_config"]["clip_qkv"] is not None:
     gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"])

examples/batched.swift/Sources/main.swift

@@ -209,7 +209,7 @@ llama_print_timings(context)
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
     let n_tokens = text.count + (add_bos ? 1 : 0)
     let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
-    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos)
+    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
     var swiftTokens: [llama_token] = []
     for i in 0 ..< tokenCount {
         swiftTokens.append(tokens[Int(i)])

examples/llava/llava-utils.h

@@ -49,9 +49,9 @@ inline bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
     return eval_tokens(ctx_llama, tokens, 1, n_past);
 }
 
-inline bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past){
+inline bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
     std::string              str2     = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, true);
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos);
     eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
     return true;
 }

examples/llava/llava.cpp

@@ -79,7 +79,13 @@ int main(int argc, char ** argv) {
 
     llama_backend_init(params.numa);
 
-    llama_model_params model_params = llama_model_default_params();
+    llama_model_params model_params              = llama_model_default_params();
+                       model_params.n_gpu_layers = params.n_gpu_layers;
+                       model_params.main_gpu     = params.main_gpu;
+                       model_params.tensor_split = params.tensor_split;
+                       model_params.use_mmap     = params.use_mmap;
+                       model_params.use_mlock    = params.use_mlock;
+
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
     if (model == NULL) {
         fprintf(stderr , "%s: error: unable to load model\n" , __func__);
@@ -91,6 +97,7 @@ int main(int argc, char ** argv) {
     ctx_params.n_ctx           = params.n_ctx < 2048 ? 2048 : params.n_ctx; // we need a longer context size to process image embeddings
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    ctx_params.seed            = params.seed;
 
     llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
 
@@ -100,7 +107,8 @@ int main(int argc, char ** argv) {
     }
 
     // make sure that the correct mmproj was used, i.e., compare apples to apples
-    int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
+    const int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
+
     if (n_img_embd != n_llama_embd) {
         printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_img_embd, n_llama_embd);
 
@@ -119,14 +127,14 @@ int main(int argc, char ** argv) {
 
     const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
 
-    // GG: are we sure that the should be a trailing whitespace at the end of this string?
-    eval_string(ctx_llama, "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER: ", params.n_batch, &n_past);
+    eval_string(ctx_llama, "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params.n_batch, &n_past, true);
     eval_image_embd(ctx_llama, image_embd, n_img_pos, params.n_batch, &n_past);
-    eval_string(ctx_llama, params.prompt.c_str(), params.n_batch, &n_past);
-    eval_string(ctx_llama, "\nASSISTANT:",        params.n_batch, &n_past);
+    eval_string(ctx_llama, (params.prompt + "\nASSISTANT:").c_str(), params.n_batch, &n_past, false);
 
     // generate the response
 
+    printf("\n");
+    printf("prompt: '%s'\n", params.prompt.c_str());
     printf("\n");
 
     for (int i = 0; i < max_tgt_len; i++) {

examples/main/main.cpp

@@ -238,7 +238,7 @@ int main(int argc, char ** argv) {
 
     if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
         LOG("tokenize the prompt\n");
-        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
     } else {
         LOG("use session tokens\n");
         embd_inp = session_tokens;
@@ -260,10 +260,10 @@ int main(int argc, char ** argv) {
     if (ctx_guidance) {
         LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
 
-        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
+        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos, true);
         LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
 
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
         LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
 
         original_prompt_len = original_inp.size();
@@ -320,8 +320,8 @@ int main(int argc, char ** argv) {
     }
 
     // prefix & suffix for instruct mode
-    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
-    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);
+    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos, true);
+    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false,   true);
 
     LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
     LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
@@ -383,6 +383,12 @@ int main(int argc, char ** argv) {
         if (!params.antiprompt.empty()) {
             for (const auto & antiprompt : params.antiprompt) {
                 LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
+                if (params.verbose_prompt) {
+                    auto tmp = ::llama_tokenize(ctx, antiprompt, false, true);
+                    for (int i = 0; i < (int) tmp.size(); i++) {
+                        LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
+                    }
+                }
             }
         }
 
@@ -392,10 +398,22 @@ int main(int argc, char ** argv) {
 
         if (!params.input_prefix.empty()) {
             LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
+            if (params.verbose_prompt) {
+                auto tmp = ::llama_tokenize(ctx, params.input_prefix, true, true);
+                for (int i = 0; i < (int) tmp.size(); i++) {
+                    LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
+                }
+            }
         }
 
         if (!params.input_suffix.empty()) {
             LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
+            if (params.verbose_prompt) {
+                auto tmp = ::llama_tokenize(ctx, params.input_suffix, false, true);
+                for (int i = 0; i < (int) tmp.size(); i++) {
+                    LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
+                }
+            }
         }
     }
     LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
@@ -717,7 +735,7 @@ int main(int argc, char ** argv) {
                 if (params.interactive) {
                     if (!params.antiprompt.empty()) {
                         // tokenize and inject first reverse prompt
-                        const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
+                        const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false, true);
                         embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
                         is_antiprompt = true;
                     }
@@ -744,8 +762,7 @@ int main(int argc, char ** argv) {
                 std::string buffer;
                 if (!params.input_prefix.empty()) {
                     LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
-                    buffer += params.input_prefix;
-                    printf("%s", buffer.c_str());
+                    printf("%s", params.input_prefix.c_str());
                 }
 
                 // color user input only
@@ -767,7 +784,6 @@ int main(int argc, char ** argv) {
                     // append input suffix if any
                     if (!params.input_suffix.empty()) {
                         LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
-                        buffer += params.input_suffix;
                         printf("%s", params.input_suffix.c_str());
                     }
 
@@ -782,10 +798,14 @@ int main(int argc, char ** argv) {
                         embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
                     }
 
-                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
+                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, false);
+                    const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
                     LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
 
+                    embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+                    embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
 
                     // instruct mode: insert response suffix
                     if (params.instruct) {