Merge 'origin/master' into hipblas

Author: SlyEcho

.devops/tools.sh

@@ -26,6 +26,8 @@ elif [[ $arg1 == '--all-in-one' || $arg1 == '-a' ]]; then
             ./quantize "$i" "${i/f16/q4_0}" q4_0
         fi
     done
+elif [[ $arg1 == '--server' || $arg1 == '-s' ]]; then
+    ./server $arg2
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "
@@ -37,4 +39,6 @@ else
     echo "              ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
     echo "  --all-in-one (-a): Execute --convert & --quantize"
     echo "              ex: \"/models/\" 7B"
+    echo "  --server (-s): Run a model on the server"
+    echo "              ex: -m /models/7B/ggml-model-q4_0.bin -c 2048 -ngl 43 -mg 1 --port 8080"
 fi

CMakeLists.txt

@@ -273,7 +273,7 @@ if (LLAMA_CUBLAS)
 
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
         if (LLAMA_CUDA_DMMV_F16)
-            set(CMAKE_CUDA_ARCHITECTURES "61") # needed for f16 CUDA intrinsics
+            set(CMAKE_CUDA_ARCHITECTURES "60;61") # needed for f16 CUDA intrinsics
         else()
             set(CMAKE_CUDA_ARCHITECTURES "52;61") # lowest CUDA 12 standard + lowest for integer intrinsics
         endif()

README.md

@@ -239,7 +239,7 @@ In order to build llama.cpp you have three different options.
 - Using `Zig`:
 
     ```bash
-    zig build -Drelease-fast
+    zig build -Doptimize=ReleaseFast
     ```
 
 ### Metal Build

examples/common.cpp

@@ -236,6 +236,24 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.mirostat_tau = std::stof(argv[i]);
+        } else if (arg == "--cfg-negative-prompt") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.cfg_negative_prompt = argv[i];
+        } else if (arg == "--cfg-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.cfg_scale = std::stof(argv[i]);
+        } else if (arg == "--cfg-smooth-factor") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.cfg_smooth_factor = std::stof(argv[i]);
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -469,6 +487,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "                        modifies the likelihood of token appearing in the completion,\n");
     fprintf(stderr, "                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
     fprintf(stderr, "                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
+    fprintf(stderr, "  --cfg-negative-prompt PROMPT \n");
+    fprintf(stderr, "                        negative prompt to use for guidance. (default: empty)\n");
+    fprintf(stderr, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
+    fprintf(stderr, "  --cfg-smooth-factor N smooth factor between old and new logits (default: %f, 1.0 = no smoothing)\n", params.cfg_smooth_factor);
     fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
     fprintf(stderr, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     fprintf(stderr, "  --no-penalize-nl      do not penalize newline token\n");
@@ -535,7 +557,7 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
     return res;
 }
 
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(const gpt_params & params) {
+struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
     auto lparams = llama_context_default_params();
 
     lparams.n_ctx        = params.n_ctx;
@@ -551,6 +573,12 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     lparams.logits_all   = params.perplexity;
     lparams.embedding    = params.embedding;
 
+    return lparams;
+}
+
+std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(const gpt_params & params) {
+    auto lparams = llama_context_params_from_gpt_params(params);
+
     llama_model * model  = llama_load_model_from_file(params.model.c_str(), lparams);
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());

examples/common.h

@@ -48,6 +48,12 @@ struct gpt_params {
     float   mirostat_tau      = 5.00f; // target entropy
     float   mirostat_eta      = 0.10f; // learning rate
 
+    // Classifier-Free Guidance
+    // https://arxiv.org/abs/2306.17806
+    std::string cfg_negative_prompt;       // string to help guidance
+    float       cfg_scale         = 1.f;   // How strong is guidance
+    float       cfg_smooth_factor = 1.f;   // Smooth factor between old and new logits
+
     std::string model             = "models/7B/ggml-model.bin"; // model path
     std::string model_alias       = "unknown"; // model alias
     std::string prompt            = "";
@@ -99,6 +105,7 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
 //
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(const gpt_params & params);
+struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
 //
 // Console utils

examples/main/main.cpp

@@ -109,10 +109,16 @@ int main(int argc, char ** argv) {
 
     llama_model * model;
     llama_context * ctx;
+    llama_context * ctx_guidance = NULL;
     g_ctx = &ctx;
 
     // load the model and apply lora adapter, if any
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (params.cfg_scale > 1.f) {
+        struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
+        ctx_guidance = llama_new_context_with_model(model, lparams);
+    }
+
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -183,15 +189,28 @@ int main(int argc, char ** argv) {
     // tokenize the prompt
     std::vector<llama_token> embd_inp;
 
-    if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
-        // Add a space in front of the first character to match OG llama tokenizer behavior
-        params.prompt.insert(0, 1, ' ');
+    // Add a space in front of the first character to match OG llama tokenizer behavior
+    params.prompt.insert(0, 1, ' ');
 
+    if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
         embd_inp = ::llama_tokenize(ctx, params.prompt, true);
     } else {
         embd_inp = session_tokens;
     }
 
+    // Tokenize negative prompt
+    std::vector<llama_token> guidance_inp;
+    int guidance_offset = 0;
+    int original_prompt_len = 0;
+    if (ctx_guidance) {
+        params.cfg_negative_prompt.insert(0, 1, ' ');
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, true);
+
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true);
+        original_prompt_len = original_inp.size();
+        guidance_offset = (int)guidance_inp.size() - original_prompt_len;
+    }
+
     const int n_ctx = llama_n_ctx(ctx);
 
     if ((int) embd_inp.size() > n_ctx - 4) {
@@ -258,6 +277,16 @@ int main(int argc, char ** argv) {
         for (int i = 0; i < (int) embd_inp.size(); i++) {
             fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]));
         }
+
+        if (ctx_guidance) {
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
+            fprintf(stderr, "%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
+            for (int i = 0; i < (int) guidance_inp.size(); i++) {
+                fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_str(ctx, guidance_inp[i]));
+            }
+        }
+
         if (params.n_keep > 0) {
         fprintf(stderr, "%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
@@ -334,11 +363,13 @@ int main(int argc, char ** argv) {
     int n_remain           = params.n_predict;
     int n_consumed         = 0;
     int n_session_consumed = 0;
+    int n_past_guidance    = 0;
 
     // the first thing we will do is to output the prompt, so set color accordingly
     console_set_color(con_st, CONSOLE_COLOR_PROMPT);
 
     std::vector<llama_token> embd;
+    std::vector<llama_token> embd_guidance;
 
     // do one empty run to warm up the model
     {
@@ -367,11 +398,12 @@ int main(int argc, char ** argv) {
             // if we run out of context:
             // - take the n_keep first tokens from the original prompt (via n_past)
             // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-            if (n_past + (int) embd.size() > n_ctx) {
+            if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
                 const int n_left = n_past - params.n_keep;
 
                 // always keep the first token - BOS
                 n_past = std::max(1, params.n_keep);
+                n_past_guidance = std::max(1, params.n_keep + guidance_offset);
 
                 // insert n_left/2 tokens at the start of embd from last_n_tokens
                 embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size());
@@ -412,6 +444,48 @@ int main(int argc, char ** argv) {
 
             // evaluate tokens in batches
             // embd is typically prepared beforehand to fit within a batch, but not always
+
+            if (ctx_guidance) {
+                int input_size = 0;
+                llama_token* input_buf = NULL;
+
+                if (n_past_guidance < (int) guidance_inp.size()) {
+                    // Guidance context should have the same data with these modifications:
+                    //
+                    // * Replace the initial prompt
+                    // * Shift everything by guidance_offset
+                    embd_guidance = guidance_inp;
+                    if (embd.begin() + original_prompt_len < embd.end()) {
+                        embd_guidance.insert(
+                            embd_guidance.end(),
+                            embd.begin() + original_prompt_len,
+                            embd.end()
+                        );
+                    }
+
+                    input_buf = embd_guidance.data();
+                    input_size = embd_guidance.size();
+                    //fprintf(stderr, "\n---------------------\n");
+                    //for (int i = 0; i < (int) embd_guidance.size(); i++) {
+                        //fprintf(stderr, "%s", llama_token_to_str(ctx, embd_guidance[i]));
+                    //}
+                    //fprintf(stderr, "\n---------------------\n");
+                } else {
+                    input_buf = embd.data();
+                    input_size = embd.size();
+                }
+
+                for (int i = 0; i < input_size; i += params.n_batch) {
+                    int n_eval = std::min(input_size - i, params.n_batch);
+                    if (llama_eval(ctx_guidance, input_buf + i, n_eval, n_past_guidance, params.n_threads)) {
+                        fprintf(stderr, "%s : failed to eval\n", __func__);
+                        return 1;
+                    }
+
+                    n_past_guidance += n_eval;
+                }
+            }
+
             for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
                 int n_eval = (int) embd.size() - i;
                 if (n_eval > params.n_batch) {
@@ -431,6 +505,7 @@ int main(int argc, char ** argv) {
         }
 
         embd.clear();
+        embd_guidance.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
             // out of user input, sample next token
@@ -473,6 +548,10 @@ int main(int argc, char ** argv) {
 
                 llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
+                if (ctx_guidance) {
+                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale, params.cfg_smooth_factor);
+                }
+
                 // Apply penalties
                 float nl_logit = logits[llama_token_nl()];
                 auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
@@ -668,6 +747,7 @@ int main(int argc, char ** argv) {
     }
 
     llama_print_timings(ctx);
+    if (ctx_guidance) { llama_free(ctx_guidance); }
     llama_free(ctx);
     llama_free_model(model);
 

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

@@ -1354,17 +1354,9 @@ struct ggml_tensor * expand(struct ggml_cgraph * g, struct ggml_tensor * t) {
         }
     }
 
-    if (t->src0) {
-        expand(g, t->src0);
-    }
-
-    if (t->src1) {
-        expand(g, t->src1);
-    }
-
-    for (int i = 0; i < GGML_MAX_OPT; ++i) {
-        if (t->opt[i]) {
-            expand(g, t->opt[i]);
+    for (int i = 0; i < GGML_MAX_SRC; ++i) {
+        if (t->src[i]) {
+            expand(g, t->src[i]);
         }
     }