Merge commit '5bf2a2771886ee86137e01dbc7492f78fb392066' into concedo_experimental

# Conflicts:
#	.devops/tools.sh
#	README.md

Author: LostRuins

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;
@@ -267,7 +285,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.lora_adapter = argv[i];
-            params.use_mmap = false;
         } else if (arg == "--lora-base") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -470,6 +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");
@@ -499,7 +520,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
     fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
-    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    fprintf(stderr, "  --lora FNAME          apply LoRA adapter\n");
     fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
     fprintf(stderr, "  -m FNAME, --model FNAME\n");
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
@@ -536,7 +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;
@@ -552,6 +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/README.md

@@ -293,5 +293,5 @@ These options provide extra functionality and customization when running the LLa
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
 -   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
--   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
+-   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model. This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

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/server/README.md

@@ -16,7 +16,7 @@ Command line options:
 -   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 -   `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 -   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
--   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
+-   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model. This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
 -   `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.

examples/server/server.cpp

@@ -632,7 +632,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
     fprintf(stderr, "  -a ALIAS, --alias ALIAS\n");
     fprintf(stderr, "                        set an alias for the model, will be added as `model` field in completion response\n");
-    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    fprintf(stderr, "  --lora FNAME          apply LoRA adapter\n");
     fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
     fprintf(stderr, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
     fprintf(stderr, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
@@ -820,7 +820,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 break;
             }
             params.lora_adapter = argv[i];
-            params.use_mmap = false;
         }
         else if (arg == "--lora-base")
         {

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]);
         }
     }