[WIP] Add Fill-In-Middle example

Makefile

@@ -1,5 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam-search tests/test-c.o
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam-search fill-in-middle tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1
@@ -449,6 +449,9 @@ baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o $(OBJS)
 beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+fill-in-middle: examples/fill-in-middle/FIM.c ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
 BUILD_TARGETS += metal
 endif

examples/fill-in-middle/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET FIM)
+add_executable(${TARGET} FIM.c)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/fill-in-middle/FIM.c

@@ -0,0 +1,165 @@
+#include <stdlib.h>
+#include <stdio.h>
+#include "../../llama.h"
+
+/*
+The FIM (Fill-In-Middle) objective is useful for generating text conditioned on a prefix and a suffix.
+For a quick summary of what's going on here, see issue #2818.
+*/
+
+
+static inline struct llama_context* codellama_create_fim_context(const char* model_path, const char** error_message) {
+    struct llama_context_params params = llama_context_default_params();
+    struct llama_model* model = llama_load_model_from_file(model_path,  params);
+    if (!model) {
+        *error_message = "Failed to load model.";
+        return NULL;
+    }
+
+    struct llama_context* context = llama_new_context_with_model(model, params);
+    if (!context) {
+        *error_message = "Failed to create context.";
+        llama_free_model(model);
+        return NULL;
+    }
+
+    return context;
+}
+
+static inline char*
+codellama_fill_in_middle(struct llama_context* ctx, const char* prefix, const char* suffix, size_t n_max_tokens, int n_threads, bool spm, const char** error_message) {
+
+    int num_tokens;
+    llama_token* tokens_end = (llama_token*)malloc(sizeof(llama_token) * n_max_tokens);
+    llama_token* tokens = tokens_end;
+    if (!tokens) {
+        *error_message = "Failed to allocate memory for tokens.";
+        return NULL;
+    }
+
+    // Append first part of prompt
+    *tokens_end++ = spm ? llama_token_suffix(ctx) : llama_token_prefix(ctx);
+    tokens_end += num_tokens = llama_tokenize(ctx, spm ? suffix : prefix, tokens_end, n_max_tokens, 1);
+    if (num_tokens < 0) {
+        *error_message = "Failed to tokenize the prompt.";
+        free(tokens);
+        return NULL;
+    }
+
+    // Append second part of prompt
+    *tokens_end++ = spm ? llama_token_prefix(ctx) : llama_token_suffix(ctx);
+    tokens_end += num_tokens = llama_tokenize(ctx, spm ? prefix : suffix, tokens_end, n_max_tokens, 1);
+    if (num_tokens < 0) {
+        *error_message = "Failed to tokenize the prompt.";
+        free(tokens);
+        return NULL;
+    }
+
+    // Append middle token
+    *tokens_end++ = llama_token_middle(ctx);
+
+    // Evaluate the LM on the prompt.
+    if (llama_eval(ctx, tokens, (int)(tokens_end - tokens), 0, n_threads)) {
+        *error_message = "Failed to evaluate the prompt.";
+        free(tokens);
+        return NULL;
+    }
+
+    // Generate tokens until n_max_tokens or the <EOT> token is generated.
+    llama_token* generated_tokens = NULL;
+    size_t num_generated_tokens = 0;
+    int vocab_size = llama_n_vocab(ctx);
+    for (size_t i = 0; i < n_max_tokens; i++) {
+        // Evaluate the LM for a single token, obtaining the logits and probabilities.
+        if (llama_eval(ctx, &generated_tokens[num_generated_tokens], 1, (int)num_generated_tokens, n_threads)) {
+            *error_message = "Failed to evaluate the prompt.";
+            free(tokens);
+            break;
+        }
+        float* logits = llama_get_logits(ctx);
+
+        // From the logits, select the most likely token.
+        float highest_log_likelihood = -1;
+        llama_token likeliest_token = -1;
+        for (llama_token token_id = 0; token_id < vocab_size; token_id++) {
+            if (logits[token_id] > highest_log_likelihood) {
+                highest_log_likelihood = logits[token_id];
+                likeliest_token = token_id;
+            }
+        }
+
+        // Don't add the token if it's <EOT>.
+        if (likeliest_token == llama_token_eot(ctx)) {
+            break;
+        }
+
+        // Append the token, so it's there for subsequent evaluations.
+        generated_tokens[num_generated_tokens++] = likeliest_token;
+    }
+
+    // Allocate memory for the final result
+    size_t result_length = 0;
+    size_t result_capacity = 4096;
+    char* result = (char*)malloc(sizeof(char) * result_capacity);
+    if (!result) {
+        *error_message = "Failed to allocate memory for result.";
+        free(tokens);
+        return NULL;
+    }
+
+    // Translate tokens to string, growing the allocation if it's too small.
+    for (size_t i = 0; i < num_generated_tokens; i++) {
+        int appended = llama_token_to_piece(ctx, generated_tokens[i], result, result_capacity - result_length);
+        if (appended < 0) {
+            // retry the token with a larger buffer
+            i--;
+            size_t new_capacity = result_capacity * 2;
+            char* new_result = (char*)realloc(result, sizeof(char) * new_capacity);
+            if (!new_result) {
+                *error_message = "Failed to allocate memory for result.";
+                free(tokens);
+                free(result);
+                return NULL;
+            }
+            result = new_result;
+            result_capacity = new_capacity;
+        }
+
+        result_length += appended;
+    }
+
+    free(tokens);
+    *error_message = NULL;
+    return result;
+}
+
+int main(int argc, char** argv) {
+    if (argc != 6) {
+        fprintf(stderr, "Usage: %s <model> <prefix> <suffix> <n_max_tokens> <n_threads>\n", argv[0]);
+        return 1;
+    }
+
+    char* model = argv[1];
+    char* prefix = argv[2];
+    char* suffix = argv[3];
+    size_t n_max_tokens = atoi(argv[4]);
+    int n_threads = atoi(argv[5]);
+    bool spm = false;
+    const char* error_message = NULL;
+
+    struct llama_context* ctx = codellama_create_fim_context(model, &error_message);
+    if (error_message) {
+        fprintf(stderr, "Error: %s\n", error_message);
+        return 1;
+    }
+
+    char* result = codellama_fill_in_middle(ctx, prefix, suffix, n_max_tokens, n_threads, spm, &error_message);
+    if (error_message) {
+        fprintf(stderr, "Error: %s\n", error_message);
+        return 1;
+    }
+
+    printf("%s%s%s\n", prefix, result, suffix);
+
+    llama_free(ctx);
+}

llama.cpp

@@ -959,6 +959,13 @@ struct llama_vocab {
 
     id linefeed_id = 13;
 
+    // codellama FIM special tokens
+    // TODO: load these from the vocabulary.
+    id special_prefix_id = 32007;
+    id special_middle_id = 32009;
+    id special_suffix_id = 32008;
+    id special_eot_id = 32010;
+
     int find_bpe_rank(std::string token_left, std::string token_right) const {
         replace_all(token_left,  " ",  "\u0120");
         replace_all(token_left,  "\n", "\u010A");
@@ -6118,6 +6125,22 @@ llama_token llama_token_nl(const struct llama_context * ctx) {
     return ctx->model.vocab.linefeed_id;
 }
 
+llama_token llama_token_prefix(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_prefix_id;
+}
+
+llama_token llama_token_middle(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_middle_id;
+}
+
+llama_token llama_token_suffix(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_suffix_id;
+}
+
+llama_token llama_token_eot(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_eot_id;
+}
+
 int llama_tokenize(
         struct llama_context * ctx,
                   const char * text,

llama.h

@@ -360,6 +360,12 @@ extern "C" {
     LLAMA_API llama_token llama_token_eos(const struct llama_context * ctx);  // end-of-sentence
     LLAMA_API llama_token llama_token_nl (const struct llama_context * ctx);  // next-line
 
+    // codellama FIM tokens
+    LLAMA_API llama_token llama_token_prefix(const struct llama_context * ctx); // Beginning of FIM prefix
+    LLAMA_API llama_token llama_token_middle(const struct llama_context * ctx); // Beginning of FIM middle
+    LLAMA_API llama_token llama_token_suffix(const struct llama_context * ctx); // Beginning of FIM suffix
+    LLAMA_API llama_token llama_token_eot   (const struct llama_context * ctx); // End of FIM middle
+
     //
     // Tokenization
     //