llama_cpp 0.8.0 → 0.9.0

data/ext/llama_cpp/src/ggml.h

@@ -401,15 +401,16 @@ extern "C" {
         GGML_OP_ALIBI,
         GGML_OP_CLAMP,
         GGML_OP_CONV_1D,
-        GGML_OP_CONV_2D,
+        GGML_OP_CONV_1D_STAGE_0,  // internal
+        GGML_OP_CONV_1D_STAGE_1,  // internal
         GGML_OP_CONV_TRANSPOSE_1D,
+        GGML_OP_CONV_2D,
+        GGML_OP_CONV_2D_STAGE_0, // internal
+        GGML_OP_CONV_2D_STAGE_1, // internal
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
 
-        GGML_OP_CONV_1D_STAGE_0,  // internal
-        GGML_OP_CONV_1D_STAGE_1,  // internal
-
         GGML_OP_UPSCALE, // nearest interpolate
 
         GGML_OP_FLASH_ATTN,
@@ -1020,9 +1021,9 @@ extern "C" {
             struct ggml_tensor  * b,
             float                 eps);
 
-    // A: n columns, m rows
-    // B: n columns, p rows  (i.e. we transpose it internally)
-    // result is m columns, p rows
+    // A: k columns, n rows => [ne03, ne02, n, k]
+    // B: k columns, m rows  (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
+    // result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
     GGML_API struct ggml_tensor * ggml_mul_mat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

data/ext/llama_cpp/src/llama.cpp

@@ -970,14 +970,15 @@ static void llama_nop(struct ggml_tensor * tensor) { // don't offload by default
     (void) tensor;
 }
 
-static std::string llama_token_to_str(const struct llama_context * ctx, llama_token token) {
+static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
     const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
         int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
-    } else {
+    }
+    else {
         result.resize(n_tokens);
     }
 
@@ -1013,8 +1014,8 @@ enum e_model {
 };
 
 static const size_t kB = 1024;
-static const size_t MB = kB*kB;
-static const size_t GB = kB*kB*kB;
+static const size_t MB = 1024*kB;
+static const size_t GB = 1024*MB;
 
 struct llama_hparams {
     bool     vocab_only;
@@ -1037,21 +1038,21 @@ struct llama_hparams {
     float f_max_alibi_bias;
 
     bool operator!=(const llama_hparams & other) const {
-        if (this->vocab_only != other.vocab_only) return true;
-        if (this->n_vocab != other.n_vocab) return true;
+        if (this->vocab_only  != other.vocab_only)  return true;
+        if (this->n_vocab     != other.n_vocab)     return true;
         if (this->n_ctx_train != other.n_ctx_train) return true;
-        if (this->n_embd != other.n_embd) return true;
-        if (this->n_head != other.n_head) return true;
-        if (this->n_head_kv != other.n_head_kv) return true;
-        if (this->n_layer != other.n_layer) return true;
-        if (this->n_rot != other.n_rot) return true;
-        if (this->n_ff != other.n_ff) return true;
+        if (this->n_embd      != other.n_embd)      return true;
+        if (this->n_head      != other.n_head)      return true;
+        if (this->n_head_kv   != other.n_head_kv)   return true;
+        if (this->n_layer     != other.n_layer)     return true;
+        if (this->n_rot       != other.n_rot)       return true;
+        if (this->n_ff        != other.n_ff)        return true;
 
         const float EPSILON = 1e-9;
 
-        if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true;
-        if (!is_float_close(this->f_norm_rms_eps, other.f_norm_rms_eps, EPSILON)) return true;
-        if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true;
+        if (!is_float_close(this->f_norm_eps,            other.f_norm_eps,            EPSILON)) return true;
+        if (!is_float_close(this->f_norm_rms_eps,        other.f_norm_rms_eps,        EPSILON)) return true;
+        if (!is_float_close(this->rope_freq_base_train,  other.rope_freq_base_train,  EPSILON)) return true;
         if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
 
         return false;
@@ -1190,17 +1191,17 @@ struct llama_vocab {
     id special_sep_id = -1;
     id special_pad_id = -1;
 
-    id linefeed_id = 13;
+    id linefeed_id       = 13;
     id special_prefix_id = 32007;
     id special_middle_id = 32009;
     id special_suffix_id = 32008;
-    id special_eot_id = 32010;
+    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");
-        replace_all(token_right, " ",  "\u0120");
-        replace_all(token_right, "\n", "\u010A");
+        GGML_ASSERT(token_left.find(" ") == std::string::npos);
+        GGML_ASSERT(token_left.find("\n") == std::string::npos);
+        GGML_ASSERT(token_right.find(" ") == std::string::npos);
+        GGML_ASSERT(token_right.find("\n") == std::string::npos);
 
         auto it = bpe_ranks.find(std::make_pair(token_left, token_right));
         if (it == bpe_ranks.end()) {
@@ -1354,10 +1355,7 @@ static bool llama_kv_cache_init(
     cache.cells.clear();
     cache.cells.resize(n_ctx);
 
-    // TODO: this should be:
-    //       cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*ggml_tensor_overhead());
-    //       change it and test that it works
-    cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
+    cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*ggml_tensor_overhead());
     memset(cache.buf.data, 0, cache.buf.size);
 
     struct ggml_init_params params;
@@ -2236,15 +2234,35 @@ static void llm_load_vocab(
     if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
         vocab.linefeed_id = llama_byte_to_token(vocab, '\n');
     } else {
-        vocab.linefeed_id = llama_tokenize_internal(vocab, "\u010A", false)[0];
+        const std::vector<int> ids = llama_tokenize_internal(vocab, "\u010A", false);
+        GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
+        vocab.linefeed_id = ids[0];
     }
 
     // special tokens
-    GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID));
-    GGUF_GET_KEY(ctx, vocab.special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID));
-    GGUF_GET_KEY(ctx, vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID));
-    GGUF_GET_KEY(ctx, vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID));
-    GGUF_GET_KEY(ctx, vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID));
+    {
+        const std::vector<std::pair<enum llm_kv, int32_t &>> special_token_types = {
+            { LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id },
+            { LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id },
+            { LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id },
+            { LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id },
+            { LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id },
+        };
+        for (const auto & it : special_token_types) {
+            const std::string & key = kv(std::get<0>(it));
+            int32_t & id = std::get<1>(it), old_id = id;
+
+            GGUF_GET_KEY(ctx, id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, key);
+            // Must be >= -1 and < vocab size. Since the key is unsigned, -1
+            // can only come from the default value, so there's no point in
+            // validating that.
+            if (size_t(id + 1) > vocab.id_to_token.size()) {
+                LLAMA_LOG_WARN("%s: bad special token: '%s' = %d, using default id %d\n",
+                    __func__, key.c_str(), id, old_id);
+                id = old_id;
+            }
+        }
+    }
 
     // build special tokens cache
     {
@@ -6101,11 +6119,10 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
 }
 
 static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
+    static const char * hex = "0123456789ABCDEF";
     switch (llama_vocab_get_type(vocab)) {
     case LLAMA_VOCAB_TYPE_SPM: {
-        char buf[7];
-        int result = snprintf(buf, sizeof(buf), "<0x%02X>", ch);
-        GGML_ASSERT(0 <= result && result < 7);
+        const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 };
         return vocab.token_to_id.at(buf);
     }
     case LLAMA_VOCAB_TYPE_BPE: {
@@ -7412,37 +7429,15 @@ void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array
     llama_sample_temp(ctx, candidates_p, temp);
 }
 
-void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty) {
-    if (last_tokens_size == 0 || penalty == 1.0f) {
-        return;
-    }
-
-    const int64_t t_start_sample_us = ggml_time_us();
-
-    for (size_t i = 0; i < candidates->size; ++i) {
-        const auto * token_iter = std::find(last_tokens, last_tokens + last_tokens_size, candidates->data[i].id);
-        if (token_iter == last_tokens + last_tokens_size) {
-            continue;
-        }
-
-        // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong.
-        // This is common fix for this problem, which is to multiply by the penalty instead of dividing.
-        if (candidates->data[i].logit <= 0) {
-            candidates->data[i].logit *= penalty;
-        } else {
-            candidates->data[i].logit /= penalty;
-        }
-    }
-
-    candidates->sorted = false;
-
-    if (ctx) {
-        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
-    }
-}
-
-void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens_p, size_t last_tokens_size, float alpha_frequency, float alpha_presence) {
-    if (last_tokens_size == 0 || (alpha_frequency == 0.0f && alpha_presence == 0.0f)) {
+void llama_sample_repetition_penalties(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   penalty_last_n,
+                           float   penalty_repeat,
+                           float   penalty_freq,
+                           float   penalty_present) {
+    if (penalty_last_n == 0 || (penalty_repeat == 1.0f && penalty_freq == 0.0f && penalty_present == 0.0f)) {
         return;
     }
 
@@ -7450,19 +7445,28 @@ void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, l
 
     // Create a frequency map to count occurrences of each token in last_tokens
     std::unordered_map<llama_token, int> token_count;
-    for (size_t i = 0; i < last_tokens_size; ++i) {
-        token_count[last_tokens_p[i]]++;
+    for (size_t i = 0; i < penalty_last_n; ++i) {
+        token_count[last_tokens[i]]++;
     }
 
     // Apply frequency and presence penalties to the candidates
     for (size_t i = 0; i < candidates->size; ++i) {
-        auto token_iter = token_count.find(candidates->data[i].id);
+        const auto token_iter = token_count.find(candidates->data[i].id);
         if (token_iter == token_count.end()) {
             continue;
         }
 
-        int count = token_iter->second;
-        candidates->data[i].logit -= float(count) * alpha_frequency + float(count > 0) * alpha_presence;
+        const int count = token_iter->second;
+
+        // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong.
+        // This is common fix for this problem, which is to multiply by the penalty instead of dividing.
+        if (candidates->data[i].logit <= 0) {
+            candidates->data[i].logit *= penalty_repeat;
+        } else {
+            candidates->data[i].logit /= penalty_repeat;
+        }
+
+        candidates->data[i].logit -= float(count) * penalty_freq + float(count > 0) * penalty_present;
     }
 
     candidates->sorted = false;
@@ -7484,14 +7488,14 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
         }
     }
 
-    const llama_token eos = llama_token_eos(ctx);
+    const llama_token eos = llama_token_eos(&ctx->model);
 
     std::vector<std::pair<std::vector<uint32_t>, llama_partial_utf8>> candidates_decoded;
     std::vector<llama_grammar_candidate>                              candidates_grammar;
 
     for (size_t i = 0; i < candidates->size; ++i) {
         const llama_token id    = candidates->data[i].id;
-        const std::string piece = llama_token_to_str(ctx, id);
+        const std::string piece = llama_token_to_piece(ctx, id);
         if (id == eos) {
             if (!allow_eos) {
                 candidates->data[i].logit = -INFINITY;
@@ -7694,7 +7698,7 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra
 void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token) {
     const int64_t t_start_sample_us = ggml_time_us();
 
-    if (token == llama_token_eos(ctx)) {
+    if (token == llama_token_eos(&ctx->model)) {
         for (const auto & stack : grammar->stacks) {
             if (stack.empty()) {
                 return;
@@ -7703,7 +7707,7 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
         GGML_ASSERT(false);
     }
 
-    const std::string piece = llama_token_to_str(ctx, token);
+    const std::string piece = llama_token_to_piece(ctx, token);
 
     // Note terminating 0 in decoded string
     const auto   decoded     = decode_utf8(piece.c_str(), grammar->partial_utf8);
@@ -8903,7 +8907,7 @@ struct llama_context * llama_new_context_with_model(
             // build worst-case graph
             int n_tokens = (int)std::min(cparams.n_ctx, cparams.n_batch);
             int n_past = cparams.n_ctx - n_tokens;
-            llama_token token = llama_token_bos(ctx); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
+            llama_token token = llama_token_bos(&ctx->model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
             ggml_cgraph * gf = llama_build_graph(*ctx, llama_batch_get_one(&token, n_tokens, n_past, 0));
 
 #ifdef GGML_USE_METAL
@@ -9664,43 +9668,44 @@ float * llama_get_embeddings(struct llama_context * ctx) {
     return ctx->embedding.data();
 }
 
-const char * llama_token_get_text(const struct llama_context * ctx, llama_token token) {
-    return ctx->model.vocab.id_to_token[token].text.c_str();
+const char * llama_token_get_text(const struct llama_model * model, llama_token token) {
+    return model->vocab.id_to_token[token].text.c_str();
 }
 
-float llama_token_get_score(const struct llama_context * ctx, llama_token token) {
-    return ctx->model.vocab.id_to_token[token].score;
+float llama_token_get_score(const struct llama_model * model, llama_token token) {
+    return model->vocab.id_to_token[token].score;
 }
 
-llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token) {
-    return ctx->model.vocab.id_to_token[token].type;
+llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token) {
+    return model->vocab.id_to_token[token].type;
 }
 
-llama_token llama_token_bos(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_bos_id;
+llama_token llama_token_bos(const struct llama_model * model) {
+    return model->vocab.special_bos_id;
 }
 
-llama_token llama_token_eos(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_eos_id;
+llama_token llama_token_eos(const struct llama_model * model) {
+    return model->vocab.special_eos_id;
 }
 
-llama_token llama_token_nl(const struct llama_context * ctx) {
-    return ctx->model.vocab.linefeed_id;
+llama_token llama_token_nl(const struct llama_model * model) {
+    return 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_prefix(const struct llama_model * model) {
+    return 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_middle(const struct llama_model * model) {
+    return 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_suffix(const struct llama_model * model) {
+    return model->vocab.special_suffix_id;
 }
 
-llama_token llama_token_eot(const struct llama_context * ctx) {
-    return ctx->model.vocab.special_eot_id;
+llama_token llama_token_eot(const struct llama_model * model) {
+    return model->vocab.special_eot_id;
 }
 
 int llama_tokenize(

data/ext/llama_cpp/src/llama.h

@@ -494,21 +494,22 @@ extern "C" {
     // Vocab
     //
 
-    LLAMA_API const char * llama_token_get_text(const struct llama_context * ctx, llama_token token);
+    LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token);
 
-    LLAMA_API float llama_token_get_score(const struct llama_context * ctx, llama_token token);
+    LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);
 
-    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token);
+    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);
 
     // Special tokens
-    LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx);  // beginning-of-sentence
-    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
+    LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
+    LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
+    LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
+
     // codellama infill tokens
-    LLAMA_API llama_token llama_token_prefix(const struct llama_context * ctx); // Beginning of infill prefix
-    LLAMA_API llama_token llama_token_middle(const struct llama_context * ctx); // Beginning of infill middle
-    LLAMA_API llama_token llama_token_suffix(const struct llama_context * ctx); // Beginning of infill suffix
-    LLAMA_API llama_token llama_token_eot   (const struct llama_context * ctx); // End of infill middle
+    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
+    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
+    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
+    LLAMA_API llama_token llama_token_eot   (const struct llama_model * model); // End of infill middle
 
     //
     // Tokenization
@@ -560,21 +561,15 @@ extern "C" {
     LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
 
     /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
-    LLAMA_API void llama_sample_repetition_penalty(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates,
-               const llama_token * last_tokens,
-                          size_t   last_tokens_size,
-                          float    penalty);
-
     /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
-    LLAMA_API void llama_sample_frequency_and_presence_penalties(
+    LLAMA_API void llama_sample_repetition_penalties(
             struct llama_context * ctx,
           llama_token_data_array * candidates,
                const llama_token * last_tokens,
-                          size_t   last_tokens_size,
-                           float   alpha_frequency,
-                           float   alpha_presence);
+                          size_t   penalty_last_n,
+                           float   penalty_repeat,
+                           float   penalty_freq,
+                           float   penalty_present);
 
     /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LLaMACpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.8.0'
+  VERSION = '0.9.0'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b1405'
+  LLAMA_CPP_VERSION = 'b1429'
 end

data/lib/llama_cpp.rb

@@ -67,9 +67,9 @@ module LLaMACpp
 
         # apply penalties
         last_n_repeat = [last_n_tokens.size, repeat_last_n, n_ctx].min
-        context.sample_repetition_penalty(candidates, last_n_tokens[-last_n_repeat..], penalty: repeat_penalty)
-        context.sample_frequency_and_presence_penalties(
-          candidates, last_n_tokens[-last_n_repeat..], frequency: frequency, presence: presence
+        context.sample_repetition_penalties(
+          candidates, last_n_tokens[-last_n_repeat..],
+          penalty_repeat: repeat_penalty, penalty_freq: frequency, penalty_present: presence
         )
 
         # temperature sampling
@@ -97,7 +97,7 @@ module LLaMACpp
 
       embd.each { |token| output << context.model.token_to_piece(token) }
 
-      break if !embd.empty? && embd[-1] == context.token_eos
+      break if !embd.empty? && embd[-1] == context.model.token_eos
     end
 
     output.join.scrub('?').strip.delete_prefix(prompt).strip

data/sig/llama_cpp.rbs

@@ -82,6 +82,16 @@ module LLaMACpp
     def desc: () -> String
     def size: () -> Integer
     def n_params: () -> Integer
+    def text: (Integer) -> String
+    def score: (Integer) -> Float
+    def type: (Integer) -> Integer
+    def token_bos: () -> Integer
+    def token_eos: () -> Integer
+    def token_nl: () -> Integer
+    def token_prefix: () -> Integer
+    def token_middle: () -> Integer
+    def token_suffix: () -> Integer
+    def token_eot: () -> Integer
   end
 
   class Timings
@@ -143,16 +153,6 @@ module LLaMACpp
 
     def initialize: (model: ::LLaMACpp::Model, params: ::LLaMACpp::ContextParams) -> void
     def embeddings: () -> Array[Float]
-    def text: (Integer) -> String
-    def score: (Integer) -> Float
-    def type: (Integer) -> Integer
-    def token_bos: () -> Integer
-    def token_eos: () -> Integer
-    def token_nl: () -> Integer
-    def token_prefix: () -> Integer
-    def token_middle: () -> Integer
-    def token_suffix: () -> Integer
-    def token_eot: () -> Integer
     def eval: (tokens: Array[Integer], n_past: Integer, ?n_tokens: Integer) -> void
     def eval_embd: (tokens: Array[Float], n_past: Integer, ?n_tokens: Integer) -> void
     def decode: (::LLaMACpp::Batch) -> void
@@ -170,8 +170,7 @@ module LLaMACpp
     def set_rng_seed: (Integer) -> void
     def load_session_file: (session_path: String) -> void
     def save_session_file: (session_path: String, session_tokens: Array[Integer]) -> void
-    def sample_repetition_penalty: (::LLaMACpp::TokenDataArray, Array[Integer], penalty: Float) -> void
-    def sample_frequency_and_presence_penalties: (::LLaMACpp::TokenDataArray, Array[Integer], frequency: Float, presence: Float) -> void
+    def sample_repetition_penalties: (::LLaMACpp::TokenDataArray, Array[Integer], penalty_repeat: Float, penalty_freq: Float, penalty_present: Float) -> void
     def sample_classifier_free_guidance: (::LLaMACpp::TokenDataArray, guidance: ::LLaMACpp::Context, scale: Float) -> void
     def sample_softmax: (::LLaMACpp::TokenDataArray) -> void
     def sample_top_k: (::LLaMACpp::TokenDataArray, k: Integer, ?min_keep: Integer) -> void

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.8.0
+  version: 0.9.0
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2023-10-21 00:00:00.000000000 Z
+date: 2023-10-28 00:00:00.000000000 Z
 dependencies: []
 description: llama_cpp.rb provides Ruby bindings for the llama.cpp.
 email: