cui-llama.rn 1.4.4 → 1.5.0

package/cpp/ops.h

@@ -0,0 +1,128 @@
+#pragma once
+
+#include "ggml.h"
+
+//
+// cache line
+//
+
+#if defined(__cpp_lib_hardware_interference_size)
+#define CACHE_LINE_SIZE std::hardware_destructive_interference_size
+#else
+#if defined(__POWER9_VECTOR__)
+#define CACHE_LINE_SIZE 128
+#elif defined(__VXE__) || defined(__VXE2__)
+#define CACHE_LINE_SIZE 256
+#else
+#define CACHE_LINE_SIZE 64
+#endif
+#endif
+
+static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void lm_ggml_compute_forward_dup(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_add(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_add1(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_acc(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_sum(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_sum_rows(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_mean(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_argmax(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_count_equal(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_repeat(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_repeat_back(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_concat(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_silu_back(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_norm(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_rms_norm(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_rms_norm_back(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_group_norm(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_l2_norm(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_out_prod(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_scale(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_set(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_cpy(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_cont(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_reshape(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_view(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_permute(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_transpose(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_get_rows(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_get_rows_back(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_diag(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_diag_mask_inf(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_diag_mask_zero(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_soft_max(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_soft_max_ext_back(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_rope(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_rope_back(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_clamp(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_conv_transpose_1d(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_im2col(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_im2col_back_f32(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_conv_transpose_2d(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_pool_1d(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_pool_2d(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_pool_2d_back(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_upscale(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_pad(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_pad_reflect_1d(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_arange(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_timestep_embedding(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_argsort(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_leaky_relu(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_flash_attn_ext(
+    const struct lm_ggml_compute_params * params,
+    const struct lm_ggml_tensor * q,
+    const struct lm_ggml_tensor * k,
+    const struct lm_ggml_tensor * v,
+    const struct lm_ggml_tensor * mask,
+    struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_flash_attn_back(
+        const struct lm_ggml_compute_params * params,
+        const bool masked,
+        struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_ssm_conv(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_ssm_scan(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_win_part(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_win_unpart(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_unary(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_get_rel_pos(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_add_rel_pos(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_rwkv_wkv6(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_rwkv_wkv7(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_gla(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_map_unary(
+    const struct lm_ggml_compute_params * params,
+    struct lm_ggml_tensor * dst,
+    const lm_ggml_unary_op_f32_t fun);
+void lm_ggml_compute_forward_map_binary(
+    const struct lm_ggml_compute_params * params,
+    struct lm_ggml_tensor * dst,
+    const lm_ggml_binary_op_f32_t fun);
+void lm_ggml_compute_forward_map_custom1_f32(
+    const struct lm_ggml_compute_params * params,
+    struct lm_ggml_tensor * dst,
+    const lm_ggml_custom1_op_f32_t fun);
+void lm_ggml_compute_forward_map_custom2_f32(
+    const struct lm_ggml_compute_params * params,
+    struct lm_ggml_tensor * dst,
+    const lm_ggml_custom2_op_f32_t fun);
+void lm_ggml_compute_forward_map_custom3_f32(
+    const struct lm_ggml_compute_params * params,
+    struct lm_ggml_tensor * dst,
+    const lm_ggml_custom3_op_f32_t fun);
+void lm_ggml_compute_forward_map_custom1(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_map_custom2(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_map_custom3(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_cross_entropy_loss(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_cross_entropy_loss_back(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+void lm_ggml_compute_forward_opt_step_adamw(const struct lm_ggml_compute_params * params, struct lm_ggml_tensor * dst);
+
+#ifdef __cplusplus
+}
+#endif

package/cpp/rn-llama.cpp

@@ -191,10 +191,10 @@ bool llama_rn_context::loadModel(common_params &params_)
     ctx = llama_init.context.get();
     if (model == nullptr)
     {
-        LOG_ERROR("unable to load model: %s", params_.model.c_str());
+        LOG_ERROR("unable to load model: %s", params_.model.path.c_str());
         return false;
     }
-    templates = common_chat_templates_from_model(model, params.chat_template);
+    templates = common_chat_templates_init(model, params.chat_template);
     n_ctx = llama_n_ctx(ctx);
 
     // We can uncomment for debugging or after this fix: https://github.com/ggerganov/llama.cpp/pull/11101
@@ -219,71 +219,46 @@ common_chat_params llama_rn_context::getFormattedChatWithJinja(
   const bool &parallel_tool_calls,
   const std::string &tool_choice
 ) const {
-  common_chat_inputs inputs;
-  inputs.messages = json::parse(messages);
-  auto useTools = !tools.empty();
-  if (useTools) {
-      inputs.tools = json::parse(tools);
-  }
-  inputs.parallel_tool_calls = parallel_tool_calls;
-  if (!tool_choice.empty()) {
-      inputs.tool_choice = tool_choice;
-  }
-  if (!json_schema.empty()) {
-      inputs.json_schema = json::parse(json_schema);
-  }
-  inputs.extract_reasoning = params.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-  inputs.stream = true;
-
-  // If chat_template is provided, create new one and use it (probably slow)
-  if (!chat_template.empty()) {
-      auto tmp = common_chat_templates_from_model(model, chat_template);
-      const common_chat_template* template_ptr = useTools && tmp.template_tool_use ? tmp.template_tool_use.get() : tmp.template_default.get();
-      if (inputs.parallel_tool_calls && !template_ptr->original_caps().supports_parallel_tool_calls) {
-          inputs.parallel_tool_calls = false;
-      }
-      return common_chat_params_init(*template_ptr, inputs);
-  } else {
-      const common_chat_template* template_ptr = useTools && templates.template_tool_use ? templates.template_tool_use.get() : templates.template_default.get();
-      if (inputs.parallel_tool_calls && !template_ptr->original_caps().supports_parallel_tool_calls) {
-          inputs.parallel_tool_calls = false;
-      }
-      return common_chat_params_init(*template_ptr, inputs);
-  }
+    common_chat_templates_inputs inputs;
+    inputs.use_jinja = true;
+    inputs.messages = common_chat_msgs_parse_oaicompat(json::parse(messages));
+    auto useTools = !tools.empty();
+    if (useTools) {
+        inputs.tools = common_chat_tools_parse_oaicompat(json::parse(tools));
+    }
+    inputs.parallel_tool_calls = parallel_tool_calls;
+    if (!tool_choice.empty()) {
+        inputs.tool_choice = common_chat_tool_choice_parse_oaicompat(tool_choice);
+    }
+    if (!json_schema.empty()) {
+        inputs.json_schema = json::parse(json_schema);
+    }
+    inputs.extract_reasoning = params.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+
+    // If chat_template is provided, create new one and use it (probably slow)
+    if (!chat_template.empty()) {
+        auto tmps = common_chat_templates_init(model, chat_template);
+        return common_chat_templates_apply(tmps.get(), inputs);
+    } else {
+        return common_chat_templates_apply(templates.get(), inputs);
+    }
 }
 
 std::string llama_rn_context::getFormattedChat(
   const std::string &messages,
   const std::string &chat_template
 ) const {
-  auto chat_json = json::parse(messages);
-
-  // Handle regular chat without tools
-  std::vector<common_chat_msg> chat_msgs;
-  for (const auto &msg : chat_json) {
-      chat_msgs.push_back({
-          msg["role"].get<std::string>(),
-          msg["content"].get<std::string>()
-      });
-  }
-
-  // If chat_template is provided, create new one and use it (probably slow)
-  if (!chat_template.empty()) {
-      auto tmp = common_chat_templates_from_model(model, chat_template);
-      return common_chat_apply_template(
-          *tmp.template_default,
-          chat_msgs,
-          true,
-          false
-      );
-  } else {
-      return common_chat_apply_template(
-          *templates.template_default,
-          chat_msgs,
-          true,
-          false
-      );
-  }
+    common_chat_templates_inputs inputs;
+    inputs.messages = common_chat_msgs_parse_oaicompat(json::parse(messages));
+    inputs.use_jinja = false;
+
+    // If chat_template is provided, create new one and use it (probably slow)
+    if (!chat_template.empty()) {
+        auto tmps = common_chat_templates_init(model, chat_template);
+        return common_chat_templates_apply(tmps.get(), inputs).prompt;
+    } else {
+        return common_chat_templates_apply(templates.get(), inputs).prompt;
+    }
 }
 
 void llama_rn_context::truncatePrompt(std::vector<llama_token> &prompt_tokens) {
@@ -358,7 +333,7 @@ void llama_rn_context::loadPrompt() {
     }
 
     // since #3228 we now have to manually manage the KV cache
-    llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+    llama_kv_self_seq_rm(ctx, 0, n_past, -1);
 
     LOG_VERBOSE("prompt ingested, n_past: %d, cached: %s, to_eval: %s",
         n_past,
@@ -388,8 +363,8 @@ completion_token_output llama_rn_context::nextToken()
         const int n_left    = n_past - params.n_keep - 1;
         const int n_discard = n_left/2;
 
-        llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-        llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+        llama_kv_self_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+        llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
         for (size_t i = params.n_keep + 1 + n_discard; i < embd.size(); i++)
         {
@@ -595,7 +570,6 @@ std::vector<float> llama_rn_context::getEmbedding(common_params &embd_params)
     float *data;
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-    printf("pooling_type: %d\n", pooling_type);
     if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
         data = llama_get_embeddings(ctx);
     } else {
@@ -644,7 +618,7 @@ std::string llama_rn_context::bench(int pp, int tg, int pl, int nr)
         }
         batch.logits[batch.n_tokens - 1] = 1; // true
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         const int64_t t_pp_start = llama_time_us();
         if (llama_decode(ctx, batch) != 0)
@@ -652,7 +626,7 @@ std::string llama_rn_context::bench(int pp, int tg, int pl, int nr)
             LOG_ERROR("llama_decode() failed during prompt", "");
         }
         const int64_t t_pp_end = llama_time_us();
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         if (is_interrupted) break;
 
@@ -676,7 +650,7 @@ std::string llama_rn_context::bench(int pp, int tg, int pl, int nr)
 
         const int64_t t_tg_end = llama_time_us();
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         const double t_pp = (t_pp_end - t_pp_start) / 1000000.0;
         const double t_tg = (t_tg_end - t_tg_start) / 1000000.0;
@@ -702,7 +676,7 @@ std::string llama_rn_context::bench(int pp, int tg, int pl, int nr)
         tg_std = 0;
     }
 
-    if (is_interrupted) llama_kv_cache_clear(ctx);
+    if (is_interrupted) llama_kv_self_clear(ctx);
     is_predicting = false;
 
     char model_desc[128];
@@ -880,8 +854,8 @@ void llama_rn_context::purge_missing_tokens(llama_context * ctx, std::vector<int
 
             //extract the unwanted tokens out from context and KV
             int diff = found - trimstart;
-            llama_kv_cache_seq_rm(ctx, 0, trimstart, trimstart + diff);
-            llama_kv_cache_seq_add(ctx, 0, trimstart + diff, -1, -diff);
+            llama_kv_self_seq_rm(ctx, 0, trimstart, trimstart + diff);
+            llama_kv_self_seq_add(ctx, 0, trimstart + diff, -1, -diff);
 
             for (size_t i = trimstart + diff; i < current_context_tokens.size() - 1; i++)
             {

package/cpp/rn-llama.h

@@ -3,8 +3,7 @@
 
 #include <sstream>
 #include <iostream>
-#include "chat.hpp"
-#include "chat-template.hpp"
+#include "chat.h"
 #include "common.h"
 #include "ggml.h"
 #include "gguf.h"
@@ -17,6 +16,7 @@
 
 namespace rnllama {
 
+
 std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token);
 
 std::string tokens_to_str(llama_context *ctx, const std::vector<llama_token>::const_iterator begin, const std::vector<llama_token>::const_iterator end);
@@ -65,7 +65,7 @@ struct llama_rn_context {
 
     llama_context *ctx = nullptr;
     common_sampler *ctx_sampling = nullptr;
-    common_chat_templates templates;
+    common_chat_templates_ptr templates;
 
     int n_ctx;