[custom] Add custom filter to use llama2c

This patch adds custom filter to use llama2c.

example pipeline
```sh
$ gst-launch-1.0 filesrc location=input.txt ! text/x-raw ! tensor_converter input-dim=3584:1:1:1 ! tensor_filter framework=custom model=libnnstreamer_customfilter_llama2.so ! filesink location=output.txt
```

model.bin and tokenizer.bin is needed.

Signed-off-by: Yelin Jeong <[email protected]>

tests/nnstreamer_example/custom_example_llama2/nnstreamer_customfilter_llama2.c

@@ -0,0 +1,177 @@
+/**
+ * NNStreamer Custom Filter Example llama2.c
+ * Copyright (C) 2024 Yelin Jeong <[email protected]>
+ *
+ * SPDX-License-Identifier: LGPL-2.1-only
+ *
+ * @file	nnstreamer_customfilter_llama2.c
+ * @date	06 Sep 2024
+ * @brief	Custom NNStreamer Filter Example llama2.c
+ * @author	Yelin Jeong <[email protected]>
+ * @bug		No known bugs except for NYI items
+ * @see     https://github.com/nnsuite/llama2.c
+ * @todo    Use property to values like temperatrue, topp
+ */
+
+
+#include <nnstreamer_plugin_api.h>
+#include <nnstreamer_util.h>
+#include <tensor_filter_custom.h>
+#include <assert.h>
+#include "api.h"
+
+#define CHECKPOINT_PATH "model.bin"
+#define TOKENIZER_PATH "tokenizer.bin"
+
+#define TEMPERATURE \
+  1.0f // 0.0 = greedy deterministic. 1.0 = original. don't set higher
+#define TOPP \
+  0.9f // top-p in nucleus sampling. 1.0 = off. 0.9 works well, but slower
+#define RNG_SEED 0ULL // seed rng with time by default
+
+/**
+ * @brief _pt_data Internal data structure
+ */
+typedef struct _pt_data {
+  Transformer *transformer;
+  Tokenizer *tokenizer;
+  Sampler *sampler;
+  char *prompt; // prompt string
+  int steps; // number of steps to run for
+
+  GstTensorsInfo info;
+} pt_data;
+
+/**
+ * @brief init callback of tensor_filter custom
+ */
+static void *
+pt_init (const GstTensorFilterProperties *prop)
+{
+  pt_data *data = (pt_data *) malloc (sizeof (pt_data));
+  Config *config;
+  UNUSED (prop);
+
+  gst_tensors_info_init (&data->info);
+
+  data->transformer = (Transformer *) malloc (sizeof (Transformer));
+
+  // build the Transformer via the model .bin file
+  build_transformer (data->transformer, (char *) CHECKPOINT_PATH);
+  config = &data->transformer->config;
+
+  if (data->steps == 0 || data->steps > config->seq_len)
+    data->steps = config->seq_len; // override to ~max length
+
+  data->tokenizer = (Tokenizer *) malloc (sizeof (Tokenizer));
+
+  // build the Tokenizer via the tokenizer .bin file
+  build_tokenizer (data->tokenizer, (char *) TOKENIZER_PATH, config->vocab_size);
+
+  data->sampler = (Sampler *) malloc (sizeof (Sampler));
+
+  // build the Sampler
+  build_sampler (data->sampler, config->vocab_size, TEMPERATURE, TOPP, RNG_SEED);
+
+  data->info.num_tensors = 1;
+  data->info.info[0].type = _NNS_UINT8;
+  data->info.info[0].dimension[0] = data->tokenizer->max_token_length * data->steps;
+
+  return data;
+}
+
+/**
+ * @brief exit callback of tensor_filter custom
+ */
+static void
+pt_exit (void *private_data, const GstTensorFilterProperties *prop)
+{
+  pt_data *data = private_data;
+  UNUSED (prop);
+  assert (private_data);
+
+  gst_tensors_info_free (&data->info);
+  free_transformer (data->transformer);
+  free_tokenizer (data->tokenizer);
+  free_sampler (data->sampler);
+
+  free (data->transformer);
+  free (data->tokenizer);
+  free (data->sampler);
+  free (data);
+}
+
+/**
+ * @brief setInputDimension callback of tensor_filter custom
+ */
+static int
+get_inputDim (void *private_data, const GstTensorFilterProperties *prop, GstTensorsInfo *info)
+{
+  UNUSED (private_data);
+  UNUSED (prop);
+  assert (info);
+
+  info->format = _NNS_TENSOR_FORMAT_SPARSE;
+  info->num_tensors = 1;
+
+  return 0;
+}
+
+/**
+ * @brief getOutputDimension callback of tensor_filter custom
+ */
+static int
+get_outputDim (void *private_data, const GstTensorFilterProperties *prop, GstTensorsInfo *info)
+{
+  pt_data *data = private_data;
+  UNUSED (prop);
+  assert (private_data);
+
+  gst_tensors_info_copy (info, &data->info);
+  return 0;
+}
+
+/**
+ * @brief invoke callback of tensor_filter custom
+ */
+static int
+pt_invoke (void *private_data, const GstTensorFilterProperties *prop,
+    const GstTensorMemory *input, GstTensorMemory *output)
+{
+  pt_data *data = private_data;
+  char *result, *prompt;
+  int len;
+
+  UNUSED (prop);
+  assert (private_data);
+  assert (input);
+  assert (output);
+
+  prompt = (char *) input[0].data;
+  len = strlen (prompt);
+
+  if (len > 1) {
+    prompt[strlen (prompt) - 1] = ' '; // Remove new line from text file
+  } else {
+    prompt = NULL;
+  }
+
+  output->size = data->tokenizer->max_token_length * data->steps;
+
+  result = get_tokens (data->transformer, data->tokenizer, data->sampler, prompt, data->steps);
+  memcpy (output[0].data, result, strlen(result));
+
+  free (result);
+  return 0;
+}
+
+static NNStreamer_custom_class NNStreamer_custom_body = {
+  .initfunc = pt_init,
+  .exitfunc = pt_exit,
+  .getInputDim = get_inputDim,
+  .getOutputDim = get_outputDim,
+  .invoke = pt_invoke,
+};
+
+/* The dyn-loaded object */
+NNStreamer_custom_class *NNStreamer_custom = &NNStreamer_custom_body;

tests/nnstreamer_example/meson.build

@@ -103,3 +103,13 @@ if tensorrt_support_is_available
     install_dir: customfilter_install_dir
   )
 endif
+
+if llama2c_support_is_available
+  library('nnstreamer_customfilter_llama2',
+    join_paths('custom_example_llama2',
+        'nnstreamer_customfilter_llama2.c'),
+    dependencies: [glib_dep, gst_dep, nnstreamer_dep, llama2c_support_deps],
+    install: get_option('install-test'),
+    install_dir: customfilter_install_dir
+  )
+endif