chat.cpp

//===----------------------------------------------------------------------===//
//
// Copyright (C) 2023 Sophgo Technologies Inc.  All rights reserved.
//
// TPU-MLIR is licensed under the 2-Clause BSD License except for the
// third-party components.
//
//===----------------------------------------------------------------------===//

#include "bmruntime_interface.h"
#include "memory.h"
#include <algorithm>
#include <assert.h>
#include <chrono>
#include <cstdlib>
#include <getopt.h>
#include <inttypes.h>
#include <iostream>
#include <numeric>
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
#include <random>
#include <stdio.h>
#include <vector>
#include "utils.h"

static const uint16_t ATTENTION_MASK = 0xF0E2;

class ChatGLM {
public:
  void init(const std::vector<int> &devid, std::string model_path);
  void deinit();
  int forward_first(std::vector<int> &tokens);
  int forward_next();
  std::vector<int> generate(std::vector<int> &history_tokens, int EOS);

  std::mt19937 sgen;
  ChatGLM() : sgen(std::random_device()()){};

private:
  void move2end(const bm_device_mem_t &kv);
  void net_launch(const std::string &net_name,
                  std::vector<bm_tensor_t> &inputs,
                  std::vector<bm_tensor_t> &outputs,
                  int stage_idx = 0);

  void head_launch(const bm_net_info_t *net, bm_device_mem_t &logits_mem);
  int greedy_search(const bm_net_info_t *net, bm_device_mem_t &logits_mem);
  int penalty_sample(const bm_net_info_t *net, bm_device_mem_t &logits_mem);

public:
  int device_num;
  int token_length;
  int SEQLEN;     // read from bmodel
  int NUM_LAYERS; // read from bmodel
  bool io_alone;
  std::vector<int> visited_tokens;
  uint16_t mask_value;

  // generation
  float temperature;
  float top_p;
  float repeat_penalty;
  int repeat_last_n;
  int max_new_tokens;
  std::string generation_mode;
  std::string prompt_mode;

private:
  std::vector<bm_handle_t> handles;
  bm_handle_t bm_handle;
  void *p_bmrt;
  std::string name_embed;
  std::string name_embed_cache;
  std::string name_lm;
  std::vector<std::string> name_blocks;
  std::vector<std::string> name_blocks_cache;
  const bm_net_info_t *net_embed;
  const bm_net_info_t *net_embed_cache;
  const bm_net_info_t *net_lm;
  std::vector<const bm_net_info_t *> net_blocks;
  std::vector<const bm_net_info_t *> net_blocks_cache;

  std::vector<bm_tensor_t> inputs_embed, inputs_embed_cache;
  std::vector<bm_tensor_t> hidden_states, hidden_states_cache;
  std::vector<bm_tensor_t> inputs_pid, next_pid;
  std::vector<bm_tensor_t> inputs_attention, next_attention;
  std::vector<std::vector<bm_tensor_t>> past_keys, past_values;
  std::vector<bm_tensor_t> present_key_cache, present_value_cache;
  std::vector<bm_tensor_t> inputs_lm, outputs_lm;
};

void ChatGLM::net_launch(const std::string &net_name,
                std::vector<bm_tensor_t> &inputs,
                std::vector<bm_tensor_t> &outputs,
                int stage_idx) {
  bool ret = bmrt_launch_tensor_ex(
    p_bmrt, net_name.c_str(), inputs.data(), inputs.size(), outputs.data(),
    outputs.size(), true, false);
  assert(ret);
  bm_thread_sync(bm_handle);
}

void ChatGLM::init(const std::vector<int> &devices, std::string model_path) {

  // request bm_handle
  device_num = devices.size();
  std::cout << "Device [ ";
  for (auto d : devices) {
    std::cout << d << " ";
  }
  std::cout << "] loading ....\n";
  for (auto d : devices) {
    bm_handle_t h;
    bm_status_t status = bm_dev_request(&h, d);
    assert(BM_SUCCESS == status);
    handles.push_back(h);
  }
  bm_handle = handles[0];

  // create bmruntime
  p_bmrt = bmrt_create_ex(handles.data(), handles.size());
  assert(NULL != p_bmrt);

  // load bmodel by file
  printf("Model[%s] loading ....\n", model_path.c_str());
  bool ret = bmrt_load_bmodel(p_bmrt, model_path.c_str());
  assert(true == ret);
  printf("Done!\n");

  // embed, lm_head
  name_embed = "embedding";
  name_embed_cache = "embedding_cache";
  name_lm = "lm_head";
  net_embed = bmrt_get_network_info(p_bmrt, name_embed.c_str());
  net_embed_cache = bmrt_get_network_info(p_bmrt, name_embed_cache.c_str());
  net_lm = bmrt_get_network_info(p_bmrt, name_lm.c_str());
  SEQLEN = net_embed->stages[0].input_shapes[0].dims[1]; // real seqlen
  auto num_nets = bmrt_get_network_number(p_bmrt);
  NUM_LAYERS = (num_nets - 3) / 2;

  // resize
  visited_tokens.resize(SEQLEN);

  // blocks
  name_blocks.resize(NUM_LAYERS);
  name_blocks_cache.resize(NUM_LAYERS);
  net_blocks.resize(NUM_LAYERS);
  net_blocks_cache.resize(NUM_LAYERS);
  for (int i = 0; i < NUM_LAYERS; i++) {
    name_blocks[i] = "block_" + std::to_string(i);
    name_blocks_cache[i] = "block_cache_" + std::to_string(i);
    net_blocks[i] = bmrt_get_network_info(p_bmrt, name_blocks[i].c_str());
    net_blocks_cache[i] =
        bmrt_get_network_info(p_bmrt, name_blocks_cache[i].c_str());
  }

  // net device mem
  hidden_states.resize(device_num);
  hidden_states_cache.resize(device_num);
  inputs_embed.resize(device_num);
  inputs_embed_cache.resize(device_num);
  int out_num = net_blocks[0]->output_num / device_num;
  int out_num_cache = net_blocks_cache[0]->output_num / device_num;
  for (int i = 0; i < device_num; ++i) {
    bmrt_tensor_with_device(
      &hidden_states[i],
      net_blocks[0]->stages[0].output_mems[0 + i * out_num],
      net_blocks[0]->output_dtypes[0 + i * out_num],
      net_blocks[0]->stages[0].output_shapes[0 + out_num]);
    bmrt_tensor_with_device(
      &hidden_states_cache[i],
      net_blocks_cache[0]->stages[0].output_mems[0 + i * out_num_cache],
      net_blocks_cache[0]->output_dtypes[0 + i * out_num_cache],
      net_blocks_cache[0]->stages[0].output_shapes[0 + out_num_cache]);
    bmrt_tensor_with_device(
      &inputs_embed[i],
      net_embed->stages[0].input_mems[i],
      net_embed->input_dtypes[i],
      net_embed->stages[0].input_shapes[i]);
    bmrt_tensor_with_device(
      &inputs_embed_cache[i],
      net_embed_cache->stages[0].input_mems[i],
      net_embed_cache->input_dtypes[i],
      net_embed_cache->stages[0].input_shapes[i]);
  }

  inputs_pid.resize(device_num);
  inputs_attention.resize(device_num);
  next_pid.resize(device_num);
  next_attention.resize(device_num);
  int in_num = net_blocks[0]->input_num / device_num;
  int in_num_cache = net_blocks_cache[0]->input_num / device_num;
  for (int i = 0; i < device_num; ++i) {
    ret = bmrt_tensor_ex(
      &inputs_pid[i], p_bmrt,
      net_blocks[0]->input_loc_devices[1 + i * in_num],
      net_blocks[0]->input_dtypes[1 + i * in_num],
      net_blocks[0]->stages[0].input_shapes[1 + i * in_num]);
    assert(true == ret);

    ret = bmrt_tensor_ex(
      &inputs_attention[i], p_bmrt,
      net_blocks[0]->input_loc_devices[2 + i * in_num],
      net_blocks[0]->input_dtypes[2 + i * in_num],
      net_blocks[0]->stages[0].input_shapes[2 + i * in_num]);
    assert(true == ret);

    ret = bmrt_tensor_ex(
      &next_pid[i], p_bmrt,
      net_blocks_cache[0]->input_loc_devices[1 + i * in_num_cache],
      net_blocks_cache[0]->input_dtypes[1 + i * in_num_cache],
      net_blocks_cache[0]->stages[0].input_shapes[1 + i * in_num_cache]);
    assert(true == ret);

    ret = bmrt_tensor_ex(
      &next_attention[i], p_bmrt,
      net_blocks_cache[0]->input_loc_devices[2 + i * in_num_cache],
      net_blocks_cache[0]->input_dtypes[2 + i * in_num_cache],
      net_blocks_cache[0]->stages[0].input_shapes[2 + i * in_num_cache]);
    assert(true == ret);
  }

  auto addr_mode = net_blocks_cache[0]->addr_mode;
  io_alone = (addr_mode == 1);
  past_keys.resize(NUM_LAYERS);
  past_values.resize(NUM_LAYERS);
  if (io_alone) {
    for (int i = 0; i < NUM_LAYERS; i++) {
      past_keys[i].resize(device_num);
      past_values[i].resize(device_num);
      auto &net = net_blocks_cache[i];
      for (int j = 0; j < device_num; j++) {
        bmrt_tensor_with_device(
          &past_keys[i][j],
          net->stages[0].input_mems[3 + j * in_num_cache],
          net->input_dtypes[3 + j * in_num_cache],
          net->stages[0].input_shapes[3 + in_num_cache]);
        bmrt_tensor_with_device(
          &past_values[i][j],
          net->stages[0].input_mems[4 + j * in_num_cache],
          net->input_dtypes[4 + j * in_num_cache],
          net->stages[0].input_shapes[4 + in_num_cache]);
      }
    }
  } else {
    for (int i = 0; i < NUM_LAYERS; i++) {
      past_keys[i].resize(device_num);
      past_values[i].resize(device_num);
      auto &net = net_blocks_cache[i];
      for (int j = 0; j < device_num; j++) {
        ret = bmrt_tensor_ex(
          &past_keys[i][j], p_bmrt,
          net->input_loc_devices[3 + j * in_num_cache],
          net->input_dtypes[3 + j * in_num_cache],
          net->stages[0].input_shapes[3 + j * in_num_cache]);
        assert(true == ret);
        ret = bmrt_tensor_ex(
          &past_values[i][j], p_bmrt,
          net->input_loc_devices[4 + j * in_num_cache],
          net->input_dtypes[4 + j * in_num_cache],
          net->stages[0].input_shapes[4 + j *in_num_cache]);
        assert(true == ret);
      }
    }
  }

  present_key_cache.resize(device_num);
  present_value_cache.resize(device_num);
  inputs_lm.resize(device_num);
  outputs_lm.resize(device_num);
  for (int i = 0; i < device_num; ++i) {
    present_key_cache[i] = past_keys[0][i];
    present_value_cache[i] = past_values[0][i];
    present_key_cache[i].shape.dims[0] = 1;
    present_value_cache[i].shape.dims[0] = 1;

    ret = bmrt_tensor_ex(&inputs_lm[i], p_bmrt, i, net_lm->input_dtypes[0],
                        net_lm->stages[0].input_shapes[0]);
    assert(true == ret);
    ret = bmrt_tensor_ex(&outputs_lm[i], p_bmrt, i, net_lm->output_dtypes[0],
                        net_lm->stages[0].output_shapes[0]);
    assert(true == ret);
  }
}

void ChatGLM::deinit() {
  for (int i = 0; i < device_num; ++i) {
    bm_free_device(handles[i], inputs_pid[i].device_mem);
    bm_free_device(handles[i], next_pid[i].device_mem);
    bm_free_device(handles[i], inputs_attention[i].device_mem);
    bm_free_device(handles[i], next_attention[i].device_mem);
    bm_free_device(handles[i], inputs_lm[i].device_mem);
    bm_free_device(handles[i], outputs_lm[i].device_mem);
  }
  if (!io_alone) {
    for (int i = 0; i < NUM_LAYERS; i++) {
      for (int j = 0; j < device_num; j++) {
        bm_free_device(handles[j], past_keys[i][j].device_mem);
        bm_free_device(handles[j], past_values[i][j].device_mem);
      }
    }
  }
  bmrt_destroy(p_bmrt);
  for (auto h : handles) {
    bm_dev_free(h);
  }
}

int ChatGLM::forward_first(std::vector<int> &tokens) {
  std::vector<int> position_id(SEQLEN, 0);
  std::vector<uint16_t> attention_mask(SEQLEN * SEQLEN, 0);
  std::copy(tokens.begin(), tokens.end(), visited_tokens.data());

  token_length = tokens.size();

  for (int i = 0; i < token_length; i++) {
    position_id[i] = i;
  }
  for (int i = 0; i < SEQLEN; i++) {
    for (int j = 0; j < SEQLEN; j++) {
      if (j <= i && i < token_length) {
      } else {
        attention_mask[i * SEQLEN + j] = ATTENTION_MASK;
      }
    }
  }

  // forward embeding
  std::vector<int> input_nums(device_num, 1);
  std::vector<void*> datas(device_num, (void*)visited_tokens.data());
  bmrt_memcpy_s2d_parallel(p_bmrt, inputs_embed.data(), datas.data(),
                          input_nums.data(), device_num);
  auto ret = bmrt_launch_tensor_ex(p_bmrt, name_embed.c_str(),
                                  inputs_embed.data(), inputs_embed.size(),
                                  hidden_states.data(), hidden_states.size(),
                                  true, false);
  assert(ret);
  bm_thread_sync(bm_handle);

  // forward blocks
  std::vector<void*> pos_id_datas(device_num, (void*)position_id.data());
  std::vector<void*> in_attn_datas(device_num, (void*)attention_mask.data());
  bmrt_memcpy_s2d_parallel(p_bmrt, inputs_pid.data(), pos_id_datas.data(),
                          input_nums.data(), device_num);
  bmrt_memcpy_s2d_parallel(p_bmrt, inputs_attention.data(),in_attn_datas.data(),
                          input_nums.data(), device_num);
  auto tmp_hidden_states = hidden_states;
  std::vector<bm_tensor_t> inputs_block;
  std::vector<bm_tensor_t> outputs_block;
  for (int i = 0; i < device_num; ++i) {
    tmp_hidden_states[i].shape = net_blocks[0]->stages[0].input_shapes[0];
    inputs_block.push_back(tmp_hidden_states[i]);
    inputs_block.push_back(inputs_pid[i]);
    inputs_block.push_back(inputs_attention[i]);
    outputs_block.push_back(tmp_hidden_states[i]);
    outputs_block.push_back(past_keys[0][i]);
    outputs_block.push_back(past_values[0][i]);
  }

  for (int i = 0; i < NUM_LAYERS; i++) {
    for (int j = 0; j < device_num; ++j) {
      outputs_block[1 + j * 3] = past_keys[i][j];
      outputs_block[2 + j * 3] = past_values[i][j];
    }
    net_launch(name_blocks[i], inputs_block, outputs_block);
  }

  // forward lmhead
  int bytes = hidden_states[0].device_mem.size / SEQLEN;
  bm_memcpy_d2d_byte(bm_handle, inputs_lm[0].device_mem, 0,
                     hidden_states[0].device_mem, (token_length - 1) * bytes,
                     bytes);
  ret = bmrt_launch_tensor_ex(p_bmrt, name_lm.c_str(), &inputs_lm[0], 1,
                              &outputs_lm[0], 1,
                              true, false);
  assert(ret);
  bm_thread_sync(bm_handle);

  int token = 0;
  bm_memcpy_d2s(bm_handle, (void *)&token, outputs_lm[0].device_mem);
  visited_tokens[token_length] = token;
  return token;
}

int ChatGLM::forward_next() {
  token_length += 1;
  int cur_token = visited_tokens[token_length - 1];

  std::vector<uint16_t> attention_mask(SEQLEN + 1, 0);
  for (int i = token_length - 1; i < SEQLEN; i++) {
    attention_mask[i] = ATTENTION_MASK;
  }
  int32_t position_id = token_length - 1;

  // embedding
  std::vector<void*> input_datas;
  std::vector<int> input_nums(device_num, 1);
  for (int i = 0; i < device_num; ++i) {
    input_datas.push_back((void*)(&cur_token));
  }
  bmrt_memcpy_s2d_parallel(p_bmrt, inputs_embed_cache.data(), input_datas.data(),
                          input_nums.data(), device_num);
  auto ret = bmrt_launch_tensor_ex(
    p_bmrt, name_embed_cache.c_str(), inputs_embed_cache.data(),
    inputs_embed_cache.size(), hidden_states_cache.data(),
    hidden_states_cache.size(), true, false);
  assert(ret);
  bm_thread_sync(bm_handle);

  // blocks
  std::vector<void*> attn_datas(device_num, attention_mask.data());
  std::vector<void*> pid_datas(device_num, &position_id);
  bmrt_memcpy_s2d_parallel(p_bmrt, next_attention.data(), attn_datas.data(),
                          input_nums.data(), device_num);
  bmrt_memcpy_s2d_parallel(p_bmrt, next_pid.data(), pid_datas.data(),
                          input_nums.data(), device_num);
  // WARNING: make inputs_lm device_num
  std::vector<bm_tensor_t> embed_1 = hidden_states_cache;
  for (int i = 0; i < device_num; ++i) {
    embed_1[i].shape = net_blocks_cache[0]->stages[0].input_shapes[0];
  }
  std::vector<bm_tensor_t> inputs_block;
  std::vector<bm_tensor_t> outputs_block;
  for (int i = 0; i < device_num; ++i) {
    inputs_block.push_back(embed_1[i]);
    inputs_block.push_back(next_pid[i]);
    inputs_block.push_back(next_attention[i]);
    inputs_block.push_back(past_keys[0][i]);
    inputs_block.push_back(past_values[0][i]);
    outputs_block.push_back(embed_1[i]);
    outputs_block.push_back(present_key_cache[i]);
    outputs_block.push_back(present_value_cache[i]);
  }

  for (int i = 0; i < NUM_LAYERS; i++) {
    for (int j = 0; j < device_num; ++j) {
      inputs_block[3 + j * 5] = past_keys[i][j];
      inputs_block[4 + j * 5] = past_values[i][j];
      int bytes = bm_mem_get_device_size(past_keys[0][j].device_mem) / SEQLEN;
      int token_offset = (token_length - 1) * bytes;
      bm_set_device_mem(&outputs_block[1 + j * 3].device_mem, bytes,
          bm_mem_get_device_addr(past_keys[i][j].device_mem) + token_offset);
      bm_set_device_mem(&outputs_block[2 + j * 3].device_mem, bytes,
          bm_mem_get_device_addr(past_values[i][j].device_mem) + token_offset);
    }
    net_launch(name_blocks_cache[i], inputs_block, outputs_block);
  }

  // forward lmhead
  ret = bmrt_launch_tensor_ex(
    p_bmrt, name_lm.c_str(), &hidden_states_cache[0], 1, &outputs_lm[0], 1,
    true, false);
  assert(ret);
  bm_thread_sync(bm_handle);

  int token = 0;
  bm_memcpy_d2s(bm_handle, (void *)&token, outputs_lm[0].device_mem);
  visited_tokens[token_length] = token;
  return token;
}

std::vector<int> ChatGLM::generate(std::vector<int> &history_tokens, int EOS) {
  if (history_tokens.empty()) {
    printf("Sorry: your question is empty!!\n");
    history_tokens.clear();
    return {};
  }

  // make sure token not too large
  if ((int)history_tokens.size() > SEQLEN - 10) {
    history_tokens.clear();
    printf("Error: your question is too large!\n");
    return {};
  }

  std::vector<int> result_tokens;
  int token = forward_first(history_tokens);
  while (token != EOS && token_length < SEQLEN) {
    result_tokens.emplace_back(token);
    token = forward_next();
  }

  return result_tokens;
}


PYBIND11_MODULE(chat, m) {
  pybind11::class_<ChatGLM>(m, "ChatGLM")
      .def(pybind11::init<>())
      .def("init", &ChatGLM::init)
      .def("forward_first", &ChatGLM::forward_first)
      .def("forward_next", &ChatGLM::forward_next)
      .def("generate", &ChatGLM::generate)
      .def("deinit", &ChatGLM::deinit)
      .def_readwrite("SEQLEN", &ChatGLM::SEQLEN) // read SEQLEN in pipeline.py
      .def_readwrite("token_length", &ChatGLM::token_length)
      .def_readwrite("temperature", &ChatGLM::temperature)
      .def_readwrite("top_p", &ChatGLM::top_p)
      .def_readwrite("repeat_penalty", &ChatGLM::repeat_penalty)
      .def_readwrite("repeat_last_n", &ChatGLM::repeat_last_n)
      .def_readwrite("max_new_tokens", &ChatGLM::max_new_tokens)
      .def_readwrite("generation_mode", &ChatGLM::generation_mode)
      .def_readwrite("prompt_mode", &ChatGLM::prompt_mode);
}