Witllm/wit/modeling_wit.py

import copy
import math
import os
import sys
import gc
from tqdm import auto as tqdm_lib
import json
from typing import Optional, Tuple, Union, Callable, List, Any, Generator
from einops import rearrange

import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch.nn import CrossEntropyLoss
from torch import nn
from safetensors.torch import load_file as safe_load_file
from safetensors.torch import save_file as safe_save_file

from qwen_generation_utils import (
    make_context,
    decode_tokens,
)

sys.path.append("..")
from tools import show
from tools import mem_tracker

# tracker = mem_tracker.MemTracker()
# tracker.track()


class RMSNorm(torch.nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        return self._norm(x.float()).type_as(x) * self.weight


class QWenAttention(nn.Module):
    def __init__(self, config, index):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.split_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.c_attn = nn.Linear(config.hidden_size, 3 * self.hidden_size)
        self.c_proj = nn.Linear(config.hidden_size, self.hidden_size, bias=not config.no_bias)
        self.attn_dropout = nn.Dropout(config.attn_dropout_prob)
        self.index = index

    def _split_heads(self, tensor, num_heads, attn_head_size):
        new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
        tensor = tensor.view(new_shape)
        return tensor

    def _merge_heads(self, tensor, num_heads, attn_head_size):
        tensor = tensor.contiguous()
        new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
        return tensor.view(new_shape)


class QWenMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        ff_dim_in = config.intermediate_size // 2
        self.w1 = nn.Linear(config.hidden_size, ff_dim_in, bias=not config.no_bias)
        self.w2 = nn.Linear(config.hidden_size, ff_dim_in, bias=not config.no_bias)
        self.c_proj = nn.Linear(ff_dim_in, config.hidden_size, bias=not config.no_bias)


class QWenBlock(nn.Module):
    def __init__(self, config, index):
        super().__init__()
        self.ln_1 = RMSNorm(
            config.hidden_size,
            eps=config.layer_norm_epsilon,
        )
        self.attn = QWenAttention(config, index)
        self.ln_2 = RMSNorm(
            config.hidden_size,
            eps=config.layer_norm_epsilon,
        )
        self.mlp = QWenMLP(config)
        self.index = index


class QWenModel(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.wte = nn.Embedding(config.vocab_size, config.hidden_size)
        self.drop = nn.Dropout(config.emb_dropout_prob)
        self.dim = config.hidden_size // config.num_attention_heads

        self.h = nn.ModuleList([QWenBlock(config, i) for i in range(config.num_hidden_layers)])
        self.ln_f = RMSNorm(
            config.hidden_size,
            eps=config.layer_norm_epsilon,
        )
        self.base = config.rotary_emb_base
        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float() / self.dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self._rotary_pos_emb_cache = None
        self._seq_len_cached = 0
        self._ntk_alpha_cached = 1.0

    def update_rotary_pos_emb_cache(self, seqlen, ntk_alpha=1.0):
        if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
            base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
            self.inv_freq = 1.0 / (
                base ** (torch.arange(0, self.dim, 2, device=self.inv_freq.device).float() / self.dim)
            )
            self._seq_len_cached = max(2 * seqlen, 16)
            self._ntk_alpha_cached = ntk_alpha
            seq = torch.arange(self._seq_len_cached, device=self.inv_freq.device)
            freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq)

            emb = torch.cat((freqs, freqs), dim=-1)
            emb = rearrange(emb, "n d -> 1 n 1 d")

            cos, sin = emb.cos(), emb.sin()
            self._rotary_pos_emb_cache = [cos, sin]


class QWenLMHeadModel(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.transformer = QWenModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        **kwargs,
    ):
        runner = QwenRunner(self)
        return runner.forwardQWen(input_ids, labels)

    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]]):
        pretrained_model_name_or_path = str(pretrained_model_name_or_path)
        resolved_archive_file = os.path.join(pretrained_model_name_or_path, "model.safetensors.index.json")
        print(f"loading weights file {resolved_archive_file}")
        with open(resolved_archive_file, "r") as f:
            index = json.loads(f.read())
        shard_filenames = sorted(set(index["weight_map"].values()))
        resolved_archive_file = [os.path.join(pretrained_model_name_or_path, f) for f in shard_filenames]
        model = cls._load_pretrained_model(resolved_archive_file)
        return model

    def _load_state_dict_into_model(self, model_to_load, state_dict, start_prefix):
        metadata = getattr(state_dict, "_metadata", None)
        state_dict = state_dict.copy()
        if metadata is not None:
            state_dict._metadata = metadata
        error_msgs = []

        def load(module: nn.Module, state_dict, prefix=""):
            local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
            args = (state_dict, prefix, local_metadata, True, [], [], error_msgs)
            if len([key for key in state_dict if key.startswith(prefix)]) > 0:
                module._load_from_state_dict(*args)

            for name, child in module._modules.items():
                if child is not None:
                    load(child, state_dict, prefix + name + ".")

        load(model_to_load, state_dict, prefix=start_prefix)
        del state_dict
        return error_msgs

    def _load_pretrained_model(cls, resolved_archive_file):
        start_prefix = ""
        model_to_load = cls
        if len(resolved_archive_file) > 1:
            resolved_archive_file = tqdm_lib.tqdm(resolved_archive_file, desc="Loading checkpoint shards")
        for shard_file in resolved_archive_file:
            state_dict = safe_load_file(shard_file)
            cls._load_state_dict_into_model(model_to_load, state_dict, start_prefix)
            del state_dict  # force memory release
            gc.collect()
        print(f"All model checkpoint weights were used when initializing {cls.__class__.__name__}.\n")
        return cls


class QwenRunner:
    def __init__(self, qwen):
        self.qwen = qwen

    @torch.no_grad()
    def Chat(
        self,
        tokenizer,
        query: str,
        query_assistant: str,
        gen_length=0,
        system: str = "You are a helpful assistant.",
        history=[],
    ):
        qwen = self.qwen
        history = copy.deepcopy(history)
        self.qwen.config.pad_token_id = tokenizer.eod_id
        self.qwen.config.eos_token_id = tokenizer.eod_id
        raw_text, context_tokens = self.prepareInput(tokenizer, query, query_assistant, history, system)
        input_ids = torch.tensor([context_tokens]).to(next(qwen.parameters()).device)
        self.unfinished_sequences = torch.ones(input_ids.shape[0], dtype=torch.long, device=input_ids.device)
        input_length = input_ids.shape[1]
        while True:
            outputs = self.forwardQWen(input_ids)
            next_token_scores = outputs[:, -1, :]

            next_token_scores = self.repetition_penalty(input_ids, next_token_scores)
            next_token_scores = self.top_p(next_token_scores)
            next_tokens = self.sample(next_token_scores)
            finish, next_tokens = self.isFinish(next_tokens)
            if finish:
                break
            input_ids = torch.cat([input_ids, next_tokens], dim=-1)
            if gen_length != 0 and (input_length + gen_length) < input_ids.shape[1]:
                break

        decoded, response, end_reason = decode_tokens(
            input_ids[0],
            tokenizer,
            raw_text_len=len(raw_text),
            context_length=len(context_tokens),
            errors="replace",
        )
        history.append((query, response))
        return input_ids[0].cpu().tolist(), history, decoded

    def _rotate_half(self, x):
        x = rearrange(x, "... (j d) -> ... j d", j=2)
        x1, x2 = x.unbind(dim=-2)
        return torch.cat((-x2, x1), dim=-1)

    def apply_rotary_pos_emb(self, t, freqs):
        rot_dim = freqs[0].shape[-1]
        cos, sin = freqs
        t_float = t.float()
        t_rot = t_float[..., :rot_dim]
        t_pass = t_float[..., rot_dim:]
        t_rot = (t_rot * cos) + (self._rotate_half(t_rot) * sin)
        return torch.cat((t_rot, t_pass), dim=-1).type_as(t)

    def split_heads(
        self,
        attention,
        hidden_states: Optional[Tuple[torch.FloatTensor]],
    ):
        atten = attention
        mixed_x_layer = atten.c_attn(hidden_states)
        query, key, value = mixed_x_layer.split(atten.split_size, dim=2)
        query = atten._split_heads(query, atten.num_heads, atten.head_dim)
        key = atten._split_heads(key, atten.num_heads, atten.head_dim)
        value = atten._split_heads(value, atten.num_heads, atten.head_dim)
        return query, key, value

    def pos_emb(self, query, key, rotary_pos_emb_list):
        rotary_pos_emb = rotary_pos_emb_list[0]
        rotary_pos_emb = [i[:, -query.shape[1] :, :, :] for i in rotary_pos_emb]
        rotary_pos_emb = (rotary_pos_emb,) * 2
        query = self.apply_rotary_pos_emb(query, rotary_pos_emb[0])
        key = self.apply_rotary_pos_emb(key, rotary_pos_emb[1])
        return query, key

    def attention(self, attention, query, key, value, causal_mask):
        query = query.permute(0, 2, 1, 3)
        key = key.permute(0, 2, 1, 3)
        value = value.permute(0, 2, 1, 3)
        attn_output = F.scaled_dot_product_attention(query, key, value, attn_mask=causal_mask).transpose(1, 2)
        context_layer = attention._merge_heads(attn_output, attention.num_heads, attention.head_dim)
        attn_output = attention.c_proj(context_layer)
        return attn_output

    def build_mask(self, query):
        size = query.size(1)
        causal_mask = torch.tril(torch.ones((size, size), dtype=torch.bool, device=query.device)).view(1, 1, size, size)
        return causal_mask

    def forwardAttention(
        self,
        attention,
        hidden_states: Optional[Tuple[torch.FloatTensor]],
        rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
    ):
        query, key, value = self.split_heads(attention, hidden_states)
        query, key = self.pos_emb(query, key, rotary_pos_emb_list)
        causal_mask = self.build_mask(query)
        return self.attention(attention, query, key, value, causal_mask)

    def forwardQWenBlock(
        self,
        block,
        hidden_states: Optional[Tuple[torch.FloatTensor]],
        rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
    ):
        layernorm_output = block.ln_1(hidden_states)

        attn_outputs = self.forwardAttention(block.attn, layernorm_output, rotary_pos_emb_list)
        attn_output = attn_outputs[0]
        layernorm_input = attn_output + hidden_states

        layernorm_output = block.ln_2(layernorm_input)
        a1 = block.mlp.w1(layernorm_output)
        a2 = block.mlp.w2(layernorm_output)
        intermediate_parallel = a1 * F.silu(a2)
        mlp_output = block.mlp.c_proj(intermediate_parallel)

        hidden_states = layernorm_input + mlp_output
        return hidden_states

    def forwardQWen(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        labels: Optional[torch.LongTensor] = None,
    ):
        transfm = self.qwen.transformer
        input_shape = input_ids.size()
        input_ids = input_ids.view(-1, input_shape[-1])
        hidden_states = transfm.wte(input_ids)
        kv_seq_len = hidden_states.size()[1]

        transfm.update_rotary_pos_emb_cache(kv_seq_len, ntk_alpha=1.0)
        cos, sin = transfm._rotary_pos_emb_cache
        rotary_pos_emb_list = [[cos[:, :kv_seq_len], sin[:, :kv_seq_len]]]

        hidden_states = transfm.drop(hidden_states)
        output_shape = input_shape + (hidden_states.size(-1),)

        for block in transfm.h:
            hidden_states = self.forwardQWenBlock(block, hidden_states, rotary_pos_emb_list=rotary_pos_emb_list)

        hidden_states = transfm.ln_f(hidden_states)
        hidden_states = hidden_states.view(output_shape)

        lm_logits = self.qwen.lm_head(hidden_states)

        loss = None
        if labels is not None:
            labels = labels.to(lm_logits.device)
            shift_labels = labels[..., 1:].contiguous().view(-1)
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_logits = shift_logits.view(-1, shift_logits.size(-1))
            mask = shift_labels < self.qwen.config.vocab_size
            shift_labels = shift_labels[mask]
            shift_logits = shift_logits[mask]
            # m = torch.max(shift_logits, 1).indices.cpu().numpy()
            # ll = shift_labels.cpu().numpy()
            loss = CrossEntropyLoss()(shift_logits, shift_labels)

        return lm_logits, loss

    def prepareInput(self, tokenizer, query, query_assistant, history, system):
        return make_context(tokenizer, query, query_assistant, history=history, system=system)

    def repetition_penalty(self, input_ids, next_token_scores):
        penalty = self.qwen.config.repetition_penalty
        score = torch.gather(next_token_scores, 1, input_ids)
        # if score < 0 then repetition penalty has to be multiplied to reduce the token probabilities
        score = torch.where(score < 0, score * penalty, score / penalty)
        next_token_scores = next_token_scores.scatter_(1, input_ids, score)
        return next_token_scores

    def top_p(self, next_token_scores):
        top_p = self.qwen.config.top_p
        filter_value = -float("Inf")
        min_tokens_to_keep = 1
        sorted_logits, sorted_indices = torch.sort(next_token_scores, descending=False)
        cumulative_probs = sorted_logits.softmax(dim=-1).cumsum(dim=-1)
        # Remove tokens with cumulative top_p above the threshold (token with 0 are kept)
        sorted_indices_to_remove = cumulative_probs <= (1 - top_p)
        # Keep at least min_tokens_to_keep
        sorted_indices_to_remove[..., -min_tokens_to_keep:] = 0
        # scatter sorted tensors to original indexing
        indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
        next_token_scores = next_token_scores.masked_fill(indices_to_remove, filter_value)
        return next_token_scores

    def sample(self, next_token_scores):
        probs = nn.functional.softmax(next_token_scores, dim=-1)
        next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
        return next_tokens

    def isFinish(self, next_tokens):
        pad_token_id = self.qwen.config.pad_token_id
        eos_token_id_tensor = torch.tensor([self.qwen.config.eos_token_id]).to(next_tokens.device)

        next_tokens = next_tokens * self.unfinished_sequences + pad_token_id * (1 - self.unfinished_sequences)
        self.unfinished_sequences = self.unfinished_sequences.mul(
            next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
        )
        return self.unfinished_sequences.max() == 0, next_tokens[:, None]