LLamaRecipes/research/long-context-llama/H2O/utils_llama.py

import math
from typing import Any, Dict, List, Optional, Tuple, Union

import pdb
import types
import torch
from torch import nn
import torch.utils.checkpoint
import torch.nn.functional as F

from transformers.models.llama.configuration_llama import LlamaConfig
from transformers.models.llama.modeling_llama import (
    LlamaAttention,
    rotate_half,
    apply_rotary_pos_emb,
    repeat_kv,
    LlamaRotaryEmbedding,
    apply_rotary_pos_emb,
    LlamaForCausalLM,
)
from cache_utils import Cache, HHCache
from transformers.utils import logging

logger = logging.get_logger(__name__)

__all__ = ["H2OLlamaForCausalLM"]

def _make_causal_mask(
    bsz: int, tgt_len: int, past_key_values_length: int, dtype: torch.dtype, device: torch.device):
    """
    Make causal mask used for bi-directional self-attention.
    """
    mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device)
    mask_cond = torch.arange(mask.size(-1), device=device)
    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
    mask = mask.to(dtype)

    if past_key_values_length > 0:
        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)

def apply_rotary_pos_emb_single(x, cos, sin, position_ids):
    # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
    cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
    sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
    cos = cos[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
    sin = sin[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
    x_embed = (x * cos) + (rotate_half(x) * sin)
    return x_embed

class H2OKVCache_LayerWise:
    def __init__(
        self,
        hh_size=4,
        recent_size=512,
        k_seq_dim=2,
        v_seq_dim=2,
    ):
        self.hh_size = hh_size
        self.recent_size = recent_size
        self.cache_size = hh_size + recent_size
        self.k_seq_dim = k_seq_dim
        self.v_seq_dim = v_seq_dim
        self.hh_score = None

    def __call__(self, past_key_values, attn_score_cache):

        self._update_hh_score(attn_score_cache)

        if past_key_values is None:
            return None
        seq_len = past_key_values[0].size(self.k_seq_dim)
        if seq_len <= self.cache_size:
            return past_key_values

        # hh-selection
        bsz, num_heads, _, head_dim = past_key_values[0].shape

        select_hh_scores = self.hh_score[:, :seq_len - self.recent_size]
        _, keep_topk = torch.topk(select_hh_scores, self.hh_size, dim=-1)
        keep_topk = keep_topk.sort().values

        # keep_recent = torch.arange(seq_len - self.recent_size, seq_len).expand(keep_topk.shape[0], 1).to(keep_topk.device)
        keep_recent = torch.arange(seq_len - self.recent_size, seq_len, device=keep_topk.device).repeat(keep_topk.shape[0], 1)
        keep_idx = torch.cat([keep_topk, keep_recent], dim=-1)

        mask = torch.zeros(self.hh_score.shape, dtype=torch.bool).to(past_key_values[0].device)
        mask = mask.scatter(-1, keep_idx, 1)

        k_hh_recent = past_key_values[0].squeeze()[mask].view(bsz, num_heads, -1, head_dim)
        v_hh_recent = past_key_values[1].squeeze()[mask].view(bsz, num_heads, -1, head_dim)

        self.hh_score= self.hh_score[mask].view(num_heads, self.cache_size)

        return (k_hh_recent, v_hh_recent)

    def evict_for_space(self, past_key_values, num_coming):
        if past_key_values is None:
            return None
        seq_len = past_key_values[0][0].size(self.k_seq_dim)
        if seq_len + num_coming <= self.cache_size:
            return past_key_values

        # hh-selection
        bsz, num_heads, _, head_dim = past_key_values[0].shape

        select_hh_scores = self.hh_score[:, :seq_len - self.recent_size + num_coming]
        _, keep_topk = torch.topk(select_hh_scores, self.hh_size, dim=-1)
        keep_topk = keep_topk.sort().values

        # keep_recent = torch.arange(seq_len - self.recent_size, seq_len).expand(keep_topk.shape[0], 1).to(keep_topk.device)
        keep_recent = torch.arange(seq_len - self.recent_size + num_coming, seq_len, device=keep_topk.device).repeat(keep_topk.shape[0], 1)
        keep_idx = torch.cat([keep_topk, keep_recent], dim=-1)

        mask = torch.zeros(self.hh_score.shape, dtype=torch.bool).to(past_key_values[0].device)
        mask = mask.scatter(-1, keep_idx, 1)

        k_hh_recent = past_key_values[0].squeeze()[mask].view(bsz, num_heads, -1, head_dim)
        v_hh_recent = past_key_values[1].squeeze()[mask].view(bsz, num_heads, -1, head_dim)

        self.hh_score= self.hh_score[mask].view(num_heads, self.cache_size)

        return (k_hh_recent, v_hh_recent)

    def _update_hh_score(self, attn_score_cache):

        num_new_tokens = attn_score_cache.shape[2]

        if self.hh_score is None:
            self.hh_score = attn_score_cache.sum(0).sum(1)
        else:
            attn_score_cache = attn_score_cache.sum(0).sum(1)
            attn_score_cache[:, :-num_new_tokens] += self.hh_score
            self.hh_score = attn_score_cache

    def _clean_scores(self):
        self.hh_score = None



def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


class H2OLlamaAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: LlamaConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.attention_dropout = config.attention_dropout
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )

        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
        self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=config.attention_bias)
        self._init_rope()

    def _init_rope(self):
        if self.config.rope_scaling is None:
            self.rotary_emb = LlamaRotaryEmbedding(
                self.head_dim,
                max_position_embeddings=self.max_position_embeddings,
                base=self.rope_theta,
            )
        else:
            scaling_type = self.config.rope_scaling["type"]
            scaling_factor = self.config.rope_scaling["factor"]
            if scaling_type == "linear":
                self.rotary_emb = LlamaLinearScalingRotaryEmbedding(
                    self.head_dim,
                    max_position_embeddings=self.max_position_embeddings,
                    scaling_factor=scaling_factor,
                    base=self.rope_theta,
                )
            elif scaling_type == "dynamic":
                self.rotary_emb = LlamaDynamicNTKScalingRotaryEmbedding(
                    self.head_dim,
                    max_position_embeddings=self.max_position_embeddings,
                    scaling_factor=scaling_factor,
                    base=self.rope_theta,
                )
            else:
                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

        if self.config.pretraining_tp > 1:
            key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
            query_slices = self.q_proj.weight.split(
                (self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
            )
            key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
            value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)

            query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
            query_states = torch.cat(query_states, dim=-1)

            key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
            key_states = torch.cat(key_states, dim=-1)

            value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
            value_states = torch.cat(value_states, dim=-1)

        else:
            query_states = self.q_proj(hidden_states)
            key_states = self.k_proj(hidden_states)
            value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        past_key_value = getattr(self, "past_key_value", past_key_value)
        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2).contiguous()

        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

        if self.config.pretraining_tp > 1:
            attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
            o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
            attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
        else:
            attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


def enable_h2ocache_forward(
    self,
    input_ids: torch.LongTensor = None,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.LongTensor] = None,
    past_key_values: Optional[List[torch.FloatTensor]] = None,
    inputs_embeds: Optional[torch.FloatTensor] = None,
    use_cache: Optional[bool] = None,
    output_attentions: Optional[bool] = None,
    output_hidden_states: Optional[bool] = None,
    return_dict: Optional[bool] = None,
    cache_position: Optional[torch.LongTensor] = None,
) -> Union[Tuple, BaseModelOutputWithPast]:
    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
    output_hidden_states = (
        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
    )
    use_cache = use_cache if use_cache is not None else self.config.use_cache
    return_dict = return_dict if return_dict is not None else self.config.use_return_dict

    if (input_ids is None) ^ (inputs_embeds is not None):
        raise ValueError(
            "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
        )

    if self.gradient_checkpointing and self.training and use_cache:
        logger.warning_once(
            "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
        )
        use_cache = False

    if inputs_embeds is None:
        inputs_embeds = self.embed_tokens(input_ids)

    past_seen_tokens = 0
    if use_cache:  # kept for BC (cache positions)
        if not isinstance(past_key_values, StaticCache):
            past_key_values = HHCache.from_legacy_cache(past_key_values)
            past_seen_tokens = past_key_values.get_seq_length()

    if cache_position is None:
        if isinstance(past_key_values, StaticCache):
            raise ValueError("cache_position is a required argument when using StaticCache.")
        cache_position = torch.arange(
            past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
        )

    if position_ids is None:
        position_ids = cache_position.unsqueeze(0)

    causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)

    # embed positions
    hidden_states = inputs_embeds

    # decoder layers
    all_hidden_states = () if output_hidden_states else None
    all_self_attns = () if output_attentions else None
    next_decoder_cache = None

    for decoder_layer in self.layers:
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if self.gradient_checkpointing and self.training:
            layer_outputs = self._gradient_checkpointing_func(
                decoder_layer.__call__,
                hidden_states,
                causal_mask,
                position_ids,
                past_key_values,
                output_attentions,
                use_cache,
                cache_position,
            )
        else:
            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_ids=position_ids,
                past_key_value=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
            )

        hidden_states = layer_outputs[0]

        if use_cache:
            next_decoder_cache = layer_outputs[2 if output_attentions else 1]

        if output_attentions:
            all_self_attns += (layer_outputs[1],)

    hidden_states = self.norm(hidden_states)

    # add hidden states from the last decoder layer
    if output_hidden_states:
        all_hidden_states += (hidden_states,)

    next_cache = None
    if use_cache:
        next_cache = (
            next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
        )
    if not return_dict:
        return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
    return BaseModelOutputWithPast(
        last_hidden_state=hidden_states,
        past_key_values=next_cache,
        hidden_states=all_hidden_states,
        attentions=all_self_attns,
    )

# TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
# KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
# (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
# `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
    if self.config._attn_implementation == "flash_attention_2":
        if attention_mask is not None and 0.0 in attention_mask:
            return attention_mask
        return None

    dtype, device = input_tensor.dtype, input_tensor.device
    min_dtype = torch.finfo(dtype).min
    sequence_length = input_tensor.shape[1]
    if hasattr(self.layers[0].self_attn, "past_key_value"):  # static cache
        target_length = self.config.max_position_embeddings
    else:  # dynamic cache
        target_length = (
            attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else cache_position[-1] + 1
        )

    causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
    if sequence_length != 1:
        causal_mask = torch.triu(causal_mask, diagonal=1)
    causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
    causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
    if attention_mask is not None:
        causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
        if attention_mask.dim() == 2:
            mask_length = attention_mask.shape[-1]
            padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
            causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
        elif attention_mask.dim() == 4:
            # backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
            # cache. In that case, the 4D attention mask attends to the newest tokens only.
            if attention_mask.shape[-2] < cache_position[0] + sequence_length:
                offset = cache_position[0]
            else:
                offset = 0
            mask_shape = attention_mask.shape
            mask_slice = (attention_mask.eq(0.0)).to(dtype=dtype) * min_dtype
            causal_mask[
                : mask_shape[0], : mask_shape[1], offset : mask_shape[2] + offset, : mask_shape[3]
            ] = mask_slice

    if (
        self.config._attn_implementation == "sdpa"
        and attention_mask is not None
        and attention_mask.device.type == "cuda"
    ):
        # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
        # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
        # Details: https://github.com/pytorch/pytorch/issues/110213
        causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

    return causal_mask














# class H2OLlamaAttention(nn.Module):
#     """Multi-headed attention from 'Attention Is All You Need' paper"""

#     def __init__(self, config: LlamaConfig):
#         super().__init__()
#         self.config = config
#         self.hidden_size = config.hidden_size
#         self.num_heads = config.num_attention_heads
#         self.head_dim = self.hidden_size // self.num_heads
#         self.num_key_value_heads = config.num_key_value_heads
#         self.num_key_value_groups = self.num_heads // self.num_key_value_heads
#         self.max_position_embeddings = config.max_position_embeddings
#         self.rope_theta = config.rope_theta

#         if (self.head_dim * self.num_heads) != self.hidden_size:
#             raise ValueError(
#                 f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
#                 f" and `num_heads`: {self.num_heads})."
#             )
#         self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
#         self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
#         self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
#         self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
#         self._init_rope()

#         self.kv_cache = H2OKVCache_LayerWise(
#             hh_size=config.hh_size,
#             recent_size=config.recent_size,
#             k_seq_dim=2,
#             v_seq_dim=2,
#         )

#     def _init_rope(self):
#         if self.config.rope_scaling is None:
#             self.rotary_emb = LlamaRotaryEmbedding(
#                 self.head_dim,
#                 max_position_embeddings=self.max_position_embeddings,
#                 base=self.rope_theta,
#             )
#         else:
#             scaling_type = self.config.rope_scaling["type"]
#             scaling_factor = self.config.rope_scaling["factor"]
#             if scaling_type == "linear":
#                 self.rotary_emb = LlamaLinearScalingRotaryEmbedding(
#                     self.head_dim,
#                     max_position_embeddings=self.max_position_embeddings,
#                     scaling_factor=scaling_factor,
#                     base=self.rope_theta,
#                 )
#             elif scaling_type == "dynamic":
#                 self.rotary_emb = LlamaDynamicNTKScalingRotaryEmbedding(
#                     self.head_dim,
#                     max_position_embeddings=self.max_position_embeddings,
#                     scaling_factor=scaling_factor,
#                     base=self.rope_theta,
#                 )
#             else:
#                 raise ValueError(f"Unknown RoPE scaling type {scaling_type}")

#     def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
#         return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

#     def _clean_cache(self):
#         self.kv_cache._clean_scores()

#     def forward(
#         self,
#         hidden_states: torch.Tensor,
#         attention_mask: Optional[torch.Tensor] = None,
#         position_ids: Optional[torch.LongTensor] = None,
#         past_key_value: Optional[Tuple[torch.Tensor]] = None,
#         output_attentions: bool = False,
#         use_cache: bool = False,
#         cache_position: Optional[torch.LongTensor] = None,
#     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:

#         bsz, q_len, _ = hidden_states.size()

#         if self.config.pretraining_tp > 1:
#             key_value_slicing = (
#                 self.num_key_value_heads * self.head_dim
#             ) // self.config.pretraining_tp
#             query_slices = self.q_proj.weight.split(
#                 (self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
#             )
#             key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
#             value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)

#             query_states = [
#                 F.linear(hidden_states, query_slices[i])
#                 for i in range(self.config.pretraining_tp)
#             ]
#             query_states = torch.cat(query_states, dim=-1)

#             key_states = [
#                 F.linear(hidden_states, key_slices[i])
#                 for i in range(self.config.pretraining_tp)
#             ]
#             key_states = torch.cat(key_states, dim=-1)

#             value_states = [
#                 F.linear(hidden_states, value_slices[i])
#                 for i in range(self.config.pretraining_tp)
#             ]
#             value_states = torch.cat(value_states, dim=-1)

#         else:
#             query_states = self.q_proj(hidden_states)
#             key_states = self.k_proj(hidden_states)
#             value_states = self.v_proj(hidden_states)

#         query_states = query_states.view(
#             bsz, q_len, self.num_heads, self.head_dim
#         ).transpose(1, 2)
#         key_states = key_states.view(
#             bsz, q_len, self.num_key_value_heads, self.head_dim
#         ).transpose(1, 2)
#         value_states = value_states.view(
#             bsz, q_len, self.num_key_value_heads, self.head_dim
#         ).transpose(1, 2)

#         # remake causal mask
#         attention_mask = _make_causal_mask(
#             bsz=bsz,
#             tgt_len=q_len,
#             past_key_values_length=past_key_value[0].shape[-2] if past_key_value is not None else 0,
#             dtype=query_states.dtype,
#             device=query_states.device,
#         )

#         kv_seq_len = key_states.shape[-2]
#         if past_key_value is not None:
#             kv_seq_len += past_key_value[0].shape[-2]

#         if not position_ids.nelement() > 1:
#             position_ids[0][0] = kv_seq_len - 1

#         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
#         ### Shift Pos: query pos is min(cache_size, idx)
#         # query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
#         query_states = apply_rotary_pos_emb_single(query_states, cos, sin, position_ids)
#         ###

#         if past_key_value is not None:
#             # reuse k, v, self_attention
#             key_states = torch.cat([past_key_value[0], key_states], dim=2)
#             value_states = torch.cat([past_key_value[1], value_states], dim=2)

#         past_key_value = (key_states, value_states) if use_cache else None

#         ### Shift Pos: key pos is the pos in cache (Rolling KV Cache and using relative pos emb)
#         key_position_ids = torch.arange(kv_seq_len, device=position_ids.device).unsqueeze(0)
#         key_states = apply_rotary_pos_emb_single(key_states, cos, sin, key_position_ids)
#         ###

#         # repeat k/v heads if n_kv_heads < n_heads
#         key_states = repeat_kv(key_states, self.num_key_value_groups)
#         value_states = repeat_kv(value_states, self.num_key_value_groups)

#         attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(
#             self.head_dim
#         )

#         if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
#             raise ValueError(
#                 f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
#                 f" {attn_weights.size()}"
#             )

#         if attention_mask is not None:
#             if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
#                 raise ValueError(
#                     f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
#                 )
#             attn_weights = attn_weights + attention_mask

#         # upcast attention to fp32
#         attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(
#             query_states.dtype
#         )

#         past_key_value = self.kv_cache(past_key_value, attn_weights.detach().clone())

#         attn_output = torch.matmul(attn_weights, value_states)

#         if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
#             raise ValueError(
#                 f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
#                 f" {attn_output.size()}"
#             )

#         attn_output = attn_output.transpose(1, 2).contiguous()
#         attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

#         if self.config.pretraining_tp > 1:
#             attn_output = attn_output.split(
#                 self.hidden_size // self.config.pretraining_tp, dim=2
#             )
#             o_proj_slices = self.o_proj.weight.split(
#                 self.hidden_size // self.config.pretraining_tp, dim=1
#             )
#             attn_output = sum(
#                 [
#                     F.linear(attn_output[i], o_proj_slices[i])
#                     for i in range(self.config.pretraining_tp)
#                 ]
#             )
#         else:
#             attn_output = self.o_proj(attn_output)

#         if not output_attentions:
#             attn_weights = None

#         return attn_output, attn_weights, past_key_value




class H2OLlamaForCausalLM(LlamaForCausalLM):
    def __init__(self, config):
        super().__init__(config)
        num_layers = len(self.model.layers)
        for layer_idx in range(num_layers):
            self.model.layers[layer_idx].self_attn = H2OLlamaAttention(config, layer_idx)

        self.model.forward = types.MethodType(enable_h2ocache_forward, self.model)