modeling_fast_slm.py

# coding=utf-8
# Copyright 2024 AI21 Labs Ltd. and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch FastSLM model."""
import inspect
import math
import copy
import warnings
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Tuple, Union
import time
from collections import OrderedDict
from functools import partial
import numpy as np
import os

import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

torch._inductor.config.max_autotune_gemm_backends = ["aten"]

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_outputs import (
    MoeCausalLMOutputWithPast,
    MoeModelOutputWithPast,
    SequenceClassifierOutputWithPast,
)
from transformers.modeling_utils import PreTrainedModel

try:
    from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
except ImportError:
    pass

from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_13
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
)
from transformers.utils.import_utils import is_torch_fx_available
from .configuration_fast_slm import FastSLMConfig
from torch.utils.checkpoint import checkpoint

import torch.distributed as dist
import math
import random

from flash_attn import flash_attn_func, flash_attn_varlen_func
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa

_flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)

from einops import rearrange, repeat, reduce, pack, unpack
from einops.layers.torch import Rearrange

from mamba_ssm.ops.selective_scan_interface import mamba_inner_fn, selective_scan_fn
from mamba_ssm.ops.triton.selective_state_update import selective_state_update

from causal_conv1d import causal_conv1d_fn, causal_conv1d_update

from .fused_mha_with_cache import fused_mha_interface

from .mamba2 import Mamba2
from mamba_ssm.utils.generation import InferenceParams
from .delta_net import Cache as fla_cache
from .delta_net import DeltaNet
import torch._dynamo
torch._dynamo.config.suppress_errors = True

from torch.cuda import CUDAGraph

logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "FastSLMConfig"



# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def _get_unpad_data(attention_mask):
    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = seqlens_in_batch.max().item()
    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
    return (
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


### Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->FastSLM
class FastSLMRMSNorm(nn.Module):
    def __init__(self, hidden_size, learnable_weight=True, eps=1e-6):
        """
        FastSLMRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        if learnable_weight:
            self.weight = nn.Parameter(torch.ones(hidden_size))
        else:
            self.weight = None
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        
        if self.weight is not None:
            return self.weight * hidden_states.to(input_dtype)
        else:
            return hidden_states.to(input_dtype)

class LlamaRotaryEmbedding(nn.Module):
    def __init__(self, config, dim, base=10000, device=None, scaling_factor=1.0):
        super().__init__()
        self.scaling_factor = scaling_factor
        self.dim = dim
        self.base = base
        self.config = config
        
        self.rope_type = config.rope_type
        
        self.factor = 2
        
        max_position_embeddings = self.config.max_position_embeddings

        if config.rope_type is None or config.rope_type == "default":
            inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
            self.max_seq_len_cached = max_position_embeddings

        elif config.rope_type == 'ntk':
            assert self.config.orig_max_position_embeddings is not None
            orig_max_position_embeddings = self.config.orig_max_position_embeddings
            
            base = base * ((self.factor * max_position_embeddings / orig_max_position_embeddings) - (self.factor - 1)) ** (self.dim / (self.dim - 2))
            inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
            
            self.max_seq_len_cached = orig_max_position_embeddings
            
        elif config.rope_type == 'dynamic_ntk':
            inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
            self.original_inv_freq = inv_freq
            self.max_seq_len_cached = self.config.orig_max_position_embeddings
                
        else:
            raise ValueError(f"Not support rope_type: {config.rope_type}")

        self.register_buffer("inv_freq", inv_freq, persistent=False)
        

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            base = self.base * ((self.factor * seq_len / self.config.orig_max_position_embeddings) - (self.factor - 1)) ** (self.dim / (self.dim - 2))
            inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
            
            self.register_buffer("inv_freq", inv_freq, persistent=False)
            self.max_seq_len_cached = seq_len

        if seq_len < self.config.orig_max_position_embeddings and self.max_seq_len_cached > self.config.orig_max_position_embeddings:  # reset
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.config.orig_max_position_embeddings
        

    @torch.no_grad()
    def forward(self, x, position_ids):
        if self.rope_type == 'dynamic_ntk':
            self._dynamic_frequency_update(position_ids, device=x.device)
            
        # x: [bs, num_attention_heads, seq_len, head_size]
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        # Force float32 since bfloat16 loses precision on long contexts
        # See https://github.com/huggingface/transformers/pull/29285
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    if q is not None:
        q_embed = (q * cos) + (rotate_half(q) * sin)

    else:
        q_embed = None
    
    if k is not None:
        k_embed = (k * cos) + (rotate_half(k) * sin)
    else:
        k_embed = None
    return q_embed, k_embed    


# Copied from transformers.models.llama.modeling_llama.repeat_kv
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 HybridMambaAttentionDynamicCache(DynamicCache):
    """
    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
    (which has a constant shape regardless of seq_len).

    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
    and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
    For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
    and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
    """

    def __init__(self, config, batch_size, dtype=torch.float16, device=None, layer_type=None):
        self.dtype = dtype
        # self.layers_block_type = config.layers_block_type
        intermediate_size = config.mamba_expand * config.hidden_size
        ssm_state_size = config.mamba_d_state
        conv_kernel_size = config.mamba_d_conv
        self.conv_states = []
        self.ssm_states = []

        self.layer_type = layer_type

        for i in range(config.num_hidden_layers):
            has_mamba_state = self.layer_type[i] == 'h' or self.layer_type[i] == 'm'
            
            if has_mamba_state:
                if hasattr(config, 'conv_dim'):
                    conv_dim = config.conv_dim[str(i)]
                else:
                    conv_dim = intermediate_size
                self.conv_states += [
                    torch.zeros(batch_size, conv_dim, conv_kernel_size, device=device, dtype=dtype)
                ]
                self.ssm_states += [
                    torch.zeros(batch_size, intermediate_size, ssm_state_size, device=device, dtype=dtype)
                ]
            else:
                self.conv_states += [torch.tensor([[]] * batch_size, device=device)]
                self.ssm_states += [torch.tensor([[]] * batch_size, device=device)]

        self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
        self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]

        self.mamba_past_length = [0 for _ in range(config.num_hidden_layers)]

    def update(
        self,
        key_states: torch.Tensor,
        value_states: torch.Tensor,
        layer_idx: int,
        cache_kwargs: Optional[Dict[str, Any]] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        # Update the cache
        if self.key_cache[layer_idx].shape[-1] == 0:
            self.key_cache[layer_idx] = key_states
            self.value_cache[layer_idx] = value_states
        else:
            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)

        return self.key_cache[layer_idx], self.value_cache[layer_idx]

    def reorder_cache(self, beam_idx: torch.LongTensor):
        """Reorders the cache for beam search, given the selected beam indices."""
        for layer_idx in range(len(self.key_cache)):
            device = self.key_cache[layer_idx].device
            self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
            device = self.value_cache[layer_idx].device
            self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))

            device = self.conv_states[layer_idx].device
            self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device))
            device = self.ssm_states[layer_idx].device
            self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))

    def get_seq_length(self, layer_idx=None) -> int:
        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
        # take any layer that contains cache and not empty tensor

        if layer_idx is None:
            max_mamba_len = max(self.mamba_past_length)
            if max_mamba_len > 0:
                return max_mamba_len
            
            else:
                max_key_len = max(cache.shape[-2] for cache in self.key_cache)
                return max_key_len

        if self.layer_type[layer_idx] == 'm':
            return self.mamba_past_length[layer_idx]

        if self.key_cache[layer_idx].shape[-1] == 0:
            return 0

        return self.key_cache[layer_idx].shape[-2]

    # def get_max_length(self) -> Optional[int]:
    #     """Returns the maximum sequence length of the cached states. Cache does not have a maximum length."""
    #     return None
    
    def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")

    @classmethod
    def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "DynamicCache":
        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
    

@dataclass
class MambaCacheParams:
    seqlen_offset: int = 0
    conv_states: Dict[int, torch.Tensor] = field(default_factory=dict)
    ssm_states: Dict[int, torch.Tensor] = field(default_factory=dict)



# Adapted from transformers.models.mistral.modeling_mistral.MistralAttention with Mistral->FastSLM
class FastSLMAttention(nn.Module):
    """
    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
    and "Generating Long Sequences with Sparse Transformers".
    """

    def __init__(self, config: FastSLMConfig, layer_idx: Optional[int] = None, input_hidden_size=None, output_hidden_size=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.hidden_size = config.hidden_size
        self.hidden_size = config.attn_hidden_size if config.attn_hidden_size > 0 else config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta

        self.kq_head_dim = config.kq_head_dim if config.kq_head_dim > 0 else self.head_dim
        self.v_head_dim = config.v_head_dim if config.v_head_dim > 0 else self.head_dim

        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.is_causal = True
        self.attention_dropout = config.attention_dropout
        
        if (self.head_dim * self.num_heads) != self.hidden_size and self.kq_head_dim == self.head_dim:
            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 if input_hidden_size is None else input_hidden_size, self.num_heads * self.kq_head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size if input_hidden_size is None else input_hidden_size, self.num_key_value_heads * self.kq_head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size if input_hidden_size is None else input_hidden_size, self.num_key_value_heads * self.v_head_dim, bias=False)

        if output_hidden_size is None:
            output_hidden_size = self.hidden_size

        self.o_proj = nn.Linear(self.num_heads * self.v_head_dim, output_hidden_size, bias=False)

        if self.config.kq_norm == "rms":
            self.k_norm = FastSLMRMSNorm(self.kq_head_dim)
            self.q_norm = FastSLMRMSNorm(self.kq_head_dim)
        elif self.config.kq_norm == "none":
            self.k_norm = None
            self.q_norm = None
        else:
            raise NotImplementedError(f"Unknown kq_norm: {self.config.kq_norm}")

        if self.config.rope:
            # print("===>  Using Rotary Position Embedding")
            self._init_rope()
            
    def _init_rope(self):
        # assert 1==0, f"max_position_embeddings: {self.max_position_embeddings}"
        self.rotary_emb = LlamaRotaryEmbedding(
            config=self.config,
            dim=self.kq_head_dim,
            base=self.rope_theta,
            device=torch.device("cuda"),
            ) 

    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 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,
            # kv_proj_last_layer = None,
            use_swa=False,
            query_states = None,
            key_states=None,
            value_states=None,
            **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        raise NotImplementedError("FastSLMAttention is an abstract class. Use one of the subclasses.")



# Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->FastSLM
class FastSLMFlashAttention2(FastSLMAttention):
    """
    FastSLM flash attention module. This module inherits from `FastSLMAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()

    def forward(
            self,
            hidden_states: torch.Tensor = None,
            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,
            # kv_proj_last_layer = None,
            use_swa=False,
            query_states = None,
            key_states=None,
            value_states=None,
            **kwargs,
    ):  
            
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )

            # overwrite attention_mask with padding_mask
            attention_mask = kwargs.pop("padding_mask")

        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.kq_head_dim).transpose(1, 2).contiguous()

        if self.q_norm is not None:
            query_states = self.q_norm(query_states)
            
        # we do kq_norm first before rope according to 
        # https://github.com/huggingface/transformers/blob/6c1d0b069de22d7ed8aa83f733c25045eea0585d/src/transformers/models/cohere/modeling_cohere.py#L568
        if self.config.rope:
            cos, sin = self.rotary_emb(hidden_states, position_ids)
            query_states, _ = apply_rotary_pos_emb(query_states, None, cos, sin)

               
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)
            
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.kq_head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.v_head_dim).transpose(1, 2)

        if self.k_norm is not None:
            key_states = self.k_norm(key_states)
        
        if self.config.rope:
            _, key_states = apply_rotary_pos_emb(None, key_states, cos, sin)

        
        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        use_sliding_windows = (
                _flash_supports_window_size
                and getattr(self.config, "sliding_window", None) is not None
                # and kv_seq_len > (self.config.sliding_window + self.config.num_memory_tokens if self.config.num_memory_tokens > 0 else self.config.sliding_window)
                and kv_seq_len > self.config.sliding_window
                and use_swa
        )

        if not _flash_supports_window_size:
            logger.warning_once(
                "The current flash attention version does not support sliding window attention, for a more memory efficient implementation"
                " make sure to upgrade flash-attn library."
            )

        swa_processed_flag = False
        if past_key_value is not None and use_cache:
            kv_layer_idx = self.layer_idx

            cache_has_contents = past_key_value.get_seq_length(kv_layer_idx) > 0
            
            if (
                    getattr(self.config, "sliding_window", None) is not None
                    # and kv_seq_len > (self.config.sliding_window + self.config.num_memory_tokens if self.config.num_memory_tokens > 0 else self.config.sliding_window)
                    and kv_seq_len > self.config.sliding_window
                    and cache_has_contents
                    and use_swa
            ):              
                slicing_tokens = 1 - self.config.sliding_window

                past_key = past_key_value[kv_layer_idx][0]
                past_value = past_key_value[kv_layer_idx][1]
                
                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
                past_value = past_value[:, :, slicing_tokens:, :].contiguous()

                past_key_value.key_cache[kv_layer_idx] = past_key
                past_key_value.value_cache[kv_layer_idx] = past_value
                                                        
                if attention_mask is not None:
                    attention_mask = attention_mask[:, slicing_tokens:]
                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
                
                swa_processed_flag = True
                            
            key_states, value_states = past_key_value.update(key_states, value_states, kv_layer_idx)

        # repeat k/v heads if n_kv_heads < n_heads
        key_states_no_repeat = key_states
        value_states_no_repeat = value_states

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
        dropout_rate = 0.0 if not self.training else self.attention_dropout

        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
        # therefore the input hidden states gets silently casted in float32. Hence, we need
        # cast them back in float16 just to be sure everything works as expected.
        input_dtype = query_states.dtype
        if input_dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = torch.get_autocast_gpu_dtype()
            # Handle the case where the model is quantized
            elif hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            else:
                target_dtype = self.q_proj.weight.dtype

            logger.warning_once(
                f"The input hidden states seems to be silently casted in float32, this might be related to"
                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
                f" {target_dtype}."
            )

            query_states = query_states.to(target_dtype)
            key_states = key_states.to(target_dtype)
            value_states = value_states.to(target_dtype)

        # Reashape to the expected shape for Flash Attention
        query_states = query_states.transpose(1, 2)  # (batch, slen, num_heads, head_dim)
        key_states = key_states.transpose(1, 2)  # (batch, slen, num_heads, head_dim)
        value_states = value_states.transpose(1, 2)  # (batch, slen, num_heads, head_dim)
                
        attn_output = self._flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            use_sliding_windows=use_sliding_windows and not swa_processed_flag,
        )

        v_dim = value_states.shape[-2] * value_states.shape[-1]
        attn_output = attn_output.reshape(-1, q_len, v_dim).contiguous()
            
        attn_output = self.o_proj(attn_output)
        
        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value, (key_states_no_repeat, value_states_no_repeat)

    def _flash_attention_forward(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            query_length,
            dropout=0.0,
            softmax_scale=None,
            use_sliding_windows=False,
    ):
        """
        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
        first unpad the input, then computes the attention scores and pad the final attention scores.

        Args:
            query_states (`torch.Tensor`):
                Input query states to be passed to Flash Attention API
            key_states (`torch.Tensor`):
                Input key states to be passed to Flash Attention API
            value_states (`torch.Tensor`):
                Input value states to be passed to Flash Attention API
            attention_mask (`torch.Tensor`):
                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                position of padding tokens and 1 for the position of non-padding tokens.
            dropout (`int`, *optional*):
                Attention dropout
            softmax_scale (`float`, *optional*):
                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
            use_sliding_windows (`bool`, *optional*):
                Whether to activate sliding window attention.
        """
        if not self._flash_attn_uses_top_left_mask:
            causal = self.is_causal
        else:
            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
            causal = self.is_causal and query_length != 1
                    
        if attention_mask is not None:
            batch_size = query_states.shape[0]
            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
                query_states, key_states, value_states, attention_mask, query_length
            )

            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens

            if not use_sliding_windows:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
        else:
            if not use_sliding_windows:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

        return attn_output

    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
                
        # On the first iteration we need to properly re-create the padding mask
        # by slicing it on the proper place
        if kv_seq_len != attention_mask.shape[-1]:
            attention_mask_num_tokens = attention_mask.shape[-1]
            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]

        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)

        if not self.training and not type(key_layer) == torch.Tensor:  ## this is for handling Mamba2 with output type <class 'mamba_ssm.ops.triton.layernorm_gated.tTensor'>
            key_layer = torch.tensor(key_layer.clone())
            value_layer = torch.tensor(value_layer.clone())
            query_layer = torch.tensor(query_layer.clone())
        
        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)

        if query_length == kv_seq_len:
            query_layer = index_first_axis(
                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
            )
            cu_seqlens_q = cu_seqlens_k
            max_seqlen_in_batch_q = max_seqlen_in_batch_k
            indices_q = indices_k
        elif query_length == 1:
            max_seqlen_in_batch_q = 1
            cu_seqlens_q = torch.arange(
                batch_size + 1, dtype=torch.int32, device=query_layer.device
            )  # There is a memcpy here, that is very bad.
            indices_q = cu_seqlens_q[:-1]
            query_layer = query_layer.squeeze(1)
        else:
            # The -q_len: slice assumes left padding.
            attention_mask = attention_mask[:, -query_length:]
            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)

        return (
            query_layer,
            key_layer,
            value_layer,
            indices_q,
            (cu_seqlens_q, cu_seqlens_k),
            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
        )



class FastSLMSDPAAttention(nn.Module):

    def __init__(self, config, layer_idx: int, reuse_kv=False):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=False
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=False
        )

        self.sliding_window = self.config.sliding_window if self.layer_idx not in self.config.global_attn_idx else None

    def forward(
        self,
        hidden_states: torch.Tensor,
        # position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_value: Optional[Cache] = None,
        **kwargs,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        # cos, sin = position_embeddings
        # query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx) # , cache_kwargs)
    
        attention_interface = ALL_ATTENTION_FUNCTIONS['flash_attention_2']

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            sliding_window=self.sliding_window,  # diff with Llama
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)

        return attn_output, attn_weights, past_key_value, (key_states, value_states)


class FastSLMFused_MHA(FastSLMAttention):
    """
    FastSLM flash attention module. This module inherits from `FastSLMAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        
        self.fused_mha_interface = fused_mha_interface
        
        # self.init_kv_cache(max_batch_size=1, max_seq_len=8000)
        

    def init_kv_cache(self, max_batch_size, max_seq_len, page_size=-1):
        if hasattr(self, 'k_cache'):
            del self.k_cache
            del self.v_cache
            
            if hasattr(self, 'page_table') and self.page_table is not None:
                del self.page_table
            
            import gc
            gc.collect()
                
            torch.cuda.empty_cache()
        
        if page_size is not None and page_size > 0:
            batch_max_pages = (max_seq_len + page_size - 1) // page_size
            cache_max_pages = (max_batch_size * max_seq_len + page_size - 1) // page_size
            self.k_cache = torch.zeros(cache_max_pages, page_size, self.num_key_value_heads, self.kq_head_dim).to(self.q_proj.weight)
            self.v_cache = torch.zeros(cache_max_pages, page_size, self.num_key_value_heads, self.v_head_dim).to(self.q_proj.weight)

            self.page_table = torch.zeros(max_batch_size, batch_max_pages, device=self.q_proj.weight.device, dtype=torch.int32)
        else:
            self.k_cache = torch.zeros(max_batch_size, max_seq_len, self.num_key_value_heads, self.kq_head_dim).to(self.q_proj.weight)
            self.v_cache = torch.zeros(max_batch_size, max_seq_len, self.num_key_value_heads, self.v_head_dim).to(self.q_proj.weight)
            
            self.page_table = None
        
        self.max_seq_len = max_seq_len
    

    def reset_kv_cache(self):
        self.k_cache = self.k_cache.zero_()
        self.v_cache = self.v_cache.zero_()
        
        if self.page_table is not None:
            self.page_table = self.page_table.zero_()
        
        
    def forward(
            self,
            hidden_states: torch.Tensor = None,
            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,
            use_swa=False,
            query_states = None,
            key_states=None,
            value_states=None,
            **kwargs,
    ):     

        # print(f"Flash Attn - layer_idx: {self.layer_idx}, attn_mask is none: {attention_mask is None}")
        # print(f"layer_idx: {self.layer_idx}, use_swq: {use_swa}")
        if not hasattr(self, 'k_cache'):
            self.init_kv_cache(max_batch_size=1, max_seq_len=8000)

        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )

            attention_mask = kwargs.pop("padding_mask")
        
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.kq_head_dim).transpose(1, 2).contiguous()

        if self.q_norm is not None:
            query_states = self.q_norm(query_states)
            
        # we do kq_norm first before rope according to 
        # https://github.com/huggingface/transformers/blob/6c1d0b069de22d7ed8aa83f733c25045eea0585d/src/transformers/models/cohere/modeling_cohere.py#L568
        if self.config.rope:
            cos, sin = self.rotary_emb(hidden_states, position_ids)
            query_states, _ = apply_rotary_pos_emb(query_states, None, cos, sin)
         
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.kq_head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.v_head_dim).transpose(1, 2)

        if self.k_norm is not None:
            key_states = self.k_norm(key_states)
        
        if self.config.rope:
            # cos, sin = self.rotary_emb(hidden_states, position_ids)
            _, key_states = apply_rotary_pos_emb(None, key_states, cos, sin)

        key_states_no_repeat = key_states
        value_states_no_repeat = value_states
        
        # Reashape to the expected shape for Flash Attention
        query_states = query_states.transpose(1, 2)  # (batch, slen, num_heads, head_dim)
        key_states = key_states.transpose(1, 2)  # (batch, slen, num_kv_heads, head_dim)
        value_states = value_states.transpose(1, 2)  # (batch, slen, num_kv_heads, head_dim)
        
        if self.k_cache.device != query_states.device:
            self.k_cache = self.k_cache.to(query_states)
            self.v_cache = self.v_cache.to(query_states)

        attn_output = self.fused_mha_interface(
            query_states,
            key_states,
            value_states,
            k_cache=self.k_cache,
            v_cache=self.v_cache,
            page_table=self.page_table,
            max_seq_len=self.max_seq_len,
            position_ids=position_ids,
        )

        v_dim = query_states.shape[-2] * value_states.shape[-1]
        attn_output = attn_output.reshape(bsz, q_len, v_dim).contiguous()
        
        attn_output = self.o_proj(attn_output)
                    
        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value, (key_states_no_repeat, value_states_no_repeat)
        

JAMBA_ATTENTION_CLASSES = {
    "flash_attention_2": FastSLMFlashAttention2,
    "fused_mha": FastSLMFused_MHA,
    "sdpa": FastSLMSDPAAttention,
}

class FastSLMMLP(nn.Module):
    def __init__(self, config: FastSLMConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.act_fn_name = config.mlp_hidden_act
        self.act_fn = ACT2FN[self.act_fn_name]
        
        if config.ffn_expand_ratio is not None:
            self.ffn_dim = int(config.ffn_expand_ratio * config.hidden_size) // 128 * 128
        else:
            self.ffn_dim = config.intermediate_size
        
        self.hidden_dim = config.hidden_size
        
        self.layer_idx = layer_idx

        if self.act_fn_name == "silu":
            self.gate_proj = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
        self.down_proj = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
        self.up_proj = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
        

    def forward(self, x):
        if self.act_fn_name == "silu":
            output = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        elif self.act_fn_name == "relu2":
            output  = self.down_proj(self.act_fn(self.up_proj(x)))
        else:
            raise NotImplementedError(f"No such hidden_act: {self.act_fn_name}")

        return output


# Adapted from transformers.models.mixtral.modeling_mixtral.MixtralSparseMoeBlock with Mistral->FastSLM
class FastSLMSparseMoeBlock(nn.Module):
    """
    This implementation is
    strictly equivalent to standard MoE with full capacity (no
    dropped tokens). It's faster since it formulates MoE operations
    in terms of block-sparse operations to accomodate imbalanced
    assignments of tokens to experts, whereas standard MoE either
    (1) drop tokens at the cost of reduced performance or (2) set
    capacity factor to number of experts and thus waste computation
    and memory on padding.
    """

    def __init__(self, config: FastSLMConfig, num_experts: int, num_experts_per_tok: int, layer_idx: int):
        super().__init__()
        self.hidden_dim = config.hidden_size
        self.ffn_dim = config.intermediate_size
        
        self.layer_idx = layer_idx

        #   these values are decided on runtime depending on the layer index
        self.num_experts = num_experts
        self.top_k = num_experts_per_tok

        if num_experts > 1:
            # expert routing
            self.router = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
        else:
            self.router = None

        self.experts = nn.ModuleList([FastSLMMLP(config, layer_idx=layer_idx) for _ in range(self.num_experts)])

    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        """ """
        if len(hidden_states.shape) == 3:
            batch_size, sequence_length, hidden_dim = hidden_states.shape
            bs_times_seq_len = batch_size * sequence_length
        elif len(hidden_states.shape) == 2: 
            assert self.num_experts == 1
            bs_times_seq_len, hidden_dim = hidden_states.shape
        else:
            batch_size, sequence_length, _, hidden_dim = hidden_states.shape
            bs_times_seq_len = batch_size * sequence_length

        if self.num_experts == 1:
            # in this case we have a single MLP block and don't need to do any routing
            final_hidden_states = self.experts[0](hidden_states)
            
            router_logits = torch.ones(
                (bs_times_seq_len, 1),
                device=hidden_states.device,
                dtype=hidden_states.dtype,
                requires_grad=hidden_states.requires_grad,
            )
            return final_hidden_states, router_logits

        # in this case we have multiple experts and need to do routing
        hidden_states = hidden_states.view(-1, hidden_dim)
        # router_logits: (batch * sequence_length, n_experts)
        router_logits = self.router(hidden_states)
        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
        # we cast back to the input dtype
        routing_weights = routing_weights.to(hidden_states.dtype)

        final_hidden_states = torch.zeros(
            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
        )

        # One hot encode the selected experts to create an expert mask
        # this will be used to easily index which expert is going to be sollicitated
        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)

        # Loop over all available experts in the model and perform the computation on each expert
        for expert_idx in range(self.num_experts):
            expert_layer = self.experts[expert_idx]
            idx, top_x = torch.where(expert_mask[expert_idx])

            if top_x.shape[0] == 0:
                continue

            # in torch it is faster to index using lists than torch tensors
            top_x_list = top_x.tolist()
            idx_list = idx.tolist()

            # Index the correct hidden states and compute the expert hidden state for
            # the current expert. We need to make sure to multiply the output hidden
            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
            current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
            current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]

            # However `index_add_` only support torch tensors for indexing so we'll use
            # the `top_x` tensor here.
            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
            
        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
        return final_hidden_states, router_logits


    


class FastSLMAttentionDecoderLayer(nn.Module):
    def __init__(self, config: FastSLMConfig, num_experts: int, layer_idx: int,):
        super().__init__()

        self.config = config
        
        self.layer_idx = layer_idx

        self.self_attn = JAMBA_ATTENTION_CLASSES[config.attn_implementation](config, layer_idx)

        if self.config.intermediate_size > 0:
            num_experts_per_tok = config.num_experts_per_tok if num_experts > 1 else 1
            self.moe = FastSLMSparseMoeBlock(config, num_experts=num_experts, num_experts_per_tok=num_experts_per_tok, layer_idx=layer_idx)
        else:
            self.moe = None

        self.input_layernorm = FastSLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.pre_moe_layernorm = FastSLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    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: Optional[bool] = False,
            output_router_logits: Optional[bool] = False,
            use_cache: Optional[bool] = False,
            use_swa=False,
            **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
                should not be returned during inference.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
        """

        if position_ids is not None and position_ids.shape[1] != hidden_states.shape[1]:
            position_ids = torch.arange(hidden_states.shape[1], device=hidden_states.device).unsqueeze(0)
                
        residual = hidden_states

        if self.input_layernorm is not None:
            hidden_states = self.input_layernorm(hidden_states)

        hidden_states, self_attn_weights, present_key_value, current_kv = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            use_swa=use_swa,
        )
        
        hidden_states = residual + hidden_states
        
        if self.moe is not None:
            residual = hidden_states
            if self.pre_moe_layernorm is not None:
                hidden_states = self.pre_moe_layernorm(hidden_states)
            hidden_states, router_logits = self.moe(hidden_states)

            hidden_states = residual + hidden_states
        else:
            router_logits = None
            
        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        if output_router_logits:
            outputs += (router_logits,)

        outputs += (current_kv,)

        return outputs



class FFNDecoderLayer(nn.Module):
    def __init__(self, config: FastSLMConfig, num_experts: int, layer_idx: int):
        super().__init__()

        self.config = config
        
        self.layer_idx = layer_idx

        num_experts_per_tok = config.num_experts_per_tok if num_experts > 1 else 1
        self.moe = FastSLMSparseMoeBlock(config, num_experts=num_experts, num_experts_per_tok=num_experts_per_tok, layer_idx=layer_idx)

        self.pre_moe_layernorm = FastSLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
            
    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: Optional[bool] = False,
            output_router_logits: Optional[bool] = False,
            use_cache: Optional[bool] = False,
            use_swa=False,
            **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
                should not be returned during inference.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
        """

        residual = hidden_states
        if self.pre_moe_layernorm is not None:
            hidden_states = self.pre_moe_layernorm(hidden_states)
        hidden_states, router_logits = self.moe(hidden_states)

        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (None,)

        if use_cache:
            outputs += (None,)

        if output_router_logits:
            outputs += (router_logits,)

        return outputs



class FastSLMMambaDecoderLayer(nn.Module):
    def __init__(self, config: FastSLMConfig, num_experts: int, layer_idx: int):
        super().__init__()

        self.config = config
        self.layer_idx = layer_idx

        self.mamba = Mamba2(config=config, layer_idx=layer_idx)
        
        self.intermediate_size = config.intermediate_size
        if self.intermediate_size > 0:
            num_experts_per_tok = config.num_experts_per_tok if num_experts > 1 else 1
            self.moe = FastSLMSparseMoeBlock(config, num_experts=num_experts, num_experts_per_tok=num_experts_per_tok, layer_idx=layer_idx)
            
        self.input_layernorm = FastSLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        
        if self.intermediate_size > 0:
            self.pre_moe_layernorm = FastSLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.pre_moe_layernorm = None
                
        self.meta_added_flag = False

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
            output_attentions: Optional[bool] = False,
            output_router_logits: Optional[bool] = False,
            use_cache: Optional[bool] = False,
            use_swa=False,
            mamba_inference_params=None,
            **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
                should not be returned during inference.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
        """
        
        if position_ids is not None and position_ids.shape[1] != hidden_states.shape[1]:
            position_ids = torch.arange(hidden_states.shape[1], device=hidden_states.device).unsqueeze(0)

        residual = hidden_states
        
        if self.input_layernorm is not None:
            hidden_states = self.input_layernorm(hidden_states)

        hidden_states, present_key_value = self.mamba(
            hidden_states=hidden_states,
            past_key_value=past_key_value,
            attention_mask=attention_mask,
            inference_params=mamba_inference_params,
        )

        attn_key_value = None

        hidden_states = residual + hidden_states
                
        if self.intermediate_size > 0:
            residual = hidden_states
            
            if self.pre_moe_layernorm is not None:
                hidden_states = self.pre_moe_layernorm(hidden_states)
                
            hidden_states, router_logits = self.moe(hidden_states)

            hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if use_cache:
            outputs += (present_key_value,)

        if output_router_logits:
            outputs += (router_logits,)
        
        outputs += (attn_key_value,)

        return outputs

    def _get_past_seqlen(self, past_key_value, seqlen):
        if past_key_value is None:
            return seqlen
        past_seqlen = past_key_value.get_seq_length(self.layer_idx)

        if past_seqlen == 0:
            return seqlen

        return past_seqlen


    
class FastSLMHybridDecoderLayer(nn.Module):
    def __init__(self, config: FastSLMConfig, num_experts: int, layer_idx: int):
        super().__init__()

        self.config = config
        
        self.layer_idx = layer_idx
                
        if config.hybrid_decoder_layer == 'mamba':
            self.mamba = Mamba2(config=config, layer_idx=layer_idx)
        if config.hybrid_decoder_layer == 'deltanet':
            ## this is to properly handle cache index
            if config.layer_types is not None:
                deltanet_idx = sum(1 for i in range(layer_idx) if config.layer_types[i] == 'deltanet')
            else:
                deltanet_idx = layer_idx
            
            self.gla = DeltaNet(hidden_size=config.hidden_size, num_heads=config.num_attention_heads, layer_idx=deltanet_idx, config=self.config)
        else:
            raise ValueError(f"Not supported: {config.hybrid_decoder_layer}")
    
        self.config = config

        if self.config.intermediate_size > 0:
            num_experts_per_tok = config.num_experts_per_tok if num_experts > 1 else 1 
            self.moe = FastSLMSparseMoeBlock(config, num_experts=num_experts, num_experts_per_tok=num_experts_per_tok, layer_idx=layer_idx)
            self.pre_moe_layernorm = FastSLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.moe = None
            self.pre_moe_layernorm = None

        self.input_layernorm = FastSLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
            

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
            output_attentions: Optional[bool] = False,
            output_router_logits: Optional[bool] = False,
            use_cache: Optional[bool] = False,
            fla_past_key_values = None,
            mamba_inference_params = None,
            use_swa=False,
            **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
                should not be returned during inference.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        if self.config.hybrid_decoder_layer == 'mamba':
            hybrid_op_hidden_states, mamba_present_key_value = self.mamba(
                hidden_states=hidden_states,
                past_key_value=past_key_value,
                attention_mask=attention_mask,
                inference_params=mamba_inference_params,
            )
            
        else:
            hybrid_op_hidden_states, _, fla_past_key_values = self.gla(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                past_key_values=fla_past_key_values,
                use_cache=use_cache,
            )
        
        self_attn_weights = self_attn_present_key_value = current_kv = None

        hidden_states = residual + hybrid_op_hidden_states
            
        if self.moe is not None:
            residual = hidden_states
            hidden_states = self.pre_moe_layernorm(hidden_states)

            hidden_states, router_logits = self.moe(hidden_states)

            hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (self_attn_present_key_value,)

        if output_router_logits:
            outputs += (router_logits,)
            
        outputs += (current_kv,)
        

        return outputs        



# Adapted from transformers.models.mistral.modeling_mistral.MistralPreTrainedModel with Mistral->FastSLM
class FastSLMPreTrainedModel(PreTrainedModel):
    config_class = FastSLMConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["FastSLMAttentionDecoderLayer", "FastSLMMambaDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, (nn.Linear, nn.Conv1d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()

    @staticmethod
    def _convert_to_standard_cache(
            past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]], batch_size: int
    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
        """
        Standardizes the format of the cache so as to match most implementations, i.e. have the seqlen as the third dim
        also for mamba layers
        """
        attn_layer_index = [k.shape == v.shape for k, v in past_key_value].index(True)
        seqlen = past_key_value[attn_layer_index][0].shape[2]
        standard_past_key_value = ()
        for k, v in past_key_value:
            if k.shape != v.shape:
                # mamba layer
                # expand doesn't use more memory, so it's fine to do it here
                standard_past_key_value += ((k.expand(-1, -1, seqlen, -1), v.expand(-1, -1, seqlen, -1)),)
            else:
                standard_past_key_value += ((k, v),)
        return standard_past_key_value

    @staticmethod
    def _convert_to_jamba_cache(
            past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]],
    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
        """
        Converts the cache to the format expected by FastSLM, i.e. dummy seqlen dimesion with size 1 for mamba layers
        """
        jamba_past_key_value = ()
        for k, v in past_key_value:
            if k.shape != v.shape:
                # mamba layer
                jamba_past_key_value += ((k[:, :, :1, :], v[:, :, :1, :]),)
            else:
                jamba_past_key_value += ((k, v),)
        return jamba_past_key_value


# Adapted from transformers.models.mistral.modeling_mistral.MistralModel with MISTRAL->JAMBA, Mistral->FastSLM
class FastSLMModel(FastSLMPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`FastSLMDecoderLayer`]

    Args:
        config: FastSLMConfig
    """

    def __init__(self, config: FastSLMConfig):
        super().__init__(config)

        config.attn_implementation = config.attn_implementation_new
        config._attn_implementation = config.attn_implementation_new

        self.config = config
        
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)

        decoder_layers = []
                        
        layer_type = []
        for i in range(config.num_hidden_layers):
            num_experts = 1
                        
            if config.layer_types[i] in ['deltanet']:
                layer_type.append('m')
                config_new = copy.deepcopy(config)
                config_new.hybrid_decoder_layer = 'deltanet'
                decoder_layer = FastSLMHybridDecoderLayer(config_new, num_experts=num_experts, layer_idx=i)
            elif config.layer_types[i] in ['m', 'm2']:
                layer_type.append('m')
                decoder_layer = FastSLMMambaDecoderLayer(config, num_experts=num_experts, layer_idx=i)
            elif config.layer_types[i] == 'a':
                layer_type.append('a')
                decoder_layer = FastSLMAttentionDecoderLayer(config, num_experts=num_experts, layer_idx=i)
            elif config.layer_types[i] == 'f':
                layer_type.append('a')
                decoder_layer = FFNDecoderLayer(config, num_experts=num_experts, layer_idx=i)
            else:
                raise ValueError(f"Unsupported layer type {config.layer_types[i]}")
            
            decoder_layers.append(decoder_layer)
            
        config.layer_type = layer_type

        if config.sliding_window is not None:
            self.sliding_window = config.sliding_window
            self.global_attn_idx = config.global_attn_idx
        else:
            self.sliding_window = None
            self.global_attn_idx = None

        if not any(isinstance(layer, FastSLMAttentionDecoderLayer) for layer in decoder_layers):
            # raise ValueError("At least one layer in the decoder must be an attention layer")
            self._attn_layer_index = []
        else:
            self._attn_layer_index = [isinstance(layer, FastSLMAttentionDecoderLayer) for layer in decoder_layers].index(
                True
            )

        if not any(isinstance(layer, FastSLMMambaDecoderLayer) for layer in decoder_layers):
            # raise ValueError("At least one layer in the decoder must be a Mamba layer")
            self._mamba_layer_index = []
        else:
            self._mamba_layer_index = [isinstance(layer, FastSLMMambaDecoderLayer) for layer in decoder_layers].index(True)

        # if (
        #         decoder_layers[self._mamba_layer_index].mamba.ssm_state_size
        #         == decoder_layers[self._mamba_layer_index].mamba.conv_kernel_size
        # ):
        #     raise ValueError("Mamba state size and convolution size must be different")

        self.layers = nn.ModuleList(decoder_layers)

        self._attn_implementation = config.attn_implementation
        
        self.final_layernorm = FastSLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        if self.config.num_memory_tokens > 0:
            self.memory_tokens = nn.Parameter(torch.randn(self.config.num_memory_tokens, self.config.hidden_size))
            
        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

        self.has_previous_state = False


    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value


    def forward(
            self,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[Union[List[torch.FloatTensor], HybridMambaAttentionDynamicCache]] = None,
            inputs_embeds: Optional[torch.FloatTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            output_router_logits: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            fla_past_key_values = None,
            mamba_inference_params = None,
    ) -> Union[Tuple, MoeModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )
        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

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

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

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(0, seq_length, dtype=torch.long, device=device
            )
            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
        else:
            if self.config.num_memory_tokens > 0 and past_key_values is not None and not self.has_previous_state:
                position_ids = position_ids.view(-1, seq_length + self.config.num_memory_tokens).long()
            else:
                position_ids = position_ids.view(-1, seq_length).long()
                
        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)
            
        ori_b, ori_n = inputs_embeds.shape[0], inputs_embeds.shape[1]

        if self.config.num_memory_tokens > 0 and (past_key_values is None or not self.has_previous_state):
            mem = repeat(self.memory_tokens, 'n d -> b n d', b = inputs_embeds.shape[0]) # prepend the memory to every segment of m by repeating the memory tokens
            inputs_embeds, mem_packed_shape = pack((mem, inputs_embeds), 'b * d')      

            if position_ids is not None and position_ids.shape[1] != inputs_embeds.shape[1]:
                position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device).unsqueeze(0)

            if attention_mask is not None and attention_mask.shape[1] < inputs_embeds.shape[1]:
                assert attention_mask.shape[1] + self.config.num_memory_tokens == inputs_embeds.shape[1]
                attention_mask = torch.cat([torch.ones(inputs_embeds.shape[0], self.config.num_memory_tokens, device=attention_mask.device), attention_mask], dim=1)
            
   
        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
            if is_padding_right:
                raise ValueError(
                    "You are attempting to perform batched generation with padding_side='right'"
                    " this may lead to unexpected behaviour for Flash Attention version of FastSLM. Make sure to "
                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                )
                
        attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None

        hidden_states = inputs_embeds

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

        for i, decoder_layer in enumerate(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,
                    attention_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    output_router_logits,
                    use_cache,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    output_router_logits=output_router_logits,
                    use_cache=use_cache,
                    use_swa=self.sliding_window is not None and i not in self.global_attn_idx,
                    fla_past_key_values=fla_past_key_values,
                    mamba_inference_params=mamba_inference_params,
                )

            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],)

            if output_router_logits:
                all_router_logits += (layer_outputs[3],)
        
        if self.final_layernorm is not None:
            hidden_states = self.final_layernorm(hidden_states)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if self.config.num_memory_tokens > 0 and (past_key_values is None or not self.has_previous_state):
            mem, hidden_states = unpack(hidden_states, mem_packed_shape, 'b * d')
            hidden_states = hidden_states[:, :ori_n, :]

        if past_key_values is not None and not self.has_previous_state:
            self.has_previous_state = True

        next_cache = None
        if use_cache:
            next_cache = next_decoder_cache

        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
                if v is not None
            )
        return MoeModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if (fla_past_key_values is None and mamba_inference_params is None) else (past_key_values, fla_past_key_values, mamba_inference_params),
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            router_logits=all_router_logits,
        )


# Adapted from transformers.models.mixtral.modeling_mixtral.MixtralForCausalLM with MIXTRAL->JAMBA, Mixtral->FastSLM
class FastSLMForCausalLM(FastSLMPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config: FastSLMConfig):
        super().__init__(config)
        self.config = config
        self.model = FastSLMModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.router_aux_loss_coef = config.router_aux_loss_coef
        self.num_experts = config.num_experts
        self.num_experts_per_tok = config.num_experts_per_tok
        # Initialize weights and apply final processing
        self.post_init()
    
    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model
        
    @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    # Ignore copy
    def 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,
            labels: Optional[torch.LongTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            output_router_logits: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            calc_logits_for_entire_prompt: Optional[bool] = True,
            fla_past_key_values = None,
            mamba_inference_params = None,
    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
        r"""
        Args:
            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

            calc_logits_for_entire_prompt (`bool`, *optional*):
                Whether or not to calculate the logits for the entire prompt, or just the last token. Only last token
                logits are needed for generation, and calculating them only for that token can save memory,
                which becomes pretty significant for long sequences.

        Returns:
        ```"""

        # print(input_ids.max())

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )

        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_router_logits=output_router_logits,
            fla_past_key_values=fla_past_key_values,
            mamba_inference_params=mamba_inference_params,
            return_dict=return_dict,
        )

        hidden_states = outputs[0]
        if calc_logits_for_entire_prompt:
            logits = self.lm_head(hidden_states)
        else:
            logits = self.lm_head(hidden_states[..., -1:, :])

        logits = logits / self.lm_head.weight.norm(p=2, dim=1)
                        
        logits = logits.float()

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)
                 
        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        # print("hidden_states.shape:", hidden_states.shape, "input_ids.shape:", input_ids.shape, "logits.shape:", logits.shape)

        return MoeCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            router_logits=outputs.router_logits,
        )

    def get_init_cache(self, max_seqlen, batch_size=1):
        past_key_values = HybridMambaAttentionDynamicCache(
            self.config, batch_size, self.dtype, device=self.device, layer_type=self.config.layer_type
        )

        mamba_inference_params = InferenceParams(max_seqlen=max_seqlen, max_batch_size=batch_size)

        fla_past_key_values = fla_cache.from_legacy_cache(None)    

        return past_key_values, fla_past_key_values, mamba_inference_params
    
    
    def init_cuda_graph_generation(
        self,
        max_new_tokens=128,
        batch_size=1,
        device=None,
    ):
        """
        Initialize CUDA graph for generation with proper cache handling and warmup.
        This function should be called once before generation to set up the graph.
        
        Args:
            max_new_tokens: Maximum number of new tokens to generate
            batch_size: Batch size for generation
            device: Device to use (defaults to model device)
            
        Returns:
            generation_state: Dictionary containing all necessary state for generation
        """
        if device is None:
            device = next(self.parameters()).device
            
        self.eval()
        
        # Initialize caches
        max_seqlen = max_new_tokens + 2048 + self.config.num_memory_tokens  # Add buffer for input
        past_key_values, fla_past_key_values, mamba_inference_params = self.get_init_cache(
            max_seqlen=max_seqlen, batch_size=batch_size
        )
        
        # Initialize KV caches for all modules
        for module in self.modules():
            if hasattr(module, 'init_kv_cache'):
                module.init_kv_cache(max_batch_size=batch_size, max_seq_len=max_seqlen)
        
        with torch.no_grad():
            # Warmup runs
            dummy_input = torch.ones((batch_size, 10), dtype=torch.long, device=device)
            for _ in range(10):
                self(dummy_input)
            
            # Prepare static tensors for CUDA graph
            static_current_input = torch.zeros((batch_size, 1), dtype=torch.long, device=device)
            static_position_ids = torch.zeros((batch_size, 1), dtype=torch.long, device=device)
            static_logits = torch.zeros((batch_size, self.config.vocab_size), device=device)
            
            # Set up for graph capture
            self.model.has_previous_state = True
            if mamba_inference_params is not None:
                mamba_inference_params.seqlen_offset = 1
            
            # Warmup runs for graph capture
            for _ in range(10):
                model_kwargs_warmup = {
                    'input_ids': static_current_input,
                    'fla_past_key_values': fla_past_key_values,
                    'mamba_inference_params': mamba_inference_params,
                    'past_key_values': past_key_values,
                    'use_cache': True,
                    'position_ids': static_position_ids,
                }
                warmup_outputs = self(**model_kwargs_warmup)
            
            # Capture CUDA graph
            generation_graph = CUDAGraph()
            with torch.cuda.graph(generation_graph):
                model_kwargs_graph = {
                    'input_ids': static_current_input,
                    'fla_past_key_values': fla_past_key_values,
                    'mamba_inference_params': mamba_inference_params,
                    'past_key_values': past_key_values,
                    'use_cache': True,
                    'position_ids': static_position_ids,
                }
                graph_outputs = self(**model_kwargs_graph)
                static_logits.copy_(graph_outputs.logits[:, -1, :])

        if fla_past_key_values is not None:
            fla_past_key_values.reset()
        
        if mamba_inference_params is not None:
            mamba_inference_params.reset(mamba_inference_params.max_seqlen, mamba_inference_params.max_batch_size)
            for key in mamba_inference_params.key_value_memory_dict:
                conv_state, ssm_state = mamba_inference_params.key_value_memory_dict[key]
                conv_state.zero_()
                ssm_state.zero_()
        
        for module in self.modules():
            if hasattr(module, 'reset_kv_cache'):
                module.reset_kv_cache()
        
        self.model.has_previous_state = False

        # Return generation state
        generation_state = {
            'generation_graph': generation_graph,
            'static_current_input': static_current_input,
            'static_position_ids': static_position_ids,
            'static_logits': static_logits,
            'past_key_values': past_key_values,
            'fla_past_key_values': fla_past_key_values,
            'mamba_inference_params': mamba_inference_params,
            'max_seqlen': max_seqlen,
            'batch_size': batch_size,
            'device': device,
        }
        
        return generation_state
    
    def generate_with_cuda_graph(
        self,
        input_ids,
        generation_state,
        max_new_tokens=128,
        temperature=1.0,
        top_k=0,
        top_p=0.9,
        eos_token_id=None,
        verbose=False,
        profiling=False,
        multi_round=False,
    ):
        """
        Generate text using pre-initialized CUDA graph state.
        
        Args:
            input_ids: Input token IDs tensor of shape (batch_size, seq_len)
            generation_state: State dictionary returned by init_cuda_graph_generation
            max_new_tokens: Maximum number of new tokens to generate
            temperature: Sampling temperature (0 for greedy)
            top_k: Top-k filtering (0 to disable)
            top_p: Top-p filtering (1.0 to disable)
            eos_token_id: End-of-sequence token ID
            pad_token_id: Padding token ID
            verbose: Whether to print generated tokens
            profiling: Whether to return timing information
            
        Returns:
            generated_ids: Tensor of shape (batch_size, input_len + generated_len)
            or decode_latency if profiling=True
        """
        self.eval()
        batch_size = input_ids.shape[0]
        device = input_ids.device
        
        # Extract state
        generation_graph = generation_state['generation_graph']
        static_current_input = generation_state['static_current_input']
        static_position_ids = generation_state['static_position_ids']
        static_logits = generation_state['static_logits']
        past_key_values = generation_state['past_key_values']
        fla_past_key_values = generation_state['fla_past_key_values']
        mamba_inference_params = generation_state['mamba_inference_params']
        
        with torch.no_grad():
            if not multi_round or mamba_inference_params.seqlen_offset == 0:
                if fla_past_key_values is not None:
                    fla_past_key_values.reset()
                
                if mamba_inference_params is not None:
                    mamba_inference_params.reset(mamba_inference_params.max_seqlen, mamba_inference_params.max_batch_size)
                    for key in mamba_inference_params.key_value_memory_dict:
                        conv_state, ssm_state = mamba_inference_params.key_value_memory_dict[key]
                        conv_state.zero_()
                        ssm_state.zero_()
                
                for module in self.modules():
                    if hasattr(module, 'reset_kv_cache'):
                        module.reset_kv_cache()
                
                self.model.has_previous_state = False
            
                # Prefill phase - process input sequence
                position_ids = torch.arange(
                    self.config.num_memory_tokens + input_ids.shape[1], dtype=torch.long, device=device
                ).unsqueeze(0).expand(batch_size, -1)

            else:
                # Prefill phase - process input sequence
                position_ids = torch.arange(
                    mamba_inference_params.seqlen_offset, mamba_inference_params.seqlen_offset + input_ids.shape[1], dtype=torch.long, device=device
                ).unsqueeze(0).expand(batch_size, -1)
            
            current_input = input_ids
            
            model_kwargs = {
                'input_ids': current_input,
                'past_key_values': past_key_values,
                'fla_past_key_values': fla_past_key_values,
                'mamba_inference_params': mamba_inference_params,
                'use_cache': True,
                'position_ids': position_ids,
            }
            
            if profiling:
                torch.cuda.synchronize()
                t1 = time.time()
            
            # Forward pass for prefill
            outputs = self(**model_kwargs)
            
            if mamba_inference_params is not None:
                if mamba_inference_params.seqlen_offset == 0:
                    mamba_inference_params.seqlen_offset = current_input.shape[1] + self.config.num_memory_tokens
                else:
                    mamba_inference_params.seqlen_offset += current_input.shape[1]

            static_position_ids.fill_(position_ids[0, -1])
            
            logits = outputs.logits[:, -1, :]  # (batch_size, vocab_size)
            generated_tokens = []
            
            # Generation loop using CUDA graph replay
            for step in range(max_new_tokens):
                # Sample next token using current logits
                if temperature == 0:
                    next_token = torch.argmax(logits, dim=-1, keepdim=True)
                else:
                    next_token = sample_token(logits, temperature=temperature, top_k=top_k, top_p=top_p)
                
                generated_tokens.append(next_token)
                
                # Check for EOS
                if not profiling and eos_token_id is not None and (next_token == eos_token_id).all():
                    if verbose:
                        print("\nEOS reached")
                    break
                
                # Update static tensors for graph replay
                static_current_input.copy_(next_token)
                static_position_ids.add_(1)
                
                # Replay the captured graph
                generation_graph.replay()
                
                if mamba_inference_params is not None:
                    mamba_inference_params.seqlen_offset += 1
                
                logits = static_logits.clone()
            
            generated_ids = torch.cat([input_ids] + generated_tokens, dim=1)
            
            if profiling:
                torch.cuda.synchronize()
                t2 = time.time()
                decode_latency = t2 - t1
                return generated_ids, decode_latency
            
            return generated_ids


    def prepare_inputs_for_generation(
            self,
            input_ids,
            past_key_values=None,
            attention_mask=None,
            inputs_embeds=None,
            output_router_logits=False,
            **kwargs,
    ):         
        if self.config.num_memory_tokens > 0:
            attention_mask = torch.cat([torch.ones(input_ids.shape[0], self.config.num_memory_tokens, device=attention_mask.device), attention_mask], dim=1)

        past_key_values = None   # Disable cache for now

        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            position_ids = position_ids[:, -input_ids.shape[1]:]
                
        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None:
            if input_ids.shape[1] == 0:
                model_inputs = {"inputs_embeds": inputs_embeds}
            else:
                inputs_embeds_new = self.model.embed_tokens(input_ids)
                model_inputs = {"inputs_embeds": torch.cat([inputs_embeds, inputs_embeds_new], dim=1)}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
            }
        )
        return model_inputs


def sample_token(logits, temperature=1.0, top_k=0, top_p=0.9):
    """
    Sample a token from logits with temperature, top-k, and top-p filtering.
    This matches the implementation in fast_slm_gen.py for consistency.
    
    Args:
        logits: Tensor of shape (batch_size, vocab_size)
        temperature: Sampling temperature
        top_k: Top-k filtering (0 to disable)
        top_p: Top-p filtering (1.0 to disable)
        
    Returns:
        next_token: Tensor of shape (batch_size, 1)
    """
    if temperature == 0:
        return torch.argmax(logits, dim=-1, keepdim=True)
    
    logits = logits / temperature
    
    # Top-k filtering - match fast_slm_gen.py implementation
    if top_k > 0:
        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
        logits.masked_fill_(indices_to_remove, float('-inf'))
    
    # Top-p filtering - match fast_slm_gen.py implementation
    if top_p < 1.0:
        sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
        
        # Remove tokens with cumulative probability above the threshold
        sorted_indices_to_remove = cumulative_probs > top_p
        # Shift the indices to the right to keep also the first token above the threshold
        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
        sorted_indices_to_remove[..., 0] = 0
        
        indices_to_remove = sorted_indices_to_remove.scatter(-1, sorted_indices, sorted_indices_to_remove)
        logits.masked_fill_(indices_to_remove, float('-inf'))
    
    probs = F.softmax(logits, dim=-1)
    return torch.multinomial(probs, num_samples=1)