# fmt: off

# Adaptation recipe lifted from Jonas et al. :>
# https://github.com/seal-rg/recurrent-pretraining/blob/main/recpre/raven_modeling_minimal.py

# coding=utf-8
# Copyright 2022 EleutherAI 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.
from typing import Callable, Optional, Union
import time

import torch
from torch import nn
import torch.nn.functional as F

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.integrations import use_kernel_forward_from_hub
from transformers.masking_utils import create_causal_mask
from transformers.modeling_layers import (
    GenericForQuestionAnswering,
    GenericForSequenceClassification,
    GenericForTokenClassification,
    GradientCheckpointingLayer,
)
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
)
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
from transformers.utils.deprecation import deprecate_kwarg
from transformers.utils.generic import check_model_inputs
from .configuration_llama import LlamaConfig


# Glue
from transformers.generation.utils import GenerateDecoderOnlyOutput


logger = logging.get_logger(__name__)


@use_kernel_forward_from_hub("RMSNorm")
class LlamaRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        LlamaRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        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)
        return self.weight * hidden_states.to(input_dtype)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


class LlamaRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: LlamaConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

        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)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class LlamaMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


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)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


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

    def __init__(self, config: LlamaConfig, layer_idx: int):
        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=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_values: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, 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_values is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        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,
            **kwargs,
        )

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


class LlamaDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: LlamaConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = LlamaAttention(config=config, layer_idx=layer_idx)

        self.mlp = LlamaMLP(config)
        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        # Self Attention
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states


@auto_docstring
class LlamaPreTrainedModel(PreTrainedModel):
    config: LlamaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["LlamaDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": LlamaDecoderLayer,
        "attentions": LlamaAttention,
    }


@auto_docstring
class LlamaModel(LlamaPreTrainedModel):
    def __init__(self, config: LlamaConfig):
        super().__init__(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)
        self.layers = nn.ModuleList(
            [LlamaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = LlamaRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

        # Initialize weights and apply final processing
        self.post_init()

    @check_model_inputs()
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position: torch.Tensor = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

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

        causal_mask = create_causal_mask(
            config=self.config,
            input_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )

        hidden_states = inputs_embeds
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


@auto_docstring
class LlamaForCausalLM(LlamaPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

    def __init__(self, config):
        super().__init__(config)
        self.model = LlamaModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

        print(f"Loading local LlamaForCausalLM!")

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from transformers import AutoTokenizer, LlamaForCausalLM

        >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        outputs: BaseModelOutputWithPast = 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,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

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

    @torch.inference_mode()
    def generate(self, input_ids: Optional[torch.LongTensor] = None, *args, **kwargs):
        """
        Custom generate entry point.
        
        If `do_mtp=True` is passed, strictly enforces MTP-only arguments and routes to `_mtp_generate`.
        Otherwise, routes to standard Hugging Face `generate`.
        """
        
        # 1. Standard Path
        if not kwargs.get("do_mtp", False):
            print("Executing standard generate() codepath.")
            return super().generate(input_ids, *args, **kwargs)

        # 2. MTP Path
        print("Executing custom MTP generation codepath!")
        
        # Handle input_ids: HF can pass it as positional (first arg) or keyword
        # We consolidate it into 'prompt' for the MTP signature
        prompt = input_ids if input_ids is not None else kwargs.pop("input_ids", None)
        if prompt is None:
            # If standard generate was called without input_ids, it might be in *args or handled deeper, 
            # but for MTP we require it explicitly.
            raise ValueError("MTP generation requires 'input_ids' to be passed.")

        # --- Argument Extraction & Strict Validation ---

        # Keys strictly allowed for MTP (will be passed to _mtp_generate)
        mtp_allowed_keys = {
            "do_mtp", "k_toks", "mask_id", "strategy", "return_mtp_result_dict", "include_prompt", "streamer"
        }
        
        # Keys from HF that we know how to map to MTP equivalents
        # max_length -> max_returned_tokens
        max_length = kwargs.pop("max_length", None)
        max_returned_tokens = kwargs.pop("max_returned_tokens", None)
        if max_returned_tokens is None and max_length is not None:
             print(f"Renaming max_length={max_length} to max_returned_tokens for MTP generation.")
             max_returned_tokens = max_length
        
        # eos_token_id -> eos_id
        eos_token_id = kwargs.pop("eos_token_id", None)
        eos_id = kwargs.pop("eos_id", None)
        if eos_id is None:
            eos_id = eos_token_id

        # Standard HF args that we SILENTLY IGNORE because they are passed automatically 
        # by the GenerationMixin but are not relevant or supported in this MTP implementation.
        ignored_hf_keys = {
            "attention_mask", "use_cache", "do_sample", "stopping_criteria", 
            "pad_token_id", "logits_processor", "max_new_tokens", "generation_config",
        }

        # Check for explicit incompatibility
        if kwargs.get("do_sample", False):
             raise ValueError("MTP generation does not support sampling (do_sample=True).")

        # Extract valid MTP args
        mtp_kwargs = {}
        for k in list(kwargs.keys()):
            if k in mtp_allowed_keys:
                mtp_kwargs[k] = kwargs.pop(k)

        # Remove ignored HF keys
        for k in list(kwargs.keys()):
            if k in ignored_hf_keys:
                kwargs.pop(k)

        # FAIL LOUDLY if anything is left in kwargs
        if kwargs:
            raise ValueError(
                f"Unsupported argument(s) passed to MTP generate: {list(kwargs.keys())}.\n"
                f"When do_mtp=True, only these args are supported: {list(mtp_allowed_keys) + ['max_returned_tokens', 'eos_id']}."
            )

        # Pre-flight checks
        if len(prompt.shape) > 1 and prompt.shape[0] > 1:
             raise NotImplementedError("MTP generation currently only supports single-example generation (no batching).")

        # Execute Unified Implementation
        # Note: We remove 'do_mtp' from kwargs before passing, as the impl doesn't need it
        mtp_kwargs.pop("do_mtp", None)
        
        return self._mtp_generate(
            prompt=prompt,
            max_returned_tokens=max_returned_tokens,
            eos_id=eos_id,
            **mtp_kwargs
        )
    
    @torch.inference_mode()
    def _mtp_generate(
        self,
        prompt: torch.Tensor,
        max_returned_tokens: int = None,
        k_toks: int = 1,
        mask_id: int = None,
        eos_id: Optional[Union[int, list]] = None,
        include_prompt: bool = True,
        streamer = None,
        strategy: Optional[list] = None,
        return_mtp_result_dict: bool = False,
    ):
        """
        Implementation of MTP generation logic.
        """
        # --- Setup Stop Tokens ---
        if isinstance(eos_id, int):
            stop_tokens = ([eos_id,],)
        elif isinstance(eos_id, list):
            assert all(isinstance(eid, int) for eid in eos_id), "If eos_id is a list, all elements must be ints."
            stop_tokens = tuple([list([eid,]) for eid in eos_id])
        elif eos_id is None:
            stop_tokens = ()
        else:
            raise ValueError(f"eos_id must be None, int, or list of lists, got {type(eos_id)}")

        # --- Validation ---
        if k_toks > 1: assert mask_id is not None, "mask_id must be provided when k_toks > 1"

        input_ids = prompt.clone()
        prompt_size = prompt.size(1)
        device = prompt.device
        
        # Get generation config (defaulting to model's if not present)
        generation_config = self.generation_config

        # --- Streaming Prompt ---
        if include_prompt:
            if streamer is not None:
                print(f"\n<BEGIN Streaming Prompt>", flush=True)
                streamer.put(input_ids)
                print(f"\n<END Streaming Prompt>", flush=True)
        
        if streamer is not None:
            print(f"\n<BEGIN Streaming Generation>", flush=True)

        stop_progress = [0] * len(stop_tokens)

        # --- Generation Loop ---
        t0_prefill = time.perf_counter()
        t1_prefill = None
        t0_gen = None
        t1_gen = None
        toks_pre_prefill = input_ids.shape[1]
        toks_post_prefill = None
        current_idx = input_ids.shape[1]
        num_fwd_evals = 0
        effective_k_values = []
        
        # Prepare kwargs for the inner loop
        model_kwargs = {}

        while current_idx + k_toks <= max_returned_tokens:

            # 0 is prefill, 1 is first step which can include compile time, then 2 is steady state
            if (t0_gen is None and num_fwd_evals == 2):  
                t1_prefill = time.perf_counter()
                t0_gen = time.perf_counter()
                toks_post_prefill = input_ids.shape[1]
            
            # Generate the token
            if k_toks > 1: 
                input_ids = self._extend_w_mask(input_ids=input_ids, k_toks=k_toks, mask_id=mask_id)

            # first step prep
            if num_fwd_evals == 0:
                model_kwargs, generation_config = self._prep_generate_args(
                    self,
                    input_ids,
                    generation_config,
                )
                assert "token_type_ids" not in model_kwargs
                assert "attention_mask" not in model_kwargs
                assert "decoder_attention_mask" not in model_kwargs

            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            next_tokens, model_outputs = self._mtp_next_tokens(
                self,
                model_inputs,
                k_toks=k_toks,
                strategy=strategy,
            )
            effective_k_values.append(next_tokens.shape[1])
        
            if streamer is not None: streamer.put(next_tokens)

            # remove the masks if any
            if k_toks > 1:
                input_ids = input_ids[:, :-(k_toks - 1)]
            input_ids = torch.cat([input_ids, next_tokens], dim=-1)
            
            # Update cache / model kwargs
            model_kwargs["past_key_values"] = model_outputs.past_key_values
            
            if strategy is None:
                if num_fwd_evals == 0:
                    # this is the end of the prefill step
                    if k_toks > 1:
                        model_kwargs["cache_position"] = torch.arange(
                            prompt_size,
                            prompt_size + k_toks + (k_toks - 1),
                            device=device,
                            dtype=torch.int64,
                        )
                    else:
                        model_kwargs["cache_position"] = torch.tensor(
                            [prompt_size], device=device, dtype=torch.int64
                        )
                else:
                    model_kwargs["cache_position"].add_(k_toks)
            else:  # we can assume that all strats produce variable number of tokens
                num_new_tokens = next_tokens.shape[1]
                if num_fwd_evals == 0:
                    # this is the end of the prefill step
                    if k_toks > 1:
                        model_kwargs["cache_position"] = torch.arange(
                            prompt_size,
                            prompt_size + num_new_tokens + (k_toks - 1),
                            device=device,
                            dtype=torch.int64,
                        )
                    else:
                        model_kwargs["cache_position"] = torch.tensor(
                            [prompt_size], device=device, dtype=torch.int64
                        )
                else:
                    if k_toks > 1:
                        recomputation_positions = model_kwargs["cache_position"][: -(k_toks - 1)]
                        previous_num_new_tokens = recomputation_positions.size(0)
                        new_start_pos = recomputation_positions[0] + previous_num_new_tokens
                        model_kwargs["cache_position"] = torch.arange(
                            new_start_pos,
                            new_start_pos + num_new_tokens + (k_toks - 1),
                            device=device,
                            dtype=torch.int64,
                        )
                    else:
                        model_kwargs["cache_position"].add_(1)
            
            current_idx += next_tokens.shape[1]
            num_fwd_evals += 1
            
            # Crop cache
            model_kwargs["past_key_values"].crop(model_kwargs["cache_position"][0])

            # Check for stop sequences
            hit_stop_seq = False
            for int_tok in next_tokens.tolist()[0]: # assuming batch size 1
                for i, seq in enumerate(stop_tokens):
                    if int_tok == seq[stop_progress[i]]:
                        stop_progress[i] += 1
                        if stop_progress[i] == len(seq):
                            hit_stop_seq = True
                            break
                    else:
                        stop_progress[i] = 0
                if hit_stop_seq:
                    break
            if hit_stop_seq:
                break

        # End of generation loop
        if streamer is not None:
            streamer.end()
            print(f"\n<END Streaming Generation>", flush=True)
        
        t1_gen = time.perf_counter()
        
        # Calculate stats
        if t1_prefill is not None and t0_gen is not None and toks_post_prefill is not None:
            t_prefill = t1_prefill - t0_prefill
            t_gen = t1_gen - t0_gen
            tokens_generated = input_ids.shape[1] - toks_post_prefill
            toks_gend_incl_prefillplus1 = input_ids.shape[1] - toks_pre_prefill
        else:
            t_prefill = t1_gen - t0_prefill
            t_gen = t1_gen - t0_prefill
            tokens_generated = input_ids.shape[1] - toks_pre_prefill
            toks_gend_incl_prefillplus1 = input_ids.shape[1] - toks_pre_prefill
            
        print(
            f"Using a total of {f'1+1+{num_fwd_evals-2}' if t0_gen is not None else f'{num_fwd_evals}'} forward evals, time for prefill plus first/compilation step {f'(1+1)' if t0_gen is not None else ''} was {t_prefill:.02f} sec, generation time was {t_gen:.02f} sec @ {tokens_generated / t_gen:.02f} tokens/sec {f'steady state' if t0_gen is not None else ''} over {tokens_generated} tokens.",
            flush=True,
        )
        print(f"Strategy used: {strategy}", flush=True)
        avg_effective_k = sum(effective_k_values) / len(effective_k_values) if effective_k_values else 0
        print(
            f"Average effective k_toks over generation: {avg_effective_k:.02f}, full array of effective k_toks: {effective_k_values}",
            flush=True,
        )

        if generation_config.return_dict_in_generate:
            raise NotImplementedError(f"Only basic return type implemented for MTP generate.")
        
        if return_mtp_result_dict:
            token_ids = input_ids if include_prompt else input_ids[:,prompt_size:]

            # testing for generation vs. aux data order integrity
            import hashlib
            leading_toks = token_ids[0,prompt_size:prompt_size+10].tolist()
            leading_toks_hash = hashlib.shake_128(str(leading_toks).encode()).hexdigest(4)
            
            mtp_result_dict = {
                "token_ids":token_ids,
                "leading_toks_hash": leading_toks_hash,
                "num_fwd_evals": num_fwd_evals,
                "t_prefill": t_prefill,
                "t_gen": t_gen,
                "tokens_generated": tokens_generated,
                "toks_gend_incl_prefillplus1": toks_gend_incl_prefillplus1,
                "avg_effective_k": avg_effective_k,
                "effective_k_values": effective_k_values,
                "tps": tokens_generated / t_gen if t_gen > 0 else 0.0,
            }
            return mtp_result_dict
        
        return input_ids if include_prompt else input_ids[:,prompt_size:]
    
    @torch.inference_mode()
    def _prep_generate_args(
        self,
        model,
        input_ids: torch.Tensor,
        generation_config = None,
        model_kwargs: dict = None,
    ):
        # Setup
        if model_kwargs is None:
            model_kwargs = {}

        if generation_config is None:
            generation_config = model.generation_config
        model_kwargs["use_cache"] = True
        if "past_key_values" in model_kwargs:
            print(f"Before _get_initial_cache_position past_key_values cache length: {model_kwargs['past_key_values'].layers[0].get_seq_length()}", flush=True)
        model_kwargs = model._get_initial_cache_position(input_ids.shape[1], input_ids.device, model_kwargs)
        if "past_key_values" in model_kwargs:
            print(f"After _get_initial_cache_position past_key_values cache length: {model_kwargs['past_key_values'].layers[0].get_seq_length()}", flush=True)
        return model_kwargs, generation_config
    
    @torch.inference_mode()
    def _top1_confidence(
        self,
        logits: torch.Tensor = None
    ):
        probs = F.softmax(logits, dim=-1) # LxV
        top1_idx = torch.argmax(probs, dim=-1) # Lx1
        top1_confs = probs[torch.arange(probs.shape[0], device=probs.device), top1_idx] # Lx1
        return top1_confs

    @torch.inference_mode()
    def _extend_w_mask(
        self,
        input_ids=None, 
        k_toks=None, 
        mask_id=None
    ):
        bsz, _ = input_ids.shape

        if k_toks - 1 > 0:
            mask_tensor = (
                torch.ones((bsz, k_toks - 1), dtype=torch.int64, device=input_ids.device)
                * mask_id
            )
            return torch.cat([input_ids, mask_tensor], dim=-1)

        return input_ids

    @torch.inference_mode()
    def _mtp_next_tokens(
        self,
        model,
        model_inputs,
        k_toks: int = 1,
        strategy: Optional[list] = None,
    ) -> torch.Tensor:
        outputs = model(**model_inputs)
        # logits = outputs.logits[:, -k_toks:, :]
        logits = outputs.logits[0, -k_toks:, :]
        if strategy is None:
            _next = torch.argmax(logits, dim=-1, keepdim=False)
        elif strategy[0] == "conf_adapt" or strategy[0] == "conf_adapt_sample@1":
            # we compute the position wise confidences using the _top1_confidence function
            top1_conf = self._top1_confidence(logits)
            # print(f"top1_conf: {top1_conf}", flush=True)
            # now we compute the position of the farthest token geq the threshold
            # but contiguously, so if we have [0.95, 0.92, 0.85, 0.97] and threshold 0.9
            # we want to get position 1 not 3, since position 2 is below the threshold
            # also being careful of situation like [0.85, 0.88, 0.95] where nothing meets the threshold
            # falling back to the first token in that case
            threshold = strategy[1]
            lt_thresh_mask = top1_conf < threshold
            # now we find the first case where the mask is true, and go back one position
            if torch.all(~lt_thresh_mask):
                # then all positions are above the threshold, we take the last position
                last_pos = k_toks - 1
            else:
                last_pos = torch.argmax(lt_thresh_mask.int()).item() - 1
                if last_pos < 0:
                    last_pos = 0

            # print(f"last_pos: {last_pos}", flush=True)

            # then we slice the logits to only keep up to that position
            logits = logits[: last_pos + 1]

            if last_pos == 0 and strategy[0] == "conf_adapt_sample@1":
                # if k is 1 this step, then draw from the distribution
                probs = torch.softmax(logits, dim=-1)
                # print(f"Sampling from probs at k={logits.size(0)}: {probs.shape}", flush=True)
                temperature = strategy[2]
                if 0.0 < temperature < 1.0:
                    probs = probs.pow(1.0 / temperature)
                    probs = probs / probs.sum(dim=-1, keepdim=True)
                else:
                    print(f"Using temperature={temperature} has no effect.", flush=True)
                _next = torch.multinomial(probs, num_samples=1).squeeze(0)
                
            else:
                _next = torch.argmax(logits, dim=-1, keepdim=False)
        elif strategy[0] == "random":
            sampling_weights = strategy[1]
            # we sample k for this step according to the provided weights
            k_toks = int(
                torch.multinomial(torch.tensor(sampling_weights), num_samples=1).item()
            )
            logits = logits[: k_toks + 1]
            _next = torch.argmax(logits, dim=-1, keepdim=False)
        else:
            raise ValueError(f"Unknown strategy: {strategy}")

        _next = _next.unsqueeze(0)  # add batch dim back
        return _next, outputs


class LlamaForSequenceClassification(GenericForSequenceClassification, LlamaPreTrainedModel): ...


class LlamaForQuestionAnswering(GenericForQuestionAnswering, LlamaPreTrainedModel):
    base_model_prefix = "transformer"  # For BC, where `transformer` was used instead of `model`


class LlamaForTokenClassification(GenericForTokenClassification, LlamaPreTrainedModel): ...


#################################### HF registration ############################################################

from transformers import AutoConfig, AutoModel, AutoModelForCausalLM

LlamaConfig.register_for_auto_class()

LlamaForCausalLM.register_for_auto_class("AutoModel")
LlamaForCausalLM.register_for_auto_class("AutoModelForCausalLM")

__all__ = [
    "LlamaForCausalLM",
    "LlamaModel",
    "LlamaPreTrainedModel",
    "LlamaForSequenceClassification",
    "LlamaForQuestionAnswering",
    "LlamaForTokenClassification",
]