from typing import Optional, Tuple, List
import flax.linen as nn
import jax
import numpy as np
import jax.numpy as jnp
from jax import lax
from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
from flax.linen.attention import dot_product_attention_weights
from flax.traverse_util import flatten_dict, unflatten_dict
from flax.linen import combine_masks, make_causal_mask
from transformers.models.gpt2.modeling_flax_gpt2 import (
    FlaxConv1D,
    FlaxGPT2MLP,
    GPT2Config,
    FlaxBaseModelOutputWithPastAndCrossAttentions,
    FlaxCausalLMOutputWithCrossAttentions,
)
from transformers.modeling_flax_utils import ACT2FN, FlaxPreTrainedModel
from typing import Callable, Literal
def print_model(flax_params, file=None):
    flat_params = flatten_dict(flax_params)
    for path, value in flat_params.items():
        name = "/".join(path)
        if hasattr(value, "shape"):
            line = f"{name} {value.shape}"
        else:
            line = f"{name} {type(value)}"
        if file:
            print(line, file=file)
        else:
            print(line)
 
def create_sinusoidal_positions(num_pos, dim):
    inv_freq = 1.0 / (10000 ** (np.arange(0, dim, 2) / dim))
    sinusoid_inp = np.einsum("i , j -> i j", np.arange(num_pos), inv_freq).astype("float32")
    sin, cos = np.sin(sinusoid_inp), np.cos(sinusoid_inp)
 
    sentinel = dim // 2 + dim % 2
    out = np.zeros((num_pos, dim))
    out[:, 0:sentinel] = sin
    out[:, sentinel:] = cos
 
    return jnp.array(out)
 
 
def rotate_every_two(tensor):
    rotate_half_tensor = jnp.stack((-tensor[:, :, :, 1::2], tensor[:, :, :, ::2]), axis=-1)
    rotate_half_tensor = rotate_half_tensor.reshape(rotate_half_tensor.shape[:-2] + (-1,))
    return rotate_half_tensor
 
 
def apply_rotary_pos_emb(tensor, sincos):
    sin_pos, cos_pos = sincos
    sin_pos = sin_pos[:, :, None, :].repeat(2, 3)
    cos_pos = cos_pos[:, :, None, :].repeat(2, 3)
    return (tensor * cos_pos) + (rotate_every_two(tensor) * sin_pos)
 
 
class LMCFlaxGPT2Attention(nn.Module):
    config: GPT2Config
    dtype: jnp.dtype = jnp.float32
    causal: bool = True
    is_cross_attention: bool = False
    def setup(self):
        config = self.config
        self.embed_dim = config.hidden_size
        self.num_heads = config.n_head
        self.head_dim = self.embed_dim // self.num_heads
        if self.config.position_embeddings == "rope":
            self.rotary_dim = self.head_dim
        if self.is_cross_attention:
            self.c_attn = FlaxConv1D(2 * self.embed_dim, dtype=self.dtype)
            self.q_attn = FlaxConv1D(self.embed_dim, dtype=self.dtype)
        else:
            self.c_attn = FlaxConv1D(3 * self.embed_dim, dtype=self.dtype)
        self.c_proj = FlaxConv1D(self.embed_dim, dtype=self.dtype)
        self.resid_dropout = nn.Dropout(rate=config.resid_pdrop)
        if self.causal:
            self.causal_mask = make_causal_mask(
                jnp.ones((1, config.max_position_embeddings), dtype="bool"), dtype="bool"
            )
        if self.config.position_embeddings == "rope":
            pos_embd_dim = self.rotary_dim or self.embed_dim
            self.embed_positions = create_sinusoidal_positions(config.max_position_embeddings, pos_embd_dim)
    def _split_heads(self, hidden_states):
        return hidden_states.reshape(hidden_states.shape[:2] + (self.num_heads, self.head_dim))
    def _merge_heads(self, hidden_states):
        return hidden_states.reshape(hidden_states.shape[:2] + (self.embed_dim,))
    @nn.compact
    def _concatenate_to_cache(self, key, value, query, attention_mask):
        """
        This function takes projected key, value states from a single input token and concatenates the states to cached
        states from previous steps. This function is slightly adapted from the official Flax repository:
        https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
        """
        # detect if we're initializing by absence of existing cache data.
        is_initialized = self.has_variable("cache", "cached_key")
        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
        if is_initialized:
            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
            # update key, value caches with our new 1d spatial slices
            cur_index = cache_index.value
            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
            key = lax.dynamic_update_slice(cached_key.value, key, indices)
            value = lax.dynamic_update_slice(cached_value.value, value, indices)
            cached_key.value = key
            cached_value.value = value
            num_updated_cache_vectors = query.shape[1]
            cache_index.value = cache_index.value + num_updated_cache_vectors
            # causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
            pad_mask = jnp.broadcast_to(
                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
            )
            attention_mask = combine_masks(pad_mask, attention_mask)
        return key, value, attention_mask
 
    def __call__(
        self,
        hidden_states,
        key_value_states: Optional[jnp.ndarray] = None,
        attention_mask=None,
        position_ids=None,
        deterministic: bool = True,
        init_cache: bool = False,
        output_attentions: bool = False,
    ):
        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None
        batch_size = hidden_states.shape[0]
        if not is_cross_attention:
            qkv_out = self.c_attn(hidden_states)
            query, key, value = jnp.split(qkv_out, 3, axis=2)
        else:
            q_out = self.q_attn(hidden_states)
            (query,) = jnp.split(q_out, 1, axis=2)
            kv_out = self.c_attn(key_value_states)
            key, value = jnp.split(kv_out, 2, axis=2)
        query = self._split_heads(query)
        key = self._split_heads(key)
        value = self._split_heads(value)
        if self.config.position_embeddings == "rope":
            sincos = jnp.take(self.embed_positions, position_ids, axis=0)
            sincos = jnp.split(sincos, 2, axis=-1)
            if self.rotary_dim is not None:
                k_rot = key[:, :, :, : self.rotary_dim]
                k_pass = key[:, :, :, self.rotary_dim :]
 
                q_rot = query[:, :, :, : self.rotary_dim]
                q_pass = query[:, :, :, self.rotary_dim :]
 
                k_rot = apply_rotary_pos_emb(k_rot, sincos)
                q_rot = apply_rotary_pos_emb(q_rot, sincos)
 
                key = jnp.concatenate([k_rot, k_pass], axis=-1)
                query = jnp.concatenate([q_rot, q_pass], axis=-1)
            else:
                key = apply_rotary_pos_emb(key, sincos)
                query = apply_rotary_pos_emb(query, sincos)
        query_length, key_length = query.shape[1], key.shape[1]
        if self.causal:
            if self.has_variable("cache", "cached_key"):
                mask_shift = self.variables["cache"]["cache_index"]
                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
                causal_mask = lax.dynamic_slice(
                    self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
                )
            else:
                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
            causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
        # combine masks if needed
        if attention_mask is not None and self.causal:
            attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
            attention_mask = combine_masks(attention_mask, causal_mask)
        elif self.causal:
            attention_mask = causal_mask
        elif attention_mask is not None:
            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
        dropout_rng = None
        if not deterministic and self.config.attn_pdrop > 0.0:
            dropout_rng = self.make_rng("dropout")
        # During fast autoregressive decoding, we feed one position at a time,
        # and cache the keys and values step by step.
        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
            key, value, attention_mask = self._concatenate_to_cache(key, value, query, attention_mask)
        # transform boolean mask into float mask
        if attention_mask is not None:
            attention_bias = lax.select(
                attention_mask > 0,
                jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
            )
        else:
            attention_bias = None
        # usual dot product attention
        attn_weights = dot_product_attention_weights(
            query, key,
            bias=attention_bias,
            dropout_rng=dropout_rng,
            dropout_rate=self.config.attn_pdrop,
            deterministic=deterministic,
            dtype=self.dtype,
            precision=None,
        )
        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value)
        attn_output = self._merge_heads(attn_output)
        attn_output = self.c_proj(attn_output)
        attn_output = self.resid_dropout(attn_output, deterministic=deterministic)
        outputs = (attn_output, attn_weights) if output_attentions else (attn_output,)
        return outputs
class LMCFlaxGPT2Router(nn.Module):
    config: GPT2Config
    dtype: jnp.dtype = jnp.float32
 
    def setup(self):
        self.n_routed_experts = self.config.num_routed_experts
        self.n_group = getattr(self.config, "n_group", 1)
        self.topk_group = getattr(self.config, "topk_group", 1)
        self.top_k = self.config.topk
        self.routed_scaling_factor = self.config.routed_scaling_factor
        self.norm_topk_prob = getattr(self.config, "norm_topk_prob", False)
        # Weight and bias for router computation
        kernel_init = jax.nn.initializers.normal(self.config.initializer_range)
        self.router_weight = self.param(
            "router_weight", kernel_init, (self.n_routed_experts, self.config.hidden_size)
        )
        self.router_bias = self.param(
            "router_bias", lambda rng, shape: jnp.zeros(shape, dtype=self.dtype), (self.n_routed_experts,)
        )
        self.e_score_correction_bias = self.param(
            "e_score_correction_bias", lambda rng, shape: jnp.zeros(shape, dtype=self.dtype), (self.n_routed_experts,)
        )
    def get_topk_indices(self, scores):
        scores_for_choice = scores + self.e_score_correction_bias[None, :]
        scores_grouped = scores_for_choice.reshape(
            -1, self.n_group, self.n_routed_experts // self.n_group
        )
        top2_scores = jax.lax.top_k(scores_grouped, 2)[0]
        group_scores = jnp.sum(top2_scores, axis=-1)
        top_group_scores, group_idx = jax.lax.top_k(group_scores, self.topk_group)
        group_mask = jnp.zeros_like(group_scores)
        group_mask = group_mask.at[
            jnp.arange(group_mask.shape[0])[:, None], group_idx
        ].set(1)
        group_mask_expanded = jnp.repeat(
            group_mask[:, :, None],
            self.n_routed_experts // self.n_group,
            axis=-1,
        )
        score_mask = group_mask_expanded.reshape(
            -1, self.n_routed_experts
        )
 
        scores_for_choice = jnp.where(
            score_mask,
            scores_for_choice,
            jnp.zeros_like(scores_for_choice),
        )
 
        topk_weights, topk_indices = jax.lax.top_k(scores_for_choice, self.top_k)
 
        return topk_indices, topk_weights
 
    def __call__(self, hidden_states):
        router_logits = jnp.matmul(hidden_states, self.router_weight.T) + self.router_bias
        scores = jax.nn.sigmoid(router_logits)
 
        if self.n_group > 1:
            topk_indices, topk_weights = self.get_topk_indices(scores)
        else:
            topk_weights, topk_indices = jax.lax.top_k(scores, self.top_k)
 
        if self.norm_topk_prob:
            denominator = jnp.sum(topk_weights, axis=-1, keepdims=True) + 1e-20
            topk_weights = topk_weights / denominator
 
        topk_weights = topk_weights * self.routed_scaling_factor
 
        return topk_indices, topk_weights
class LMCFlaxGPT2MoE(nn.Module):
    config: GPT2Config
    dtype: jnp.dtype = jnp.float32
    def setup(self):
        self.n_experts = self.config.num_routed_experts
        self.n_shared_experts = self.config.num_shared_experts
        self.top_k = self.config.topk
        self.router = LMCFlaxGPT2Router(config=self.config, dtype=self.dtype)
        self.experts = [
            FlaxGPT2MLP(
                config=self.config,
                intermediate_size=self.config.n_inner,
                dtype=self.dtype,
            )
            for _ in range(self.n_experts)
        ]
        self.shared_experts = FlaxGPT2MLP(
            config=self.config,
            intermediate_size=self.config.n_inner* self.n_shared_experts,
            dtype=self.dtype,
        )
 
    def __call__(self, hidden_states, deterministic: bool = True):
        residual = hidden_states
        orig_shape = hidden_states.shape
        if self.n_experts == 0 or self.top_k == 0:
            hidden_states = self.shared_experts(hidden_states, deterministic=deterministic)
            return hidden_states
        hidden_states_flat = hidden_states.reshape(-1, hidden_states.shape[-1])
        topk_indices, topk_weights = self.router(hidden_states_flat)
        expert_outputs = []
        for expert in self.experts:
            output = expert(hidden_states_flat, deterministic=deterministic)  # [n_tokens, hidden_dim]
            expert_outputs.append(output)
        expert_outputs = jnp.stack(expert_outputs, axis=1)  # [n_tokens, n_experts, hidden_dim]
        # Build routing mask:
        routing_mask = jax.nn.one_hot(topk_indices, self.n_experts, dtype=self.dtype)
        routing_mask = routing_mask.sum(axis=1) > 0   # [n_tokens, n_experts]
        # Compute weights per expert
        weights_per_expert = jax.nn.one_hot(topk_indices, self.n_experts, dtype=self.dtype)
        weights_per_expert = (weights_per_expert * topk_weights[..., None]).sum(axis=1)
        # Mask out experts not routed
        weights_per_expert = weights_per_expert * routing_mask.astype(self.dtype)
        # Multiply and sum
        weighted_expert_outputs = expert_outputs * weights_per_expert[..., None]  # [n_tokens, n_experts, hidden_dim]
        final_output = weighted_expert_outputs.sum(axis=1)  # [n_tokens, hidden_dim]
        # Add shared expert
        shared_output = self.shared_experts(residual, deterministic=deterministic)
        final_output = final_output.reshape(orig_shape) + shared_output
        return final_output
class LMCFlaxGPT2Block(nn.Module):
    config: GPT2Config
    dtype: jnp.dtype = jnp.float32
    def setup(self):
        hidden_size = self.config.hidden_size
        inner_dim = self.config.n_inner if self.config.n_inner is not None else 4 * hidden_size
        self.ln_1 = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
        self.attn = LMCFlaxGPT2Attention(self.config, dtype=self.dtype)
        self.ln_2 = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
        if self.config.add_cross_attention:
            self.crossattention = LMCFlaxGPT2Attention(
                config=self.config, dtype=self.dtype, causal=False, is_cross_attention=True
            )
            self.ln_cross_attn = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
        self.moe = LMCFlaxGPT2MoE(self.config, dtype=self.dtype)
    def __call__(
        self,
        hidden_states,
        attention_mask=None,
        position_ids=None,
        encoder_hidden_states: Optional[jnp.ndarray] = None,
        encoder_attention_mask: Optional[jnp.ndarray] = None,
        deterministic: bool = True,
        init_cache: bool = False,
        output_attentions: bool = False,
    ):
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
        attn_outputs = self.attn(
            hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            deterministic=deterministic,
            init_cache=init_cache,
            output_attentions=output_attentions,
        )
        # residual connection
        attn_output = attn_outputs[0]  # output_attn: a, (attentions)
        outputs = attn_outputs[1:]
        # residual connection
        hidden_states = attn_output + residual
        # Cross-Attention Block
        if encoder_hidden_states is not None:
            # add one self-attention block for cross-attention
            if not hasattr(self, "crossattention"):
                raise ValueError(
                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
                    "cross-attention layers by setting `config.add_cross_attention=True`"
                )
            residual = hidden_states
            hidden_states = self.ln_cross_attn(hidden_states)
            cross_attn_outputs = self.crossattention(
                hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                position_ids=position_ids,
                deterministic=deterministic,
                output_attentions=output_attentions,
            )
            attn_output = cross_attn_outputs[0]
            # residual connection
            hidden_states = residual + attn_output
            outputs = outputs + cross_attn_outputs[1:]  # add cross attentions if we output attention weights
        residual = hidden_states
        hidden_states = self.ln_2(hidden_states)
        feed_forward_hidden_states = self.moe(hidden_states, deterministic=deterministic)
        # residual connection
        hidden_states = residual + feed_forward_hidden_states
        outputs = (hidden_states,) + outputs
        return outputs
 
 
 
class LMCFlaxGPT2PreTrainedModel(FlaxPreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """
 
    config_class = GPT2Config
    base_model_prefix = "transformer"
    module_class: nn.Module = None
 
    def __init__(
        self,
        config: GPT2Config,
        input_shape: Tuple = (1, 1),
        seed: int = 0,
        dtype: jnp.dtype = jnp.float32,
        _do_init: bool = True,
        **kwargs,
    ):
        module = self.module_class(config=config, dtype=dtype, **kwargs)
        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
 
    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
        # init input tensors
        input_ids = jnp.zeros(input_shape, dtype="i4")
        attention_mask = jnp.ones_like(input_ids)
        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
        params_rng, dropout_rng = jax.random.split(rng)
        rngs = {"params": params_rng, "dropout": dropout_rng}
 
        if self.config.add_cross_attention:
            encoder_hidden_states = jnp.zeros(input_shape + (self.config.n_embd,))
            encoder_attention_mask = attention_mask
            module_init_outputs = self.module.init(
                rngs,
                input_ids,
                attention_mask,
                position_ids,
                encoder_hidden_states,
                encoder_attention_mask,
                return_dict=False,
            )
        else:
            module_init_outputs = self.module.init(rngs, input_ids, attention_mask, position_ids, return_dict=False)
 
        random_params = module_init_outputs["params"]
 
        if params is not None:
            random_params = flatten_dict(unfreeze(random_params))
            params = flatten_dict(unfreeze(params))
            for missing_key in self._missing_keys:
                params[missing_key] = random_params[missing_key]
            self._missing_keys = set()
            return freeze(unflatten_dict(params))
        else:
            return random_params
 
    def init_cache(self, batch_size, max_length):
        r"""
        Args:
            batch_size (`int`):
                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
            max_length (`int`):
                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
                cache.
        """
        # init input variables to retrieve cache
        input_ids = jnp.ones((batch_size, max_length))
        attention_mask = jnp.ones_like(input_ids)
        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
 
        init_variables = self.module.init(
            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
        )
        return unfreeze(init_variables["cache"])
 
    def __call__(
        self,
        input_ids,
        attention_mask=None,
        position_ids=None,
        encoder_hidden_states: Optional[jnp.ndarray] = None,
        encoder_attention_mask: Optional[jnp.ndarray] = None,
        params: Optional[dict] = None,
        past_key_values: Optional[dict] = None,
        dropout_rng: jax.random.PRNGKey = None,
        train: bool = False,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ):
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.return_dict
 
        if encoder_hidden_states is not None and encoder_attention_mask is None:
            batch_size, sequence_length = encoder_hidden_states.shape[:2]
            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
 
        batch_size, sequence_length = input_ids.shape
 
        if position_ids is None:
            if past_key_values is not None:
                raise ValueError("Make sure to provide `position_ids` when passing `past_key_values`.")
 
            position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
 
        if attention_mask is None:
            attention_mask = jnp.ones((batch_size, sequence_length))
 
        # Handle any PRNG if needed
        rngs = {}
        if dropout_rng is not None:
            rngs["dropout"] = dropout_rng
 
        inputs = {"params": params or self.params}
 
        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be changed by FlaxGPT2Attention module
        if past_key_values:
            inputs["cache"] = past_key_values
            mutable = ["cache"]
        else:
            mutable = False
 
        outputs = self.module.apply(
            inputs,
            jnp.array(input_ids, dtype="i4"),
            jnp.array(attention_mask, dtype="i4"),
            jnp.array(position_ids, dtype="i4"),
            encoder_hidden_states,
            encoder_attention_mask,
            not train,
            False,
            output_attentions,
            output_hidden_states,
            return_dict,
            rngs=rngs,
            mutable=mutable,
        )
 
        # add updated cache to model output
        if past_key_values is not None and return_dict:
            outputs, past_key_values = outputs
            outputs["past_key_values"] = unfreeze(past_key_values["cache"])
            return outputs
        elif past_key_values is not None and not return_dict:
            outputs, past_key_values = outputs
            outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
 
        return outputs
 
 
class LMCFlaxGPT2BlockCollection(nn.Module):
    config: GPT2Config
    dtype: jnp.dtype = jnp.float32
 
    def setup(self):
        self.blocks = [
            LMCFlaxGPT2Block(self.config.lmc_config, name=str(i), dtype=self.dtype) if i in self.config.lmc_layer_indices
            else LMCFlaxGPT2Block(self.config, name=str(i), dtype=self.dtype)
            for i in range(self.config.num_hidden_layers)
        ]
 
    def __call__(
        self,
        hidden_states,
        attention_mask=None,
        position_ids=None,
        encoder_hidden_states: Optional[jnp.ndarray] = None,
        encoder_attention_mask: Optional[jnp.ndarray] = None,
        deterministic: bool = True,
        init_cache: bool = False,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
    ):
        all_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None
        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
 
        for block in self.blocks:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
 
            layer_outputs = block(
                hidden_states,
                attention_mask,
                position_ids=position_ids,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                deterministic=deterministic,
                init_cache=init_cache,
                output_attentions=output_attentions,
            )
            hidden_states = layer_outputs[0]
 
            if output_attentions:
                all_attentions += (layer_outputs[1],)
 
                if encoder_hidden_states is not None:
                    all_cross_attentions += (layer_outputs[2],)
 
        # this contains possible `None` values - `FlaxGPT2Module` will filter them out
        outputs = (hidden_states, all_hidden_states, all_attentions, all_cross_attentions)
 
        return outputs
 
 
class LMCFlaxGPT2Module(nn.Module):
    config: GPT2Config
    dtype: jnp.dtype = jnp.float32
 
    def setup(self):
        self.embed_dim = self.config.hidden_size
 
        self.wte = nn.Embed(
            self.config.vocab_size,
            self.embed_dim,
            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
            dtype=self.dtype,
        )
        if self.config.position_embeddings == "learnable":
            self.wpe = nn.Embed(
                self.config.max_position_embeddings,
                self.embed_dim,
                embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
                dtype=self.dtype,
            )
        if self.config.position_embeddings == "sinusoidal":
            self.wpe_table = create_sinusoidal_positions(self.config.max_position_embeddings, self.embed_dim)
        self.dropout = nn.Dropout(rate=self.config.embd_pdrop)
        self.h = LMCFlaxGPT2BlockCollection(self.config, dtype=self.dtype)
        self.ln_f = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
 
    def __call__(
        self,
        input_ids,
        attention_mask,
        position_ids,
        encoder_hidden_states: Optional[jnp.ndarray] = None,
        encoder_attention_mask: Optional[jnp.ndarray] = None,
        deterministic=True,
        init_cache: bool = False,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
    ):
        input_embeds = self.wte(input_ids.astype("i4"))
        if self.config.position_embeddings == "learnable":
            position_embeds = self.wpe(position_ids.astype("i4"))
            hidden_states = input_embeds + position_embeds
        if self.config.position_embeddings == "sinusoidal":
            position_embeds = jnp.take(self.wpe_table, position_ids, axis=0)
            hidden_states = input_embeds + position_embeds
        if self.config.position_embeddings == "rope":
            hidden_states = input_embeds
        hidden_states = self.dropout(hidden_states, deterministic=deterministic)
        outputs = self.h(
            hidden_states,
            attention_mask,
            position_ids=position_ids,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            deterministic=deterministic,
            init_cache=init_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
 
        hidden_states = outputs[0]
        hidden_states = self.ln_f(hidden_states)
 
        if output_hidden_states:
            all_hidden_states = outputs[1] + (hidden_states,)
            outputs = (hidden_states, all_hidden_states) + outputs[2:]
        else:
            outputs = (hidden_states,) + outputs[1:]
 
        if not return_dict:
            return tuple(v for v in outputs if v is not None)
 
        return FlaxBaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            hidden_states=outputs[1],
            attentions=outputs[2],
            cross_attentions=outputs[3],
        )
class LMCFlaxGPT2Model(LMCFlaxGPT2PreTrainedModel):
    module_class = LMCFlaxGPT2Module
class LMCFlaxGPT2LMHeadModule(nn.Module):
    config: GPT2Config
    dtype: jnp.dtype = jnp.float32
 
    def setup(self):
        self.transformer = LMCFlaxGPT2Module(self.config, dtype=self.dtype)
        self.lm_head = nn.Dense(
            self.config.vocab_size,
            use_bias=False,
            dtype=self.dtype,
            kernel_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
        )
    def __call__(
        self,
        input_ids,
        attention_mask,
        position_ids,
        encoder_hidden_states: Optional[jnp.ndarray] = None,
        encoder_attention_mask: Optional[jnp.ndarray] = None,
        deterministic: bool = True,
        init_cache: bool = False,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
    ):
        outputs = self.transformer(
            input_ids,
            attention_mask,
            position_ids,
            encoder_hidden_states,
            encoder_attention_mask,
            deterministic=deterministic,
            init_cache=init_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
 
        hidden_states = outputs[0]
 
        if self.config.tie_word_embeddings:
            shared_kernel = self.transformer.variables["params"]["wte"]["embedding"].T
            lm_logits = self.lm_head.apply({"params": {"kernel": shared_kernel}}, hidden_states)
        else:
            lm_logits = self.lm_head(hidden_states)
        if not return_dict:
            return (lm_logits,) + outputs[1:]
        return FlaxCausalLMOutputWithCrossAttentions(
            logits=lm_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            cross_attentions=outputs.cross_attentions,
        )
 
 
 
class LMCFlaxGPT2LMHeadModel(LMCFlaxGPT2PreTrainedModel):
    module_class = LMCFlaxGPT2LMHeadModule
 
    def prepare_inputs_for_generation(self, input_ids, max_length, attention_mask: Optional[jax.Array] = None):
        # initializing the cache
        batch_size, seq_length = input_ids.shape
 
        past_key_values = self.init_cache(batch_size, max_length)
        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
        # But since GPT2 uses a causal mask, those positions are masked anyways.
        # Thus we can create a single static attention_mask here, which is more efficient for compilation
        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
        if attention_mask is not None:
            position_ids = attention_mask.cumsum(axis=-1) - 1
            extended_attention_mask = lax.dynamic_update_slice(
                extended_attention_mask, attention_mask.astype("i4"), (0, 0)
            )
        else:
            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
 
        return {
            "past_key_values": past_key_values,
            "attention_mask": extended_attention_mask,
            "position_ids": position_ids,
        }
 
    def update_inputs_for_generation(self, model_outputs, model_kwargs):
        model_kwargs["past_key_values"] = model_outputs.past_key_values
        model_kwargs["position_ids"] = model_kwargs["position_ids"][:, -1:] + 1
        return model_kwargs