"""Modeling for the Q-50M architecture (replica of q-project/Q-50M-Base).

Self-contained implementation of the exact architecture described in the
model card: Mistral-shaped 512-wide 10-layer decoder, GQA 8q/2kv, per-head
QK-Norm, RoPE in every layer (configurable NoPE), scalar content-dependent
gates (gate = multiplier * sigmoid(Wx)) on attention and MLP residual
branches, SwiGLU MLP, tied input/output embeddings, final RMSNorm.

Follows the transformers v5 modeling pattern (DynamicCache, create_causal_mask,
GradientCheckpointingLayer, GenerationMixin).
"""

from collections.abc import Callable

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.masking_utils import create_causal_mask
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs

try:
    from .configuration_q50m import Q50MConfig
except ImportError:  # standalone use inside an HF model directory
    from configuration_q50m import Q50MConfig


def _rotate_half(x: torch.Tensor) -> torch.Tensor:
    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):
    # cos/sin: (batch, seq, head_dim)
    cos = cos.unsqueeze(1)
    sin = sin.unsqueeze(1)
    return q * cos + _rotate_half(q) * sin, k * cos + _rotate_half(k) * sin


class Q50MRMSNorm(nn.Module):
    def __init__(self, hidden_size: int, eps: float = 1e-6) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        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)).to(input_dtype)

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


class Q50MRotaryEmbedding(nn.Module):
    def __init__(self, config: Q50MConfig, device=None):
        super().__init__()
        self.config = config
        inv_freq = 1.0 / (
            config.rope_theta
            ** (torch.arange(0, config.head_dim, 2, dtype=torch.float32, device=device) / config.head_dim)
        )
        self.inv_freq = nn.Buffer(inv_freq, persistent=True)

    @torch.no_grad()
    def forward(self, x, position_ids):
        inv_freq_expanded = (
            self.inv_freq[None, :, None].expand(position_ids.shape[0], -1, 1)
            .to(dtype=torch.float32, device=x.device)
        )
        position_ids_expanded = position_ids[:, None, :].float()
        freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
        emb = torch.cat((freqs, freqs), dim=-1)
        return emb.cos().to(dtype=x.dtype), emb.sin().to(dtype=x.dtype)


class Q50MScalarGate(nn.Module):
    """Content-dependent scalar gate: output = branch * multiplier * sigmoid(W x)."""

    def __init__(self, hidden_size: int, multiplier: float = 2.0):
        super().__init__()
        self.projection = nn.Linear(hidden_size, 1, bias=False)
        self.multiplier = multiplier

    def forward(self, branch: torch.Tensor, residual_input: torch.Tensor) -> torch.Tensor:
        return branch * self.multiplier * torch.sigmoid(self.projection(residual_input))

    def reset_parameters(self):
        nn.init.zeros_(self.projection.weight)


def _repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    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: torch.Tensor | None,
    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:
        attn_weights = attn_weights + attention_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


def _sdpa_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    """Local SDPA wrapper. HF's generic SDPA interface falls back to the math
    backend for GQA shapes (q heads != kv heads), which materializes
    (batch, heads, seq, seq) fp32 attention scores. Expanding kv first keeps
    the fused flash / memory-efficient kernels eligible."""
    n_rep = getattr(module, "num_key_value_groups", 1)
    if n_rep > 1:
        key = _repeat_kv(key, n_rep)
        value = _repeat_kv(value, n_rep)
    is_causal = (
        attention_mask is None
        and getattr(module, "is_causal", False)
        and query.shape[2] > 1
    )
    attn_output = nn.functional.scaled_dot_product_attention(
        query,
        key,
        value,
        attn_mask=attention_mask,
        dropout_p=dropout,
        is_causal=is_causal,
        scale=scaling,
    )
    return attn_output.transpose(1, 2).contiguous(), None


class Q50MAttention(nn.Module):
    def __init__(self, config: Q50MConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = config.head_dim
        self.num_heads = config.num_attention_heads
        self.num_kv_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_kv_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True
        self.use_rope = layer_idx not in config.nope_layers

        inner = self.num_heads * self.head_dim
        kv_inner = self.num_kv_heads * self.head_dim
        self.q_proj = nn.Linear(config.hidden_size, inner, bias=False)
        self.k_proj = nn.Linear(config.hidden_size, kv_inner, bias=False)
        self.v_proj = nn.Linear(config.hidden_size, kv_inner, bias=False)
        self.o_proj = nn.Linear(inner, config.hidden_size, bias=False)
        self.q_norm = Q50MRMSNorm(self.head_dim, config.rms_norm_eps) if config.qk_norm else None
        self.k_norm = Q50MRMSNorm(self.head_dim, config.rms_norm_eps) if config.qk_norm else None
        self.output_gate = (
            Q50MScalarGate(config.hidden_size, config.gate_multiplier)
            if config.attention_scalar_gate
            else None
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_mask: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)
        q = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        k = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        v = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        if self.q_norm is not None:
            q = self.q_norm(q)
            k = self.k_norm(k)
        if self.use_rope:
            cos, sin = position_embeddings
            q, k = _apply_rotary_pos_emb(q, k, cos, sin)

        if past_key_values is not None:
            k, v = past_key_values.update(k, v, self.layer_idx)

        if self.config._attn_implementation == "sdpa":
            attention_interface: Callable = _sdpa_attention_forward
        else:
            attention_interface = ALL_ATTENTION_FUNCTIONS.get_interface(
                self.config._attn_implementation, _eager_attention_forward
            )
        attn_output, attn_weights = attention_interface(
            self,
            q, k, v,
            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)
        if self.output_gate is not None:
            attn_output = self.output_gate(attn_output, hidden_states)
        return attn_output, attn_weights


class Q50MMLP(nn.Module):
    def __init__(self, config: Q50MConfig):
        super().__init__()
        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
        self.output_gate = (
            Q50MScalarGate(config.hidden_size, config.gate_multiplier)
            if config.mlp_scalar_gate
            else None
        )

    def forward(self, hidden_states):
        output = self.down_proj(F.silu(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
        if self.output_gate is not None:
            output = self.output_gate(output, hidden_states)
        return output


class Q50MDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Q50MConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = Q50MAttention(config, layer_idx)
        self.mlp = Q50MMLP(config)
        self.input_layernorm = Q50MRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Q50MRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        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,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        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


class Q50MPreTrainedModel(PreTrainedModel):
    config_class = Q50MConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Q50MDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_sdpa = True
    _supports_flash_attn = False
    _supports_flex_attn = False


class Q50MModel(Q50MPreTrainedModel):
    def __init__(self, config: Q50MConfig):
        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(
            [Q50MDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Q50MRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Q50MRotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        self.post_init()

    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        rotary_position_ids: torch.LongTensor | None = 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 = self.embed_tokens(input_ids)

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

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

        # Unpadded, uncached SDPA can use its native causal flag, including
        # under torch.compile where HF otherwise materializes a dense mask.
        causal_mask = None if (
            attention_mask is None and past_key_values is None and standard_positions
            and self.config._attn_implementation == "sdpa"
        ) else create_causal_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )

        hidden_states = inputs_embeds
        # Slotted Zarya keeps source positions for RoPE but attends causally in
        # the permuted sequence order. Never let HF infer packed-sequence masks
        # from those nonmonotonic rotary positions.
        position_embeddings = self.rotary_emb(
            hidden_states, position_ids=position_ids if rotary_position_ids is None else rotary_position_ids)

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            if self.gradient_checkpointing and self.training:
                hidden_states = self._gradient_checkpointing_func(
                    decoder_layer.forward,
                    hidden_states,
                    causal_mask,
                    position_ids,
                    past_key_values,
                    use_cache,
                    position_embeddings,
                )
            else:
                hidden_states = decoder_layer(
                    hidden_states,
                    attention_mask=causal_mask,
                    position_embeddings=position_embeddings,
                    position_ids=position_ids,
                    past_key_values=past_key_values,
                    use_cache=use_cache,
                    **kwargs,
                )

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


class Q50MForCausalLM(Q50MPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}

    def __init__(self, config: Q50MConfig):
        super().__init__(config)
        self.model = Q50MModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        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 forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        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,
            **kwargs,
        )
        hidden_states = outputs.last_hidden_state
        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:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss = F.cross_entropy(
                shift_logits.view(-1, shift_logits.size(-1)),
                shift_labels.view(-1),
                ignore_index=-100,
            )
        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
        )