modeling_agora.py

"""
Agora model implementation.
Architecture: Decoder-only transformer with GQA, RoPE, SiLU/SwiGLU MLP, RMSNorm.
Compatible with Hugging Face Transformers ≥ 4.40.
"""

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

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

from transformers import PreTrainedModel
from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
)
from transformers.utils import logging

from .configuration_agora import AgoraConfig

logger = logging.get_logger(__name__)


# ── RMSNorm ──────────────────────────────────────────────────────────────────

class AgoraRMSNorm(nn.Module):
    def __init__(self, hidden_size: int, eps: float = 1e-5):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        variance = x.pow(2).mean(-1, keepdim=True)
        x = x * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * x


# ── Rotary Position Embeddings ────────────────────────────────────────────────

def rotate_half(x: torch.Tensor) -> torch.Tensor:
    x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
    return torch.cat([-x2, x1], dim=-1)


def apply_rotary_pos_emb(
    q: torch.Tensor,
    k: torch.Tensor,
    cos: torch.Tensor,
    sin: torch.Tensor,
    position_ids: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
    cos = cos[position_ids].unsqueeze(1)  # [bs, 1, seq, dim]
    sin = sin[position_ids].unsqueeze(1)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class AgoraRotaryEmbedding(nn.Module):
    def __init__(self, dim: int, max_position_embeddings: int = 4096, base: float = 10000.0):
        super().__init__()
        self.dim = dim
        self.max_position_embeddings = max_position_embeddings
        self.base = base
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self._build_cache(max_position_embeddings)

    def _build_cache(self, seq_len: int):
        t = torch.arange(seq_len, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(t, self.inv_freq)
        emb = torch.cat([freqs, freqs], dim=-1)
        self.register_buffer("cos_cached", emb.cos()[None, None], persistent=False)
        self.register_buffer("sin_cached", emb.sin()[None, None], persistent=False)

    def forward(self, x: torch.Tensor, seq_len: int):
        if seq_len > self.max_position_embeddings:
            self._build_cache(seq_len)
        return (
            self.cos_cached[:, :, :seq_len, ...].to(x.dtype),
            self.sin_cached[:, :, :seq_len, ...].to(x.dtype),
        )


# ── MLP (SwiGLU / SiLU gate) ─────────────────────────────────────────────────

class AgoraMLP(nn.Module):
    def __init__(self, config: AgoraConfig):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        bias = config.mlp_bias

        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.up_proj   = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=bias)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))


# ── Grouped-Query Attention ───────────────────────────────────────────────────

def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """Expand key/value heads to match query head count."""
    if n_rep == 1:
        return hidden_states
    bs, num_kv, seq_len, head_dim = hidden_states.shape
    return (
        hidden_states[:, :, None, :, :]
        .expand(bs, num_kv, n_rep, seq_len, head_dim)
        .reshape(bs, num_kv * n_rep, seq_len, head_dim)
    )


class AgoraAttention(nn.Module):
    def __init__(self, config: AgoraConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.num_kv_heads = config.num_key_value_heads
        self.head_dim = config.head_dim
        self.num_kv_groups = self.num_heads // self.num_kv_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.attention_dropout = config.attention_dropout
        bias = config.attention_bias

        self.q_proj = nn.Linear(self.hidden_size, self.num_heads    * self.head_dim, bias=bias)
        self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=bias)
        self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=bias)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size,     bias=bias)

        self.rotary_emb = AgoraRotaryEmbedding(
            self.head_dim,
            max_position_embeddings=self.max_position_embeddings,
            base=self.rope_theta,
        )

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

        q = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads,    self.head_dim).transpose(1, 2)
        k = self.k_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
        v = self.v_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = k.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        cos, sin = self.rotary_emb(v, seq_len=kv_seq_len)
        q, k = apply_rotary_pos_emb(q, k, cos.squeeze(0).squeeze(0), sin.squeeze(0).squeeze(0), position_ids)

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

        k = repeat_kv(k, self.num_kv_groups)
        v = repeat_kv(v, self.num_kv_groups)

        scale = math.sqrt(self.head_dim)
        attn_weights = torch.matmul(q, k.transpose(2, 3)) / scale

        if attention_mask is not None:
            attn_weights = attn_weights + attention_mask

        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype)

        if self.training and self.attention_dropout > 0.0:
            attn_weights = F.dropout(attn_weights, p=self.attention_dropout)

        attn_output = torch.matmul(attn_weights, v)
        attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, q_len, -1)
        attn_output = self.o_proj(attn_output)

        return attn_output, (attn_weights if output_attentions else None), past_key_value


# ── Decoder Layer ─────────────────────────────────────────────────────────────

class AgoraDecoderLayer(nn.Module):
    def __init__(self, config: AgoraConfig, layer_idx: int):
        super().__init__()
        self.self_attn = AgoraAttention(config, layer_idx=layer_idx)
        self.mlp = AgoraMLP(config)
        self.input_layernorm = AgoraRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = AgoraRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

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

        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
        )
        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

        outputs = (hidden_states,)
        if output_attentions:
            outputs += (self_attn_weights,)
        if use_cache:
            outputs += (present_key_value,)
        return outputs


# ── Base Model ────────────────────────────────────────────────────────────────

class AgoraPreTrainedModel(PreTrainedModel):
    config_class = AgoraConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["AgoraDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]

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


class AgoraModel(AgoraPreTrainedModel):
    def __init__(self, config: AgoraConfig):
        super().__init__(config)
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
        self.layers = nn.ModuleList(
            [AgoraDecoderLayer(config, layer_idx=i) for i in range(config.num_hidden_layers)]
        )
        self.norm = AgoraRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.gradient_checkpointing = False
        self.post_init()

    def get_input_embeddings(self):
        return self.embed_tokens

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

    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[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        output_attentions    = output_attentions    if output_attentions    is not None else self.config.output_attentions
        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        use_cache            = use_cache            if use_cache            is not None else self.config.use_cache
        return_dict          = return_dict          if return_dict          is not None else self.config.use_return_dict

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("Specify either input_ids or inputs_embeds, not both.")
        if input_ids is not None:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("input_ids or inputs_embeds must be provided.")

        past_key_values_length = 0
        if use_cache:
            if not isinstance(past_key_values, Cache):
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
            past_key_values_length = past_key_values.get_seq_length()

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(
                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
            ).unsqueeze(0)

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

        # Build causal mask
        attention_mask = self._prepare_4d_causal_attention_mask(
            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
        )

        hidden_states = inputs_embeds
        all_hidden_states = () if output_hidden_states else None
        all_self_attns   = () if output_attentions else None
        next_decoder_cache = None

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

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states, attention_mask, position_ids, None, output_attentions, False,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                )

            hidden_states = layer_outputs[0]
            if use_cache:
                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None

        if not return_dict:
            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )

    # Minimal causal mask helper (avoids importing private HF utils)
    def _prepare_4d_causal_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
        bsz, tgt_len = input_shape
        dtype, device = inputs_embeds.dtype, inputs_embeds.device
        src_len = tgt_len + past_key_values_length

        # Causal mask
        mask = torch.full((tgt_len, src_len), torch.finfo(dtype).min, device=device)
        mask_cond = torch.arange(mask.size(-1), device=device)
        mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(0), 1), 0)
        mask = mask[None, None, :, :].expand(bsz, 1, tgt_len, src_len).to(dtype)

        if attention_mask is not None:
            pad_mask = (1.0 - attention_mask[:, None, None, :].to(dtype)) * torch.finfo(dtype).min
            mask = mask + pad_mask[:, :, :, :src_len]

        return mask


# ── Causal LM Head ────────────────────────────────────────────────────────────

class AgoraForCausalLM(AgoraPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config: AgoraConfig):
        super().__init__(config)
        self.model   = AgoraModel(config)
        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, v): self.model.embed_tokens = v
    def get_output_embeddings(self): return self.lm_head
    def set_output_embeddings(self, v): self.lm_head = v
    def set_decoder(self, d): self.model = d
    def get_decoder(self): return self.model

    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[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states).float()

        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss = CrossEntropyLoss()(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1))

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

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

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs):
        if past_key_values is not None:
            past_len = past_key_values.get_seq_length() if isinstance(past_key_values, Cache) else past_key_values[0][0].shape[2]
            input_ids = input_ids[:, past_len:]

        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values is not None:
                position_ids = position_ids[:, -input_ids.shape[1]:]

        model_inputs = {"input_ids": input_ids} if inputs_embeds is None else {"inputs_embeds": inputs_embeds}
        model_inputs.update({
            "position_ids": position_ids,
            "past_key_values": past_key_values,
            "use_cache": kwargs.get("use_cache"),
            "attention_mask": attention_mask,
        })
        return model_inputs

    @staticmethod
    def _reorder_cache(past_key_values, beam_idx):
        return tuple(
            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
            for layer_past in past_key_values
        )