modeling_step3p5.py

# Copyright 2025 The LLAMA4 and HuggingFace Inc. team. All rights reserved.
#
#
# 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 dataclasses import dataclass
from typing import Callable, Optional, Tuple, Union

import torch
import torch.nn as 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.masking_utils import (create_causal_mask,
                                        create_sliding_window_causal_mask)
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutputWithPast, ModelOutput
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, can_return_tuple, logging

from .configuration_step3p5 import Step3p5Config

logger = logging.get_logger(__name__)

__all__ = ["Step3p5Model", "Step3p5ForCausalLM"]

class Step3p5RotaryEmbedding(nn.Module):

    def __init__(self, config: Step3p5Config, device=None, layer_idx=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        self.layer_idx = layer_idx
        if config.rope_parameters is not None:
            self.rope_type = config.rope_parameters.get(
                "rope_type", config.rope_parameters.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

        partial_rotary_factors = getattr(config, "partial_rotary_factors",
                                         None)
        if partial_rotary_factors is not None:
            config.partial_rotary_factor = partial_rotary_factors[
                self.layer_idx]
        else:
            config.partial_rotary_factor = 1.0

        self.rope_theta = config.rope_theta
        if isinstance(config.rope_theta, list):
            self.rope_theta = config.rope_theta.copy()
            config.rope_theta = self.rope_theta[self.layer_idx]

        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
        config.rope_theta = self.rope_theta

    @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().to(x.device)

        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.
    """
    rotary_dim = cos.shape[-1]
    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]

    # Apply rotary embeddings on the first half or full tensor
    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)

    # Concatenate back to full shape
    q_embed = torch.cat([q_embed, q_pass], dim=-1)
    k_embed = torch.cat([k_embed, k_pass], dim=-1)
    return q_embed, k_embed


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)


# Adapted from transformers.models.llama.modeling_llama.eager_attention_forward -> llama4 doesn't cast attn weights to fp32
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,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)
    # breakpoint()
    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)
    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

@dataclass
class Step3p5CausalLMOutputWithPast(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Language modeling loss (for next-token prediction).
    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
        `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    """

    loss: Optional[torch.FloatTensor] = None
    last_hidden_state: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
    past_key_values: Optional[list[torch.FloatTensor]] = None
    hidden_states: Optional[tuple[torch.FloatTensor]] = None
    attentions: Optional[tuple[torch.FloatTensor]] = None


class Step3p5MLP(nn.Module):

    def __init__(self, config, intermediate_size=None, swiglu_limit=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = intermediate_size if intermediate_size is not None else config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size,
                                   self.intermediate_size,
                                   bias=False)
        self.up_proj = nn.Linear(self.hidden_size,
                                 self.intermediate_size,
                                 bias=False)
        self.down_proj = nn.Linear(self.intermediate_size,
                                   self.hidden_size,
                                   bias=False)
        self.act_fn = ACT2FN["silu"]
        self.limit = swiglu_limit

    def forward(self, x):
        up = self.up_proj(x)
        gate = self.act_fn(self.gate_proj(x))
        if self.limit is not None:
            gate = gate.clamp(min=None, max=self.limit)
            up = up.clamp(min=-self.limit, max=self.limit)

        return self.down_proj(gate * up)


def sigmoid_routing_function(gating_output: torch.Tensor, topk: int,
                             renormalize: bool):
    gating_output = gating_output.float()
    gate_prob = torch.sigmoid(gating_output)
    gate_prob = gate_prob / gate_prob.sum(dim=-1, keepdim=True)
    topk_prob, indices = torch.topk(gate_prob, k=topk, dim=1)
    expert_topk_weight = topk_prob
    if renormalize:
        expert_topk_weight = expert_topk_weight / torch.sum(
            expert_topk_weight, dim=-1, keepdim=True)
    return expert_topk_weight, indices


def softmax_routing_function(gating_output: torch.Tensor, top_k: int,
                             renormalize: bool):
    gating_output = gating_output.float()
    gate_prob = torch.softmax(gating_output, dim=-1)
    gate_prob = gate_prob / gate_prob.sum(dim=-1, keepdim=True)
    topk_prob, indices = torch.topk(gate_prob, k=top_k, dim=1)
    expert_topk_weight = topk_prob
    if renormalize:
        expert_topk_weight = expert_topk_weight / torch.sum(
            expert_topk_weight, dim=-1, keepdim=True)
    return expert_topk_weight, indices.to(torch.int32)


class MoELinear(nn.Module):

    def __init__(self, num_experts, in_features, out_features):
        super().__init__()
        self.num_experts = num_experts
        self.in_features = in_features
        self.out_features = out_features
        self.weight = nn.Parameter(
            torch.empty(num_experts, out_features, in_features))

    def forward(self, x, expert_id):
        x = F.linear(x.float(), self.weight[expert_id].float())
        return x


class Step3p5MoEMLP(nn.Module):

    def __init__(self, config, swiglu_limit=None):
        super().__init__()
        self.num_experts = config.moe_num_experts
        self.top_k = config.moe_top_k
        self.hidden_size = config.hidden_size
        self.moe_intermediate_size = config.moe_intermediate_size

        self.use_moe_router_bias = config.use_moe_router_bias
        if self.use_moe_router_bias:
            self.router_bias = nn.Parameter(torch.zeros(config.moe_num_experts,
                                                        dtype=torch.float32),
                                            requires_grad=False)
            self.custom_routing_function = self.router_bias_func
        elif config.moe_router_activation == "sigmoid":
            self.custom_routing_function = sigmoid_routing_function
        else:
            self.custom_routing_function = None
        self.need_fp32_gate = config.need_fp32_gate
        self.routed_scaling_factor = getattr(config,
                                             "moe_router_scaling_factor", 1.0)
        
        # gating
        self.gate = nn.Linear(self.hidden_size, self.num_experts, bias=False)
            
        self.act_fn = ACT2FN["silu"]
        self.limit = swiglu_limit

        self.up_proj = MoELinear(self.num_experts, self.hidden_size,
                                 self.moe_intermediate_size)
        self.gate_proj = MoELinear(self.num_experts, self.hidden_size,
                                   self.moe_intermediate_size)
        self.down_proj = MoELinear(self.num_experts,
                                   self.moe_intermediate_size,
                                   self.hidden_size)

    def router_bias_func(self, gating_output: torch.Tensor, topk: int,
                         renormalize: bool):
        gate_prob = torch.sigmoid(gating_output.float())
        gate_prob_with_bias = gate_prob + self.router_bias.unsqueeze(0)
        _, indices = torch.topk(gate_prob_with_bias, k=topk, dim=1)
        topk_prob = torch.gather(gate_prob, 1, indices)
        expert_topk_weight = topk_prob
        if renormalize:
            expert_topk_weight = expert_topk_weight / (
                torch.sum(expert_topk_weight, dim=-1, keepdim=True) + 1e-20)
        return expert_topk_weight, indices

    def get_expert_output(self, inputs: torch.Tensor, expert_id):
        #if self.limit is None:
        up = self.up_proj(inputs, expert_id)
        gate = self.act_fn(self.gate_proj(inputs, expert_id))
        if self.limit is not None:
            gate = gate.clamp(min=None, max=self.limit)
            up = up.clamp(min=-self.limit, max=self.limit)

        return self.down_proj(gate * up, expert_id)

    def forward(self, hidden_states):
        """ """
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        if self.need_fp32_gate:
            router_logits = torch.matmul(hidden_states.to(torch.float32), self.gate.weight.t().to(torch.float32))
        else:
            # router_logits: (batch * sequence_length, n_experts)
            router_logits = self.gate(hidden_states)
        
        if self.custom_routing_function:
            routing_weights, selected_experts = self.custom_routing_function(
                router_logits, self.top_k, renormalize=True)
        else:
            routing_weights = F.softmax(router_logits,
                                        dim=1,
                                        dtype=torch.float)
            routing_weights, selected_experts = torch.topk(routing_weights,
                                                           self.top_k,
                                                           dim=-1)

        routing_weights = routing_weights * self.routed_scaling_factor

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

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

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

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

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


class Step3p5RMSNorm(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        eps: float = 1e-5,
    ) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        dtype = x.dtype
        x = x.float()
        variance = x.pow(2).mean(dim=-1, keepdim=True)
        normed = x * torch.rsqrt(variance + self.variance_epsilon)
        normed = normed * (self.weight.float() + 1)
        return normed.to(dtype)
class Step3p5Attention(nn.Module):

    def __init__(self, config: Step3p5Config, layer_idx):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.num_attention_heads = config.num_attention_heads
        self.num_key_value_heads = config.num_attention_groups

        layer_types = getattr(config, "layer_types", [])
        if layer_types:
            enable_sliding_window = layer_types[
                self.layer_idx] == "sliding_attention"
        else:
            enable_sliding_window = self.layer_idx % 2 == 0
        
        if hasattr(config, "yarn_only_types") and layer_types[
                self.layer_idx] not in config.yarn_only_types:
            config.rope_parameters = None
        else:
            config.rope_parameters = getattr(config, "rope_scaling", None)

        self.sliding_window = config.sliding_window
        if enable_sliding_window:
            self.num_attention_heads = config.attention_other_setting[
                "num_attention_heads"]
            self.num_key_value_heads = config.attention_other_setting[
                "num_attention_groups"]

        if self.sliding_window is not None and enable_sliding_window:
            self.sliding_window = (self.sliding_window)
        else:
            self.sliding_window = None
        self.head_dim = getattr(config, "head_dim",
                        config.hidden_size // self.num_attention_heads)
        self.num_key_value_groups = self.num_attention_heads // self.num_key_value_heads

        self.rotary_emb = Step3p5RotaryEmbedding(config, layer_idx=layer_idx)

        self.q_size = self.num_attention_heads * self.head_dim
        self.kv_size = self.num_key_value_heads * self.head_dim
        self.scaling = self.head_dim**-0.5

        self.q_proj = nn.Linear(config.hidden_size, self.q_size, bias=False)
        self.k_proj = nn.Linear(config.hidden_size, self.kv_size, bias=False)
        self.v_proj = nn.Linear(config.hidden_size, self.kv_size, bias=False)
        self.o_proj = nn.Linear(self.q_size, config.hidden_size, bias=False)
        self.q_norm = Step3p5RMSNorm(self.head_dim,
                                    eps=config.rms_norm_eps)
        self.k_norm = Step3p5RMSNorm(self.head_dim,
                                    eps=config.rms_norm_eps)

        self.use_head_wise_attn_gate = config.use_head_wise_attn_gate
        if self.use_head_wise_attn_gate:
            self.g_proj = nn.Linear(config.hidden_size,
                                    self.num_attention_heads,
                                    bias=False)

        self.use_rope = True
        use_rope_layers = getattr(config, "use_rope_layers", None)
        if use_rope_layers:
            self.use_rope = use_rope_layers[self.layer_idx]

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor],
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor],
               Optional[Tuple[torch.Tensor]]]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_norm(
            self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
        key_states = self.k_norm(
            self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(
            1, 2)
        if self.use_head_wise_attn_gate:
            gate_states = self.g_proj(hidden_states)
        cos, sin = self.rotary_emb(hidden_states, position_ids)

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

        # query_states, key_states = apply_rotary_pos_emb(query_norm_states, key_norm_states, cos, sin)
        if past_key_value is not None:
            # sin and cos are specific to RoPE models; position_ids needed for the static cache
            cache_kwargs = {
                "sin": sin,
                "cos": cos,
                "cache_position": cache_position
            }
            key_states, value_states = past_key_value.update(
                key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = eager_attention_forward
        # TODO: considering FP8；
        # RuntimeError: Expected attn_mask dtype to be bool or float or to match query dtype,
        # but got attn_mask.dtype: long int and  query.dtype: c10::BFloat16 instead.
        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,
            sliding_window=self.sliding_window,  # main diff with Llama
            **kwargs,
        )
        attn_output = attn_output.reshape(*input_shape, -1)
        if self.use_head_wise_attn_gate:
            output = attn_output.view(
                *attn_output.shape[:-1], self.num_attention_heads,
                self.head_dim) * gate_states.unsqueeze(-1).sigmoid()
            attn_output = output.view(*attn_output.shape)
        attn_output = self.o_proj(attn_output)

        return attn_output, attn_weights


class Step3p5DecoderLayer(GradientCheckpointingLayer):

    def __init__(self, config, layer_idx):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.layer_idx = layer_idx
        self.self_attn = Step3p5Attention(config, layer_idx)
        self.attention_type = config.layer_types[layer_idx]

        moe_layers_enum = getattr(config, "moe_layers_enum", None)
        if moe_layers_enum is not None:
            moe_layers_idx = [
                int(i) for i in moe_layers_enum.strip().split(',')
            ]
        else:
            moe_layers_idx = [i for i in range(1, config.num_hidden_layers)]
        self.is_moe_layer = layer_idx in moe_layers_idx
        self.use_moe = False

        if config.swiglu_limits_shared and config.swiglu_limits_shared[
                layer_idx] is not None and config.swiglu_limits_shared[
                    layer_idx] != 0:
            swiglu_limit_shared = config.swiglu_limits_shared[layer_idx]
        else:
            swiglu_limit_shared = None
        if config.swiglu_limits and config.swiglu_limits[
                layer_idx] is not None and config.swiglu_limits[layer_idx] != 0:
            swiglu_limit = config.swiglu_limits[layer_idx]
        else:
            swiglu_limit = None
        if self.is_moe_layer:
            self.moe = Step3p5MoEMLP(config, swiglu_limit=swiglu_limit)  #
            self.share_expert = Step3p5MLP(
                config,
                intermediate_size=config.share_expert_dim,
                swiglu_limit=swiglu_limit_shared)
            self.use_moe = True
        else:
            self.mlp = Step3p5MLP(config,
                                 intermediate_size=config.intermediate_size,
                                 swiglu_limit=swiglu_limit_shared)

        self.input_layernorm = Step3p5RMSNorm(
            config.hidden_size,
            eps=config.rms_norm_eps)
        self.post_attention_layernorm = Step3p5RMSNorm(
            config.hidden_size,
            eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[tuple[torch.Tensor]] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> torch.FloatTensor:
        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_value=past_key_value,
            cache_position=cache_position,
            **kwargs,
        )
        hidden_states = residual + hidden_states
        
        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        if self.use_moe:
            share_output = self.share_expert(hidden_states)
            moe_output = self.moe(hidden_states)
            ffn_output = moe_output + share_output
        else:
            ffn_output = self.mlp(hidden_states)
        if isinstance(ffn_output, tuple):
            hidden_states, _ = ffn_output
        else:
            hidden_states = ffn_output

        hidden_states = residual + hidden_states
        return hidden_states


class Step3p5PreTrainedModel(PreTrainedModel):
    # Link this model family to its configuration class so PreTrainedModel.from_pretrained
    # can load the config instead of failing with a NoneType error.
    config_class = Step3p5Config
    supports_gradient_checkpointing = True
    _skip_keys_device_placement = ["past_key_values"]
    _keys_to_ignore_on_load_unexpected = [
        r"model\.layers\.45\.*",
        r"model\.layers\.46\.*",
        r"model\.layers\.47\.*"
    ]
    _supports_flash_attn = False
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_static_cache = True
    _supports_attention_backend = True


class Step3p5Model(Step3p5PreTrainedModel, GenerationMixin):
    _no_split_modules = ["Step3p5DecoderLayer"]
    base_model_prefix = "model"
    _tied_weights_keys = ["lm_head.weight"]
    config: Step3p5Config
    def __init__(self, config: Step3p5Config):
        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([
            Step3p5DecoderLayer(config, layer_idx)
            for layer_idx in range(config.num_hidden_layers)
        ])
        self.norm = Step3p5RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.gradient_checkpointing = False
        self.has_sliding_layers = "sliding_attention" in self.config.layer_types

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

    def get_input_embeddings(self, input_ids):
        return self.embed_tokens(input_ids)

    @can_return_tuple
    def forward(
        self,
        input_ids: 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,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> 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 None) ^ (inputs_embeds is not None):
            raise ValueError(
                "You must specify exactly one of input_ids or inputs_embeds")

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

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

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache()

        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.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)

        hidden_states = inputs_embeds

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            # Prepare mask arguments
            mask_kwargs = {
                "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,
            }
            # Create the masks
            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
            }

            # The sliding window alternating layers are not always activated depending on the config
            if self.has_sliding_layers:
                causal_mask_mapping[
                    "sliding_attention"] = create_sliding_window_causal_mask(
                        **mask_kwargs)

        # # create position embeddings to be shared across the decoder layers
        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        for decoder_layer in self.layers[:self.config.num_hidden_layers]:
            if output_hidden_states:
                all_hidden_states += (hidden_states, )

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[
                    decoder_layer.attention_type],
                position_ids=position_ids,
                past_key_value=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                **kwargs,
            )

            hidden_states = layer_outputs

        hidden_states = self.norm(hidden_states)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )


class Step3p5ForCausalLM(Step3p5PreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    config: Step3p5Config

    def __init__(self, config: Step3p5Config):
        super().__init__(config)
        self.model = Step3p5Model(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.get_input_embeddings()

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

    def get_output_embeddings(self):
        return self.model.get_output_embeddings()

    def set_output_embeddings(self, new_embeddings):
        self.model.set_output_embeddings(new_embeddings)

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

    def get_decoder(self):
        return self.model.get_decoder()
    
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        num_patches=None,
        patch_pixel_values=None,
        patch_newline_mask=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,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Union[tuple, Step3p5CausalLMOutputWithPast]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
        Example:
        ```python
        >>> from transformers import AutoTokenizer, Llama4ForCausalLM
        >>> model = Llama4ForCausalLM.from_pretrained("meta-llama4/Llama4-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama4/Llama4-2-7b-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."
        ```"""

        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)
        # breakpoint()
        outputs = self.model(
            input_ids=input_ids,
            num_patches=num_patches,
            patch_pixel_values=patch_pixel_values,
            patch_newline_mask=patch_newline_mask,
            position_ids=position_ids,
            attention_mask=attention_mask,
            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,
            cache_position=cache_position,
            **kwargs,
        )
        hidden_states = outputs.last_hidden_state
        logits = self.lm_head(hidden_states)

        return Step3p5CausalLMOutputWithPast(logits=logits, )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        attention_mask=None,
        cache_position=None,
        logits_to_keep=None,
        **kwargs,
    ):

        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            logits_to_keep=logits_to_keep,
            **kwargs,
        )

        if cache_position[0] == 0:
            # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
            # Otherwise we need pixel values to be passed to model
            model_inputs["pixel_values"] = pixel_values

        return model_inputs

    def _fix_state_dict_key_on_load(self, key: str) -> tuple[str, bool]:
        if key.startswith("language_model."):
            return key[len("language_model."):], True

        return key, False