modeling_brumby.py

# coding=utf-8
# Copyright 2024 HuggingFace Inc. team.
# Copyright (c) 2025, NVIDIA CORPORATION. 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.
"""PyTorch Brumby model."""

from typing import Callable, Optional, Union, Any, Dict, Tuple

import torch
from torch import nn

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache, StaticCache, SlidingWindowCache
from transformers.generation import GenerationMixin
from transformers.integrations import use_kernel_forward_from_hub
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple
from transformers.utils.deprecation import deprecate_kwarg
from transformers.utils.generic import check_model_inputs
from .configuration_brumby import BrumbyConfig


try:
    from retention.triton import power_retention, power_retention_inference
    from retention._utils import compute_expanded_dim
except ImportError:
    raise ImportError("Retention is required by the Brumby model. Please install it with `pip install retention`.")


class PowerAttentionDynamicCache(DynamicCache):
    """
    A dynamic cache that encompasses 2 sets of caches:
    1. attention cache with a short seq_len (determined by the chunk_size parameter):
       - key_cache: [batch_size, num_heads, chunk_size, head_dim]
       - value_cache: [batch_size, num_heads, chunk_size, head_dim]
       - gating_cache: [batch_size, num_heads, chunk_size]
    2. fixed-size state-based cache:
       - state: [batch_size, num_heads, state_dim, head_dim]
       - sum_of_keys: [batch_size, num_heads, state_dim]

    where state_dim is determined by the power of expansion for power attention.
    """
    def __init__(self, config: BrumbyConfig, batch_size: int, dtype=torch.bfloat16, device=None):
        super().__init__()
        self.config = config
        self.batch_size = batch_size
        self.device = device
        self.dtype = dtype
        self.chunk_size = config.chunk_size
        self.head_dim = config.hidden_size // config.num_attention_heads
        self.p = config.p
        self.state_dim = compute_expanded_dim(self.head_dim, deg=self.p)
        
        self.states = [None for _ in range(config.num_hidden_layers)]
        self.sum_of_keys = [None for _ in range(config.num_hidden_layers)]
        self.key_cache = [torch.tensor([[]] * batch_size, device=device, dtype=dtype) for _ in range(config.num_hidden_layers)]
        self.value_cache = [torch.tensor([[]] * batch_size, device=device, dtype=dtype) for _ in range(config.num_hidden_layers)]
        self.gate_cache = [torch.tensor([[]] * batch_size, device=device, dtype=torch.float32) for _ in range(config.num_hidden_layers)]

    def clean_cache(self, layer_idx: int) -> None:
        self.key_cache[layer_idx] = torch.tensor([[]] * self.batch_size, device=self.device, dtype=self.dtype)
        self.value_cache[layer_idx] = torch.tensor([[]] * self.batch_size, device=self.device, dtype=self.dtype)
        self.gate_cache[layer_idx] = torch.tensor([[]] * self.batch_size, device=self.device, dtype=torch.float32)

    def update_cache(
        self,
        key_states: torch.Tensor,
        value_states: torch.Tensor,
        gate_states: torch.Tensor,
        layer_idx: int,
        cache_kwargs: Optional[Dict[str, Any]] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        if self.key_cache[layer_idx].shape[-1] == 0:
            self.key_cache[layer_idx] = key_states
            self.value_cache[layer_idx] = value_states
            self.gate_cache[layer_idx] = gate_states
        else:
            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)
            self.gate_cache[layer_idx] = torch.cat([self.gate_cache[layer_idx], gate_states], dim=2)
        return self.key_cache[layer_idx], self.value_cache[layer_idx], self.gate_cache[layer_idx], self.states[layer_idx], self.sum_of_keys[layer_idx]

    def reorder_cache(self, beam_idx: torch.LongTensor):
        """Reorders the cache for beam search, given the selected beam indices."""
        for layer_idx in range(len(self.key_cache)):
            device = self.key_cache[layer_idx].device
            self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
            device = self.value_cache[layer_idx].device
            self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))
            device = self.gate_cache[layer_idx].device
            self.gate_cache[layer_idx] = self.gate_cache[layer_idx].index_select(0, beam_idx.to(device))
            device = self.states[layer_idx].device
            self.states[layer_idx] = self.states[layer_idx].index_select(0, beam_idx.to(device))
            device = self.sum_of_keys[layer_idx].device
            self.sum_of_keys[layer_idx] = self.sum_of_keys[layer_idx].index_select(0, beam_idx.to(device))

    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
        # take any layer that contains cache and not empty tensor
        if layer_idx is None:
            layer_idx = 0
        if layer_idx >= len(self.key_cache):
            return 0
        # Check if the cache for this layer is empty
        if self.key_cache[layer_idx].numel() == 0:
            return 0
        return self.key_cache[layer_idx].shape[-2]

    def to_legacy_cache(self) -> tuple[tuple[torch.Tensor, torch.Tensor]]:
        raise NotImplementedError("PowerAttentionDynamicCache does not have a legacy cache equivalent.")

    @classmethod
    def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "PowerAttentionDynamicCache":
        raise NotImplementedError("PowerAttentionDynamicCache does not have a legacy cache equivalent.")

    def update_state(self, layer_idx: int, new_state: torch.Tensor, new_sum_of_keys: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        self.states[layer_idx] = new_state
        self.sum_of_keys[layer_idx] = new_sum_of_keys
        return self.states[layer_idx], self.sum_of_keys[layer_idx]

    def reset(self):
        self.states.zero_()
        self.sum_of_keys.zero_()
        self.key_cache.zero_()
        self.value_cache.zero_()
        self.gate_cache.zero_()



@use_kernel_forward_from_hub("RMSNorm")
class BrumbyRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        BrumbyRMSNorm is equivalent to T5LayerNorm
        """
        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)

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


class BrumbyMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=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[config.hidden_act]

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


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


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

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

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

    return attn_output, attn_weights


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

    def __init__(self, config: BrumbyConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True
        self.use_exp = config.use_exp
        self.prefill_chunk_size = config.prefill_chunk_size
        self.chunk_size = config.chunk_size
        self.switch_over_seq_len = config.switch_over_seq_len

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.g_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.q_norm = BrumbyRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
        self.k_norm = BrumbyRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
        self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_values: Optional[PowerAttentionDynamicCache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, Optional[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)
        gate_states = self.g_proj(hidden_states).view(hidden_shape[:-1]).transpose(1, 2)
        gate_states = nn.functional.logsigmoid(gate_states.to(torch.float32))

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

        if past_key_values is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states, gate_states, state, sum_of_keys = past_key_values.update_cache(key_states, value_states, gate_states, self.layer_idx, cache_kwargs)

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

            attn_output, attn_weights = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask,
                dropout=0.0 if not self.training else self.attention_dropout,
                scaling=self.scaling,
                sliding_window=self.sliding_window,  # diff with Llama
                **kwargs,
            )
        
        elif query_states.shape[2] == 1:
            key_len = key_states.shape[2]
            attn_output, state, sum_of_keys = power_retention_inference(
                query_states.transpose(1, 2),
                key_states.transpose(1, 2),
                value_states.transpose(1, 2),
                gate_states.transpose(1, 2),
                initial_state=state,
                sum_of_keys=sum_of_keys,
                deg=2,
                scale=self.scaling,
                switch_over_seq_len=self.chunk_size,
            )
            if self.chunk_size is not None and key_len >= self.chunk_size:
                past_key_values.clean_cache(self.layer_idx)
                past_key_values.update_state(self.layer_idx, state, sum_of_keys)

            attn_weights = None

        else:
            key_len = key_states.shape[2]
            attn_output = power_retention(
                query_states.transpose(1, 2),
                key_states.transpose(1, 2),
                value_states.transpose(1, 2),
                gate_states.transpose(1, 2),
                deg=2,
                scale=self.scaling,
                chunk_size=self.prefill_chunk_size, # enable chunked prefilling by default
                switch_over_seq_len=self.switch_over_seq_len,
            )
            attn_weights = None


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


class BrumbyDecoderLayer(nn.Module):
    def __init__(self, config: BrumbyConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

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

        self.mlp = BrumbyMLP(config)
        self.input_layernorm = BrumbyRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = BrumbyRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attention_type = config.layer_types[layer_idx]

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

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


@auto_docstring
class BrumbyPreTrainedModel(PreTrainedModel):
    config: BrumbyConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["BrumbyDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": BrumbyDecoderLayer,
        "attentions": BrumbyAttention,
    }


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

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

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

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

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

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

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


@auto_docstring
class BrumbyModel(BrumbyPreTrainedModel):
    def __init__(self, config: BrumbyConfig):
        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(
            [BrumbyDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = BrumbyRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = BrumbyRotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        self.has_sliding_layers = "sliding_attention" in self.config.layer_types

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

    @check_model_inputs
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[PowerAttentionDynamicCache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = 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:
            raise ValueError("Brumby requires an initialized `PowerAttentionDynamicCache` to return a cache. None was provided, so no cache will be returned.")

        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)

        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions=False
        )

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

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

        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
        )

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool = False,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and past_key_values is not None:
                is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
                if is_padding_right:
                    raise ValueError(
                        "You are attempting to perform batched generation with padding_side='right'"
                        " this may lead to unexpected behaviour for Flash Attention version of Qwen3. Make sure to "
                        " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                    )
            if attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)
        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if (
            self.config._attn_implementation == "sdpa"
            and not (using_static_cache or using_sliding_window_cache)
            and not output_attentions
        ):
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                sliding_window=self.config.sliding_window,
                is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        # SlidingWindowCache or StaticCache
        if using_sliding_window_cache or using_static_cache:
            target_length = past_key_values.get_max_cache_shape()
        # DynamicCache or no cache
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
            config=self.config,
            past_key_values=past_key_values,
        )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type in ["cuda", "xpu"]
            and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        device: torch.device,
        cache_position: torch.Tensor,
        batch_size: int,
        config: BrumbyConfig,
        past_key_values: Cache,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            device (`torch.device`):
                The device to place the 4D attention mask on.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
            config (`Qwen3Config`):
                The model's configuration class
            past_key_values (`Cache`):
                The cache class that is being used currently to generate
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
            )
            diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            if config.sliding_window is not None:
                # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
                # the check is needed to verify is current checkpoint was trained with sliding window or not
                if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
                    sliding_attend_mask = torch.arange(target_length, device=device) <= (
                        cache_position.reshape(-1, 1) - config.sliding_window
                    )
                    diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
            causal_mask *= diagonal_attend_mask
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                if attention_mask.shape[-1] > target_length:
                    attention_mask = attention_mask[:, :target_length]
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )
        return causal_mask


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

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

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

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        **kwargs,
    ):
        # Copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
        # Overwitten -- uses `past_key_values` as opposed to `past_key_values`
        empty_past_kv = past_key_values is None
        
        # First pass
        if not isinstance(past_key_values, PowerAttentionDynamicCache):
            past_key_values = PowerAttentionDynamicCache(self.config, input_ids.shape[0], self.dtype, device=self.device)

        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
        # Exception 1: when passing input_embeds, input_ids may be missing entries
        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
        # Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case.
        #              (we can't check exception 3 while compiling)
        if not empty_past_kv:
            if (
                inputs_embeds is not None  # Exception 1
                or cache_position[-1] >= input_ids.shape[1]  # Exception 3
            ):
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]
        else:
            past_key_values = PowerAttentionDynamicCache(
                self.config, input_ids.shape[0], self.dtype, device=self.device
            )

        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if not empty_past_kv:
                position_ids = position_ids[:, -input_ids.shape[1] :]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and empty_past_kv:
            # TODO(pjin): workaround fix for properly extending inputs_embeds;
            # longer term, may be better handled elsewhere in .generate().
            if input_ids is not None and inputs_embeds.shape[1] < input_ids.shape[1]:
                new_token_embeds = self.get_input_embeddings()(input_ids[:,inputs_embeds.shape[1]:])
                inputs_embeds = torch.cat([inputs_embeds, new_token_embeds], dim=1)
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "cache_position": cache_position,
            }
        )
        return model_inputs

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

        >>> model = BrumbyForCausalLM.from_pretrained("Qwen/Brumby-8B")
        >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Brumby-8B")

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

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

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

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

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


__all__ = [
    "BrumbyForCausalLM",
    "BrumbyPreTrainedModel",
    "BrumbyModel",
]