#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/nemotron_h/modular_nemotron_h.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_nemotron_h.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
# 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.


import math
from collections.abc import Callable

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

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...integrations import (
    use_experts_implementation,
    use_kernel_forward_from_hub,
    use_kernel_func_from_hub_with_fallback,
    use_kernelized_func,
)
from ...integrations.accelerate import force_accelerate_hooks
from ...masking_utils import create_causal_mask, create_recurrent_attention_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...models.zamba2.modeling_zamba2 import Zamba2RMSNormGated
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
from ...utils.generic import merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from .configuration_nemotron_h import NemotronHConfig


# Helper methods for segment sum computation


def pad_tensor_by_size(input_tensor: torch.Tensor, pad_size: int):
    """
    Padding x tensor with `pad_size` on the seq_len dim (dim=1)

    Assumes that we only have tensors of either size 4 or 3
    """
    pad_shape = (0, 0, 0, 0, 0, pad_size, 0, 0) if len(input_tensor.shape) == 4 else (0, 0, 0, pad_size, 0, 0)

    return torch.nn.functional.pad(input_tensor, pad_shape, mode="constant", value=0)


def reshape_into_chunks(input_tensor, pad_size, chunk_size):
    """
    Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
    simultaneously splitting it into chunk sequences.

    Assumes that we only have tensors of either size 4 or 3
    """
    # [bsz, seq_len, ...] -> [bsz, seq_len multiple of chunk_size, ...]
    input_tensor = pad_tensor_by_size(input_tensor, pad_size)

    if len(input_tensor.shape) == 3:
        # [bsz, seq_len multiple of chunk_size, num_heads] -> [bsz, -1, chunk_size, num_heads]
        return input_tensor.reshape(input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2])
    else:
        # [bsz, seq_len multiple of chunk_size, num_heads, head_dim or state_size] -> [bsz, -1, chunk_size, num_heads, head_dim or state_size]
        return input_tensor.reshape(
            input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2], input_tensor.shape[3]
        )


def segment_sum(input_tensor):
    """
    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
    """
    chunk_size = input_tensor.size(-1)
    # 1. expand input tensor to have an additional dimension and repeat along that dimension
    # [..., chunk_size] -> [..., chunk_size, chunk_size]
    input_tensor = input_tensor[..., None].expand(*input_tensor.size(), chunk_size)
    # 2. create a lower triangular mask with the diagonal set to 0 to 0 out elements above diag
    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=-1)
    input_tensor = input_tensor.masked_fill(~mask, 0)
    # 3. compute actual cumsum
    tensor_segsum = torch.cumsum(input_tensor, dim=-2)

    # 4. apply mask to keep only the lower triangular part of the cumulative sum result (incl diagonal this time)
    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=0)
    tensor_segsum = tensor_segsum.masked_fill(~mask, -torch.inf)
    return tensor_segsum


def apply_mask_to_padding_states(hidden_states, attention_mask):
    """
    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
    """
    # NOTE: attention mask is a 2D boolean tensor
    if attention_mask is not None:
        dtype = hidden_states.dtype
        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

    return hidden_states


@use_kernel_func_from_hub_with_fallback("causal_conv1d_update", "causal_conv1d")
def causal_conv1d_update(
    hidden_states: torch.Tensor,
    conv_state: torch.Tensor,
    weight: nn.Parameter,
    bias: nn.Parameter | None = None,
    activation: str | None = None,
):
    _, hidden_size, seq_len = hidden_states.shape
    state_len = conv_state.shape[-1]

    hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
    conv_state.copy_(hidden_states_new[:, :, -state_len:])
    out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size)
    out = out[:, :, -seq_len:]
    if activation is not None:
        out = ACT2FN[activation](out)
    return out.to(hidden_states.dtype)


@use_kernel_func_from_hub_with_fallback("causal_conv1d_fn", "causal_conv1d")
def causal_conv1d_fn(
    hidden_states: torch.Tensor,
    weight: nn.Parameter,
    bias: nn.Parameter | None = None,
    activation: str | None = None,
    **kwargs,
):
    _, hidden_size, seq_len = hidden_states.shape
    padding = weight.shape[-1] - 1

    out = F.conv1d(
        hidden_states.to(weight.dtype),
        weight=weight.unsqueeze(1),
        bias=bias,
        padding=padding,
        groups=hidden_size,
    )[:, :, :seq_len]
    if activation is not None:
        out = ACT2FN[activation](out)
    return out.to(hidden_states.dtype)


@use_kernel_func_from_hub_with_fallback("mamba_split_conv1d_scan_combined", "mamba_ssm")
def mamba2_split_conv1d_scan_combined(
    zxbcdt: torch.Tensor,
    conv1d_weight: torch.Tensor,
    conv1d_bias: torch.Tensor | None,
    dt_bias: torch.Tensor,
    A: torch.Tensor,
    D: torch.Tensor,
    chunk_size: int,
    initial_states: torch.Tensor | None = None,
    dt_limit: tuple[float, float] = (0.0, float("inf")),
    return_final_states: bool = False,
    activation: str = "silu",
    rmsnorm_weight: torch.Tensor | None = None,
    rmsnorm_eps: float = 1e-6,
    outproj_weight: torch.Tensor | None = None,
    outproj_bias: torch.Tensor | None = None,
    headdim: int | None = None,
    ngroups: int = 1,
    norm_before_gate: bool = True,
    **kwargs,
):
    return None


@use_kernel_func_from_hub_with_fallback("selective_state_update", "mamba_ssm")
def mamba2_selective_state_update(
    state: torch.Tensor,
    hidden_states: torch.Tensor,
    dt: torch.Tensor,
    A: torch.Tensor,
    B: torch.Tensor,
    C: torch.Tensor,
    D: torch.Tensor | None = None,
    dt_bias: torch.Tensor | None = None,
    dt_softplus: bool = False,
    z: torch.Tensor | None = None,
    **kwargs,
):
    batch_size, num_heads, head_dim = hidden_states.shape
    num_groups = B.shape[1]
    state_size = B.shape[-1]

    if dt_bias is not None:
        dt = dt + dt_bias.to(dt.dtype)
    if dt_softplus:
        dt = F.softplus(dt)
    dt = dt[..., None]

    # Discretize A
    dA = torch.exp(dt.float() * A.float()).to(device=state.device)

    # Discretize B
    B = B.reshape(batch_size, num_groups, 1, state_size)
    B = B.expand(batch_size, num_groups, num_heads // num_groups, state_size).contiguous()
    B = B.reshape(batch_size, num_heads, 1, state_size)
    dB = dt * B

    # Discretize x into dB
    dBx = (dB * hidden_states[..., None]).to(device=state.device)

    # State calculation
    ssm_states = state * dA + dBx
    state.copy_(ssm_states.to(state.dtype))

    # Subsequent output
    C = C.reshape(batch_size, num_groups, 1, state_size)
    C = C.expand(batch_size, num_groups, num_heads // num_groups, state_size).contiguous()
    C = C.reshape(batch_size, num_heads, state_size)

    # Reshape ssm_states to merge the first two dimensions
    ssm_states = ssm_states.to(device=C.device, dtype=C.dtype)
    ssm_states_reshaped = ssm_states.view(batch_size * num_heads, head_dim, state_size)
    C_reshaped = C.view(batch_size * num_heads, state_size, 1)
    out = torch.bmm(ssm_states_reshaped, C_reshaped)
    out = out.view(batch_size, num_heads, head_dim)

    # D skip connection
    if D is not None:
        out = (out + hidden_states * D).to(out.dtype)

    if z is not None:
        out = out * F.silu(z)

    return out.to(hidden_states.dtype)


@use_kernel_func_from_hub_with_fallback("mamba_chunk_scan_combined", "mamba_ssm")
def mamba2_chunk_scan(
    hidden_states: torch.Tensor,
    dt: torch.Tensor,
    A: torch.Tensor,
    B: torch.Tensor,
    C: torch.Tensor,
    chunk_size: int,
    D: torch.Tensor | None = None,
    dt_bias: torch.Tensor | None = None,
    initial_states: torch.Tensor | None = None,
    dt_softplus: bool = False,
    dt_limit: tuple[float, float] = (0.0, float("inf")),
    return_final_states: bool = False,
    **kwargs,
):
    batch_size, sequence_length, num_heads, head_dim = hidden_states.shape
    num_groups = B.shape[2]

    if dt_bias is not None:
        dt = dt + dt_bias.to(dt.dtype)
    if dt_softplus:
        dt = F.softplus(dt)
    dt = torch.clamp(dt, min=dt_limit[0], max=dt_limit[1])

    hidden_states = hidden_states.float()
    B = B.float().repeat_interleave(num_heads // num_groups, dim=2, output_size=num_heads)
    C = C.float().repeat_interleave(num_heads // num_groups, dim=2, output_size=num_heads)

    pad_size = (chunk_size - sequence_length % chunk_size) % chunk_size
    D_residual = None
    if D is not None:
        D_residual = D[..., None] * pad_tensor_by_size(hidden_states, pad_size)

    # Discretize x and A
    hidden_states = hidden_states * dt[..., None].float()
    A = A.to(hidden_states.dtype) * dt.float()

    # Rearrange into blocks/chunks
    hidden_states, A, B, C = [reshape_into_chunks(tensor, pad_size, chunk_size) for tensor in (hidden_states, A, B, C)]

    A = A.permute(0, 3, 1, 2)
    A_cumsum = torch.cumsum(A, dim=-1)

    # 1. Compute the output for each intra-chunk (diagonal blocks)
    # This is the analog of a causal mask
    L = torch.exp(segment_sum(A))

    # Contraction of C and B to get G (attention-weights like)
    G = (C[:, :, :, None, :, :] * B[:, :, None, :, :, :]).sum(dim=-1)

    # Compute M, equivalent to applying attention mask to weights
    M = (G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]).sum(dim=-1)

    # Compute Y_diag (apply to values)
    Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(dim=3)

    # 2. Compute the state for each intra-chunk
    # (right term of low-rank factorization of off-diagonal blocks; B terms)
    decay_states = torch.exp(A_cumsum[:, :, :, -1:] - A_cumsum)
    B_decay = B * decay_states.permute(0, -2, -1, 1)[..., None]
    states = (B_decay[..., None, :] * hidden_states[..., None]).sum(dim=2)

    # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
    # (middle term of factorization of off-diag blocks; A terms)
    previous_states = (
        initial_states[:, None].to(dtype=states.dtype, device=states.device)
        if initial_states is not None
        else torch.zeros_like(states[:, :1])
    )
    states = torch.cat([previous_states, states], dim=1)
    decay_chunk = torch.exp(segment_sum(F.pad(A_cumsum[:, :, :, -1], (1, 0)))).transpose(1, 3)
    new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
    states, final_state = new_states[:, :-1], new_states[:, -1]

    # 4. Compute state -> output conversion per chunk
    # (left term of low-rank factorization of off-diagonal blocks; C terms)
    state_decay_out = torch.exp(A_cumsum)
    C_times_states = C[..., None, :] * states[:, :, None, ...]
    Y_off = C_times_states.sum(-1) * state_decay_out.permute(0, 2, 3, 1)[..., None]

    # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)
    output = Y_diag + Y_off
    output = output.reshape(batch_size, -1, num_heads, head_dim)

    if D_residual is not None:
        output = output + D_residual

    # Cutting off padded chunks
    if pad_size > 0:
        output = output[:, :sequence_length]

    if return_final_states:
        return output, final_state

    return output


@use_kernelized_func(
    [
        causal_conv1d_fn,
        causal_conv1d_update,
        mamba2_split_conv1d_scan_combined,
        mamba2_selective_state_update,
        mamba2_chunk_scan,
    ]
)
class NemotronHMamba2Mixer(nn.Module):
    """
    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
    and is why Mamba is called **selective** state spaces)
    """

    def __init__(self, config: NemotronHConfig, layer_idx: int | None = None, initialize_mixer_weights: bool = True):
        super().__init__()
        self.config = config
        self.num_heads = config.mamba_num_heads
        self.hidden_size = config.hidden_size
        self.ssm_state_size = config.ssm_state_size
        self.conv_kernel_size = config.conv_kernel
        self.intermediate_size = config.mamba_num_heads * config.mamba_head_dim
        self.layer_idx = layer_idx
        self.use_conv_bias = config.use_conv_bias
        self.activation = config.mamba_hidden_act
        self.act = ACT2FN[config.mamba_hidden_act]

        self.n_groups = config.n_groups
        self.head_dim = config.mamba_head_dim
        self.chunk_size = config.chunk_size
        # No upper limit
        self.time_step_limit = (config.time_step_min, float("inf"))

        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size

        self.conv1d = nn.Conv1d(
            in_channels=self.conv_dim,
            out_channels=self.conv_dim,
            bias=config.use_conv_bias,
            kernel_size=self.conv_kernel_size,
            groups=self.conv_dim,
            padding=self.conv_kernel_size - 1,
        )
        # projection of the input hidden states
        projection_size = self.intermediate_size + self.conv_dim + self.num_heads
        self.in_proj = nn.Linear(
            self.hidden_size,
            projection_size,
            bias=config.use_bias,
        )
        # selective projection used to make dt, B and C input dependent

        # time step projection (discretization)
        self.dt_bias = nn.Parameter(torch.empty(self.num_heads))

        # S4D real initialization. These are not discretized!
        # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
        self.A_log = nn.Parameter(torch.empty(self.num_heads))
        self.norm = Zamba2RMSNormGated(
            self.intermediate_size, group_size=self.intermediate_size // self.n_groups, eps=config.layer_norm_epsilon
        )
        self.D = nn.Parameter(torch.empty(self.num_heads))
        if initialize_mixer_weights and self.dt_bias.device.type != "meta":
            self.init_nemotron_h_mamba2_weights()
        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)

        self.layer_type = config.layer_types[layer_idx]
        self.use_mem_eff_path = True

    @torch.no_grad()
    def init_nemotron_h_mamba2_weights(self):
        A = torch.arange(1, self.num_heads + 1, device=self.A_log.device, dtype=torch.float32)
        init.copy_(self.A_log, torch.log(A))
        init.ones_(self.D)
        init.ones_(self.dt_bias)

    @force_accelerate_hooks("conv1d")
    def forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs,
    ):
        batch_size, seq_len, _ = hidden_states.shape
        dtype = hidden_states.dtype
        use_precomputed_states = cache_params is not None and cache_params.has_previous_state(self.layer_idx)

        # 1. Gated MLP's linear projection
        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
        projected_states = self.in_proj(hidden_states)

        A = -torch.exp(self.A_log.float())
        fused_kwargs = kwargs | {"dt_limit": self.time_step_limit}
        if self.training and cache_params is None:
            fused_output = mamba2_split_conv1d_scan_combined(
                projected_states,
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                self.dt_bias,
                A,
                D=self.D,
                chunk_size=self.chunk_size,
                activation=self.activation,
                rmsnorm_weight=self.norm.weight,
                rmsnorm_eps=self.norm.variance_epsilon,
                outproj_weight=self.out_proj.weight,
                outproj_bias=self.out_proj.bias,
                headdim=self.head_dim,
                ngroups=self.n_groups,
                norm_before_gate=False,
                return_final_states=False,
                **fused_kwargs,
            )

            # Only kernels can use this shortcircuit, fallback to normal torch otherwise
            if fused_output is not None:
                return fused_output

        gate, hidden_states_B_C, dt = projected_states.split(
            [self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
        )

        if use_precomputed_states:
            conv_state = cache_params.layers[self.layer_idx].conv_states[0]
            recurrent_state = cache_params.layers[self.layer_idx].recurrent_states[0]

        # 2. Convolution sequence transformation
        hidden_states_B_C = hidden_states_B_C.transpose(1, 2)
        if use_precomputed_states and seq_len == 1 and not cache_params.layers[self.layer_idx].record_past:
            hidden_states_B_C = causal_conv1d_update(
                hidden_states_B_C,
                conv_state,
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                activation=self.activation,
            )
        else:
            if cache_params is not None:
                hidden_states_B_C = cache_params.update_conv_state(
                    hidden_states_B_C,
                    self.layer_idx,
                    conv_kernel_size=self.conv_kernel_size,
                )

            hidden_states_B_C = causal_conv1d_fn(
                hidden_states_B_C,
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                activation=self.activation,
                **kwargs,
            )

            if cache_params is not None:
                hidden_states_B_C = hidden_states_B_C[:, :, -seq_len:]

        # 3. SSM transformation
        hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C.transpose(1, 2), attention_mask)
        hidden_states, B, C = torch.split(
            hidden_states_B_C,
            [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size],
            dim=-1,
        )

        # Recurrent form
        if use_precomputed_states and seq_len == 1:
            hidden_states = hidden_states.view(batch_size, self.num_heads, self.head_dim)
            dt = dt.transpose(1, 2).expand(-1, -1, self.head_dim)
            A = A[:, None, None].expand(-1, self.head_dim, self.ssm_state_size)
            B = B.view(batch_size, self.n_groups, self.ssm_state_size)
            C = C.view(batch_size, self.n_groups, self.ssm_state_size)
            D = self.D[:, None].expand(-1, self.head_dim)
            dt_bias = self.dt_bias[:, None].expand(-1, self.head_dim)

            scan_output = mamba2_selective_state_update(
                recurrent_state,
                hidden_states,
                dt,
                A,
                B,
                C,
                D,
                z=None,
                dt_bias=dt_bias,
                dt_softplus=True,
                **kwargs,
            )
            scan_output = scan_output.view(batch_size, 1, -1)

        # Chunk form
        else:
            output_final_state = cache_params is not None
            scan_result = mamba2_chunk_scan(
                hidden_states.view(batch_size, seq_len, self.num_heads, self.head_dim),
                dt,
                A,
                B.view(batch_size, seq_len, self.n_groups, self.ssm_state_size),
                C.view(batch_size, seq_len, self.n_groups, self.ssm_state_size),
                chunk_size=self.chunk_size,
                D=self.D,
                z=None,
                return_final_states=output_final_state,
                dt_bias=self.dt_bias,
                dt_softplus=True,
                initial_states=recurrent_state if use_precomputed_states else None,
                dt_limit=self.time_step_limit,
                **kwargs,
            )

            if output_final_state:
                scan_output, final_state = scan_result
                cache_params.update_recurrent_state(final_state, layer_idx=self.layer_idx)
            else:
                scan_output = scan_result

            scan_output = scan_output.reshape(batch_size, seq_len, -1)

        scan_output = self.norm(scan_output, gate)

        # 4. Final linear projection
        contextualized_states = self.out_proj(scan_output.to(dtype))
        return contextualized_states


@use_kernel_forward_from_hub("RMSNorm")
class NemotronHRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        NemotronHRMSNorm 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 NemotronHMLP(nn.Module):
    def __init__(self, config, intermediate_size=None, **kwargs):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = intermediate_size or config.intermediate_size
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
        self.act_fn = ACT2FN[config.mlp_hidden_act]

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


@use_experts_implementation(has_gate=False)
class NemotronHExperts(nn.Module):
    """
    Collection of expert weights stored as 3D tensors.

    **Architecture Note**: Unlike Mixtral or DeepSeek which use gated MLPs,
    NemotronH uses a standard MLP architecture with only up_proj and down_proj
    """

    def __init__(self, config):
        super().__init__()
        self.num_experts = config.n_routed_experts
        self.hidden_dim = config.hidden_size
        self.intermediate_dim = config.moe_intermediate_size

        # Determine input/output dimension based on whether latent projection is used
        input_dim = config.moe_latent_size if config.moe_latent_size is not None else config.hidden_size

        # All expert weights stored as 3D tensors: (num_experts, out_dim, in_dim)
        # up_proj: (num_experts, intermediate_dim, input_dim)
        self.up_proj = nn.Parameter(torch.empty(self.num_experts, self.intermediate_dim, input_dim))
        # down_proj: (num_experts, input_dim, intermediate_dim)
        self.down_proj = nn.Parameter(torch.empty(self.num_experts, input_dim, self.intermediate_dim))

        self.act_fn = ACT2FN[config.mlp_hidden_act]

    def forward(self, hidden_states: torch.Tensor, top_k_index: torch.Tensor, top_k_weights: torch.Tensor):
        final_hidden_states = torch.zeros_like(hidden_states, dtype=top_k_weights.dtype)

        # Create expert mask to identify which tokens go to which experts
        with torch.no_grad():
            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
            expert_mask = expert_mask.permute(2, 1, 0)  # (num_experts, num_experts_per_tok, num_tokens)
            # Only iterate over experts that have at least one token assigned
            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero().squeeze(-1)

        for expert_idx in expert_hit:
            expert_idx = expert_idx.item()
            # Find which tokens are routed to this expert
            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])

            if token_idx.numel() == 0:
                continue

            # Get input for this expert
            current_state = hidden_states[token_idx]

            # Expert computation: down_proj(act_fn(up_proj(x)))
            # No gating mechanism unlike Mixtral which uses: down_proj(act_fn(gate_proj(x)) * up_proj(x))
            current_hidden_states = torch.nn.functional.linear(current_state, self.up_proj[expert_idx])
            current_hidden_states = self.act_fn(current_hidden_states)
            current_hidden_states = torch.nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])

            # Apply routing weights
            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]

            # Accumulate into final output
            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))

        return final_hidden_states.to(hidden_states.dtype)


class NemotronHMoE(nn.Module):
    """
    Mixture-of-Experts (MoE) module for NemotronH.

    Unique architectures:
    - Uses non-gated MLP experts (NemotronHExperts) instead of gated experts
    - Adds optional latent projection for computational efficiency
    """

    def __init__(self, config, layer_idx: int | None = None):
        super().__init__()
        self.config = config

        # Replace with NemotronH-specific experts (non-gated MLP architecture)
        self.experts = NemotronHExperts(config)
        self.gate = NemotronHTopkRouter(config)

        # Override shared_experts to use NemotronHMLP with correct intermediate size
        self.shared_experts = NemotronHMLP(config=config, intermediate_size=config.moe_shared_expert_intermediate_size)

        # NemotronH-specific latent projection layers
        if config.moe_latent_size is not None:
            self.fc1_latent_proj = nn.Linear(config.hidden_size, config.moe_latent_size, bias=config.mlp_bias)
            self.fc2_latent_proj = nn.Linear(config.moe_latent_size, config.hidden_size, bias=config.mlp_bias)
        else:
            self.fc1_latent_proj = nn.Identity()
            self.fc2_latent_proj = nn.Identity()

    def forward(self, hidden_states) -> torch.Tensor:
        residuals = hidden_states
        orig_shape = hidden_states.shape
        _, topk_weights, topk_indices = self.gate(hidden_states)
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])

        # NemotronH-specific: latent projection
        hidden_states = self.fc1_latent_proj(hidden_states)
        hidden_states = self.experts(hidden_states, topk_indices, topk_weights)
        hidden_states = self.fc2_latent_proj(hidden_states)

        hidden_states = hidden_states.view(*orig_shape)
        hidden_states = hidden_states + self.shared_experts(residuals)
        return hidden_states


class NemotronHTopkRouter(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))
        self.routed_scaling_factor = config.routed_scaling_factor
        self.num_group = config.n_group
        self.topk_group = config.topk_group
        self.norm_topk_prob = config.norm_topk_prob
        self.e_score_correction_bias = nn.Buffer(torch.zeros(self.num_experts))

    def forward(self, hidden_states):
        hidden_states = hidden_states.view(-1, self.hidden_dim)
        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
        scores = router_logits.sigmoid()
        scores_for_choice = scores + self.e_score_correction_bias
        group_scores = (
            scores_for_choice.view(-1, self.num_group, self.num_experts // self.num_group)
            .topk(2, dim=-1)[0]
            .sum(dim=-1)
        )
        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
        group_mask = torch.zeros_like(group_scores)
        group_mask.scatter_(1, group_idx, 1)
        score_mask = (
            group_mask.unsqueeze(-1)
            .expand(-1, self.num_group, self.num_experts // self.num_group)
            .reshape(-1, self.num_experts)
        )
        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), float("-inf"))
        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
        topk_weights = scores.gather(1, topk_indices)
        if self.norm_topk_prob:
            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
            topk_weights /= denominator
        topk_weights = topk_weights * self.routed_scaling_factor
        return router_logits, topk_weights, topk_indices


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)


@use_kernel_forward_from_hub("rotary_pos_emb")
def apply_rotary_pos_emb(q, k, cos, sin, 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.
        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: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

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

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

    return attn_output, attn_weights


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

    def __init__(self, config: NemotronHConfig, 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.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = 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 past_key_values is not None:
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        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,
            **kwargs,
        )

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


MIXER_TYPES = {
    "linear_attention": NemotronHMamba2Mixer,
    "full_attention": NemotronHAttention,
    "moe": NemotronHMoE,
    "mlp": NemotronHMLP,
}


class NemotronHBlock(GradientCheckpointingLayer):
    """
    A single transformer block in the NemotronH model.

    This block can contain different types of mixers (Mamba, Attention, MLP, or MoE)
    depending on the configuration. Each block applies pre-normalization followed by
    the mixer, then adds a residual connection.

    Args:
        config (`NemotronHConfig`):
            Model configuration specifying the block architecture.
        layer_idx (`int`):
            Index of this block in the model. Used to determine the block type from
            `config.layers_block_type[layer_idx]`.
    """

    def __init__(self, config, layer_idx):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.norm = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

        self.block_type = config.layers_block_type[layer_idx]
        self.mixer = MIXER_TYPES[self.block_type](config, layer_idx=layer_idx)

    def forward(
        self,
        hidden_states,
        past_key_values: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        use_cache: bool | None = False,
        **kwargs: Unpack[TransformersKwargs],
    ):
        residual = hidden_states
        hidden_states = self.norm(hidden_states.to(dtype=self.norm.weight.dtype))

        if self.block_type == "linear_attention":
            hidden_states = self.mixer(hidden_states, cache_params=past_key_values, attention_mask=attention_mask)
        elif self.block_type == "full_attention":
            hidden_states, _ = self.mixer(
                hidden_states=hidden_states,
                past_key_values=past_key_values,
                attention_mask=attention_mask,
                position_ids=position_ids,
                use_cache=use_cache,
                **kwargs,
            )
        else:
            hidden_states = self.mixer(hidden_states)

        hidden_states = residual + hidden_states

        return hidden_states


class NemotronHPreTrainedModel(PreTrainedModel):
    config: NemotronHConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["NemotronHBlock"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _is_stateful = True
    _can_compile_fullgraph = True
    _can_record_outputs = {
        "hidden_states": NemotronHBlock,
        "attentions": NemotronHAttention,
    }
    _keep_in_fp32_modules_strict = [
        "e_score_correction_bias",
    ]
    _keys_to_ignore_on_load_unexpected = [r"mtp.*"]

    @torch.no_grad()
    def _init_weights(self, module):
        """Initialize the weights."""
        super()._init_weights(module)
        if isinstance(module, NemotronHMamba2Mixer):
            # Only re-initialise params that were NOT loaded from a checkpoint.
            # `_is_hf_initialized` is set by `from_pretrained` on each loaded
            # parameter; without this guard a post-load safety pass of
            # `_init_weights` would overwrite checkpoint values of
            # A_log / D / dt_bias with fresh random draws.
            if not getattr(module.A_log, "_is_hf_initialized", False):
                A = torch.arange(1, self.config.mamba_num_heads + 1)
                init.copy_(module.A_log, torch.log(A))
            if not getattr(module.D, "_is_hf_initialized", False):
                init.ones_(module.D)
            if not getattr(module.dt_bias, "_is_hf_initialized", False):
                dt = torch.exp(
                    torch.rand(self.config.mamba_num_heads)
                    * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
                    + math.log(self.config.time_step_min)
                ).clamp(min=self.config.time_step_floor)

                # # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
                inv_dt = dt + torch.log(-torch.expm1(-dt))
                with torch.no_grad():
                    init.copy_(module.dt_bias, inv_dt)
        elif isinstance(module, NemotronHTopkRouter):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            init.zeros_(module.e_score_correction_bias)
        elif isinstance(module, NemotronHExperts):
            # Initialize expert weights
            init.normal_(module.up_proj, mean=0.0, std=self.config.initializer_range)
            init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range)

        if isinstance(module, nn.Linear):
            if module.bias is not None:
                if not getattr(module.bias, "_is_hf_initialized", False):
                    init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            init.normal_(module.weight, std=self.config.initializer_range)

        if self.config.rescale_prenorm_residual:
            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
            #
            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
            for name, p in module.named_parameters():
                if name == "out_proj.weight":
                    # Skip checkpoint-loaded weights so a post-load safety
                    # pass of `_init_weights` doesn't silently overwrite them.
                    if getattr(p, "_is_hf_initialized", False):
                        continue
                    # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
                    # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
                    init.kaiming_uniform_(p, a=math.sqrt(5))
                    with torch.no_grad():
                        p_new = p / math.sqrt(self.config.num_hidden_layers)
                        init.copy_(p, p_new)


class NemotronHModel(NemotronHPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList([NemotronHBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])

        self.norm_f = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.embeddings

    def set_input_embeddings(self, new_embeddings):
        self.embeddings = new_embeddings

    @merge_with_config_defaults
    @capture_outputs
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):  # ^ is python for xor
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

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

        hidden_states = inputs_embeds

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

        # Under a compilable cache, `generate()` precomputes per-pattern masks and hands them in as a dict;
        # otherwise we build them here.
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            # Prepare mask arguments
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
            }
            # Create the masks
            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
                "linear_attention": create_recurrent_attention_mask(**mask_kwargs),
            }

        for layer_idx, mixer_block in enumerate(self.layers):
            hidden_states = mixer_block(
                hidden_states,
                attention_mask=causal_mask_mapping.get(mixer_block.block_type),
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                **kwargs,
            )

        hidden_states = self.norm_f(hidden_states)

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


# Adapted from transformers.models.jamba.modeling_jamba.JambaForCausalLM with Jamba->NemotronH, JAMBA->NEMOTRON_H
class NemotronHForCausalLM(NemotronHPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {}

    def __init__(self, config: NemotronHConfig):
        super().__init__(config)
        self.model = NemotronHModel(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()

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs,
    ) -> tuple | 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, NemotronHForCausalLM

        >>> model = NemotronHForCausalLM.from_pretrained("Zyphra/NemotronH-7B-v1")
        >>> tokenizer = AutoTokenizer.from_pretrained("Zyphra/NemotronH-7B-v1")

        >>> 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 = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

        hidden_states = outputs[0]
        # 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, :]).float()

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

        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, **kwargs):
        kwargs["logits_to_keep"] = self.config.num_logits_to_keep
        model_inputs = super().prepare_inputs_for_generation(input_ids, **kwargs)
        return model_inputs

    @staticmethod
    def create_masks_for_generate(config, inputs_embeds, attention_mask, past_key_values, position_ids=None, **_):
        # Nemotron-H layer_types include non-attention block types (moe / mlp) that the default dispatch
        # table doesn't enumerate, so we return both masks the forward needs as a dict.
        mask_kwargs = {
            "config": config.get_text_config(),
            "inputs_embeds": inputs_embeds,
            "attention_mask": attention_mask,
            "past_key_values": past_key_values,
            "position_ids": position_ids,
        }
        return {
            "full_attention": create_causal_mask(**mask_kwargs),
            "linear_attention": create_recurrent_attention_mask(**mask_kwargs),
        }


__all__ = ["NemotronHPreTrainedModel", "NemotronHModel", "NemotronHForCausalLM"]