modeling_nemotron_h.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           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 contextlib
import math
from collections.abc import Callable
from typing import Any

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

import copy

from transformers import initialization as init
from transformers.activations import ACT2FN
from transformers.generation import GenerationMixin
from transformers.integrations import (
    lazy_load_kernel,
    use_kernel_forward_from_hub,
    use_kernel_func_from_hub,
    use_kernelized_func,
)
from transformers.masking_utils import create_causal_mask
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.models.zamba2.modeling_zamba2 import Zamba2RMSNormGated
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling, logging
from transformers.utils.generic import merge_with_config_defaults
from transformers.utils.import_utils import resolve_internal_import
from transformers.utils.output_capturing import capture_outputs
from .configuration_nemotron_h import NemotronHConfig


logger = logging.get_logger(__name__)


class NemotronHHybridDynamicCache:
    """
    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
    (which has a constant shape regardless of seq_len).

    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
    and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
    For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
    and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
    """

    is_compileable = False

    def __init__(
        self, config: NemotronHConfig, batch_size: int, dtype: torch.dtype = torch.float16, device: str | None = None
    ):
        self.dtype = dtype
        self.layers_block_type = config.layers_block_type
        self.has_previous_state = False
        self.intermediate_size = int(config.mamba_num_heads * config.mamba_head_dim)
        self.ssm_state_size = config.ssm_state_size
        self.conv_kernel_size = config.conv_kernel
        self.n_mamba_heads = config.mamba_num_heads
        self.transformer_layers = []
        self._modules = {}
        self._parameters = {}
        self._buffers = {}
        self.conv_states = {}
        self.ssm_states = {}
        for i in range(config.num_hidden_layers):
            if self.layers_block_type[i] == "mamba":
                # Only allocate mamba cache for mamba layers
                self.conv_states[i] = torch.zeros(
                    batch_size,
                    self.intermediate_size + 2 * config.n_groups * self.ssm_state_size,
                    self.conv_kernel_size,
                    device=device,
                    dtype=dtype,
                )
                self.ssm_states[i] = torch.zeros(
                    batch_size,
                    self.n_mamba_heads,
                    config.mamba_head_dim,
                    self.ssm_state_size,
                    device=device,
                    dtype=dtype,
                )
            else:
                # For attention and moe layers, use empty tensors
                self.conv_states[i] = torch.tensor([[]] * batch_size, device=device)
                self.ssm_states[i] = torch.tensor([[]] * batch_size, device=device)

            if self.layers_block_type[i] == "attention":
                self.transformer_layers.append(i)
        self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
        self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]

    def __len__(self):
        return len(self.key_cache)

    def update(
        self,
        key_states: torch.Tensor,
        value_states: torch.Tensor,
        layer_idx: int,
        cache_kwargs: dict[str, Any] | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Update the cache
        if self.key_cache[layer_idx].shape[-1] == 0:
            self.key_cache[layer_idx] = key_states
            self.value_cache[layer_idx] = value_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)

        return self.key_cache[layer_idx], self.value_cache[layer_idx]

    def reorder_cache(self, beam_idx: torch.LongTensor):
        """Reorders the cache for beam search, given the selected beam indices."""
        if self.get_seq_length() > 0:
            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.conv_states[layer_idx].device
                self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device))
                device = self.ssm_states[layer_idx].device
                self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))

    def get_seq_length(self, layer_idx: int | None = 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
        layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
        if len(self.key_cache) <= layer_idx or self.key_cache[layer_idx].numel() == 0:
            return 0
        return self.key_cache[layer_idx].shape[-2]

    def get_mask_sizes(self, query_length, layer_idx: int) -> tuple[int, int]:
        """Return the length and offset of the cache, used to generate the mask"""
        # transformers >= 5.6 passes `query_length` as int; earlier versions passed a `cache_position` tensor.
        if isinstance(query_length, torch.Tensor):
            query_length = query_length.shape[0]
        kv_offset = 0
        kv_length = self.get_seq_length(layer_idx) + query_length
        return kv_length, kv_offset

    def update_conv_state(
        self, layer_idx: int, new_conv_state: torch.Tensor, cache_position: torch.LongTensor
    ) -> torch.Tensor:
        conv_state = self.conv_states[layer_idx]
        cache_position = cache_position.clamp(0, self.conv_kernel_size - 1)

        conv_state = conv_state.roll(shifts=-1, dims=-1)
        conv_state[:, :, cache_position] = new_conv_state.to(conv_state.device)
        self.conv_states[layer_idx].zero_()
        self.conv_states[layer_idx] += conv_state
        return self.conv_states[layer_idx]

    def reset(self):
        self.conv_states.zero_()
        self.ssm_states.zero_()


# 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


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):
        super().__init__()
        self.config = config
        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.use_mem_eff_path = True

        self.n_groups = config.n_groups
        self.head_dim = config.mamba_head_dim
        self.num_heads = config.mamba_num_heads
        self.chunk_size = config.chunk_size

        self.time_step_limit = config.time_step_limit
        self.time_step_min = config.time_step_min
        self.time_step_max = config.time_step_max

        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)
        # instantiate once and copy inv_dt in init_weights of PretrainedModel
        self.dt_bias = nn.Parameter(torch.ones(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
        A = torch.arange(1, self.num_heads + 1)
        self.A_log = nn.Parameter(torch.log(A))

        self.norm = Zamba2RMSNormGated(
            self.intermediate_size, group_size=self.intermediate_size // self.n_groups, eps=config.layer_norm_epsilon
        )
        self.D = nn.Parameter(torch.ones(self.num_heads))

        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)

        global causal_conv1d_update, causal_conv1d_fn
        causal_conv1d = lazy_load_kernel("causal-conv1d")
        causal_conv1d_update = getattr(causal_conv1d, "causal_conv1d_update", None)
        causal_conv1d_fn = getattr(causal_conv1d, "causal_conv1d_fn", None)

        global selective_state_update, mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
        mamba_ssm = lazy_load_kernel("mamba-ssm")
        selective_state_update = resolve_internal_import(
            mamba_ssm, chained_path="ops.triton.selective_state_update.selective_state_update"
        )
        mamba_chunk_scan_combined = resolve_internal_import(
            mamba_ssm, chained_path="ops.triton.ssd_combined.mamba_chunk_scan_combined"
        )
        mamba_split_conv1d_scan_combined = resolve_internal_import(
            mamba_ssm, chained_path="ops.triton.ssd_combined.mamba_split_conv1d_scan_combined"
        )

        global is_fast_path_available
        is_fast_path_available = all(
            (
                selective_state_update,
                mamba_chunk_scan_combined,
                mamba_split_conv1d_scan_combined,
                causal_conv1d_fn,
                causal_conv1d_update,
            )
        )

        if not is_fast_path_available:
            logger.warning_once(
                "The fast path is not available because one of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)`"
                " is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and"
                " https://github.com/Dao-AILab/causal-conv1d"
            )

    def cuda_kernels_forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: NemotronHHybridDynamicCache | None = None,
        attention_mask: torch.Tensor | None = None,
    ):
        # set up dimensions for reshapes later

        batch_size, seq_len, _ = hidden_states.shape
        groups_time_state_size = self.n_groups * self.ssm_state_size
        d_to_remove = 2 * self.intermediate_size + 2 * self.n_groups * self.ssm_state_size + self.num_heads

        # getting projected states from cache if it exists
        if cache_params is not None and cache_params.has_previous_state:
            in_projected_states = self.in_proj(hidden_states.squeeze(1))  # (B 2D)
            d_mlp = (in_projected_states.shape[-1] - d_to_remove) // 2
            split_projection_dim = [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads]
            _, _, gate, hidden_states_B_C, dt = torch.split(in_projected_states, split_projection_dim, dim=-1)

            hidden_states_B_C = causal_conv1d_update(
                hidden_states_B_C,
                cache_params.conv_states[self.layer_idx],
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                self.activation,
            )

            hidden_states, B, C = torch.split(
                hidden_states_B_C,
                [self.intermediate_size, groups_time_state_size, groups_time_state_size],
                dim=-1,
            )
            A = -torch.exp(self.A_log.float())  # (nheads,)

            A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
            dt = dt[:, :, None].expand(-1, -1, self.head_dim)
            dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
            D = self.D[:, None, ...].expand(-1, self.head_dim)
            B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups)
            C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups)
            hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim)
            hidden_states = selective_state_update(
                cache_params.ssm_states[self.layer_idx],
                hidden_states_reshaped,
                dt,
                A,
                B,
                C,
                D,
                z=None,
                dt_bias=dt_bias,
                dt_softplus=True,
            )
            hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim)
            hidden_states = self.norm(hidden_states, gate)
            out = self.out_proj(hidden_states)[:, None, ...]
        # if no cache is found, calling the kernel
        else:
            if attention_mask is not None and not torch.all(attention_mask == 1):
                # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
                dtype = hidden_states.dtype
                hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
            # 1. Gated MLP's linear projection
            projected_states = self.in_proj(hidden_states)
            A = -torch.exp(self.A_log.float())  # (num_heads) or (intermediate_size, state_size)
            dt_limit_kwargs = {} if self.time_step_limit is None else {"dt_limit": self.time_step_limit}
            if attention_mask is not None:
                input_not_masked = torch.all(attention_mask == 1)
            else:
                input_not_masked = True

            if self.use_mem_eff_path and self.training and cache_params is None and input_not_masked:
                out, ssm_state = mamba_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,
                    seq_idx=None,
                    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=True,
                    **dt_limit_kwargs,
                )

            else:
                gate, hidden_states_B_C, time_step = torch.split(
                    projected_states,
                    [self.intermediate_size, self.conv_dim, self.num_heads],
                    dim=-1,
                )

                # 1D Convolution
                if cache_params is not None:
                    hidden_states_B_C_t = hidden_states_B_C.transpose(1, 2)
                    conv_state = nn.functional.pad(
                        hidden_states_B_C_t, (self.conv_kernel_size - hidden_states_B_C_t.shape[-1], 0)
                    )
                    cache_params.conv_states[self.layer_idx].copy_(conv_state)
                if causal_conv1d_fn is None or self.activation not in ["silu", "swish"]:
                    hidden_states_B_C = self.act(
                        self.conv1d(hidden_states_B_C.transpose(1, 2)).transpose(1, 2)[:, :seq_len]
                    )  # (B, L, self.d_inner + 2 * ngroups * d_state)
                else:
                    hidden_states_B_C = causal_conv1d_fn(
                        x=hidden_states_B_C.transpose(1, 2),
                        weight=self.conv1d.weight.squeeze(1),
                        bias=self.conv1d.bias,
                        activation=self.activation,
                    ).transpose(1, 2)[:, :seq_len]
                hidden_states, B, C = torch.split(
                    hidden_states_B_C,
                    [self.intermediate_size, groups_time_state_size, groups_time_state_size],
                    dim=-1,
                )
                if attention_mask is not None and not torch.all(attention_mask == 1):
                    # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
                    dtype = hidden_states.dtype
                    hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
                scan_output, ssm_state = mamba_chunk_scan_combined(
                    hidden_states.view(batch_size, seq_len, -1, self.head_dim),
                    time_step,
                    A,
                    B.view(batch_size, seq_len, self.n_groups, -1),
                    C.view(batch_size, seq_len, self.n_groups, -1),
                    chunk_size=self.chunk_size,
                    D=self.D,
                    z=None,
                    seq_idx=None,
                    return_final_states=True,
                    dt_bias=self.dt_bias,
                    dt_softplus=True,
                    **dt_limit_kwargs,
                )
                if ssm_state is not None and cache_params is not None:
                    cache_params.ssm_states[self.layer_idx].copy_(ssm_state)
                scan_output = scan_output.view(batch_size, seq_len, -1)
                # Multiply "gate" branch and apply extra normalization layer
                scan_output = self.norm(scan_output, gate)
                out = self.out_proj(scan_output)
        return out

    # fmt: off
    def torch_forward(self, input_states, cache_params: NemotronHHybridDynamicCache | None=None, attention_mask: torch.Tensor | None=None):
        batch_size, seq_len, _ = input_states.shape
        dtype = input_states.dtype
        # Gated MLP's linear projection
        if cache_params is not None and cache_params.has_previous_state:
            projected_states = self.in_proj(input_states.squeeze(1))
        else:
            if attention_mask is not None:
                # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
                input_states = (input_states * attention_mask[:, :, None]).to(dtype)
            projected_states = self.in_proj(input_states)
        d_mlp = (projected_states.shape[-1] - 2 * self.intermediate_size - 2 * self.n_groups * self.ssm_state_size- self.num_heads) // 2
        _, _, gate, hidden_states, dt = projected_states.split(
                [d_mlp, d_mlp, self.intermediate_size,  self.conv_dim, self.num_heads], dim=-1
        )

        # Convolution sequence transformation
        if cache_params is not None:
            ssm_state = cache_params.ssm_states[self.layer_idx].clone()
            ssm_state = ssm_state.to(hidden_states.device)
            if cache_params.has_previous_state:
                gate = gate.unsqueeze(1)
                conv_state = cache_params.conv_states[self.layer_idx]                   # [batch, intermediate_size, conv_kernel_size]
                conv_state = torch.roll(conv_state, shifts=-1, dims=-1)
                # handle batched generation - states are copied through
                conv_state[:, :, -1] = hidden_states[:, 0, :] if hidden_states.ndim == 3 else hidden_states
                cache_params.conv_states[self.layer_idx].copy_(conv_state)
                hidden_states = torch.sum(conv_state.to(projected_states.device) * self.conv1d.weight[:, 0, :], dim=-1)
                if self.use_conv_bias:
                    hidden_states += self.conv1d.bias
                hidden_states = self.act(hidden_states).to(dtype)[:, None, ...]         # [batch, 1, intermediate_size] : decoding
            else:
                hidden_states = hidden_states.transpose(1,2)
                conv_state = nn.functional.pad(
                    hidden_states,
                    (self.conv_kernel_size - hidden_states.shape[-1], 0)
                )
                cache_params.conv_states[self.layer_idx].copy_(conv_state)
                hidden_states = self.act(self.conv1d(hidden_states).transpose(1,2))[:, :seq_len, :]     # [batch, intermediate_size, seq_len]
                if attention_mask is not None:
                    dtype = hidden_states.dtype
                    # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
                    hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
        else:
            ssm_state = torch.zeros(
                (batch_size, self.num_heads, self.head_dim, self.ssm_state_size),
                device=hidden_states.device, dtype=dtype
            )
            hidden_states = self.act(self.conv1d(hidden_states.transpose(1, 2))[..., :seq_len].transpose(1, 2))
        hidden_states, B, C = torch.split(hidden_states, [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size], dim=-1)
        A = -torch.exp(self.A_log.float())                            # [num_heads]
        if cache_params is not None and cache_params.has_previous_state:
            # Note: there is no need to pad parameter matrices here, as there is just one new token
            # for batched generation
            dt = dt[:, None, ...] if dt.ndim == 2 else dt[:, 0, :][:, None, ...]
            dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim)
            # [num_heads] -> [num_heads, head_dim]
            dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim)

            dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype))
            dt = torch.clamp(dt, self.time_step_min) #, self.time_step_max)
            A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
            # [bsz, num_heads, head_dim, state_size]
            dA = torch.exp(dt[..., None] * A)

            # Discretize B
            # [bsz, n_groups * state_size] -> [bsz, n_groups, 1, state_size] ->
            # -> [bsz, n_groups, group to head repetition factor, state_size] -> [bsz, num_heads, state_size]
            B = B.reshape(batch_size, self.n_groups, -1)[..., None, :]
            B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous()
            B = B.reshape(batch_size, -1, B.shape[-1])
            # [bsz, num_heads, head_dim, state_size]
            dB = dt[..., None] * B[..., None, :]

            # Discretize x into dB
            # [bsz, intermediate_size] -> [bsz, num_heads, head_dim]
            hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim)
            dBx = dB * hidden_states[..., None]

            # State calculation
            cache_params.ssm_states[self.layer_idx].copy_(
                cache_params.ssm_states[self.layer_idx] * dA + dBx
            )

            # Subsequent output
            # [bsz, n_groups * state_size] -> [bsz, num_heads, state_size]
            C = C.reshape(batch_size, self.n_groups, -1)[..., None, :]
            C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous()
            C = C.reshape(batch_size, -1, C.shape[-1])
            # [bsz, num_heads, head_dim]

            ssm_states = cache_params.ssm_states[self.layer_idx].to(C.dtype)  # Shape: [b, h, d, n]
            # Reshape ssm_states to merge the first two dimensions
            ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size)  # Shape: [b*h, d, n]
            C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1)  # Shape: [b*h, n, 1]
            y = torch.bmm(ssm_states_reshaped, C_reshaped)
            y = y.view(batch_size, self.num_heads, self.head_dim)

            # D skip connection
            # [num_heads] -> [num_heads, head_dim]
            D = self.D[..., None].expand(self.D.shape[0], self.head_dim)
            y = (y + hidden_states * D).to(y.dtype)

            # [bsz, num_heads, head_dim] -> [bsz, 1, intermediate_size]
            y = y.reshape(batch_size, -1)[:, None, ...]
        else:
            # begin ssd naive implementation without einsums
            dt = nn.functional.softplus(dt + self.dt_bias)
            dt = torch.clamp(dt, self.time_step_min)
            hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float()
            B = B.reshape(batch_size, seq_len,  -1, self.ssm_state_size).float()
            C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
            B = B.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads)
            C = C.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads)
            pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size

            D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size)

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

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


            # [bsz, -1, chunk_size, num_heads] -> [bsz, num_heads, -1, chunk_size]
            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))

            # First, contraction of C and B to get G (attention-weights like)
            G_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, : ,:]  # shape: (b, c, l, s, h, n)
            G = G_intermediate.sum(dim=-1)  # shape: (b, c, l, s, h)


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

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

            # (right term of low-rank factorization of off-diagonal blocks; B terms)

            decay_states = torch.exp(A_cumsum[:, :, :, -1:] - A_cumsum)
            B_decay_contraction = B * decay_states.permute(0, 2, 3, 1)[..., None]
            # permute back B * decay states
            states = (B_decay_contraction.permute(0, 1, 3, 2, 4)[..., None]  * hidden_states.permute(0, 1, 3, 2, 4)[..., None, :]).sum(dim=3).permute(0, 1, 2, 4, 3)
            if cache_params is not None and cache_params.has_previous_state:
                previous_states = cache_params.ssm_states[self.layer_idx][:, None, ...]
            else:
                previous_states = torch.zeros_like(states[:, :1])
            states = torch.cat([previous_states, states], dim=1)
            decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))

            states_permuted = states.permute(0, 2, 1, 3, 4)
            result = (decay_chunk[..., None, None] * states_permuted[:, :, None, ...]).sum(dim=2)
            new_states = result.permute(0, 2, 1, 3, 4)
            states, ssm_state = new_states[:, :-1], new_states[:, -1]

            # 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)
            # compute Yoff
            C_times_states = (C[..., None, :] * states[:, :, None, ...])
            state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1)
            Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None])
            # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)

            y = Y_diag + Y_off
            # [bsz, -1, self.chunk_size, num_heads, head_dim] -> [bsz, (padded) seq_len, num_heads, head_dim]
            y = y.reshape(batch_size, -1, self.num_heads, self.head_dim)

            y = y + D_residual
            # Cutting off padded chunks
            if pad_size > 0:
                y = y[:, :seq_len, :, :]
            y = y.reshape(batch_size, seq_len, -1)
            if ssm_state is not None and cache_params is not None:
                cache_params.ssm_states[self.layer_idx].copy_(ssm_state)

        scan_output = self.norm(y, gate)

        # end ssd naive

        # 4. Final linear projection
        contextualized_states = self.out_proj(scan_output.to(dtype))  # [batch, seq_len, hidden_size]
        return contextualized_states
    # fmt: on

    def forward(
        self,
        hidden_states,
        cache_params: NemotronHHybridDynamicCache | None = None,
        attention_mask: torch.Tensor | None = None,
    ):
        if is_fast_path_available and "cuda" in self.in_proj.weight.device.type and not is_torchdynamo_compiling():
            return self.cuda_kernels_forward(hidden_states, cache_params, attention_mask)

        return self.torch_forward(hidden_states, cache_params, attention_mask)


@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):
        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)))


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

    - Experts are stored as an nn.ModuleList of NemotronHMLP, to match the per-expert checkpoint
      format (`experts.<idx>.up_proj.weight` / `experts.<idx>.down_proj.weight`).
    - Optional latent projection wraps the experts.
    """

    def __init__(self, config, layer_idx: int | None = None):
        super().__init__()
        self.config = config
        self.n_routed_experts = config.n_routed_experts
        self.n_group = config.n_group
        self.topk_group = config.topk_group
        self.norm_topk_prob = config.norm_topk_prob
        self.routed_scaling_factor = config.routed_scaling_factor
        self.top_k = config.num_experts_per_tok

        self.gate = NemotronHTopkRouter(config)

        # Optional latent projection; `moe_latent_size` is absent on older configs.
        moe_latent_size = getattr(config, "moe_latent_size", None)
        if moe_latent_size is not None:
            self.fc1_latent_proj = nn.Linear(config.hidden_size, moe_latent_size, bias=config.mlp_bias)
            self.fc2_latent_proj = nn.Linear(moe_latent_size, config.hidden_size, bias=config.mlp_bias)
            expert_input_dim = moe_latent_size
        else:
            self.fc1_latent_proj = nn.Identity()
            self.fc2_latent_proj = nn.Identity()
            expert_input_dim = config.hidden_size

        # Shallow-copy config so each expert's NemotronHMLP runs in `expert_input_dim` dim without
        # mutating the top-level config.
        expert_config = copy.copy(config)
        expert_config.hidden_size = expert_input_dim
        self.experts = nn.ModuleList(
            [NemotronHMLP(expert_config, intermediate_size=config.moe_intermediate_size) for _ in range(self.n_routed_experts)]
        )
        self.shared_experts = NemotronHMLP(config=config, intermediate_size=config.moe_shared_expert_intermediate_size)

    def route_tokens_to_experts(self, router_logits):
        router_logits = router_logits.sigmoid()
        router_logits_for_choice = router_logits + self.gate.e_score_correction_bias
        group_scores = (
            router_logits_for_choice.view(-1, self.n_group, self.n_routed_experts // self.n_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.n_group, self.n_routed_experts // self.n_group)
            .reshape(-1, self.n_routed_experts)
        )
        scores_for_choice = router_logits_for_choice.masked_fill(~score_mask.bool(), 0.0)
        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
        topk_weights = router_logits.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 topk_indices, topk_weights

    def forward(self, hidden_states):
        residuals = hidden_states
        orig_shape = hidden_states.shape
        router_logits = self.gate(hidden_states)
        topk_indices, topk_weights = self.route_tokens_to_experts(router_logits)
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])

        # Route each token through its top-k experts (ModuleList version).
        expert_inputs = self.fc1_latent_proj(hidden_states)
        expert_outputs = torch.zeros_like(expert_inputs, dtype=topk_weights.dtype)

        with torch.no_grad():
            expert_mask = torch.nn.functional.one_hot(topk_indices, num_classes=self.n_routed_experts)
            expert_mask = expert_mask.permute(2, 1, 0)  # (num_experts, top_k, num_tokens)
            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()
            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
            if token_idx.numel() == 0:
                continue
            current_state = expert_inputs[token_idx]
            current_hidden_states = self.experts[expert_idx](current_state)
            current_hidden_states = current_hidden_states * topk_weights[token_idx, top_k_pos, None]
            expert_outputs.index_add_(0, token_idx, current_hidden_states.to(expert_outputs.dtype))

        expert_outputs = expert_outputs.to(expert_inputs.dtype)
        hidden_states = self.fc2_latent_proj(expert_outputs)

        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.config = config
        self.n_routed_experts = config.n_routed_experts

        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size)))
        self.register_buffer("e_score_correction_bias", torch.zeros(self.n_routed_experts))

    def forward(self, hidden_states):
        hidden_states = hidden_states.view(-1, self.config.hidden_size)
        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
        return router_logits


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_func_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: NemotronHHybridDynamicCache | None = None,
        cache_position: torch.LongTensor | 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, {"cache_position": cache_position}
            )

        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 = {
    "mamba": NemotronHMamba2Mixer,
    "attention": NemotronHAttention,
    "moe": NemotronHMoE,
}


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: NemotronHHybridDynamicCache | None = None,
        cache_position: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        use_cache: bool | None = False,
        **kwargs: Unpack[TransformersKwargs],
    ):
        if hidden_states.device.type == "cuda":
            # Use cuda stream to avoid NaN when using multiple GPUs, which is caused by multi-GPU synchronization issue.
            # Mamba might launch on the default cuda stream that not strictly respect the current Pytorch cuda stream.
            # This leads to kernel reading uninitialized memory before the data transfer is complete.
            stream_context = torch.cuda.stream(torch.cuda.default_stream(hidden_states.device))
        else:
            stream_context = contextlib.nullcontext()

        with stream_context:
            residual = hidden_states
            hidden_states = self.norm(hidden_states.to(dtype=self.norm.weight.dtype))

            if self.block_type == "mamba":
                hidden_states = self.mixer(hidden_states, cache_params=past_key_values, attention_mask=attention_mask)
            elif self.block_type == "attention":
                hidden_states, _ = self.mixer(
                    hidden_states=hidden_states,
                    past_key_values=past_key_values,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    user_cache=use_cache,
                    cache_position=cache_position,
                    **kwargs,
                )
            else:
                hidden_states = self.mixer(hidden_states)

            hidden_states = residual + hidden_states
            return hidden_states


class NemotronHPreTrainedModel(PreTrainedModel):
    config: NemotronHConfig
    base_model_prefix = "backbone"
    _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_record_outputs = {
        "hidden_states": NemotronHBlock,
        "attentions": NemotronHAttention,
    }
    _keep_in_fp32_modules_strict = [
        "e_score_correction_bias",
    ]
    _tied_weights_keys = {}
    _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):
            # Initialize A_log and D parameters
            A = torch.arange(1, self.config.mamba_num_heads + 1)
            init.copy_(module.A_log, torch.log(A))
            init.ones_(module.D)

            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)
            module.dt_bias._no_reinit = True
        elif isinstance(module, NemotronHTopkRouter):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            init.zeros_(module.e_score_correction_bias)
        # Experts are now an nn.ModuleList of NemotronHMLP; their nn.Linear submodules are handled by
        # the generic branch below, so no special init branch is needed.

        if isinstance(module, nn.Linear):
            if module.bias is not None:
                if not getattr(module.bias, "_no_reinit", 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":
                    # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
                    # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
                    # We need to reinit p since this code could be called multiple times
                    # Having just p *= scale would repeatedly scale it down
                    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)

        # Legacy checkpoints may use "embedding." (singular); rewrite to "embeddings." on load.
        self._register_load_state_dict_pre_hook(self.load_hook)

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

    def load_hook(self, state_dict, prefix, *args):
        for k in list(state_dict.keys()):
            if "embedding." in k:
                state_dict[k.replace("embedding.", "embeddings.")] = state_dict.pop(k)

    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: NemotronHHybridDynamicCache | None = None,
        use_cache: bool | None = None,
        cache_position: torch.LongTensor | 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 = NemotronHHybridDynamicCache(
                config=self.config,
                batch_size=inputs_embeds.shape[0],
                dtype=inputs_embeds.dtype,
                device=inputs_embeds.device,
            )

        hidden_states = inputs_embeds

        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 + hidden_states.shape[1], device=hidden_states.device
            )
        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        causal_mask = create_causal_mask(
            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,
        )
        mamba_mask = self._update_mamba_mask(attention_mask, cache_position)

        # Map block types to their corresponding masks
        block_type_to_mask = {
            "mamba": mamba_mask,
            "attention": causal_mask,
            "moe": None,
        }

        for layer_idx, mixer_block in enumerate(self.layers):
            layer_mask = block_type_to_mask[mixer_block.block_type]

            hidden_states = mixer_block(
                hidden_states,
                attention_mask=layer_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                cache_position=cache_position,
                **kwargs,
            )

        hidden_states = self.norm_f(hidden_states)

        if past_key_values is not None and not past_key_values.has_previous_state:
            past_key_values.has_previous_state = True

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

    def _update_mamba_mask(self, attention_mask, cache_position):
        """
        No need for zeroing states when
            1. Cached forward
            2. Attending to all inputs
        """
        mamba_mask = attention_mask
        if (cache_position is not None and cache_position[0] > 0) or (
            attention_mask is not None and torch.all(attention_mask == 1)
        ):
            mamba_mask = None
        return mamba_mask


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

    @classmethod
    def _supports_default_dynamic_cache(cls) -> bool:
        # This model supplies its own `NemotronHHybridDynamicCache` via `prepare_inputs_for_generation`.
        # Preventing transformers from injecting a plain `DynamicCache` keeps our mixer code
        # (which reads `cache_params.conv_states[layer_idx]` / `.has_previous_state`) happy.
        return False

    def __init__(self, config):
        super().__init__(config)
        # Attribute name matches the checkpoint key prefix ("backbone.").
        self.backbone = 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()

    def get_input_embeddings(self):
        return self.backbone.get_input_embeddings()

    def set_input_embeddings(self, new_embeddings):
        return self.backbone.set_input_embeddings(new_embeddings)

    @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: NemotronHHybridDynamicCache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        cache_position: torch.LongTensor | 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.backbone(
            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[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,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        is_first_iteration=False,
        **kwargs,
    ):
        # Overwritten -- has a unique cache type, `NemotronHHybridDynamicCache`

        if past_key_values is None:
            past_key_values = NemotronHHybridDynamicCache(
                self.config, input_ids.shape[0], dtype=self.dtype, device=self.device
            )

        kwargs["logits_to_keep"] = self.config.num_logits_to_keep
        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            cache_position=cache_position,
            position_ids=position_ids,
            use_cache=use_cache,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        return model_inputs


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