Add Kolibri 1

@@ -737,6 +737,8 @@         title: jina_embeddings_v3       - local: model_doc/kimi_linear         title: KimiLinear+      - local: model_doc/kolibri1+        title: Kolibri1       - local: model_doc/laguna         title: Laguna       - local: model_doc/led

@@ -0,0 +1,48 @@+<!--Copyright 2026 The HuggingFace Team. All rights reserved.++Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with+the License. You may obtain a copy of the License at++http://www.apache.org/licenses/LICENSE-2.0++Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the+specific language governing permissions and limitations under the License.++⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be+rendered properly in your Markdown viewer.++-->+*This model was contributed to Hugging Face Transformers on 2026-10-12.*++# Kolibri1++[Kolibri 1](https://huggingface.co/Aleph-Alpha/Kolibri-1) is Aleph Alpha's mixture-of-experts reasoning model with a focus on German and English. Each of its 50 layers routes every token to 6 of 384 experts with sigmoid routing and adds one shared expert. Four sliding-window attention layers alternate with one full-attention layer, and RoPE is applied only in the sliding-window layers.++```python+from transformers import AutoModelForCausalLM, AutoTokenizer++tokenizer = AutoTokenizer.from_pretrained("Aleph-Alpha/Kolibri-1-BF16")+model = AutoModelForCausalLM.from_pretrained("Aleph-Alpha/Kolibri-1-BF16", device_map="auto")++messages = [{"role": "user", "content": "Erkläre kurz, was ein Mixture-of-Experts-Modell ist."}]+inputs = tokenizer.apply_chat_template(+    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True+).to(model.device)+outputs = model.generate(**inputs, max_new_tokens=256)+print(tokenizer.decode(outputs[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True))+```++## Kolibri1Config++[[autodoc]] Kolibri1Config++## Kolibri1Model++[[autodoc]] Kolibri1Model+    - forward++## Kolibri1ForCausalLM++[[autodoc]] Kolibri1ForCausalLM+    - forward

@@ -1965,6 +1965,13 @@ def _build_checkpoint_conversion_mapping():     mapping["exaone_moe"] = mapping["qwen2_moe"].copy()     mapping["exaone_moe"] += [WeightRenaming("mlp.e_score_correction_bias", "mlp.gate.e_score_correction_bias")] +    mapping["kolibri1"] = mapping["qwen2_moe"].copy()+    mapping["kolibri1"] += [+        WeightRenaming("moe.router.expert_bias", "mlp.gate.e_score_correction_bias"),+        WeightRenaming("post_attn_norm", "post_self_attn_layernorm"),+        WeightRenaming("post_ffn_norm", "post_mlp_layernorm"),+    ]+     mapping["mimo_v2_flash"] = mapping["qwen2_moe"].copy()     mapping["mimo_v2_flash"] += [         WeightRenaming("self_attn.attention_sink_bias", "self_attn.sinks"),

@@ -247,6 +247,7 @@     from .jina_embeddings_v3 import *     from .kimi_k25 import *     from .kimi_linear import *+    from .kolibri1 import *     from .kosmos2 import *     from .kosmos2_5 import *     from .kyutai_speech_to_text import *

@@ -347,6 +347,7 @@         ("kimi_k25", "Kimi_K25Config"),         ("kimi_k25_vision", "Kimi_K25VisionConfig"),         ("kimi_linear", "KimiLinearConfig"),+        ("kolibri1", "Kolibri1Config"),         ("kosmos-2", "Kosmos2Config"),         ("kosmos-2.5", "Kosmos2_5Config"),         ("kosmos_2_5_text_model", "Kosmos2_5TextConfig"),

@@ -289,6 +289,7 @@ class _BaseModelWithGenerate(PreTrainedModel, GenerationMixin):         ("kimi_k25", "Kimi_K25Model"),         ("kimi_k25_vision", "Kimi_K25VisionModel"),         ("kimi_linear", "KimiLinearModel"),+        ("kolibri1", "Kolibri1Model"),         ("kosmos-2", "Kosmos2Model"),         ("kosmos-2.5", "Kosmos2_5Model"),         ("kyutai_speech_to_text", "KyutaiSpeechToTextModel"),@@ -806,6 +807,7 @@ class _BaseModelWithGenerate(PreTrainedModel, GenerationMixin):         ("jamba", "JambaForCausalLM"),         ("jetmoe", "JetMoeForCausalLM"),         ("kimi_linear", "KimiLinearForCausalLM"),+        ("kolibri1", "Kolibri1ForCausalLM"),         ("laguna", "LagunaForCausalLM"),         ("lfm2", "Lfm2ForCausalLM"),         ("lfm2_moe", "Lfm2MoeForCausalLM"),

@@ -0,0 +1,33 @@+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨+#           This file was automatically generated from the modules in `src/transformers/models/kolibri1`.+#           Do NOT edit this file manually as any edits will be overwritten by auto-generation of the file.+#                       A module is picked up once it exports a name through `__all__`.+#                   Regenerate the file with: `python utils/check_inits.py --fix_and_overwrite`+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.+#+# Licensed under the Apache License, Version 2.0 (the "License");+# you may not use this file except in compliance with the License.+# You may obtain a copy of the License at+#+#     http://www.apache.org/licenses/LICENSE-2.0+#+# Unless required by applicable law or agreed to in writing, software+# distributed under the License is distributed on an "AS IS" BASIS,+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.+# See the License for the specific language governing permissions and+# limitations under the License.+from typing import TYPE_CHECKING++from ...utils import _LazyModule+from ...utils.import_utils import define_import_structure+++if TYPE_CHECKING:+    from .configuration_kolibri1 import *+    from .modeling_kolibri1 import *+else:+    import sys++    _file = globals()["__file__"]+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

@@ -0,0 +1,112 @@+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨+#           This file was automatically generated from src/transformers/models/kolibri1/modular_kolibri1.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_kolibri1.py file directly. One of our CI enforces this.+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.+#+# Licensed under the Apache License, Version 2.0 (the "License");+# you may not use this file except in compliance with the License.+# You may obtain a copy of the License at+#+#     http://www.apache.org/licenses/LICENSE-2.0+#+# Unless required by applicable law or agreed to in writing, software+# distributed under the License is distributed on an "AS IS" BASIS,+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.+# See the License for the specific language governing permissions and+# limitations under the License.++from huggingface_hub.dataclasses import strict++from ...configuration_utils import PreTrainedConfig+from ...modeling_rope_utils import RopeParameters+from ...utils import auto_docstring+++@auto_docstring(checkpoint="Aleph-Alpha/Kolibri-1")+@strict+class Kolibri1Config(PreTrainedConfig):+    r"""+    ```python+    >>> from transformers import Kolibri1Model, Kolibri1Config++    >>> configuration = Kolibri1Config()+    >>> model = Kolibri1Model(configuration)+    >>> configuration = model.config+    ```+    """++    model_type = "kolibri1"+    keys_to_ignore_at_inference = ["past_key_values"]++    attribute_map = {+        "num_experts": "num_local_experts",+    }+    base_model_tp_plan = {+        "layers.*.self_attn.q_proj": "colwise",+        "layers.*.self_attn.k_proj": "colwise",+        "layers.*.self_attn.v_proj": "colwise",+        "layers.*.self_attn.q_norm": "replicated_with_grad_allreduce",+        "layers.*.self_attn.k_norm": "replicated_with_grad_allreduce",+        "layers.*.self_attn.o_proj": "rowwise",+        "layers.*.mlp.experts.gate_up_proj": "packed_colwise",+        "layers.*.mlp.experts.down_proj": "rowwise",+        "layers.*.mlp.experts": "moe_tp_experts",+        "layers.*.mlp.shared_experts.gate_proj": "colwise",+        "layers.*.mlp.shared_experts.up_proj": "colwise",+        "layers.*.mlp.shared_experts.down_proj": "rowwise",+    }+    # Token dispatch by default, so `ep_size` can exceed `tp_size` (bounded by `num_key_value_heads`)+    base_model_ep_plan = {+        "layers.*.mlp.experts.gate_up_proj": "grouped_gemm",+        "layers.*.mlp.experts.down_proj": "grouped_gemm",+        "layers.*.mlp.experts": "ep_dispatch_experts",+    }+    base_model_pp_plan = {+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),+        "norm": (["hidden_states"], ["hidden_states"]),+    }++    vocab_size: int = 128000+    hidden_size: int = 2560+    num_hidden_layers: int = 50+    num_attention_heads: int = 48+    num_key_value_heads: int = 4+    hidden_act: str = "silu"+    max_position_embeddings: int = 262144+    initializer_range: float = 0.02+    rms_norm_eps: float = 1e-6+    use_cache: bool = True+    tie_word_embeddings: bool = False+    rope_parameters: RopeParameters | dict | None = None+    sliding_window: int | None = 513+    attention_dropout: float | int = 0.0+    moe_intermediate_size: int = 512+    num_experts_per_tok: int = 6+    num_experts: int = 384+    norm_topk_prob: bool = False+    output_router_logits: bool = False+    pad_token_id: int | None = None+    bos_token_id: int | None = None+    eos_token_id: int | list[int] | None = None+    head_dim: int = 128+    shared_expert_intermediate_size: int = 512+    layer_types: list[str] | None = None++    def __post_init__(self, **kwargs):+        if self.layer_types is None:+            self.layer_types = [+                "full_attention" if (i + 1) % 5 == 0 else "sliding_attention" for i in range(self.num_hidden_layers)+            ]+        super().__post_init__(**kwargs)++    def validate_architecture(self):+        super().validate_architecture()+        if "sliding_attention" in self.layer_types and (self.sliding_window is None or self.sliding_window <= 0):+            raise ValueError(f"Sliding-attention layers need a positive `sliding_window`, got {self.sliding_window}.")+++__all__ = ["Kolibri1Config"]

@@ -0,0 +1,599 @@+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨+#           This file was automatically generated from src/transformers/models/kolibri1/modular_kolibri1.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_kolibri1.py file directly. One of our CI enforces this.+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.+#+# Licensed under the Apache License, Version 2.0 (the "License");+# you may not use this file except in compliance with the License.+# You may obtain a copy of the License at+#+#     http://www.apache.org/licenses/LICENSE-2.0+#+# Unless required by applicable law or agreed to in writing, software+# distributed under the License is distributed on an "AS IS" BASIS,+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.+# See the License for the specific language governing permissions and+# limitations under the License.++from 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+from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask+from ...modeling_flash_attention_utils import FlashAttentionKwargs+from ...modeling_layers import GradientCheckpointingLayer+from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast+from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel+from ...processing_utils import Unpack+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torch_greater_or_equal+from ...utils.generic import merge_with_config_defaults+from ...utils.output_capturing import OutputRecorder, capture_outputs+from .configuration_kolibri1 import Kolibri1Config+++@use_kernel_forward_from_hub("RMSNorm")+class Kolibri1RMSNorm(nn.Module):+    def __init__(self, hidden_size, eps: float = 1e-6) -> None:+        """+        Kolibri1RMSNorm 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}"+++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+++class Kolibri1Attention(nn.Module):+    def __init__(self, config: Kolibri1Config, layer_idx: int):+        super().__init__()+        self.config = config+        self.layer_idx = layer_idx+        self.num_attention_heads = config.num_attention_heads+        self.num_key_value_heads = config.num_key_value_heads+        self.hidden_size = config.hidden_size+        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.attention_dropout = config.attention_dropout+        self.is_causal = True+        self.scaling = self.head_dim**-0.5+        self.sliding_window = config.sliding_window+        layer_type = config.layer_types[layer_idx] if hasattr(config, "layer_types") else None+        self.is_sliding = layer_type == "sliding_attention"++        self.q_proj = nn.Linear(self.hidden_size, self.num_attention_heads * self.head_dim, bias=False)+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)+        self.o_proj = nn.Linear(self.num_attention_heads * self.head_dim, self.hidden_size, bias=False)++        self.q_norm = Kolibri1RMSNorm(self.head_dim, eps=config.rms_norm_eps)+        self.k_norm = Kolibri1RMSNorm(self.head_dim, eps=config.rms_norm_eps)++    def forward(+        self,+        hidden_states: torch.Tensor,+        position_embeddings: tuple[torch.Tensor, torch.Tensor],+        attention_mask: torch.Tensor | None = None,+        past_key_values: Cache | None = None,+        **kwargs: Unpack[TransformersKwargs],+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[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)++        # We use QK-norm+        query_states = self.q_norm(query_states)+        key_states = self.k_norm(key_states)++        cos, sin = position_embeddings+        # We use global NoPE for hybrid attention model+        if self.sliding_window is None or self.is_sliding:+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)++        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,+            sliding_window=self.sliding_window if self.is_sliding else None,+            **kwargs,+        )++        attn_output = attn_output.reshape(*input_shape, -1).contiguous()+        attn_output = self.o_proj(attn_output)+        return attn_output, attn_weights+++class Kolibri1MLP(nn.Module):+    def __init__(self, config, intermediate_size=None):+        super().__init__()+        self.config = config+        self.hidden_size = config.hidden_size+        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)+        self.act_fn = ACT2FN[config.hidden_act]++    def forward(self, x):+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))+        return down_proj+++@use_experts_implementation+class Kolibri1Experts(nn.Module):+    """Collection of expert weights stored as 3D tensors."""++    def __init__(self, config):+        super().__init__()+        self.num_experts = config.num_experts+        self.hidden_dim = config.hidden_size+        self.intermediate_dim = config.moe_intermediate_size+        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))+        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))+        self.act_fn = ACT2FN[config.hidden_act]++    def forward(+        self,+        hidden_states: torch.Tensor,+        top_k_index: torch.Tensor,+        top_k_weights: torch.Tensor,+    ) -> torch.Tensor:+        final_hidden_states = torch.zeros_like(hidden_states)+        with torch.no_grad():+            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts + 1)+            expert_mask = expert_mask.permute(2, 1, 0)+            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()++        for expert_idx in expert_hit:+            expert_idx = expert_idx[0]+            if expert_idx == self.num_experts:+                continue+            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])+            current_state = hidden_states[token_idx]+            gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)+            current_hidden_states = self.act_fn(gate) * up+            current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])+            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]+            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))++        return final_hidden_states+++class Kolibri1TopKRouter(nn.Module):+    def __init__(self, config):+        super().__init__()+        self.top_k = config.num_experts_per_tok+        self.num_experts = config.num_experts+        self.norm_topk_prob = config.norm_topk_prob+        self.hidden_dim = config.hidden_size+        self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))+        self.e_score_correction_bias = nn.Buffer(torch.zeros(self.num_experts))++    def forward(self, hidden_states):+        hidden_states = hidden_states.reshape(-1, self.hidden_dim)+        if (+            hidden_states.is_cuda+            and hidden_states.dtype == self.weight.dtype+            and not torch.is_grad_enabled()+            and is_torch_greater_or_equal("2.8.0")+        ):+            # Faster inference (e.g. in vLLM): fp32 logits from a bf16 GEMM instead of an fp32 one.+            # `out_dtype` is CUDA-only, has no autograd and needs matching dtypes+            router_logits = torch.mm(hidden_states, self.weight.t(), out_dtype=torch.float32)+        else:+            router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))+        router_indices = torch.topk(router_logits + self.e_score_correction_bias, self.top_k, dim=-1)[1]+        router_scores = router_logits.gather(1, router_indices).sigmoid()+        if self.norm_topk_prob:+            router_scores = router_scores / router_scores.sum(dim=-1, keepdim=True)+        return router_logits, router_scores, router_indices+++class Kolibri1SparseMoeBlock(nn.Module):+    """+    A mixed expert module containing shared experts.+    """++    def __init__(self, config: Kolibri1Config):+        super().__init__()+        self.config = config+        self.experts = Kolibri1Experts(config)+        self.gate = Kolibri1TopKRouter(config)+        self.shared_experts = Kolibri1MLP(config, intermediate_size=config.shared_expert_intermediate_size)++    def forward(self, hidden_states: torch.Tensor) -> 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])+        hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape)+        hidden_states = hidden_states + self.shared_experts(residuals)+        return hidden_states+++class Kolibri1DecoderLayer(GradientCheckpointingLayer):+    def __init__(self, config: Kolibri1Config, layer_idx: int):+        super().__init__()+        self.hidden_size = config.hidden_size+        self.self_attn = Kolibri1Attention(config=config, layer_idx=layer_idx)+        self.mlp = Kolibri1SparseMoeBlock(config)+        self.input_layernorm = Kolibri1RMSNorm(config.hidden_size, eps=config.rms_norm_eps)+        self.post_attention_layernorm = Kolibri1RMSNorm(config.hidden_size, eps=config.rms_norm_eps)+        self.post_self_attn_layernorm = Kolibri1RMSNorm(config.hidden_size, eps=config.rms_norm_eps)+        self.post_mlp_layernorm = Kolibri1RMSNorm(config.hidden_size, eps=config.rms_norm_eps)++    def forward(+        self,+        hidden_states: torch.Tensor,+        attention_mask: torch.Tensor | None = None,+        position_ids: torch.LongTensor | None = None,+        past_key_values: Cache | None = None,+        use_cache: bool | None = False,+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,+        **kwargs: Unpack[FlashAttentionKwargs],+    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:+        residual = hidden_states+        hidden_states = self.input_layernorm(hidden_states)++        hidden_states, _ = self.self_attn(+            hidden_states=hidden_states,+            attention_mask=attention_mask,+            position_ids=position_ids,+            past_key_values=past_key_values,+            use_cache=use_cache,+            position_embeddings=position_embeddings,+            **kwargs,+        )++        hidden_states = self.post_self_attn_layernorm(hidden_states)+        hidden_states = residual + hidden_states++        residual = hidden_states+        hidden_states = self.post_attention_layernorm(hidden_states)+        hidden_states = self.mlp(hidden_states)+        hidden_states = self.post_mlp_layernorm(hidden_states)+        hidden_states = residual + hidden_states+        return hidden_states+++@auto_docstring+class Kolibri1PreTrainedModel(PreTrainedModel):+    config: Kolibri1Config+    base_model_prefix = "model"+    supports_gradient_checkpointing = True+    _no_split_modules = ["Kolibri1DecoderLayer"]+    _skip_keys_device_placement = ["past_key_values"]+    _supports_flash_attn = True+    _supports_sdpa = True+    _supports_flex_attn = True++    _can_compile_fullgraph = True+    _supports_attention_backend = True+    _can_record_outputs = {+        "router_logits": OutputRecorder(Kolibri1TopKRouter, index=0),+        "hidden_states": Kolibri1DecoderLayer,+        "attentions": Kolibri1Attention,+    }+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]++    @torch.no_grad()+    def _init_weights(self, module):+        super()._init_weights(module)+        std = self.config.initializer_range+        if isinstance(module, Kolibri1Experts):+            init.normal_(module.gate_up_proj, mean=0.0, std=std)+            init.normal_(module.down_proj, mean=0.0, std=std)+        elif isinstance(module, Kolibri1TopKRouter):+            init.normal_(module.weight, mean=0.0, std=std)+        if isinstance(module, Kolibri1TopKRouter):+            init.zeros_(module.e_score_correction_bias)+++class Kolibri1RotaryEmbedding(nn.Module):+    def __init__(self, config: Kolibri1Config):+        super().__init__()+        self.max_seq_len_cached = config.max_position_embeddings+        self.original_max_seq_len = config.max_position_embeddings++        self.config = config++        self.rope_type = self.config.rope_parameters["rope_type"]+        rope_init_fn: Callable = self.compute_default_rope_parameters+        if self.rope_type != "default":+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]+        inv_freq, self.attention_scaling = rope_init_fn(self.config)++        self.inv_freq = nn.Buffer(inv_freq, persistent=False)+        self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False)++    @staticmethod+    def compute_default_rope_parameters(config: Kolibri1Config, **kwargs) -> tuple[torch.Tensor, float]:+        """+        Computes the inverse frequencies according to the original RoPE implementation+        Args:+            config ([`~transformers.PreTrainedConfig`]):+                The model configuration.+        Returns:+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).+        """+        base = config.rope_parameters["rope_theta"]+        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads++        attention_factor = 1.0  # Unused in this type of RoPE+        # Compute the inverse frequencies+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))+        return inv_freq, attention_factor++    @torch.no_grad()+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)+    def forward(self, x, position_ids):+        freqs = position_ids[..., None].float() * self.inv_freq.to(device=x.device, dtype=torch.float)+        emb = torch.cat((freqs, freqs), dim=-1)+        cos = emb.cos() * self.attention_scaling+        sin = emb.sin() * self.attention_scaling++        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)+++@auto_docstring+class Kolibri1Model(Kolibri1PreTrainedModel):+    def __init__(self, config: Kolibri1Config):+        super().__init__(config)+        self.padding_idx = config.pad_token_id+        self.vocab_size = config.vocab_size++        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)+        self.layers = nn.ModuleList(+            [Kolibri1DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]+        )+        self.norm = Kolibri1RMSNorm(config.hidden_size, eps=config.rms_norm_eps)+        self.rotary_emb = Kolibri1RotaryEmbedding(config=config)+        self.gradient_checkpointing = False++        # Initialize weights and apply final processing+        self.post_init()++    @merge_with_config_defaults+    @capture_outputs+    @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,+        use_cache: bool | None = None,+        **kwargs: Unpack[TransformersKwargs],+    ) -> MoeModelOutputWithPast:+        if (input_ids is None) ^ (inputs_embeds is not None):+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")++        if inputs_embeds is None:+            inputs_embeds = self.embed_tokens(input_ids)++        if use_cache and past_key_values is None:+            past_key_values = DynamicCache(config=self.config)++        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(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens+            position_ids = position_ids.unsqueeze(0)++        # It may already have been prepared by e.g. `generate`+        if not isinstance(causal_mask_mapping := attention_mask, dict):+            # Prepare mask arguments+            mask_kwargs = {+                "config": self.config,+                "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),+                "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs),+            }++        hidden_states = inputs_embeds+        position_embeddings = self.rotary_emb(hidden_states, position_ids)++        for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):+            hidden_states = decoder_layer(+                hidden_states,+                attention_mask=causal_mask_mapping[self.config.layer_types[i]],+                position_ids=position_ids,+                past_key_values=past_key_values,+                use_cache=use_cache,+                position_embeddings=position_embeddings,+                **kwargs,+            )++        hidden_states = self.norm(hidden_states)+        return MoeModelOutputWithPast(+            last_hidden_state=hidden_states,+            past_key_values=past_key_values if use_cache else None,+        )+++@auto_docstring+class Kolibri1ForCausalLM(Kolibri1PreTrainedModel, GenerationMixin):+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}+    _tp_plan = {"lm_head": "colwise_gather_output"}+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}+    _fsdp_plan = {"lm_head": "keep_full_weight"}++    def __init__(self, config):+        super().__init__(config)+        self.model = Kolibri1Model(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: Unpack[TransformersKwargs],+    ) -> MoeCausalLMOutputWithPast:+        r"""+        Example:++        ```python+        >>> from transformers import AutoTokenizer, Kolibri1ForCausalLM++        >>> model = Kolibri1ForCausalLM.from_pretrained("Aleph-Alpha/Kolibri-1")+        >>> tokenizer = AutoTokenizer.from_pretrained("Aleph-Alpha/Kolibri-1")++        >>> prompt = "Hey, are you conscious? Can you talk to me?"+        >>> inputs = tokenizer(prompt, return_tensors="pt")++        >>> # Generate+        >>> generate_ids = model.generate(**inputs, max_new_tokens=30)+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]+        ```"""+        outputs: MoeModelOutputWithPast = 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.last_hidden_state+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep+        logits = self.lm_head(hidden_states[:, slice_indices, :])++        loss = None+        if labels is not None:+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)++        return MoeCausalLMOutputWithPast(+            loss=loss,+            logits=logits,+            past_key_values=outputs.past_key_values,+            hidden_states=outputs.hidden_states,+            attentions=outputs.attentions,+            router_logits=outputs.router_logits,+        )+++__all__ = ["Kolibri1ForCausalLM", "Kolibri1Model", "Kolibri1PreTrainedModel"]

@@ -0,0 +1,196 @@+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.+#+# Licensed under the Apache License, Version 2.0 (the "License");+# you may not use this file except in compliance with the License.+# You may obtain a copy of the License at+#+#     http://www.apache.org/licenses/LICENSE-2.0+#+# Unless required by applicable law or agreed to in writing, software+# distributed under the License is distributed on an "AS IS" BASIS,+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.+# See the License for the specific language governing permissions and+# limitations under the License.++import torch+import torch.nn.functional as F+from huggingface_hub.dataclasses import strict+from torch import nn++from ... import initialization as init+from ...configuration_utils import PreTrainedConfig+from ...utils import auto_docstring, is_torch_greater_or_equal+from ..deepseek_v3.modeling_deepseek_v3 import DeepseekV3ForCausalLM, DeepseekV3MoE+from ..exaone4.modeling_exaone4 import Exaone4Attention+from ..glm4.modeling_glm4 import Glm4DecoderLayer+from ..qwen2_moe.modeling_qwen2_moe import Qwen2MoeModel+from ..qwen3_moe.configuration_qwen3_moe import Qwen3MoeConfig+from ..qwen3_moe.modeling_qwen3_moe import (+    Qwen3MoeExperts,+    Qwen3MoeMLP,+    Qwen3MoePreTrainedModel,+    Qwen3MoeTopKRouter,+)+++@auto_docstring(checkpoint="Aleph-Alpha/Kolibri-1")+@strict+class Kolibri1Config(Qwen3MoeConfig):+    r"""+    ```python+    >>> from transformers import Kolibri1Model, Kolibri1Config++    >>> configuration = Kolibri1Config()+    >>> model = Kolibri1Model(configuration)+    >>> configuration = model.config+    ```+    """++    model_type = "kolibri1"+    base_model_tp_plan = {+        "layers.*.self_attn.q_proj": "colwise",+        "layers.*.self_attn.k_proj": "colwise",+        "layers.*.self_attn.v_proj": "colwise",+        "layers.*.self_attn.q_norm": "replicated_with_grad_allreduce",+        "layers.*.self_attn.k_norm": "replicated_with_grad_allreduce",+        "layers.*.self_attn.o_proj": "rowwise",+        "layers.*.mlp.experts.gate_up_proj": "packed_colwise",+        "layers.*.mlp.experts.down_proj": "rowwise",+        "layers.*.mlp.experts": "moe_tp_experts",+        "layers.*.mlp.shared_experts.gate_proj": "colwise",+        "layers.*.mlp.shared_experts.up_proj": "colwise",+        "layers.*.mlp.shared_experts.down_proj": "rowwise",+    }+    # Token dispatch by default, so `ep_size` can exceed `tp_size` (bounded by `num_key_value_heads`)+    base_model_ep_plan = {+        "layers.*.mlp.experts.gate_up_proj": "grouped_gemm",+        "layers.*.mlp.experts.down_proj": "grouped_gemm",+        "layers.*.mlp.experts": "ep_dispatch_experts",+    }++    vocab_size: int = 128000+    hidden_size: int = 2560+    num_hidden_layers: int = 50+    num_attention_heads: int = 48+    head_dim: int = 128+    max_position_embeddings: int = 262144+    sliding_window: int | None = 513+    moe_intermediate_size: int = 512+    shared_expert_intermediate_size: int = 512+    num_experts_per_tok: int = 6+    num_experts: int = 384+    layer_types: list[str] | None = None++    intermediate_size = AttributeError()+    use_sliding_window = AttributeError()+    decoder_sparse_step = AttributeError()+    mlp_only_layers = AttributeError()+    attention_bias = AttributeError()+    router_aux_loss_coef = AttributeError()++    def __post_init__(self, **kwargs):+        if self.layer_types is None:+            self.layer_types = [+                "full_attention" if (i + 1) % 5 == 0 else "sliding_attention" for i in range(self.num_hidden_layers)+            ]+        PreTrainedConfig.__post_init__(self, **kwargs)++    def validate_architecture(self):+        PreTrainedConfig.validate_architecture(self)+        if "sliding_attention" in self.layer_types and (self.sliding_window is None or self.sliding_window <= 0):+            raise ValueError(f"Sliding-attention layers need a positive `sliding_window`, got {self.sliding_window}.")+++class Kolibri1Attention(Exaone4Attention):+    def __init__(self, config: Kolibri1Config, layer_idx: int):+        super().__init__(config, layer_idx)+        del self.sliding_window_pattern+++class Kolibri1MLP(Qwen3MoeMLP):+    pass+++class Kolibri1Experts(Qwen3MoeExperts):+    pass+++class Kolibri1TopKRouter(Qwen3MoeTopKRouter):+    def __init__(self, config):+        super().__init__(config)+        self.e_score_correction_bias = nn.Buffer(torch.zeros(self.num_experts))++    def forward(self, hidden_states):+        hidden_states = hidden_states.reshape(-1, self.hidden_dim)+        if (+            hidden_states.is_cuda+            and hidden_states.dtype == self.weight.dtype+            and not torch.is_grad_enabled()+            and is_torch_greater_or_equal("2.8.0")+        ):+            # Faster inference (e.g. in vLLM): fp32 logits from a bf16 GEMM instead of an fp32 one.+            # `out_dtype` is CUDA-only, has no autograd and needs matching dtypes+            router_logits = torch.mm(hidden_states, self.weight.t(), out_dtype=torch.float32)+        else:+            router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))+        router_indices = torch.topk(router_logits + self.e_score_correction_bias, self.top_k, dim=-1)[1]+        router_scores = router_logits.gather(1, router_indices).sigmoid()+        if self.norm_topk_prob:+            router_scores = router_scores / router_scores.sum(dim=-1, keepdim=True)+        return router_logits, router_scores, router_indices+++class Kolibri1SparseMoeBlock(DeepseekV3MoE):+    def __init__(self, config: Kolibri1Config):+        super().__init__(config)+        self.gate = Kolibri1TopKRouter(config)+        self.shared_experts = Kolibri1MLP(config, intermediate_size=config.shared_expert_intermediate_size)+++class Kolibri1DecoderLayer(Glm4DecoderLayer):+    def __init__(self, config: Kolibri1Config, layer_idx: int):+        super().__init__(config, layer_idx)+        self.mlp = Kolibri1SparseMoeBlock(config)+++class Kolibri1PreTrainedModel(Qwen3MoePreTrainedModel):+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]++    @torch.no_grad()+    def _init_weights(self, module):+        super()._init_weights(module)+        if isinstance(module, Kolibri1TopKRouter):+            init.zeros_(module.e_score_correction_bias)+++class Kolibri1Model(Qwen2MoeModel):+    pass+++class Kolibri1ForCausalLM(DeepseekV3ForCausalLM):+    def forward(self, **super_kwargs):+        r"""+        Example:++        ```python+        >>> from transformers import AutoTokenizer, Kolibri1ForCausalLM++        >>> model = Kolibri1ForCausalLM.from_pretrained("Aleph-Alpha/Kolibri-1")+        >>> tokenizer = AutoTokenizer.from_pretrained("Aleph-Alpha/Kolibri-1")++        >>> prompt = "Hey, are you conscious? Can you talk to me?"+        >>> inputs = tokenizer(prompt, return_tensors="pt")++        >>> # Generate+        >>> generate_ids = model.generate(**inputs, max_new_tokens=30)+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]+        ```"""+        return super().forward(**super_kwargs)+++__all__ = [+    "Kolibri1Config",+    "Kolibri1ForCausalLM",+    "Kolibri1Model",+    "Kolibri1PreTrainedModel",+]

GitHub publishes no diff for this file. View it on GitHub ↗

@@ -0,0 +1,152 @@+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.+#+# Licensed under the Apache License, Version 2.0 (the "License");+# you may not use this file except in compliance with the License.+# You may obtain a copy of the License at+#+#     http://www.apache.org/licenses/LICENSE-2.0+#+# Unless required by applicable law or agreed to in writing, software+# distributed under the License is distributed on an "AS IS" BASIS,+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.+# See the License for the specific language governing permissions and+# limitations under the License.+"""Testing suite for the PyTorch Kolibri1 model."""++import unittest++from huggingface_hub.errors import StrictDataclassClassValidationError++from transformers import AutoTokenizer, is_torch_available+from transformers.testing_utils import (+    Expectations,+    require_torch,+    require_torch_accelerator_memory,+    require_torch_gpu,+    slow,+    torch_device,+)++from ...causal_lm_tester import CausalLMModelTest, CausalLMModelTester+from ...test_memory_cleanup_mixin import MemoryCleanupMixin+++if is_torch_available():+    import torch++    from transformers import Kolibri1Config, Kolibri1ForCausalLM, Kolibri1Model+++class Kolibri1ModelTester(CausalLMModelTester):+    if is_torch_available():+        base_model_class = Kolibri1Model++    def __init__(self, parent):+        super().__init__(parent=parent)+        self.layer_types = ["sliding_attention", "full_attention"]+++@require_torch+class Kolibri1ModelTest(CausalLMModelTest, unittest.TestCase):+    model_tester_class = Kolibri1ModelTester++    def test_sliding_window_required(self):+        with self.assertRaises(StrictDataclassClassValidationError):+            Kolibri1Config(sliding_window=None)++    def test_no_load_balancing_loss(self):+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()+        model = Kolibri1ForCausalLM(config).to(torch_device).eval()+        input_ids = input_dict["input_ids"]++        with torch.no_grad():+            plain = model(input_ids, labels=input_ids)+            with_router_logits = model(input_ids, labels=input_ids, output_router_logits=True)++        self.assertIsNone(with_router_logits.aux_loss)+        self.assertEqual(len(with_router_logits.router_logits), config.num_hidden_layers)+        torch.testing.assert_close(with_router_logits.loss, plain.loss)++    @require_torch_gpu+    def test_router_with_activations_in_another_dtype(self):+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()+        model = Kolibri1ForCausalLM(config).to(torch_device, torch.bfloat16).eval()+        inputs_embeds = model.get_input_embeddings()(input_dict["input_ids"]).float()++        with torch.no_grad(), torch.autocast(torch_device, dtype=torch.bfloat16):+            logits = model(inputs_embeds=inputs_embeds).logits++        self.assertFalse(logits.isnan().any())+++@slow+@require_torch_accelerator_memory(memory=180)+class Kolibri1IntegrationTest(MemoryCleanupMixin, unittest.TestCase):+    TEST_MODEL_ID = "Aleph-Alpha/Kolibri-1-BF16"++    @classmethod+    def setUpClass(cls):+        super().setUpClass()+        # Dropped again by the mixin's `tearDownClass`.+        cls.model = None++    @classmethod+    def get_model(cls):+        if cls.model is None:+            cls.model = Kolibri1ForCausalLM.from_pretrained(+                cls.TEST_MODEL_ID, device_map="auto", experts_implementation="eager"+            )+        return cls.model++    def generate_reply(self, content, max_new_tokens):+        tokenizer = AutoTokenizer.from_pretrained(self.TEST_MODEL_ID)+        model = self.get_model()+        inputs = tokenizer.apply_chat_template(+            [{"role": "user", "content": content}],+            add_generation_prompt=True,+            enable_thinking=False,+            return_tensors="pt",+            return_dict=True,+        ).to(model.device)+        with torch.no_grad():+            generated_ids = model.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)+        return tokenizer.decode(generated_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)++    def test_model_logits(self):+        input_ids = [452, 22090, 493, 1678, 2459]+        model = self.get_model()+        input_ids = torch.tensor([input_ids]).to(model.device)+        with torch.no_grad():+            out = model(input_ids).logits.float().cpu()++        # fmt: off+        EXPECTED_MEAN = Expectations(+            {+                ("cuda", None): torch.tensor([[-3.6648, -4.0749, -3.8838, -4.3631, -3.9150]]),+            }+        ).get_expectation()+        EXPECTED_SLICE = Expectations(+            {+                ("cuda", None): torch.tensor(+                    [-2.0000, -7.0938, -8.3750, -9.5000, -8.6875, -8.6875, -8.9375, -7.9062, -4.7188, 1.8594]+                ),+            }+        ).get_expectation()+        # fmt: on+        torch.testing.assert_close(out.mean(-1), EXPECTED_MEAN, atol=1e-2, rtol=1e-2)+        torch.testing.assert_close(out[0, 0, :10], EXPECTED_SLICE, atol=1e-2, rtol=1e-2)++    def test_model_generation(self):+        EXPECTED_TEXT = (+            "Ein Mixture-of-Experts-Modell ist ein neuronales Netz, das aus mehreren spezialisierten Experten "+            "besteht, von denen bei jeder Vorhersage nur ein kleiner Teil aktiviert wird, um Effizienz und "+            "Skalierbarkeit zu erhöhen."+        )+        text = self.generate_reply("Erkläre in einem Satz, was ein Mixture-of-Experts-Modell ist.", 40)+        self.assertEqual(EXPECTED_TEXT, text)++    def test_model_recall_beyond_sliding_window(self):+        filler = "Der Himmel ist blau. Das Gras ist grün. Die Sonne scheint. " * 150+        prompt = f"Merke dir: Der Geheimcode lautet 4721.\n\n{filler}\nWie lautet der Geheimcode? Antworte nur mit der Zahl."+        text = self.generate_reply(prompt, 10)+        self.assertEqual("4721", text)