Update README.md

.eval_results/swe-bench_pro.yaml

@@ -1,7 +0,0 @@
-- dataset:
-    id: ScaleAI/SWE-bench_Pro
-    task_id: SWE_Bench_Pro
-  value: 44.5
-  source:
-    url: https://huggingface.co/poolside/Laguna-XS.2
-    name: Model Card

.eval_results/swe-bench_verified.yaml

@@ -1,7 +0,0 @@
-- dataset:
-    id: SWE-bench/SWE-bench_Verified
-    task_id: swe_bench_%_resolved
-  value: 68.2
-  source:
-    url: https://huggingface.co/poolside/Laguna-XS.2
-    name: Model Card

.eval_results/terminal-bench-2.0.yaml

@@ -1,7 +0,0 @@
-- dataset:
-    id: harborframework/terminal-bench-2.0
-    task_id: terminalbench_2
-  value: 30.1
-  source:
-    url: https://huggingface.co/poolside/Laguna-XS.2
-    name: Model Card

LICENSE.md

@@ -1,202 +0,0 @@
-
-                                 Apache License
-                           Version 2.0, January 2004
-                        http://www.apache.org/licenses/
-
-   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
-
-   1. Definitions.
-
-      "License" shall mean the terms and conditions for use, reproduction,
-      and distribution as defined by Sections 1 through 9 of this document.
-
-      "Licensor" shall mean the copyright owner or entity authorized by
-      the copyright owner that is granting the License.
-
-      "Legal Entity" shall mean the union of the acting entity and all
-      other entities that control, are controlled by, or are under common
-      control with that entity. For the purposes of this definition,
-      "control" means (i) the power, direct or indirect, to cause the
-      direction or management of such entity, whether by contract or
-      otherwise, or (ii) ownership of fifty percent (50%) or more of the
-      outstanding shares, or (iii) beneficial ownership of such entity.
-
-      "You" (or "Your") shall mean an individual or Legal Entity
-      exercising permissions granted by this License.
-
-      "Source" form shall mean the preferred form for making modifications,
-      including but not limited to software source code, documentation
-      source, and configuration files.
-
-      "Object" form shall mean any form resulting from mechanical
-      transformation or translation of a Source form, including but
-      not limited to compiled object code, generated documentation,
-      and conversions to other media types.
-
-      "Work" shall mean the work of authorship, whether in Source or
-      Object form, made available under the License, as indicated by a
-      copyright notice that is included in or attached to the work
-      (an example is provided in the Appendix below).
-
-      "Derivative Works" shall mean any work, whether in Source or Object
-      form, that is based on (or derived from) the Work and for which the
-      editorial revisions, annotations, elaborations, or other modifications
-      represent, as a whole, an original work of authorship. For the purposes
-      of this License, Derivative Works shall not include works that remain
-      separable from, or merely link (or bind by name) to the interfaces of,
-      the Work and Derivative Works thereof.
-
-      "Contribution" shall mean any work of authorship, including
-      the original version of the Work and any modifications or additions
-      to that Work or Derivative Works thereof, that is intentionally
-      submitted to Licensor for inclusion in the Work by the copyright owner
-      or by an individual or Legal Entity authorized to submit on behalf of
-      the copyright owner. For the purposes of this definition, "submitted"
-      means any form of electronic, verbal, or written communication sent
-      to the Licensor or its representatives, including but not limited to
-      communication on electronic mailing lists, source code control systems,
-      and issue tracking systems that are managed by, or on behalf of, the
-      Licensor for the purpose of discussing and improving the Work, but
-      excluding communication that is conspicuously marked or otherwise
-      designated in writing by the copyright owner as "Not a Contribution."
-
-      "Contributor" shall mean Licensor and any individual or Legal Entity
-      on behalf of whom a Contribution has been received by Licensor and
-      subsequently incorporated within the Work.
-
-   2. Grant of Copyright License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      copyright license to reproduce, prepare Derivative Works of,
-      publicly display, publicly perform, sublicense, and distribute the
-      Work and such Derivative Works in Source or Object form.
-
-   3. Grant of Patent License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      (except as stated in this section) patent license to make, have made,
-      use, offer to sell, sell, import, and otherwise transfer the Work,
-      where such license applies only to those patent claims licensable
-      by such Contributor that are necessarily infringed by their
-      Contribution(s) alone or by combination of their Contribution(s)
-      with the Work to which such Contribution(s) was submitted. If You
-      institute patent litigation against any entity (including a
-      cross-claim or counterclaim in a lawsuit) alleging that the Work
-      or a Contribution incorporated within the Work constitutes direct
-      or contributory patent infringement, then any patent licenses
-      granted to You under this License for that Work shall terminate
-      as of the date such litigation is filed.
-
-   4. Redistribution. You may reproduce and distribute copies of the
-      Work or Derivative Works thereof in any medium, with or without
-      modifications, and in Source or Object form, provided that You
-      meet the following conditions:
-
-      (a) You must give any other recipients of the Work or
-          Derivative Works a copy of this License; and
-
-      (b) You must cause any modified files to carry prominent notices
-          stating that You changed the files; and
-
-      (c) You must retain, in the Source form of any Derivative Works
-          that You distribute, all copyright, patent, trademark, and
-          attribution notices from the Source form of the Work,
-          excluding those notices that do not pertain to any part of
-          the Derivative Works; and
-
-      (d) If the Work includes a "NOTICE" text file as part of its
-          distribution, then any Derivative Works that You distribute must
-          include a readable copy of the attribution notices contained
-          within such NOTICE file, excluding those notices that do not
-          pertain to any part of the Derivative Works, in at least one
-          of the following places: within a NOTICE text file distributed
-          as part of the Derivative Works; within the Source form or
-          documentation, if provided along with the Derivative Works; or,
-          within a display generated by the Derivative Works, if and
-          wherever such third-party notices normally appear. The contents
-          of the NOTICE file are for informational purposes only and
-          do not modify the License. You may add Your own attribution
-          notices within Derivative Works that You distribute, alongside
-          or as an addendum to the NOTICE text from the Work, provided
-          that such additional attribution notices cannot be construed
-          as modifying the License.
-
-      You may add Your own copyright statement to Your modifications and
-      may provide additional or different license terms and conditions
-      for use, reproduction, or distribution of Your modifications, or
-      for any such Derivative Works as a whole, provided Your use,
-      reproduction, and distribution of the Work otherwise complies with
-      the conditions stated in this License.
-
-   5. Submission of Contributions. Unless You explicitly state otherwise,
-      any Contribution intentionally submitted for inclusion in the Work
-      by You to the Licensor shall be under the terms and conditions of
-      this License, without any additional terms or conditions.
-      Notwithstanding the above, nothing herein shall supersede or modify
-      the terms of any separate license agreement you may have executed
-      with Licensor regarding such Contributions.
-
-   6. Trademarks. This License does not grant permission to use the trade
-      names, trademarks, service marks, or product names of the Licensor,
-      except as required for reasonable and customary use in describing the
-      origin of the Work and reproducing the content of the NOTICE file.
-
-   7. Disclaimer of Warranty. Unless required by applicable law or
-      agreed to in writing, Licensor provides the Work (and each
-      Contributor provides its Contributions) on an "AS IS" BASIS,
-      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
-      implied, including, without limitation, any warranties or conditions
-      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
-      PARTICULAR PURPOSE. You are solely responsible for determining the
-      appropriateness of using or redistributing the Work and assume any
-      risks associated with Your exercise of permissions under this License.
-
-   8. Limitation of Liability. In no event and under no legal theory,
-      whether in tort (including negligence), contract, or otherwise,
-      unless required by applicable law (such as deliberate and grossly
-      negligent acts) or agreed to in writing, shall any Contributor be
-      liable to You for damages, including any direct, indirect, special,
-      incidental, or consequential damages of any character arising as a
-      result of this License or out of the use or inability to use the
-      Work (including but not limited to damages for loss of goodwill,
-      work stoppage, computer failure or malfunction, or any and all
-      other commercial damages or losses), even if such Contributor
-      has been advised of the possibility of such damages.
-
-   9. Accepting Warranty or Additional Liability. While redistributing
-      the Work or Derivative Works thereof, You may choose to offer,
-      and charge a fee for, acceptance of support, warranty, indemnity,
-      or other liability obligations and/or rights consistent with this
-      License. However, in accepting such obligations, You may act only
-      on Your own behalf and on Your sole responsibility, not on behalf
-      of any other Contributor, and only if You agree to indemnify,
-      defend, and hold each Contributor harmless for any liability
-      incurred by, or claims asserted against, such Contributor by reason
-      of your accepting any such warranty or additional liability.
-
-   END OF TERMS AND CONDITIONS
-
-   APPENDIX: How to apply the Apache License to your work.
-
-      To apply the Apache License to your work, attach the following
-      boilerplate notice, with the fields enclosed by brackets "[]"
-      replaced with your own identifying information. (Don't include
-      the brackets!)  The text should be enclosed in the appropriate
-      comment syntax for the file format. We also recommend that a
-      file or class name and description of purpose be included on the
-      same "printed page" as the copyright notice for easier
-      identification within third-party archives.
-
-   Copyright 2026 Poolside
-
-   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.

README.md

@@ -1,18 +1,19 @@
 ---
-library_name: transformers
+library_name: vllm
 inference: false
+base_model:
+- poolside/Laguna-XS.2-base
 extra_gated_description: >-
   To learn more about how we process your personal data, please read our <a
-  href="https://poolside.ai/legal/privacy">Privacy Policy</a>.
+  href="https://poolside.ai/privacy">Privacy Policy</a>.
 tags:
 - laguna-xs.2
-- vllm
 license: apache-2.0
 pipeline_tag: text-generation
 ---
 
 <p align="center">
-  <img alt="poolside-banner" src="https://poolside.ai/assets/laguna/laguna-xs2-banner.svg" width="800px">
+  <img alt="poolside-banner" src="">
 </p>
 
 <p align="center">
@@ -27,12 +28,14 @@ pipeline_tag: text-generation
 Laguna XS.2 is a 33B total parameter Mixture-of-Experts model with 3B activated parameters per token designed for agentic coding and long-horizon work on a local machine. It uses Sliding Window Attention with per-head gating in 30 out of 40 layers for fast inference and low KV cache requirements.
 
 > [!NOTE]
-> For more details on how we trained this model, including on data automixing and async off-policy agent RL, check out our [release blog post](https://poolside.ai/blog/laguna-a-deeper-dive).
+> This is the instruct model with native reasoning support and interleaved thinking. For the base model, see [Laguna XS.2-base](https://huggingface.co/poolside/Laguna-XS.2-base).
+
+For more details on how we trained this model, including on data automixing and async off-policy agent RL, check out our [release blog post](https://poolside.ai/blog/laguna-a-deeper-dive).
 
 ## Highlights
 - **Mixed SWA and global attention layout**: Laguna XS.2 uses sigmoid gating with per-layer rotary scales, enabling mixed SWA (Sliding Window Attention) and global attention layers in a 3:1 ratio (across 40 total layers)
-- **KV cache in FP8**: KV cache quantized to FP8, reducing memory per token
-- **Native reasoning support**: Interleaved thinking between tool calls with support for enabling and disabling thinking per-request
+- **KV cache in FP8**: All quantization formats use a KV cache quantized to FP8, reducing memory per token
+- **Native reasoning support**: Interleaved thinking enabled by default
 - **Local-ready**: At 33B total parameters and 3B activated, Laguna XS.2 is compact enough to run on a Mac with 36 GB of RAM. [Available on Ollama](https://ollama.com/library/laguna-xs.2)
 - **Apache 2.0 license**: Use and modify freely for commercial and non-commercial purposes
 
@@ -40,7 +43,7 @@ Laguna XS.2 is a 33B total parameter Mixture-of-Experts model with 3B activated
 
 ## Model overview
 
-- Training: pre-training, post-training and reinforcement learning stages
+- Training: pre-training, post-training and reinforcement learning stages (instruct)
 - Number of parameters: 33B total with 3B activated per token
 - Optimizer: Muon
 - Layers: 40 layers (10 layers with global attention, 30 layers with sliding window attention)
@@ -48,41 +51,40 @@ Laguna XS.2 is a 33B total parameter Mixture-of-Experts model with 3B activated
 - Sliding Window: 512 tokens
 - Modality: text-to-text
 - Context window: 131,072 tokens
-- Reasoning support: interleaved thinking with preserved thinking
+- Reasoning support: thinking default enabled; interleaved thinking with preserved thinking supported
 
 ## Benchmark results
 
-<p align="center">
-  <img alt="benchmarks" src="https://poolside.ai/assets/laguna/laguna-xs2-chart.svg" width="800px">
-</p>
+[Placeholder for chart SVG]
+
+We evaluate Laguna XS.2 with thinking enabled in our agent harness, pool (see the Usage section below to download and run locally), across all benchmarks. For other models, we use the best available publicly-reported score; if not available, we calculate baselines using OpenHands (SWE-bench family) or Terminus 2 (Terminal-Bench 2.0) using the settings below.
 
-| Model                     | Size (total params.) | SWE-bench Verified | SWE-bench Multilingual | SWE-bench Pro (Public Dataset) | Terminal-Bench 2.0 |
-|---------------------------|----------------------|--------------------|------------------------|--------------------------------|--------------------|
-| **Laguna XS.2**           | 33B                  | 68.2%              | 62.4%                  | 44.5%                          | 30.1%              |
-| Devstral Small 2          | 24B dense            | 68.0%              | 55.7%                  | -                              | 22.5%              |
-| Gemma 4 31B IT            | 31B dense            | 52.0%              | 51.7%                  | 35.7%                          | 42.9%              |
-| Qwen3.5-35B-A3B           | 35B                  | 69.2%              | 60.3%                  | 44.6%                          | 40.5%              |
-| Qwen3.6-35B-A3B           | 35B                  | 73.4%              | 67.2%                  | 49.5%                          | 51.5%              |
-| Claude Haiku 4.5          | -                    | 73.3%              | -                      | 39.5%                          | 29.8%              |
-| GPT-5.4 Nano              | -                    | -                  | -                      | 52.4%                          | 46.3%              |
+| Model                     | Size (total params.) | SWE-bench Pro (Public Dataset) | SWE-bench Verified | SWE-bench Multilingual | Terminal-Bench 2.0 |
+|---------------------------|----------------------|--------------------------------|--------------------|------------------------|--------------------|
+| **Laguna XS.2**           | 33B                  | 44.5%                          | 68.2%              | 62.4%                  | 30.1%              |
+| Devstral Small 2          | 24B dense            | -                              | 68.0%              | 55.7%                  | 22.5%              |
+| Gemma 4 31B IT            | 31B dense            | 35.7%                          | 52.0%              | 51.7%                  | 42.9%              |
+| Qwen3.5-35B-A3B           | 35B                  | 44.6%                          | 69.2%              | 60.3%                  | 40.5%              |
+| GPT-5.4 Nano              | -                    | 52.4%                          | -                  | -                      | 46.3%              |
+| Qwen3.6-27B               | 27B dense            | 53.2%                          | 77.2%              | 71.3%                  | 59.3%              |
 
-*We used the highest publicly-referenced scores for all comparison models across each benchmark. In almost all cases these were official scores published in release blog posts or equivalent, with the exception of Gemma 4 31B IT where the highest published scores were [reported by the Qwen team](https://qwen.ai/blog?id=qwen3.6-35b-a3b) and Claude Haiku 4.5 where the highest published (verified) scores for SWE-bench Pro and Terminal-Bench 2.0 are from their respective official leaderboards.*
+*We used the highest publicly-referenced scores for all comparison models across each benchmark. In all cases these were official scores published in release blog posts or equivalent, with the exception of Gemma 4 31B IT where the highest published scores were [reported by the Qwen team](https://qwen.ai/blog?id=qwen3.6-35b-a3b).*
 
 <details>
 <summary>Expand for benchmarking methodology</summary>
 
 All benchmarking for Laguna XS.2 was completed using the Laude Institute’s Harbor Framework with our [agent harness](https://github.com/poolsideai/pool), using a maximum of 500 steps and sandboxed execution using 8 GB RAM/2 CPUs (with the exception of Terminal-Bench 2.0; see below). The same sampling parameters were used for all benchmarking: temperature=0.7 and top_k=20.  Some base task images and verifiers were patched to fix infrastructure reliability issues inherent in task setup, such as rate limits on third-party dependencies in external registries used by the verifier. More details outlining these updates and other findings will follow in a future technical blog post.
 
+- SWE-bench Pro: mean pass@1 averaged over 3 runs.
 - SWE-bench Verified: mean pass@1 averaged over 4 runs.
 - SWE-bench Multilingual: mean pass@1 averaged over 7 runs.
-- SWE-bench Pro: mean pass@1 averaged over 3 runs.
 - Terminal-Bench 2.0: mean pass@1 averaged over 5 runs. 48GB RAM/32 CPUs.
 
 </details>
 
 ## Usage
 
-Laguna XS.2 has launch-day support in vLLM and Transformers, and TRT-LLM thanks to the support of the team at NVIDIA.
+Laguna XS.2 has launch-day support in vLLM and Transformers, and TRT-LLM and SGLang thanks to the support of the team at NVIDIA.
 
 The fastest way to get started is with our API, directly or using OpenRouter.
 
@@ -105,6 +107,8 @@ Launch and *Log in with Poolside* to get a free API key.
 pool
 ```
 
+[Placeholder for screenshot]
+
 Use in any [ACP client](https://agentclientprotocol.com/get-started/clients). Configure Zed and JetBrains automatically:
 
 ```shell
@@ -114,127 +118,35 @@ pool acp setup --editor zed|jetbrains
 Use pool with Ollama with one-command setup:
 
 ```shell
-ollama pull laguna-xs.2
-ollama launch pool --model laguna-xs.2
+ollama pull laguna.xs-2
+ollama launch pool --model laguna.xs-2
 ```
 
+(requires Ollama 0.20.8 or later)
+
 #### Feedback and issues
 
 Submit feedback with `/feedback` and read the [full documentation on GitHub](https://github.com/poolsideai/pool).
 
-### Local deployment
+*By downloading and using pool, you agree to the Poolside [End User License Agreement (EULA)](https://poolside.ai/legal/eula).*
 
-Laguna XS.2 is supported in vLLM and Transformers, and TRT-LLM thanks to the support of the team at NVIDIA. Use Laguna-XS.2 with Ollama (with MLX support) and the mlx-lm framework for the best experience on your local machine.
+### Local deployment
 
-#### vLLM
+[vLLM, Transformers v5, TRT-LLM, SGLang, ...]
 
-Serve Laguna XS.2 locally with vLLM and query it from any OpenAI-compatible client (see [Controlling reasoning](#controlling-reasoning) for tool calls, streaming, and reasoning extraction):
+Thanks to support from Ollama and the mlx-lm team...
 
-> [!NOTE]
-> Laguna XS.2 support has been merged into vLLM ([vllm-project/vllm#41129](https://github.com/vllm-project/vllm/pull/41129)) and will ship in the next release. Until then, install a nightly wheel:
+[Device frameworks: Ollama, mlx-lm, ...]
 
-```shell
-pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
-
-VLLM_USE_DEEP_GEMM=0 vllm serve \
-    --model poolside/Laguna-XS.2 \
-    --tool-call-parser poolside_v1 \
-    --reasoning-parser poolside_v1 \
-    --enable-auto-tool-choice \
-    --served-model-name laguna \
-    --default-chat-template-kwargs '{"enable_thinking": true}'
-```
+#### vLLM
 
-See the [vLLM recipes page](https://recipes.vllm.ai/poolside/Laguna-XS.2) for additional deployment guidance.
+[...]
 
 #### Transformers
 
-Laguna XS.2 is supported in Transformers `v5.7.0` and later ([huggingface/transformers#45673](https://github.com/huggingface/transformers/pull/45673)).
-
-```python
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-model_id = "poolside/Laguna-XS.2"
-
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(
-    model_id,
-    dtype=torch.bfloat16,
-    device_map="auto",
-)
-
-messages = [
-    {"role": "user", "content": "Write a Python retry wrapper with exponential backoff."},
-]
-
-# Reasoning is on by default; pass enable_thinking=False to skip the <think> block.
-inputs = tokenizer.apply_chat_template(
-    messages,
-    add_generation_prompt=True,
-    return_tensors="pt",
-    enable_thinking=True,
-).to(model.device)
-
-outputs = model.generate(
-    inputs,
-    max_new_tokens=1024,
-    do_sample=True,
-    temperature=0.7,
-    top_k=20,
-)
-
-response = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
-print(response)
-```
-
-#### TRT-LLM
-
-> [!NOTE]
-> Requires building TensorRT-LLM from the upstream PR that adds Laguna XS.2 support
-> ([NVIDIA/TensorRT-LLM#13559](https://github.com/NVIDIA/TensorRT-LLM/pull/13559)).
-> Once that PR merges, the same code will work on a released `tensorrt-llm` wheel.
-
-Laguna XS.2's `configuration_laguna.py` imports a few `transformers >= 4.58` symbols.
-TRT-LLM currently pins `transformers 4.57`, so the PR ships a `laguna_minimal_overlay.sh` script that symlinks the checkpoint and patches only the config file with a compat shim. Load TRT-LLM against the **overlay directory**, not the original checkpoint.
-
-```shell
-# 1. Check out the PR branch and build TRT-LLM from source (see the TensorRT-LLM build docs).
-git clone https://github.com/NVIDIA/TensorRT-LLM.git && cd TensorRT-LLM
-git fetch origin pull/13559/head:laguna && git checkout laguna
-
-# 2. Download the checkpoint.
-huggingface-cli download poolside/Laguna-XS.2 --local-dir ~/models/Laguna-XS.2
-
-# 3. Build the transformers-4.57 compat overlay (echoes the overlay path).
-OVERLAY=$(bash laguna_minimal_overlay.sh ~/models/Laguna-XS.2)
-```
-
-```python
-from tensorrt_llm import LLM, SamplingParams
-
-llm = LLM(
-    model=OVERLAY,             # overlay path, not the original checkpoint
-    trust_remote_code=True,
-    tensor_parallel_size=1,
-)
-
-sampling = SamplingParams(max_tokens=1024, temperature=0.7, top_k=20)
-out = llm.generate(["Write a Python retry wrapper with exponential backoff."], sampling)
-print(out[0].outputs[0].text)
-```
-
-Or serve with an OpenAI-compatible endpoint:
-
-```shell
-trtllm-serve "$OVERLAY" --port 8000 --trust-remote-code
-```
+[...]
 
-The same recipe works for the [FP8](https://huggingface.co/poolside/Laguna-XS.2-FP8) and [NVFP4](https://huggingface.co/poolside/Laguna-XS.2-NVFP4) variants: quantization is detected automatically from `quantization_config`, no extra flags required.
-
-#### Ollama
-
-Visit [Ollama's model library](https://ollama.com/library/laguna-xs.2) to pull to your local machine.
+#### [Other frameworks]
 
 ## Controlling reasoning
 
@@ -277,8 +189,8 @@ response = client.chat.completions.create(
 reasoning, content, tool_calls = "", "", []
 for chunk in response:
   delta = chunk.choices[0].delta
-  if hasattr(delta, "reasoning_content") and delta.reasoning_content:
-    reasoning += delta.reasoning_content
+  if hasattr(delta, "reasoning") and delta.reasoning:
+    reasoning += delta.reasoning
   if hasattr(delta, "content") and delta.content:
     content += delta.content
   if hasattr(delta, "tool_calls") and delta.tool_calls:
@@ -296,7 +208,7 @@ print(f"Reasoning: {reasoning}\nContent: {content}\nTool calls: {tool_calls}\n")
 messages.append({
   "role": "assistant",
   "content": content,
-  "reasoning_content": reasoning,
+  "reasoning": reasoning,
   "tool_calls": [{"id": tc["id"], "type": "function", "function": tc["function"]} for tc in tool_calls]
 })
 
@@ -358,10 +270,6 @@ For agentic coding use cases, we recommend enabling thinking and preserving reas
 
 ## License
 
-This model is licensed under the [Apache 2.0 License](https://huggingface.co/poolside/Laguna-XS.2/blob/main/LICENSE.md).
-
-## Intended and Responsible Use 
-
-Laguna XS.2 is designed for software engineering and agentic coding use cases, and you are responsible for confirming that it is appropriate for your intended application. Laguna XS.2 is subject to the [Apache 2.0 License](https://huggingface.co/poolside/Laguna-XS.2/blob/main/LICENSE.md), and should be used consistently with Poolside's [Acceptable Use Policy](https://poolside.ai/legal/acceptable-use-policy). We advise against circumventing Laguna XS.2 safety guardrails without implementing substantially equivalent mitigations appropriate for your use case.
+This model is licensed under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0.txt).
 
-Please report security vulnerabilities or safety concerns to [security@poolside.ai](mailto:security@poolside.ai).
+You must not use this model in a manner that infringes, misappropriates, or otherwise violates any third party’s rights, including intellectual property rights.

chat_template.jinja

@@ -1,12 +1,12 @@
-{#- Iteration on laguna_glm_thinking_v5/chat_template.jinja -#}
-{#- Adds a default system message (used when no system message is provided in `messages`). -#}
+{#- Copied from laguna_glm_thinking_v4/chat_template.jinja -#}
+{#- Removes prefix that references <think> token, and replaces message.reasoning_content reference with message.reasoning -#}
 {{- "〈|EOS|〉" -}}
 {%- set enable_thinking = enable_thinking | default(false) -%}
 {%- set render_assistant_messages_raw = render_assistant_messages_raw | default(false) -%}
 {%- set add_generation_prompt = add_generation_prompt | default(false) -%}
 
 {#- ───── header (system message) ───── -#}
-{%- set system_message = "You are a helpful, conversationally-fluent assistant made by Poolside. You are here to be helpful to users through natural language conversations." -%}
+{%- set system_message = "" -%}
 {%- if messages and messages[0].role == "system" -%}
   {%- set system_message = messages[0].content -%}
 {%- endif -%}

config.json

@@ -49,8 +49,7 @@
       "rope_type": "default",
       "rope_theta": 10000.0,
       "partial_rotary_factor": 1.0
-    },
-    "original_max_position_embeddings": 4096
+    }
   },
   "layer_types": [
     "full_attention",
@@ -180,4 +179,14 @@
     64,
     64,
     64
-  ]}
+  ],
+  "compression_config": {
+    "mode": null,
+    "group_size": 32,
+    "eps": 1e-05,
+    "filter_fqns": [
+      "output"
+    ],
+    "recompute_fake_quantize": false
+  }
+}

configuration_laguna.py

@@ -1,4 +1,5 @@
-# Copyright 2026 Poolside and the HuggingFace Inc. team. All rights reserved.
+# ruff: noqa
+# Copyright 2025 Poolside and 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.
@@ -11,44 +12,79 @@
 # 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 Any, Literal
-
-from huggingface_hub.dataclasses import strict
-
 from transformers.configuration_utils import PreTrainedConfig
 from transformers.modeling_rope_utils import RopeParameters
-from transformers.utils import auto_docstring
 
 
-@auto_docstring(checkpoint="poolside/laguna-XS.2")
-@strict
 class LagunaConfig(PreTrainedConfig):
     r"""
-    partial_rotary_factor (`float`, *optional*):
-        Fraction of ``head_dim`` to rotate. Folded into each ``rope_parameters[layer_type]``
-        entry by ``__post_init__``.
-    num_attention_heads_per_layer (`list[int]`, *optional*):
-        Per-layer override for ``num_attention_heads``. Length must equal ``num_hidden_layers``.
-    mlp_layer_types (`list[str]`, *optional*):
-        Per-layer MLP type — ``"dense"`` or ``"sparse"``. Length must equal
-        ``num_hidden_layers``. Defaults to first layer dense, rest sparse.
-    moe_routed_scaling_factor (`float`, *optional*, defaults to 1.0):
-        Scalar applied to routed-expert output before combining with the shared-expert output.
-    moe_apply_router_weight_on_input (`bool`, *optional*, defaults to `False`):
-        Whether to apply router weights to the MoE input rather than the output. Not supported
-        in transformers yet; ``True`` will raise a ``NotImplementedError`` for now.
-    moe_router_logit_softcapping (`float`, *optional*, defaults to 0.0):
-        Scaling factor when applying tanh softcapping on the logits of the MoE router logits.
-
-    Example:
-
-    ```python
-    >>> from transformers import LagunaModel, LagunaConfig
-
-    >>> configuration = LagunaConfig()
-    >>> model = LagunaModel(configuration)
-    >>> configuration = model.config
-    ```
+    Configuration class for Laguna model.
+
+    Laguna is Poolside's MoE architecture with:
+    - Attention output gating (softplus gate)
+    - Sigmoid routing instead of softmax
+    - No QKV bias
+    - Explicit head_dim parameter
+
+    Args:
+        head_dim (`int`, *optional*, defaults to 128):
+            Dimension of attention heads. Laguna uses explicit head_dim rather than
+            computing it from hidden_size // num_attention_heads.
+        qkv_bias (`bool`, *optional*, defaults to `False`):
+            Whether to add bias to QKV projections. Laguna uses no QKV bias.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to add bias to attention output projection. Laguna uses no attention bias.
+        gating (`bool`, *optional*, defaults to `True`):
+            Whether to use softplus output gating on attention. When True, a g_proj linear
+            layer is added and attn_output = attn_output * softplus(g_proj(x)).
+        sliding_window (`int`, *optional*):
+            Sliding window attention size. Used by layers whose type in ``layer_types``
+            is ``"sliding_attention"``. When ``None``, all layers use full attention.
+        layer_types (`list[str]`, *optional*):
+            Per-layer attention type. Each element should be ``"sliding_attention"`` or
+            ``"global_attention"``. Length must equal ``num_hidden_layers``. When ``None``,
+            all layers default to global attention.
+        swa_attention_sink_enabled (`bool`, *optional*, defaults to `False`):
+            Whether to enable learnable attention sinks on sliding-window attention layers.
+            When enabled, a per-head bias parameter is added that allows the model to attend
+            to position 0 even when it falls outside the sliding window.
+        swa_rope_parameters (`RopeParameters`, *optional*):
+            Separate RoPE configuration for sliding-window attention layers. When ``None``,
+            SWA layers use the same RoPE as global attention layers.
+        vocab_size (`int`, *optional*, defaults to 100352):
+            Vocabulary size of the Laguna model.
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 8192):
+            Dimension of the MLP representations for dense layers.
+        num_hidden_layers (`int`, *optional*, defaults to 48):
+            Number of hidden layers in the Transformer.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            Number of key-value heads for GQA.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            Maximum sequence length.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-6):
+            Epsilon for RMSNorm layers.
+        num_experts (`int`, *optional*, defaults to 256):
+            Number of routed experts.
+        num_experts_per_tok (`int`, *optional*, defaults to 16):
+            Number of experts selected per token (top-k).
+        moe_intermediate_size (`int`, *optional*, defaults to 1024):
+            Intermediate size of routed experts.
+        shared_expert_intermediate_size (`int`, *optional*, defaults to 1024):
+            Intermediate size of the shared expert.
+        norm_topk_prob (`bool`, *optional*, defaults to `True`):
+            Whether to normalize top-k routing probabilities.
+        decoder_sparse_step (`int`, *optional*, defaults to 1):
+            Frequency of MoE layers (1 = every layer is MoE after mlp_only_layers).
+        mlp_only_layers (`list[int]`, *optional*, defaults to `[0]`):
+            Layer indices that use dense MLP instead of MoE.
+        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
+            Auxiliary loss coefficient for load balancing.
+        rope_parameters (`RopeParameters`, *optional*):
+            RoPE configuration. Defaults to rope_theta=500000.0.
     """
 
     model_type = "laguna"
@@ -57,19 +93,11 @@ class LagunaConfig(PreTrainedConfig):
         "layers.*.self_attn.q_proj": "colwise",
         "layers.*.self_attn.k_proj": "colwise",
         "layers.*.self_attn.v_proj": "colwise",
-        "layers.*.self_attn.g_proj": "colwise",
+        "layers.*.self_attn.g_proj": "colwise",  # Laguna-specific gating projection
         "layers.*.self_attn.o_proj": "rowwise",
-        "layers.*.self_attn.q_norm": "replicated_with_grad_allreduce",
-        "layers.*.self_attn.k_norm": "replicated_with_grad_allreduce",
         "layers.*.mlp.gate_proj": "colwise",
         "layers.*.mlp.up_proj": "colwise",
         "layers.*.mlp.down_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",
     }
     base_model_pp_plan = {
         "embed_tokens": (["input_ids"], ["inputs_embeds"]),
@@ -77,137 +105,83 @@ class LagunaConfig(PreTrainedConfig):
         "norm": (["hidden_states"], ["hidden_states"]),
     }
 
-    # Qwen2Moe-inherited defaults we want to override for Laguna's typical shape.
-    vocab_size: int = 100352
-    hidden_size: int = 2048
-    intermediate_size: int = 8192
-    num_hidden_layers: int = 40
-    num_attention_heads: int = 48
-    num_key_value_heads: int = 8
-    hidden_act: str = "silu"
-    max_position_embeddings: int = 131072
-    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 = None
-    attention_dropout: float | int = 0.0
-    moe_intermediate_size: int = 512
-    shared_expert_intermediate_size: int = 512
-    num_experts_per_tok: int = 8
-    num_experts: int = 256
-    output_router_logits: bool = False
-    router_aux_loss_coef: float = 0.001
-    layer_types: list[str] | None = None
-    pad_token_id: int | None = None
-    bos_token_id: int | None = None
-    eos_token_id: int | list[int] | None = None
-
-    # Laguna-specific attention
-    head_dim: int = 128
-    attention_bias: bool = False
-    partial_rotary_factor: float | None = None
-    num_attention_heads_per_layer: list[int] | None = None
-    # Laguna-specific MoE
-    mlp_layer_types: list[str] | None = None
-    moe_routed_scaling_factor: float = 1.0
-    moe_apply_router_weight_on_input: bool = False
-    moe_router_logit_softcapping: float = 0.0
-
-    def __post_init__(self, **kwargs):
-        if self.layer_types is None:
-            self.layer_types = ["full_attention"] * self.num_hidden_layers
-        if self.mlp_layer_types is None:
-            self.mlp_layer_types = ["dense"] + ["sparse"] * (self.num_hidden_layers - 1)
-        if self.num_attention_heads_per_layer is None:
-            self.num_attention_heads_per_layer = [self.num_attention_heads] * self.num_hidden_layers
-
-        default_rope_params: dict[Literal["full_attention", "sliding_attention"], dict[str, Any]] = {
-            "full_attention": {"rope_type": "default", "rope_theta": 500000.0},
-            "sliding_attention": {"rope_type": "default", "rope_theta": 10000.0},
-        }
-        if self.rope_parameters is None:
-            self.rope_parameters = default_rope_params
-
-        self._normalize_rope_parameters()
-        # Skip ``Qwen2MoeConfig.__post_init__`` — it references ``mlp_only_layers`` /
-        # ``use_sliding_window`` / ``max_window_layers`` which Laguna drops above.
-        super().__post_init__(**kwargs)
-
-    def _normalize_rope_parameters(self):
-        """Coerce ``rope_parameters`` to the nested ``{layer_type: {...}}`` shape.
-
-        Accepts an already-nested dict as-is, or a flat dict that gets broadcast to every
-        layer type. A top-level ``partial_rotary_factor`` is folded into each sub-dict as
-        a default.
-        """
-        layer_types = set(self.layer_types)
-        rope_params = self.rope_parameters or {}
-        is_nested = isinstance(rope_params, dict) and any(k in layer_types for k in rope_params)
-        if is_nested:
-            nested = {lt: dict(rope_params.get(lt, {})) for lt in layer_types}
-        else:
-            nested = {lt: dict(rope_params) for lt in layer_types}
-
-        if self.partial_rotary_factor is not None:
-            for params in nested.values():
-                params.setdefault("partial_rotary_factor", self.partial_rotary_factor)
-
-        for params in nested.values():
-            params.setdefault("rope_type", "default")
-
-        self.rope_parameters = nested
-        # Null the top-level field now that its value lives in each sub-dict — otherwise
-        # ``standardize_rope_params`` would overwrite per-type values with the global one.
-        self.partial_rotary_factor = None
-
-    def convert_rope_params_to_dict(self, **kwargs):
-        # No need to handle BC for new models, because they have no old-format `rope_scaling`
-        return kwargs
-
-    def _validate_yarn_rope_parameters(self, rope_parameters: dict, ignore_keys=None):
-        """Override: parent reads ``self.rope_parameters["original_max_position_embeddings"]``
-        for its post-hoc factor sanity-check, which works for flat rope configs but raises
-        ``KeyError`` when ``self.rope_parameters`` is the Laguna/Gemma3-style per-layer-type
-        map (its keys are layer types like ``"full_attention"``). Fix locally by reading
-        from the per-call ``rope_parameters`` dict that ``validate_rope`` already passes in.
-        """
-        # Delegate to parent for the shared checks by temporarily swapping in a flat
-        # ``self.rope_parameters`` that has the key the parent expects. Cheapest way to
-        # share the parent's logic without reimplementing it here.
-        flat = getattr(self, "rope_parameters", None)
+    def __init__(
+        self,
+        vocab_size: int = 100352,
+        hidden_size: int = 2048,
+        intermediate_size: int = 8192,
+        num_hidden_layers: int = 48,
+        num_attention_heads: int = 32,
+        num_key_value_heads: int = 8,
+        head_dim: int = 128,
+        qkv_bias: bool = False,
+        attention_bias: bool = False,
+        gating: bool = True,
+        hidden_act: str = "silu",
+        max_position_embeddings: int = 4096,
+        initializer_range: float = 0.02,
+        rms_norm_eps: float = 1e-6,
+        use_cache: bool = True,
+        tie_word_embeddings: bool = False,
+        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
+        attention_dropout: float = 0.0,
+        sliding_window: int | None = None,
+        layer_types: list[str] | None = None,
+        swa_attention_sink_enabled: bool = False,
+        swa_rope_parameters: RopeParameters | None = None,
+        num_experts: int = 256,
+        num_experts_per_tok: int = 16,
+        moe_intermediate_size: int = 1024,
+        shared_expert_intermediate_size: int = 1024,
+        norm_topk_prob: bool = True,
+        decoder_sparse_step: int = 1,
+        mlp_only_layers: list[int] | None = None,
+        router_aux_loss_coef: float = 0.001,
+        output_router_logits: bool = False,
+        **kwargs,
+    ):
+        # Default mlp_only_layers: first layer is dense (moe_first_k_dense_replace=1)
+        if mlp_only_layers is None:
+            mlp_only_layers = [0]
+
+        # Default rope_parameters with Laguna's theta
+        if rope_parameters is None:
+            rope_parameters = {"rope_type": "default", "rope_theta": 500000.0}
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.qkv_bias = qkv_bias
+        self.attention_bias = attention_bias
+        self.gating = gating
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
         self.rope_parameters = rope_parameters
-        try:
-            super()._validate_yarn_rope_parameters(rope_parameters, ignore_keys=ignore_keys)
-        finally:
-            self.rope_parameters = flat
-
-    def validate_architecture(self):
-        """Part of ``@strict``-powered validation."""
-        if self.moe_apply_router_weight_on_input:
-            raise NotImplementedError(
-                "moe_apply_router_weight_on_input=True is not yet supported in the "
-                "transformers implementation of Laguna."
-            )
-        if (
-            self.num_attention_heads_per_layer is not None
-            and len(self.num_attention_heads_per_layer) != self.num_hidden_layers
-        ):
-            raise ValueError(
-                f"num_attention_heads_per_layer length ({len(self.num_attention_heads_per_layer)}) "
-                f"must equal num_hidden_layers ({self.num_hidden_layers})."
-            )
-        if len(self.layer_types) != self.num_hidden_layers:
-            raise ValueError(
-                f"layer_types length ({len(self.layer_types)}) "
-                f"must equal num_hidden_layers ({self.num_hidden_layers})."
-            )
-        if len(self.mlp_layer_types) != self.num_hidden_layers:
-            raise ValueError(
-                f"mlp_layer_types length ({len(self.mlp_layer_types)}) "
-                f"must equal num_hidden_layers ({self.num_hidden_layers})."
-            )
+        self.attention_dropout = attention_dropout
+        # Sliding window attention arguments
+        self.sliding_window = sliding_window
+        self.layer_types = layer_types
+        self.swa_attention_sink_enabled = swa_attention_sink_enabled
+        self.swa_rope_parameters = swa_rope_parameters
+        # MoE arguments
+        self.num_experts = num_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_intermediate_size = moe_intermediate_size
+        self.shared_expert_intermediate_size = shared_expert_intermediate_size
+        self.norm_topk_prob = norm_topk_prob
+        self.decoder_sparse_step = decoder_sparse_step
+        self.mlp_only_layers = mlp_only_layers
+        self.router_aux_loss_coef = router_aux_loss_coef
+        self.output_router_logits = output_router_logits
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
 
 
 __all__ = ["LagunaConfig"]

modeling_laguna.py

@@ -1,4 +1,5 @@
-# Copyright 2026 Poolside and the HuggingFace Inc. team. All rights reserved.
+# ruff: noqa
+# Copyright 2025 Poolside and 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.
@@ -12,34 +13,37 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from collections.abc import Callable
 from typing import Optional
+from collections.abc import Callable
 
 import torch
 import torch.nn.functional as F
 from torch import nn
-
 from transformers import initialization as init
+from transformers.utils import auto_docstring, can_return_tuple, is_grouped_mm_available
+from transformers.generation import GenerationMixin
 from transformers.activations import ACT2FN
 from transformers.cache_utils import Cache, DynamicCache
-from transformers.generation import GenerationMixin
-from transformers.integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernelized_func
-from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
-from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
-from transformers.modeling_layers import GradientCheckpointingLayer
-from transformers.modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
-from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.integrations import (
+    use_kernelized_func,
+    use_kernel_func_from_hub,
+    use_kernel_forward_from_hub,
+)
+from transformers.masking_utils import create_causal_mask
+from transformers.utils.generic import OutputRecorder, TransformersKwargs, maybe_autocast, check_model_inputs
 from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import MoeModelOutputWithPast, MoeCausalLMOutputWithPast
 from transformers.processing_utils import Unpack
-from transformers.utils import auto_docstring, can_return_tuple
-from transformers.utils.generic import TransformersKwargs, maybe_autocast
-from transformers.utils.output_capturing import OutputRecorder, capture_outputs
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+
 from .configuration_laguna import LagunaConfig
 
 
 @use_kernel_forward_from_hub("RMSNorm")
 class LagunaRMSNorm(nn.Module):
-    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
+    def __init__(self, hidden_size, eps=1e-6):
         """
         LagunaRMSNorm is equivalent to T5LayerNorm
         """
@@ -47,7 +51,7 @@ class LagunaRMSNorm(nn.Module):
         self.weight = nn.Parameter(torch.ones(hidden_size))
         self.variance_epsilon = eps
 
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+    def forward(self, hidden_states):
         input_dtype = hidden_states.dtype
         hidden_states = hidden_states.to(torch.float32)
         variance = hidden_states.pow(2).mean(-1, keepdim=True)
@@ -61,35 +65,27 @@ class LagunaRMSNorm(nn.Module):
 class LagunaRotaryEmbedding(nn.Module):
     inv_freq: torch.Tensor  # fix linting for `register_buffer`
 
-    def __init__(self, config: LagunaConfig, device=None, layer_type=None):
+    def __init__(self, config: LagunaConfig, device=None):
         super().__init__()
         self.max_seq_len_cached = config.max_position_embeddings
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
 
-        self.layer_types = list(set(config.layer_types))
-        self.rope_type = {}
-        for layer_type in self.layer_types:
-            rope_params = self.config.rope_parameters[layer_type]
-            if rope_params is None:
-                continue
+        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, device)
 
-            self.rope_type[layer_type] = rope_params["rope_type"]
-            rope_init_fn: Callable = self.compute_default_rope_parameters
-            if self.rope_type[layer_type] != "default":
-                rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type[layer_type]]
-            curr_inv_freq, curr_attention_scaling = rope_init_fn(self.config, device, layer_type=layer_type)
-            self.register_buffer(f"{layer_type}_inv_freq", curr_inv_freq, persistent=False)
-            self.register_buffer(f"{layer_type}_original_inv_freq", curr_inv_freq.clone(), persistent=False)
-            setattr(self, f"{layer_type}_attention_scaling", curr_attention_scaling)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
         config: LagunaConfig | None = None,
         device: Optional["torch.device"] = None,
         seq_len: int | None = None,
-        layer_type: str | None = None,
     ) -> tuple["torch.Tensor", float]:
         """
         Computes the inverse frequencies according to the original RoPE implementation
@@ -100,18 +96,14 @@ class LagunaRotaryEmbedding(nn.Module):
                 The device to use for initialization of the inverse frequencies.
             seq_len (`int`, *optional*):
                 The current sequence length. Unused for this type of RoPE.
-            layer_type (`str`, *optional*):
-                The current layer type if the model has different RoPE parameters per type.
-                Should not be used unless `config.layer_types is not None`
-        Returns:
+
+        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[layer_type]["rope_theta"]
-        # key difference to gemma3: partial rope
-        partial_rotary_factor = config.rope_parameters[layer_type].get("partial_rotary_factor", 1.0)
-        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
-        dim = int(head_dim * partial_rotary_factor)
+        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
 
@@ -123,19 +115,16 @@ class LagunaRotaryEmbedding(nn.Module):
 
     @torch.no_grad()
     @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
-    def forward(self, x, position_ids, layer_type=None):
-        inv_freq = getattr(self, f"{layer_type}_inv_freq")
-        attention_scaling = getattr(self, f"{layer_type}_attention_scaling")
-
-        inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
-            cos = emb.cos() * attention_scaling
-            sin = emb.sin() * attention_scaling
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
 
         return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
 
@@ -157,97 +146,71 @@ class LagunaMLP(nn.Module):
 
 
 class LagunaTopKRouter(nn.Module):
+    """Laguna MoE router using sigmoid scoring (not softmax)."""
+
     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.Parameter(torch.zeros(config.num_experts), requires_grad=False)
-        self.router_logit_softcapping = config.moe_router_logit_softcapping
 
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
         hidden_states = hidden_states.reshape(-1, self.hidden_dim)
-        router_logits = F.linear(hidden_states, self.weight).float()
-        # Optional logits softcapping
-        if self.router_logit_softcapping > 0.0:
-            router_logits = torch.tanh(router_logits / self.router_logit_softcapping) * self.router_logit_softcapping
-        # Sigmoid instead of softmax normalization
-        routing_scores = torch.sigmoid(router_logits)
-
-        scores_for_selection = routing_scores + self.e_score_correction_bias.to(routing_scores.dtype)
-        _, selected_experts = torch.topk(scores_for_selection, self.top_k, dim=-1)
-        routing_weights = routing_scores.gather(-1, selected_experts)
-        routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
+        router_logits = F.linear(hidden_states, self.weight)
+        # Laguna-specific: sigmoid routing in float32 for precision
+        routing_weights = torch.sigmoid(router_logits.float())
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        if self.norm_topk_prob:
+            routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
         routing_weights = routing_weights.to(hidden_states.dtype)
-
         return router_logits, routing_weights, selected_experts
 
 
-@use_experts_implementation
-class LagunaExperts(nn.Module):
-    """Collection of expert weights stored as 3D tensors."""
+class LagunaSparseMoeBlock(nn.Module):
+    """Laguna MoE block using sigmoid router, per-expert MLPs, and a shared expert."""
 
     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)
-            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 LagunaSparseMoeBlock(nn.Module):
-    def __init__(self, config: LagunaConfig):
-        super().__init__()
-        self.experts = LagunaExperts(config)
+        self.top_k = config.num_experts_per_tok
         self.gate = LagunaTopKRouter(config)
-        self.shared_experts = LagunaMLP(config, intermediate_size=config.shared_expert_intermediate_size)
-        self.routed_scaling_factor = config.moe_routed_scaling_factor
+        self.experts = nn.ModuleList(
+            [LagunaMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
+        )
+        self.shared_expert = LagunaMLP(config, intermediate_size=config.shared_expert_intermediate_size)
+        self.shared_expert_gate = (
+            nn.Linear(config.hidden_size, 1, bias=False) if getattr(config, "moe_shared_gate", False) else None
+        )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         batch_size, sequence_length, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
-        shared_output = self.shared_experts(hidden_states)
 
+        shared_expert_output = self.shared_expert(hidden_states)
+        if self.shared_expert_gate is not None:
+            shared_expert_output = shared_expert_output * torch.sigmoid(self.shared_expert_gate(hidden_states))
+
+        # Routed experts
         _, routing_weights, selected_experts = self.gate(hidden_states)
-        hidden_states = self.experts(hidden_states, selected_experts, routing_weights)
-        # Additional scaling
-        hidden_states = hidden_states * self.routed_scaling_factor
-        hidden_states = hidden_states + shared_output
+        final_hidden_states = torch.zeros_like(hidden_states)
 
-        hidden_states = hidden_states.reshape(batch_size, sequence_length, hidden_dim)
-        return hidden_states
+        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts)
+        expert_mask = expert_mask.permute(2, 1, 0)
+
+        for expert_idx in range(self.num_experts):
+            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
+            if token_idx.shape[0] == 0:
+                continue
+            current_state = hidden_states[token_idx]
+            current_hidden_states = self.experts[expert_idx](current_state)
+            current_hidden_states = current_hidden_states * routing_weights[token_idx, top_k_pos, None]
+            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
+
+        final_hidden_states = final_hidden_states + shared_expert_output
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states
 
 
 def rotate_half(x):
@@ -257,12 +220,10 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-# Adapted from transformers.models.glm.modular_glm.apply_rotary_pos_emb
+@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.
 
-    Removes the interleaving of cos and sin from GLM
-
     Args:
         q (`torch.Tensor`): The query tensor.
         k (`torch.Tensor`): The key tensor.
@@ -275,24 +236,15 @@ def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
             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:
+
+    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)
-
-    # Keep half or full tensor for later concatenation
-    rotary_dim = cos.shape[-1]
-    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
-    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
-
-    # Apply rotary embeddings on the first half or full tensor
-    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
-    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)
-
-    # Concatenate back to full shape
-    q_embed = torch.cat([q_embed, q_pass], dim=-1)
-    k_embed = torch.cat([k_embed, k_pass], dim=-1)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
     return q_embed, k_embed
 
 
@@ -323,7 +275,8 @@ def eager_attention_forward(
 
     attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
     if attention_mask is not None:
-        attn_weights = attn_weights + attention_mask
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
 
     attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
@@ -333,39 +286,33 @@ def eager_attention_forward(
     return attn_output, attn_weights
 
 
+# Laguna attention is identical to Qwen2MoE attention except:
+# - No QKV bias
+# - Explicit head_dim from config
+# - Output gating: attn_output = attn_output * softplus(g_proj(hidden_states))
+# - No sliding window (full attention only)
 @use_kernelized_func(apply_rotary_pos_emb)
 class LagunaAttention(nn.Module):
-    """Afmoe-style SWA/GQA attention with Laguna-specific gating and per-layer head count."""
-
-    def __init__(self, config: LagunaConfig, layer_idx: int, num_heads: int):
+    def __init__(self, config: LagunaConfig, layer_idx: int):
         super().__init__()
-        # Number of heads is controlled via `config.num_attention_heads_per_layer` which is passed from the parent for the specific layer
-        self.num_heads = num_heads
         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 = self.num_heads // config.num_key_value_heads
+        self.head_dim = config.head_dim
+        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
 
-        # Per-layer head count: rebuild q_proj and o_proj using self.num_heads (parent uses config.num_attention_heads).
-        self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
-        self.k_proj = nn.Linear(
-            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.v_proj = nn.Linear(
-            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.o_proj = nn.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=config.attention_bias)
-        # Parent LlamaAttention already sets: layer_idx, num_heads, num_key_value_heads, num_key_value_groups, head_dim
-        # We only add Laguna-specific attributes
-        self.is_local_attention = config.layer_types[layer_idx] == "sliding_attention"
-        self.sliding_window = config.sliding_window if self.is_local_attention else None
-
-        self.q_norm = LagunaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
-        self.k_norm = LagunaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
-        self.g_proj = nn.Linear(config.hidden_size, self.num_heads, bias=False)
+        # Laguna: no QKV bias, explicit head_dim
+        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * config.head_dim, bias=False)
+        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * config.head_dim, bias=False)
+        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * config.head_dim, bias=False)
+        self.o_proj = nn.Linear(config.num_attention_heads * config.head_dim, config.hidden_size, bias=False)
+        # Laguna-specific: gating projection
+        self.g_proj = nn.Linear(config.hidden_size, config.num_attention_heads * config.head_dim, bias=False)
+        # QK normalization (RMSNorm applied per-head after reshape, before RoPE)
+        self.q_norm = LagunaRMSNorm(config.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = LagunaRMSNorm(config.head_dim, eps=config.rms_norm_eps)
 
     def forward(
         self,
@@ -373,28 +320,36 @@ class LagunaAttention(nn.Module):
         position_embeddings: tuple[torch.Tensor, torch.Tensor],
         attention_mask: torch.Tensor | None,
         past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
         **kwargs: Unpack[FlashAttentionKwargs],
     ) -> 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)
-        key_states = self.k_proj(hidden_states).view(hidden_shape)
-        value_states = self.v_proj(hidden_states).view(hidden_shape)
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(hidden_shape).transpose(1, 2)
+        key_states = key_states.view(hidden_shape).transpose(1, 2)
+        value_states = value_states.view(hidden_shape).transpose(1, 2)
 
-        query_states = self.q_norm(query_states).transpose(1, 2)
-        key_states = self.k_norm(key_states).transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
+        # QK normalization (applied per-head before RoPE)
+        query_states = self.q_norm(query_states)
+        key_states = self.k_norm(key_states)
 
         cos, sin = position_embeddings
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
         if past_key_values is not None:
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
-        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
-            self.config._attn_implementation, eager_attention_forward
-        )
         attn_output, attn_weights = attention_interface(
             self,
             query_states,
@@ -403,30 +358,37 @@ class LagunaAttention(nn.Module):
             attention_mask,
             dropout=0.0 if not self.training else self.attention_dropout,
             scaling=self.scaling,
-            sliding_window=self.sliding_window,
             **kwargs,
         )
 
         attn_output = attn_output.reshape(*input_shape, -1).contiguous()
 
+        # Laguna-specific: apply gating BEFORE o_proj
+        # gate values are computed from original hidden_states, applied in attention dimension
         gate = F.softplus(self.g_proj(hidden_states).float()).to(attn_output.dtype)
-        attn_output = (attn_output.view(*input_shape, -1, self.head_dim) * gate.unsqueeze(-1)).view(*input_shape, -1)
+        attn_output = attn_output * gate
 
         attn_output = self.o_proj(attn_output)
+
         return attn_output, attn_weights
 
 
 class LagunaDecoderLayer(GradientCheckpointingLayer):
+    """Laguna decoder layer with gated attention and sigmoid-routed MoE."""
+
     def __init__(self, config: LagunaConfig, layer_idx: int):
         super().__init__()
-        self.hidden_size = config.hidden_size
-        self.self_attn = LagunaAttention(config, layer_idx, config.num_attention_heads_per_layer[layer_idx])
-        if config.mlp_layer_types[layer_idx] == "sparse":
+        self.self_attn = LagunaAttention(config, layer_idx)
+        # Use MoE or dense MLP based on layer configuration
+        if (layer_idx not in config.mlp_only_layers) and (
+            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
+        ):
             self.mlp = LagunaSparseMoeBlock(config)
         else:
             self.mlp = LagunaMLP(config, intermediate_size=config.intermediate_size)
         self.input_layernorm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.post_attention_layernorm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.hidden_size = config.hidden_size
 
     def forward(
         self,
@@ -435,6 +397,7 @@ class LagunaDecoderLayer(GradientCheckpointingLayer):
         position_ids: torch.LongTensor | None = None,
         past_key_values: Cache | None = None,
         use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
         position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
         **kwargs: Unpack[TransformersKwargs],
     ) -> torch.Tensor:
@@ -447,6 +410,7 @@ class LagunaDecoderLayer(GradientCheckpointingLayer):
             position_ids=position_ids,
             past_key_values=past_key_values,
             use_cache=use_cache,
+            cache_position=cache_position,
             position_embeddings=position_embeddings,
             **kwargs,
         )
@@ -470,8 +434,9 @@ class LagunaPreTrainedModel(PreTrainedModel):
     _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
-
-    _can_compile_fullgraph = True
+    _can_compile_fullgraph = (
+        is_grouped_mm_available()
+    )  # https://huggingface.co/docs/transformers/experts_interface#torchcompile
     _supports_attention_backend = True
     _can_record_outputs = {
         "router_logits": OutputRecorder(LagunaTopKRouter, index=0),
@@ -483,24 +448,10 @@ class LagunaPreTrainedModel(PreTrainedModel):
     def _init_weights(self, module):
         super()._init_weights(module)
         std = self.config.initializer_range
-        if isinstance(module, LagunaExperts):
-            init.normal_(module.gate_up_proj, mean=0.0, std=std)
-            init.normal_(module.down_proj, mean=0.0, std=std)
-        elif isinstance(module, LagunaTopKRouter):
-            init.normal_(module.weight, mean=0.0, std=std)
         if isinstance(module, LagunaTopKRouter):
-            torch.nn.init.zeros_(module.e_score_correction_bias)
-        elif isinstance(module, LagunaRotaryEmbedding):
-            for layer_type in module.layer_types:
-                rope_init_fn = module.compute_default_rope_parameters
-                if module.rope_type[layer_type] != "default":
-                    rope_init_fn = ROPE_INIT_FUNCTIONS[module.rope_type[layer_type]]
-                curr_inv_freq, _ = rope_init_fn(module.config, layer_type=layer_type)
-                init.copy_(getattr(module, f"{layer_type}_inv_freq"), curr_inv_freq)
-                init.copy_(getattr(module, f"{layer_type}_original_inv_freq"), curr_inv_freq)
+            init.normal_(module.weight, mean=0.0, std=std)
 
 
-@auto_docstring
 class LagunaModel(LagunaPreTrainedModel):
     def __init__(self, config: LagunaConfig):
         super().__init__(config)
@@ -518,8 +469,7 @@ class LagunaModel(LagunaPreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    @capture_outputs
-    @auto_docstring
+    @check_model_inputs
     def forward(
         self,
         input_ids: torch.LongTensor | None = None,
@@ -528,50 +478,49 @@ class LagunaModel(LagunaPreTrainedModel):
         past_key_values: Cache | None = None,
         inputs_embeds: torch.FloatTensor | None = None,
         use_cache: bool | None = None,
+        cache_position: torch.LongTensor | 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 use_cache and past_key_values is None:
+            past_key_values = DynamicCache(config=self.config)
+
         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 cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
 
         if position_ids is None:
-            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)
-
-        if not isinstance(causal_mask_mapping := attention_mask, dict):
-            mask_kwargs = {
-                "config": self.config,
-                "inputs_embeds": inputs_embeds,
-                "attention_mask": attention_mask,
-                "past_key_values": past_key_values,
-                "position_ids": position_ids,
-            }
-            mask_creation_functions = {
-                "full_attention": lambda: create_causal_mask(**mask_kwargs),
-                "sliding_attention": lambda: create_sliding_window_causal_mask(**mask_kwargs),
-            }
-            causal_mask_mapping = {}
-            for layer_type in set(self.config.layer_types):
-                causal_mask_mapping[layer_type] = mask_creation_functions[layer_type]()
+            position_ids = cache_position.unsqueeze(0)
+
+        # Laguna uses full attention only (no sliding window)
+        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,
+        )
 
         hidden_states = inputs_embeds
-        position_embeddings = {}
-        for layer_type in set(self.config.layer_types):
-            position_embeddings[layer_type] = self.rotary_emb(hidden_states, position_ids, layer_type)
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
 
-        for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
             hidden_states = decoder_layer(
                 hidden_states,
-                attention_mask=causal_mask_mapping[self.config.layer_types[i]],
-                position_embeddings=position_embeddings[self.config.layer_types[i]],
+                attention_mask=causal_mask,
                 position_ids=position_ids,
                 past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
                 **kwargs,
             )
 
@@ -579,7 +528,7 @@ class LagunaModel(LagunaPreTrainedModel):
 
         return MoeModelOutputWithPast(
             last_hidden_state=hidden_states,
-            past_key_values=past_key_values if use_cache else None,
+            past_key_values=past_key_values,
         )
 
 
@@ -609,7 +558,8 @@ def load_balancing_loss_func(
             The attention_mask used in forward function
             shape [batch_size X sequence_length] if not None.
 
-    Returns:
+    Returns
+    -------
         The auxiliary loss.
     """
     if gate_logits is None or not isinstance(gate_logits, tuple):
@@ -668,7 +618,7 @@ def load_balancing_loss_func(
 @auto_docstring
 class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
     _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
-    _tp_plan = {"lm_head": "colwise_gather_output"}
+    _tp_plan = {"lm_head": "colwise_rep"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
 
     def __init__(self, config):
@@ -695,15 +645,17 @@ class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
         labels: torch.LongTensor | None = None,
         use_cache: bool | None = None,
         output_router_logits: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
         logits_to_keep: int | torch.Tensor = 0,
         **kwargs: Unpack[TransformersKwargs],
     ) -> MoeCausalLMOutputWithPast:
         r"""
-        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+        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]`.
         """
+        # TODO (Joe) add example here after we got rid of the stale mistral example
 
         output_router_logits = (
             output_router_logits if output_router_logits is not None else self.config.output_router_logits
@@ -718,6 +670,7 @@ class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
             inputs_embeds=inputs_embeds,
             use_cache=use_cache,
             output_router_logits=output_router_logits,
+            cache_position=cache_position,
             **kwargs,
         )
 
@@ -738,8 +691,8 @@ class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
                 self.num_experts_per_tok,
                 attention_mask,
             )
-            if labels is not None:
-                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+            if labels is not None and isinstance(aux_loss, torch.Tensor):
+                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)
 
         return MoeCausalLMOutputWithPast(
             loss=loss,