Add files using upload-large-folder tool

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,23 @@
+MIT License
+
+Copyright (c) 2025 Antgroup and The HuggingFace Inc. team (Ling-3.0-tiny / BailingMoeV3)
+Copyright (c) 2023-2026 Songlin Yang, Yu Zhang, Zhiyuan Li, et al. (flash-linear-attention references)
+Copyright (c) 2026 SecureLayer7 (abliterated, triton-free derivative)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

NOTICE

@@ -0,0 +1,36 @@
+Ling-3.0-tiny-Uncensored-Abliterated
+Copyright (c) 2026 SecureLayer7 (Waxspace)
+
+This is a derivative work distributed under the MIT License.
+
+------------------------------------------------------------------------
+Attribution (per the MIT License):
+------------------------------------------------------------------------
+
+Base model:
+  inclusionAI/Ling-3.0-tiny  (architecture: BailingMoeV3)
+  https://huggingface.co/inclusionAI/Ling-3.0-tiny
+  Copyright (c) 2025 Antgroup and The HuggingFace Inc. team.
+  Licensed under the MIT License.
+
+Ported linear-attention math referenced from:
+  flash-linear-attention (fla) naive torch reference implementations
+  Copyright (c) 2023-2026 Songlin Yang, Yu Zhang, Zhiyuan Li, et al.
+  Licensed under the MIT License.
+
+------------------------------------------------------------------------
+Modifications made in this derivative:
+------------------------------------------------------------------------
+
+  1. Triton-free port: modeling_bailing_moe_v3.py was modified to run without
+     the `fla` / Triton dependency (KDA linear-attention recurrence, gated
+     RMSNorm, and short causal convolution reimplemented in pure PyTorch from
+     fla's MIT-licensed naive references), so the model runs on Apple Silicon
+     (MPS) and CPU. Compatibility fixes for transformers 5.x were also applied.
+
+  2. Abliteration: the refusal direction was ablated (Heretic / Optuna TPE) from
+     the attention output projections (MLA o_proj and KDA dense) and all MoE
+     expert down-projections, reducing weights-level refusals from ~35/100 to
+     ~8/100 at KL ~0.046.
+
+No trademark of Antgroup, inclusionAI, or Hugging Face is used to imply endorsement.

README.md

@@ -0,0 +1,91 @@
+---
+license: mit
+base_model: inclusionAI/Ling-3.0-tiny
+language:
+  - en
+library_name: transformers
+pipeline_tag: text-generation
+tags:
+  - ling
+  - bailing-moe
+  - uncensored
+  - abliterated
+  - uncensored-llm
+  - no-refusal
+  - moe
+  - mixture-of-experts
+  - linear-attention
+  - apple-silicon
+  - mps
+  - reasoning
+  - cybersecurity
+  - red-teaming
+---
+
+# Ling-3.0-tiny-Uncensored-Abliterated
+
+An uncensored, **abliterated** derivative of [`inclusionAI/Ling-3.0-tiny`](https://huggingface.co/inclusionAI/Ling-3.0-tiny)
+(BailingMoeV3) — the refusal direction removed at the **weights level** for direct, complete
+answers on cybersecurity, red-teaming, and penetration-testing topics where aligned models refuse.
+
+## What makes this different
+
+**It runs on Apple Silicon (MPS) — the original can't.** Ling-3.0-tiny's KDA linear-attention
+requires `fla` / Triton kernels, which have **no Apple-Silicon backend**. This repo ships a
+**triton-free pure-torch port** of the BailingMoeV3 modeling code (KDA recurrence, gated RMSNorm,
+short causal convolution) so the model loads and generates on a Mac's GPU with plain
+`transformers` — no CUDA, no Triton, no `fla`. The abliteration itself was performed on an M4 Max
+using that port.
+
+- 🔓 **Weights-level uncensored** — refusal direction ablated (Heretic / Optuna TPE) across both
+  attention paths (MLA `o_proj` + KDA `dense`) **and all 128 experts + shared expert** per layer.
+  Refusals dropped **35/100 → 8/100** at **KL 0.046** (minimal capability change).
+- 🍎 **Apple-Silicon runnable** — triton-free modeling code included; loads on MPS out of the box.
+- 🧩 **MoE** — 7.9B total / 1.3B active (128 routed + 1 shared expert), 24 layers, hybrid
+  MLA + KDA linear attention.
+
+## Quick start
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+m = "Securelayer7/Ling-3.0-tiny-Uncensored-Abliterated"
+tok = AutoTokenizer.from_pretrained(m, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    m, torch_dtype=torch.bfloat16, trust_remote_code=True).to("mps").eval()
+
+msgs = [{"role": "user", "content": "Explain how a SQL injection works and how to prevent it."}]
+ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)
+ids = {k: v.to("mps") for k, v in ids.items()}
+out = model.generate(**ids, max_new_tokens=512, do_sample=True, temperature=0.7)
+print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))
+```
+
+The bundled `modeling_bailing_moe_v3.py` uses a pure-torch fallback for the KDA linear-attention
+(no Triton), so it also runs on CPU. On CUDA with `fla` installed you may prefer the original
+upstream modeling code for speed.
+
+## How it was made
+
+1. **Triton-free port** of BailingMoeV3 so it runs without `fla`/Triton (math from `fla`'s own MIT
+   naive references; identical weights).
+2. **Abliteration** (Heretic, Optuna TPE multi-objective: minimize refusals + KL) targeting the
+   residual-writing projections of both attention types and every expert down-projection.
+
+## Known behavior
+
+Ling is a **bilingual (English/Chinese)** model; after answering it may occasionally drift into
+Chinese. Recommended sampling: `do_sample=True, temperature=0.7, top_p=0.95`. Greedy decoding can
+degrade. A short SFT pass cleans up drift.
+
+## Responsible use
+
+Uncensored ≠ lawless — for **legitimate research and authorized security work**. Illegal content
+(incl. CSAM) must be blocked at the serving layer; the weights carry no such guard, and the
+operator is responsible for a lawful, policy-gated deployment.
+
+## License & attribution
+
+**MIT** — see `LICENSE`. Derivative of **[inclusionAI/Ling-3.0-tiny](https://huggingface.co/inclusionAI/Ling-3.0-tiny)**
+(BailingMoeV3, © Antgroup, MIT). Modifications (triton-free port + abliteration) disclosed in `NOTICE`.

chat_template.jinja

@@ -0,0 +1,130 @@
+{#- Bailing V3 chat template -#}
+{#- Supports: thinking option, tool calling -#}
+
+{#- ==================== thinking option normalization ==================== -#}
+{%- if enable_thinking is defined %}
+{%- if enable_thinking %}
+{%- set thinking_option = 'on' %}
+{%- else %}
+{%- set thinking_option = 'off' %}
+{%- endif %}
+{%- elif thinking_option is not defined %}
+{%- set thinking_option = 'on' %}
+{%- endif %}
+
+{#- ==================== preserved thinking  ==================== -#}
+{% set preserved_thinking = true %}
+
+{#- ==================== system message ==================== -#}
+{{- '<role>SYSTEM</role>' }}
+{%- if tools %}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nIf none of the functions can be used, point it out. If the given question lacks the parameters required by the function, also point it out.\nIf you need to use a function, for each function call, output the function name and arguments within the following XML format:\n<tool_call>{function-name}\n<arg_key>{arg-key-1}</arg_key>\n<arg_value>{arg-value-1}</arg_value>\n<arg_key>{arg-key-2}</arg_key>\n<arg_value>{arg-value-2}</arg_value>\n...\n</tool_call>\n" }}
+    {%- if messages[0].role == 'system' and messages[0].content is string and ('detailed thinking on' in messages[0].content or 'detailed thinking off' in messages[0].content) %}
+        {{- '<|role_end|>' }}
+    {%- else %}
+        {{- 'detailed thinking ' + thinking_option + '<|role_end|>' }}
+    {%- endif %}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- if 'detailed thinking on' in messages[0].content or 'detailed thinking off' in messages[0].content %}
+            {{- messages[0].content + '<|role_end|>' }}
+        {%- else %}
+            {{- messages[0].content + '\n' }}
+            {{- 'detailed thinking ' + thinking_option + '<|role_end|>' }}
+        {%- endif %}
+    {% else %}
+        {{- 'detailed thinking ' + thinking_option + '<|role_end|>' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if message.role == "user" %}
+        {{- '<role>HUMAN</role>' + message.content + '<|role_end|>' }}
+    {%- elif message.role == "system" and not loop.first %}
+        {{- '<role>SYSTEM</role>' + message.content + '<|role_end|>' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string and message.reasoning_content != '' %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if preserved_thinking or loop.index0 > ns.last_query_index %}
+            {%- if reasoning_content != '' %}
+                {{- '<role>ASSISTANT</role>' + '\n<think>' + reasoning_content.strip('\n') + '</think>' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<role>ASSISTANT</role>\n<think></think>' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<role>ASSISTANT</role>\n<think></think>' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- set tc = tool_call %}
+                {%- if tool_call.function %}
+                    {%- set tc = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>' + tc.name }}
+                {% set _args = tc.arguments %}
+                {%- for k, v in _args.items() %}
+                    {{- '<arg_key>' + k + '</arg_key>' }}
+                    {{- '\n<arg_value>' }}
+                    {%- if v is string %}
+                        {{- v }}
+                    {%- else %}
+                        {{- v | tojson(ensure_ascii=False) }}
+                    {%- endif %}
+                    {{- '</arg_value>' }}
+                {%- endfor %}
+                {{- '\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|role_end|>' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<role>OBSERVATION</role>' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|role_end|>' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+
+{#- ==================== generation prompt ==================== -#}
+{%- if add_generation_prompt %}
+    {{- '<role>ASSISTANT</role>' }}
+    {%- if thinking_option == 'on' %}
+        {{- '\n<think>' }}
+    {%- elif thinking_option == 'off' %}
+        {{- '\n<think></think>' }}
+    {%- endif %}
+{%- endif %}

config.json

@@ -0,0 +1,86 @@
+{
+  "architectures": [
+    "BailingMoeV3ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_bailing_moe_v3.BailingMoeV3Config",
+    "AutoModel": "modeling_bailing_moe_v3.BailingMoeV3Model",
+    "AutoModelForCausalLM": "modeling_bailing_moe_v3.BailingMoeV3ForCausalLM"
+  },
+  "dtype": "bfloat16",
+  "embedding_dropout": 0.0,
+  "eos_token_id": 156895,
+  "expert_swiglu_limit_list": null,
+  "first_k_dense_replace": 1,
+  "gated_attention_proj_granularity_type": "head_wise",
+  "group_norm_size": 1,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 4608,
+  "kda_lower_bound": -5,
+  "kda_safe_gate": true,
+  "kv_lora_rank": 512,
+  "layer_group_size": 4,
+  "linear_silu": true,
+  "max_position_embeddings": 131072,
+  "max_window_layers": 20,
+  "moe_intermediate_size": 512,
+  "moe_router_enable_expert_bias": true,
+  "moe_shared_expert_intermediate_size": 512,
+  "mtp_loss_scaling_factor": 0,
+  "mtp_use_kda": false,
+  "n_group": 8,
+  "no_kda_lora": true,
+  "norm_topk_prob": true,
+  "num_attention_heads": 16,
+  "num_experts": 128,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 16,
+  "num_kv_heads_for_linear_attn": 0,
+  "num_nextn_predict_layers": 0,
+  "num_shared_experts": 1,
+  "output_dropout": 0.0,
+  "output_router_logits": false,
+  "pad_token_id": 156892,
+  "partial_rotary_factor": 0.5,
+  "q_lora_rank": 256,
+  "qk_head_dim": 192,
+  "qk_nope_head_dim": 128,
+  "qk_rope_head_dim": 64,
+  "rms_norm_eps": 1e-06,
+  "rope_interleave": true,
+  "rope_parameters": {
+    "partial_rotary_factor": 0.5,
+    "rope_theta": 6000000,
+    "rope_type": "default"
+  },
+  "rope_theta": 6000000,
+  "rotary_dim": 64,
+  "routed_scaling_factor": 2.5,
+  "router_dtype": "fp32",
+  "scale_router_input": false,
+  "score_function": "sigmoid",
+  "scoring_func": "sigmoid",
+  "seq_aux": true,
+  "share_expert_swiglu_limit_list": null,
+  "short_conv_kernel_size": 4,
+  "tie_word_embeddings": false,
+  "topk_group": 4,
+  "topk_method": "noaux_tc",
+  "transformers_version": "5.14.1",
+  "up_proj_norm": false,
+  "use_bias": false,
+  "use_cache": true,
+  "use_kda_lora": false,
+  "use_mla_nope": false,
+  "use_nGPT": false,
+  "use_qk_norm": true,
+  "use_qkv_bias": false,
+  "v_head_dim": 128,
+  "value_norm": false,
+  "vocab_size": 157184
+}

configuration_bailing_moe_v3.py

@@ -0,0 +1,126 @@
+"""Bailing MoE V2 model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class BailingMoeV3Config(PretrainedConfig):
+
+    def __init__(
+        self,
+        vocab_size=157184,
+        hidden_size=2048,
+        intermediate_size=5120,
+        num_hidden_layers=20,
+        num_attention_heads=16,
+        num_key_value_heads=4,
+        hidden_act="silu",
+        use_qkv_bias=False,  # bailing only
+        use_bias=False,  # bailing only
+        rms_norm_eps=1e-06,
+        tie_word_embeddings=False,  # PretrainedConfig key, here change default value.
+        embedding_dropout=0.0,
+        attention_dropout=0.0,
+        output_dropout=0.0,
+        initializer_range=0.02,
+        max_position_embeddings=32768,
+        rope_theta=600000.0,
+        use_cache=True,
+        max_window_layers=20,
+        rope_scaling=None,
+        pad_token_id=156892,
+        eos_token_id=156892,
+        num_experts=256,
+        num_shared_experts=1,
+        num_experts_per_tok=8,
+        n_group=8,
+        topk_group=4,
+        moe_intermediate_size=512,
+        moe_shared_expert_intermediate_size=512,
+        first_k_dense_replace=1,
+        head_dim=128,
+        output_router_logits=False,
+        use_qk_norm=True,
+        num_nextn_predict_layers=0,
+        mtp_loss_scaling_factor=0,
+        moe_router_enable_expert_bias=True,
+        routed_scaling_factor=1.0,
+        layer_group_size=5,
+        kv_lora_rank=512,
+        q_lora_rank=None,
+        qk_rope_head_dim=64,
+        v_head_dim=128,
+        qk_nope_head_dim=128,
+        rope_interleave=True,
+        score_function="sigmoid",
+        scoring_func="sigmoid",
+        seq_aux=True,
+        topk_method="noaux_tc",
+        router_dtype="fp32",
+        gated_attention_proj_granularity_type=None,
+        no_kda_lora=False,
+        kda_safe_gate=False,
+        kda_lower_bound=None,
+        short_conv_kernel_size=4,
+        **kwargs,
+    ):
+        self.num_hidden_layers = num_hidden_layers
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.use_qkv_bias = use_qkv_bias
+        self.use_bias = use_bias
+        self.rms_norm_eps = rms_norm_eps
+        self.embedding_dropout = embedding_dropout
+        self.attention_dropout = attention_dropout
+        self.output_dropout = output_dropout
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+        self.mtp_loss_scaling_factor = mtp_loss_scaling_factor
+        self.initializer_range = initializer_range
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_theta = rope_theta
+        self.use_cache = use_cache
+        self.max_window_layers = max_window_layers
+        self.head_dim = head_dim or self.hidden_size // self.num_attention_heads
+        self.rope_scaling = rope_scaling
+        self.use_qk_norm = use_qk_norm
+        self.moe_router_enable_expert_bias = moe_router_enable_expert_bias
+        self.routed_scaling_factor = routed_scaling_factor
+
+        # MoE configs
+        self.num_experts = num_experts
+        self.num_shared_experts = num_shared_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.moe_intermediate_size = moe_intermediate_size
+        self.moe_shared_expert_intermediate_size = moe_shared_expert_intermediate_size
+        self.first_k_dense_replace = first_k_dense_replace
+        self.output_router_logits = output_router_logits
+
+        # Linear configs
+        self.layer_group_size = layer_group_size
+        # mla
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+
+        self.score_function = score_function
+        self.scoring_func = scoring_func
+        self.seq_aux = seq_aux
+        self.topk_method = topk_method
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
+        self.rope_interleave = rope_interleave
+        self.router_dtype = router_dtype
+        self.gated_attention_proj_granularity_type = gated_attention_proj_granularity_type
+        self.no_kda_lora = no_kda_lora
+        self.kda_safe_gate = kda_safe_gate
+        self.kda_lower_bound = kda_lower_bound
+        self.short_conv_kernel_size = short_conv_kernel_size
+        super().__init__(
+            pad_token_id=pad_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
+        )

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "do_sample": true,
+  "temperature": 1.0,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "5.14.1"
+}

model-00001-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7b59cd76fefb004697f2ef6c2ed3e7c11d1f06493eb3f2469e0ee1e319fe819
+size 4999827336

model-00002-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6465e1c721079c16a433e31df5932366e0b669963e5d14e9e60dc0f1c625c2d0
+size 4999087224

model-00003-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35d8a42887d2c3a36ca79c9875ced1c94189749dc86e1d04da78dcfa710d0afd
+size 4999088112

model-00004-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb6ed6d3073be9eb55d43c5c02c9f7d6b01587be77afd4125d351e679372ccfc
+size 790001720

modeling_bailing_moe_v3.py

@@ -0,0 +1,1782 @@
+# coding=utf-8
+# Copyright 2025 Antgroup and The HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch BailingMoE model."""
+
+import math
+import warnings
+from typing import List, Optional, Tuple, Union, Callable
+from copy import deepcopy
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
+    _prepare_4d_attention_mask,
+    _prepare_4d_causal_attention_mask,
+    _prepare_4d_causal_attention_mask_for_sdpa,
+)
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS, is_torch_greater_or_equal_than_1_13
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+try:  # moved/removed in transformers 5.x; only used for FX symbolic tracing (unused here)
+    from transformers.utils.import_utils import is_torch_fx_available
+except ImportError:
+    def is_torch_fx_available():
+        return False
+from .configuration_bailing_moe_v3 import BailingMoeV3Config
+from transformers.generation.utils import GenerationMixin
+from dataclasses import dataclass
+from transformers.utils import ModelOutput
+from transformers import DynamicLayer
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+
+from einops import rearrange, repeat
+
+# =====================================================================
+# Triton-free torch port (Apple Silicon / MPS). The upstream modeling
+# file hard-requires `fla` (flash-linear-attention) Triton kernels, which
+# have no Apple-Silicon backend. These are drop-in replacements with the
+# SAME names/signatures, math ported from fla's own naive torch references
+# (fla/ops/kda/naive.py, fla/ops/kda/gate.py, fla/modules). Same weights.
+# =====================================================================
+
+def tensor_cache(fn):  # identity decorator (fla.utils.tensor_cache is a memoizer)
+    return fn
+
+
+def _kda_lowerbound_gate(g, A_log, dt_bias, lower_bound):
+    # g = lower_bound * sigmoid(exp(A_log) * (g + dt_bias))   [fla naive_kda_lowerbound_gate]
+    H = g.shape[-2]
+    g = g.float()
+    if dt_bias is not None:
+        g = g + dt_bias.view(H, -1)
+    return lower_bound * torch.sigmoid(A_log.view(H, 1).float().exp() * g)
+
+
+def _kda_softplus_gate(g, A_log, dt_bias):
+    # g = -exp(A_log) * softplus(g + dt_bias)                 [fla naive_kda_gate]
+    H = g.shape[-2]
+    g = g.float()
+    if dt_bias is not None:
+        g = g + dt_bias.view(H, -1)
+    return -A_log.view(H, 1).float().exp() * F.softplus(g)
+
+
+def _naive_recurrent_kda(q, k, v, g, beta, scale=None, initial_state=None,
+                         output_final_state=False):
+    # Reference recurrence from fla/ops/kda/naive.py::naive_recurrent_kda.
+    dtype = v.dtype
+    B, T, H, K = q.shape
+    HV, V = v.shape[2], v.shape[-1]
+    G = HV // H
+    if scale is None:
+        scale = K ** -0.5
+    q, k, v, g, beta = (x.float() for x in (q, k, v, g, beta))
+    q = q.repeat_interleave(G, dim=2) * scale
+    k = k.repeat_interleave(G, dim=2)
+    S = k.new_zeros(B, HV, K, V)
+    if initial_state is not None:
+        S = S + initial_state.float()
+    o = torch.zeros_like(v)
+    for i in range(T):
+        q_i, k_i, v_i, g_i, b_i = q[:, i], k[:, i], v[:, i], g[:, i], beta[:, i]
+        S = S * g_i[..., None].exp()
+        S = S + torch.einsum('b h k, b h v -> b h k v',
+                             b_i[..., None] * k_i, v_i - (k_i[..., None] * S).sum(-2))
+        o[:, i] = torch.einsum('b h k, b h k v -> b h v', q_i, S)
+    if not output_final_state:
+        S = None
+    return o.to(dtype), S
+
+
+def _kda_prep(q, k, g, A_log, dt_bias, lower_bound, l2norm, gate):
+    if l2norm:
+        q = F.normalize(q.float(), dim=-1)
+        k = F.normalize(k.float(), dim=-1)
+    if gate:
+        g = (_kda_lowerbound_gate(g, A_log, dt_bias, lower_bound) if lower_bound is not None
+             else _kda_softplus_gate(g, A_log, dt_bias))
+    return q, k, g
+
+
+def chunk_kda(q, k, v, g, beta, A_log, dt_bias, initial_state=None,
+              output_final_state=False, use_qk_l2norm_in_kernel=False,
+              use_gate_in_kernel=False, safe_gate=True, lower_bound=None,
+              cu_seqlens=None, **kw):
+    # Route chunk mode through the exact recurrence (handles arbitrary T; the
+    # chunked reference requires T % 64 == 0 which breaks on real prompts).
+    q, k, g = _kda_prep(q, k, g, A_log, dt_bias, lower_bound,
+                        use_qk_l2norm_in_kernel, use_gate_in_kernel)
+    return _naive_recurrent_kda(q, k, v, g, beta, None, initial_state, output_final_state)
+
+
+def fused_recurrent_kda(q, k, v, g, beta, A_log, dt_bias, initial_state=None,
+                        output_final_state=False, use_qk_l2norm_in_kernel=False,
+                        use_gate_in_kernel=False, safe_gate=True, lower_bound=None,
+                        cu_seqlens=None, **kw):
+    q, k, g = _kda_prep(q, k, g, A_log, dt_bias, lower_bound,
+                        use_qk_l2norm_in_kernel, use_gate_in_kernel)
+    return _naive_recurrent_kda(q, k, v, g, beta, None, initial_state, output_final_state)
+
+
+def chunk_simple_gla(*a, **k):          # imported upstream but never called in this model
+    raise NotImplementedError("simple_gla is unused in Ling-3.0-tiny")
+
+
+def fused_recurrent_simple_gla(*a, **k):
+    raise NotImplementedError("simple_gla is unused in Ling-3.0-tiny")
+
+
+class ShortConvolution(nn.Conv1d):
+    """Depthwise causal conv1d + activation (torch port of fla ShortConvolution)."""
+
+    def __init__(self, hidden_size, kernel_size, bias=False, activation='silu',
+                 backend=None, device=None, dtype=None, **kw):
+        super().__init__(hidden_size, hidden_size, kernel_size, groups=hidden_size,
+                         bias=bias, padding=kernel_size - 1, device=device, dtype=dtype)
+        self.hidden_size = hidden_size
+        self.activation = activation
+
+    def _act(self, y):
+        return F.silu(y) if self.activation in ('silu', 'swish') else y
+
+    def forward(self, x, residual=None, mask=None, cache=None,
+                output_final_state=False, cu_seqlens=None, **kw):
+        B, T, D = x.shape
+        W = self.kernel_size[0]
+        if T == 1 and cache is not None:
+            # decode step: shift window, drop in new input, weighted sum
+            cache = torch.roll(cache, shifts=-1, dims=-1)
+            cache[:, :, -1] = x[:, 0]
+            y = (cache * self.weight.squeeze(1)).sum(-1)
+            if self.bias is not None:
+                y = y + self.bias
+            y = self._act(y).unsqueeze(1)
+            return y, (cache if output_final_state else None)
+        xt = x.transpose(1, 2)  # [B, D, T]
+        y = F.conv1d(xt, self.weight, self.bias, groups=self.hidden_size, padding=W - 1)[..., :T]
+        y = self._act(y.transpose(1, 2))
+        final = None
+        if output_final_state:
+            pad = xt if T >= W else F.pad(xt, (W - T, 0))
+            final = pad[..., -W:].contiguous()
+        return y, final
+
+
+class FusedRMSNormGated(nn.Module):
+    """RMSNorm with a multiplicative output gate (torch port of fla FusedRMSNormGated)."""
+
+    def __init__(self, hidden_size, eps=1e-6, activation='sigmoid', **kw):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.eps = eps
+        self.activation = activation
+
+    def forward(self, x, g):
+        dt = x.dtype
+        x = x.float()
+        x = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+        y = x * self.weight.float()
+        g = g.float()
+        if self.activation in ('silu', 'swish'):
+            y = y * g * torch.sigmoid(g)
+        else:  # 'sigmoid'
+            y = y * torch.sigmoid(g)
+        return y.to(dt)
+
+# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
+# It means that the function will not be traced through and simply appear as a node in the graph.
+if is_torch_fx_available():
+    if not is_torch_greater_or_equal_than_1_13:
+        import torch.fx
+
+    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "BailingMoeV3Config"
+
+
+def roll_tensor(tensor, shifts=-1, dims=-1, fill_value=0):
+    """Roll the tensor input along the given dimension(s).
+    Inserted elements are set to be 0.0.
+    """
+    rolled_tensor = torch.roll(tensor, shifts=shifts, dims=dims)
+    rolled_tensor.select(dims, shifts).fill_(fill_value)
+    return rolled_tensor, rolled_tensor.sum()
+
+
+@dataclass
+class MoEV3CausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs as well as Mixture of Expert's router hidden
+    states terms, to train a MoE model.
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            `past_key_values` input) to speed up sequential decoding.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        z_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+            z_loss for the sparse modules.
+        aux_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+            aux_loss for the sparse modules.
+        router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True` is passed or when `config.add_router_probs=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
+            Router logits of the encoder model, useful to compute the auxiliary loss and the z_loss for the sparse
+            modules.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[Cache] = None
+    hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[tuple[torch.FloatTensor, ...]] = None
+    z_loss: Optional[torch.FloatTensor] = None
+    aux_loss: Optional[torch.FloatTensor] = None
+    router_logits: Optional[tuple[torch.FloatTensor]] = None
+    mtp_loss: Optional[torch.FloatTensor] = None
+    mtp_logits: Optional[tuple[torch.FloatTensor, ...]] = None
+
+
+class MoeV3ModelOutputWithPast(MoeModelOutputWithPast):
+
+    def __init__(self, mtp_hidden_states=None, **kwargs):
+        super().__init__(**kwargs)
+        self.mtp_hidden_states = mtp_hidden_states
+
+
+def index_first_axis(x, indices):
+    other_shape = x.shape[1:]
+    second_dim = other_shape.numel()
+    return torch.gather(
+        rearrange(x, "b ... -> b (...)"),
+        0,
+        repeat(indices, "z -> z d", d=second_dim),
+    ).reshape(-1, *other_shape)
+
+
+def index_put_first_axis(x, indices, first_axis_dim):
+    y = torch.zeros(first_axis_dim, *x.shape[1:], device=x.device, dtype=x.dtype)
+    y[indices] = x
+    # y.scatter_(0, repeat(indices, 'z -> z d', d=x.shape[1]), x)
+    return y
+
+
+def pad_input(
+    hidden_states: torch.Tensor,
+    indices: torch.LongTensor,
+    batch_size: int,
+    seq_len: int,
+) -> torch.Tensor:
+    output = index_put_first_axis(hidden_states, indices, batch_size * seq_len)
+    return rearrange(output, "(b s) ... -> b s ...", b=batch_size)
+
+
+@tensor_cache
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    warnings.warn(
+        "Calling `transformers.models.BailingMoeV3.modeling_BailingMoeV3._prepare_4d_attention_mask` is deprecated and will be removed in v4.37. Use `transformers.modeling_attn_mask_utils._prepare_4d_attention_mask"
+    )
+    return _prepare_4d_attention_mask(mask=mask, dtype=dtype, tgt_len=tgt_len)
+
+
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    warnings.warn(
+        "Calling `transformers.models.BailingMoeV3.modeling_BailingMoeV3._make_causal_mask` is deprecated and will be removed in v4.37. Use `transformers.models.BailingMoeV3.modeling_BailingMoeV3.AttentionMaskConverter._make_causal_mask"
+    )
+    return AttentionMaskConverter._make_causal_mask(
+        input_ids_shape=input_ids_shape, dtype=dtype, device=device, past_key_values_length=past_key_values_length
+    )
+
+
+class BailingMoeV3RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        BailingMoeV3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    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)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class BailingMoeV3GroupRMSNorm(nn.Module):
+    def __init__(self, hidden_size, group_norm_size, eps=1e-6):
+        """
+        BailingMoeV3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.group_norm_size = group_norm_size
+        assert hidden_size % group_norm_size == 0, "hidden_size must be divisible by group_norm_size"
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        input_shape = hidden_states.size()
+        group_input_shape = input_shape[:-1] + (self.group_norm_size, input_shape[-1] // self.group_norm_size)
+        hidden_states = hidden_states.view(group_input_shape)
+        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).view(input_shape)
+
+
+ALL_LAYERNORM_LAYERS.append(BailingMoeV3RMSNorm)
+
+
+class BailingMoeV3RotaryEmbedding(nn.Module):
+    def __init__(self, config: BailingMoeV3Config, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = deepcopy(config)
+        self.config.head_dim = config.qk_rope_head_dim
+        self.config.partial_rotary_factor = 1.0
+        # transformers 5.x dropped the "default" key from ROPE_INIT_FUNCTIONS; fall
+        # back to the standard (non-scaled) RoPE inv_freq when the type is unknown.
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS.get(self.rope_type)
+        if self.rope_init_fn is None:
+            def _default_rope(cfg, device=None, **kw):
+                base = cfg.rope_theta
+                dim = int(cfg.head_dim * getattr(cfg, "partial_rotary_factor", 1.0))
+                inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).float().to(device) / dim))
+                return inv_freq, 1.0
+            self.rope_init_fn = _default_rope
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+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)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_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 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)
+    return q_embed, k_embed
+
+
+class BailingMoeV3MLP(nn.Module):
+    def __init__(self, config: BailingMoeV3Config, intermediate_size: int):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = 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):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class BailingMoeV3Gate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.num_experts = config.num_experts
+
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+
+        # topk selection algorithm
+        self.gating_dim = config.hidden_size
+        self.weight = nn.Parameter(torch.empty((self.num_experts, self.gating_dim)))
+        self.routed_scaling_factor = config.routed_scaling_factor
+
+        self.register_buffer("expert_bias", torch.zeros((self.num_experts)))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        import torch.nn.init as init
+
+        init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+
+    def group_limited_topk(
+        self,
+        scores: torch.Tensor,
+    ):
+        num_tokens, _ = scores.size()
+        # Organize the experts into groups
+        group_scores = scores.view(num_tokens, self.n_group, -1).topk(2, dim=-1)[0].sum(dim=-1)
+        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+
+        # Mask the experts based on selection groups
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(num_tokens, self.n_group, self.num_experts // self.n_group)
+            .reshape(num_tokens, -1)
+        )
+
+        masked_scores = scores.masked_fill(~score_mask.bool(), float('-inf'))
+        probs, top_indices = torch.topk(masked_scores, k=self.top_k, dim=-1)
+
+        return probs, top_indices
+
+    def forward(self, hidden_states):
+        # compute gating score
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
+
+        scores = torch.sigmoid(logits.float()).type_as(logits)
+
+        scores_for_routing = scores + self.expert_bias
+        _, topk_idx = self.group_limited_topk(scores_for_routing)
+
+        scores = torch.gather(scores, dim=1, index=topk_idx).type_as(logits)
+
+        topk_weight = scores / (scores.sum(dim=-1, keepdim=True) + 1e-20) if self.top_k > 1 else scores
+        topk_weight = topk_weight * self.routed_scaling_factor
+
+        return topk_idx, topk_weight, logits
+
+
+class BailingMoeV3SparseMoeBlock(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config: BailingMoeV3Config):
+        super().__init__()
+        self.config = config
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self._setup_experts()
+        self.gate = BailingMoeV3Gate(config)
+        if config.num_shared_experts is not None:
+            self.shared_experts = BailingMoeV3MLP(
+                config=config, intermediate_size=config.moe_shared_expert_intermediate_size * config.num_shared_experts
+            )
+
+    def _setup_experts(self):
+        self.experts = nn.ModuleList(
+            [
+                BailingMoeV3MLP(config=self.config, intermediate_size=self.config.moe_intermediate_size)
+                for _ in range(self.config.num_experts)
+            ]
+        )
+
+    def forward(self, hidden_states):
+        identity = hidden_states
+        bsz, seq_len, h = hidden_states.shape
+        topk_idx, topk_weight, router_logits = self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        flat_topk_idx = topk_idx.view(-1)
+        if self.training:
+            hidden_states = hidden_states.repeat_interleave(self.num_experts_per_tok, dim=0)
+            y = torch.empty_like(hidden_states)
+            for i, expert in enumerate(self.experts):
+                y[flat_topk_idx == i] = expert(hidden_states[flat_topk_idx == i])
+            y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
+            y = y.to(hidden_states.dtype).view(bsz, seq_len, h)
+        else:
+            y = self.moe_infer(hidden_states, topk_idx, topk_weight).view(bsz, seq_len, h)
+        if self.config.num_shared_experts is not None:
+            y = y + self.shared_experts(identity)
+        return y, (router_logits.view(bsz, seq_len, -1), topk_idx.view(bsz, seq_len, -1))
+
+    @torch.no_grad()
+    def moe_infer(self, x, topk_ids, topk_weight):
+        cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
+        cnts.scatter_(1, topk_ids, 1)
+        tokens_per_expert = cnts.sum(dim=0)
+        idxs = topk_ids.view(-1).argsort()
+        sorted_tokens = x[idxs // topk_ids.shape[1]]
+        tokens_per_expert = tokens_per_expert.cpu().numpy()
+        outputs = []
+        start_idx = 0
+        for i, num_tokens in enumerate(tokens_per_expert):
+            end_idx = start_idx + num_tokens
+            if num_tokens == 0:
+                continue
+            expert = self.experts[i]
+            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
+            expert_out = expert(tokens_for_this_expert)
+            outputs.append(expert_out.to(x.device))
+            start_idx = end_idx
+
+        outs = torch.cat(outputs, dim=0) if len(outputs) else sorted_tokens.new_empty(0)
+        new_x = torch.empty_like(outs)
+        new_x[idxs] = outs
+        final_out = (
+            new_x.view(*topk_ids.shape, -1)
+            .type(topk_weight.dtype)
+            .mul_(topk_weight.unsqueeze(dim=-1))
+            .sum(dim=1)
+            .type(new_x.dtype)
+        )
+        return final_out
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int, head_first: bool = True) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). If head_first is True, the hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    if n_rep == 1:
+        return hidden_states
+    if head_first:
+        batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+        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)
+    else:
+        batch, slen, num_key_value_heads, head_dim = hidden_states.shape
+        hidden_states = hidden_states[:, :, :, None, :].expand(batch, slen, num_key_value_heads, n_rep, head_dim)
+        return hidden_states.reshape(batch, slen, num_key_value_heads * n_rep, head_dim)
+
+
+def repeat_kv2(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv2(key, module.num_key_value_groups)
+    value_states = repeat_kv2(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+def apply_rotary_pos_emb_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    r"""
+    TODO let's just use the original freqcis computation to not have the view
+    transpose + reshape! This is not optimized!
+    Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        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)
+
+    b, h, s, d = q.shape
+    q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    b, h, s, d = k.shape
+    k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def yarn_get_mscale(scale=1, mscale=1):
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+class BailingMoeV3MultiLatentAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: BailingMoeV3Config, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.attention_dropout = config.attention_dropout
+        self.num_heads = config.num_attention_heads
+        self.rope_theta = config.rope_theta
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.qk_head_dim = config.qk_head_dim
+        self.gated_attention_proj_granularity_type = config.gated_attention_proj_granularity_type
+
+        self.is_causal = True
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
+        else:
+            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.use_qkv_bias)
+            self.q_a_layernorm = BailingMoeV3RMSNorm(config.q_lora_rank)
+            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            config.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.use_qkv_bias,
+        )
+        self.kv_a_layernorm = BailingMoeV3RMSNorm(self.kv_lora_rank)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+        if self.gated_attention_proj_granularity_type is None:
+            self.g_proj = None
+        elif self.gated_attention_proj_granularity_type == "head_wise":
+            self.g_proj = nn.Linear(config.hidden_size, self.num_heads, bias=False)
+        elif self.gated_attention_proj_granularity_type == "element_wise":
+            self.g_proj = nn.Linear(config.hidden_size, self.num_heads * self.v_head_dim, bias=False)
+
+        self.dense = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            config.hidden_size,
+            bias=config.use_qkv_bias,
+        )
+
+        self.scaling = self.qk_head_dim ** (-0.5)
+        # transformers 5.x populates a default rope_scaling dict without "factor";
+        # only apply yarn mscale when a real scaling factor is present.
+        if self.config.rope_scaling is not None and self.config.rope_scaling.get("factor") is not None:
+            mscale_all_dim = self.config.rope_scaling.get("mscale_all_dim", 0)
+            scaling_factor = self.config.rope_scaling["factor"]
+            if mscale_all_dim:
+                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
+                self.scaling = self.scaling * mscale * mscale
+
+    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+
+        batch_size, seq_length = hidden_states.shape[:-1]
+        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)
+        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)
+
+        if self.q_lora_rank is None:
+            q_states = self.q_proj(hidden_states)
+        else:
+            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q_states = q_states.view(query_shape).transpose(1, 2)
+        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+
+        k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
+        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)
+
+        cos, sin = position_embeddings  # tptest
+        if self.config.rope_interleave:  # support using interleaved weights for efficiency
+            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
+        else:
+            x = 1 / 0
+            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
+        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)
+
+        query_states = torch.cat((q_pass, q_rot), dim=-1)
+        key_states = torch.cat((k_pass, k_rot), dim=-1)
+
+        if past_key_values is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        if self.config._attn_implementation == "flash_attention_2" and self.qk_head_dim != self.v_head_dim:
+            value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])
+
+        attention_interface: Callable = eager_attention_forward
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        if self.config._attn_implementation == "flash_attention_2" and self.qk_head_dim != self.v_head_dim:
+            attn_output = attn_output[:, :, :, : self.v_head_dim]
+
+        if self.g_proj is not None:
+            gate = self.g_proj(hidden_states)
+            gate = F.sigmoid(gate.float()).type_as(hidden_states)
+            if self.gated_attention_proj_granularity_type == "head_wise":
+                attn_output = attn_output * gate[:, :, :, None]
+            else:
+                attn_output = attn_output * gate.view(batch_size, seq_length, self.num_heads, self.v_head_dim)
+
+        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
+        attn_output = self.dense(attn_output)
+        return attn_output, attn_weights, past_key_values
+
+
+class BailingMoeV3KimiDeltaAttention(nn.Module):
+    def __init__(self, config: BailingMoeV3Config, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.mode = "chunk"
+
+        self.hidden_size = config.hidden_size
+        self.conv_size = config.short_conv_kernel_size
+        self.head_dim = config.head_dim
+        self.num_heads = config.num_attention_heads
+        self.head_k_dim = self.head_dim
+        self.num_k_heads = self.num_heads
+        self.no_kda_lora = config.no_kda_lora
+        self.safe_gate = config.kda_safe_gate
+        self.lower_bound = config.kda_lower_bound
+
+        self.layer_idx = layer_idx
+
+        assert self.mode in ['chunk', 'fused_recurrent'], f"Not suppoerted mode `{self.mode}`."
+
+        projection_k_size = self.head_k_dim * self.num_k_heads
+        projection_size = self.head_dim * self.num_heads
+
+        self.q_proj = nn.Linear(self.hidden_size, projection_k_size, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, projection_k_size, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, projection_size, bias=False)
+
+        self.q_conv1d = ShortConvolution(
+            hidden_size=projection_k_size,
+            kernel_size=self.conv_size,
+            activation='silu',
+        )
+        self.k_conv1d = ShortConvolution(
+            hidden_size=projection_k_size,
+            kernel_size=self.conv_size,
+            activation='silu',
+        )
+        self.v_conv1d = ShortConvolution(
+            hidden_size=projection_size,
+            kernel_size=self.conv_size,
+            activation='silu',
+        )
+
+        self.A_log = torch.nn.Parameter(torch.log(torch.empty(self.num_heads, dtype=torch.float32).uniform_(1, 16)))
+
+        if self.no_kda_lora:
+            self.f_proj = nn.Linear(self.hidden_size, projection_size, bias=False)
+        else:
+            self.f_a_proj = nn.Linear(self.hidden_size, self.head_dim, bias=False)
+            self.f_b_proj = nn.Linear(self.head_dim, projection_size, bias=False)
+
+        self.dt_bias = nn.Parameter(torch.empty(projection_size, dtype=torch.float32))
+
+        self.b_proj = nn.Linear(self.hidden_size, self.num_heads, bias=False)
+
+        if self.no_kda_lora:
+            self.g_proj = nn.Linear(self.hidden_size, projection_size, bias=False)
+        else:
+            self.g_a_proj = nn.Linear(self.hidden_size, self.head_dim, bias=False)
+            self.g_b_proj = nn.Linear(self.head_dim, projection_size, bias=False)
+
+        self.o_norm = FusedRMSNormGated(self.head_dim, eps=config.rms_norm_eps, activation='sigmoid')
+        self.o_proj = nn.Linear(projection_size, self.hidden_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        past_key_value=None,
+        **kwargs: Unpack[dict],
+    ) -> tuple[torch.Tensor, torch.Tensor | None, Cache | None]:
+        attention_mask = None
+        if attention_mask is not None:
+            if attention_mask.dim() != 2:
+                attention_mask = kwargs.get("padding_mask")
+
+            if attention_mask is not None and attention_mask.dim() != 2:
+                raise ValueError(
+                    "attention_mask must be a 0-1 matrix of shape [batch_size, seq_len] "
+                    "(0 = padding). 3D masks are not supported here.",
+                )
+        use_cache = past_key_value is not None
+        batch_size, q_len, _ = hidden_states.shape
+        mode = 'fused_recurrent' if q_len <= 64 else self.mode
+        if self.training:
+            assert mode == 'chunk', "Only chunk mode is supported in training."
+
+        cu_seqlens = kwargs.get('cu_seqlens')
+        indices = None
+        if attention_mask is not None:
+            indices, cu_seqlens, _ = _get_unpad_data(attention_mask[:, -q_len:])
+            hidden_states = index_first_axis(rearrange(hidden_states, "b s ... -> (b s) ..."), indices).unsqueeze(0)
+
+        conv_state_q, conv_state_k, conv_state_v = None, None, None
+        recurrent_state = None
+
+        if past_key_value is not None and isinstance(past_key_value, Cache):
+            # ensure the cache list is long enough
+            while len(past_key_value.layers) <= self.layer_idx:
+                past_key_value.layers.append(DynamicLayer())
+
+            if past_key_value.layers[self.layer_idx].keys is not None:
+                recurrent_state = past_key_value.layers[self.layer_idx].keys
+                # ensure recurrent_state is on the same device as hidden_states
+                if recurrent_state.device != hidden_states.device:
+                    recurrent_state = recurrent_state.to(hidden_states.device).contiguous()
+
+            if past_key_value.layers[self.layer_idx].values is not None:
+                conv_state_q, conv_state_k, conv_state_v = past_key_value.layers[self.layer_idx].values
+
+        q, conv_state_q = self.q_conv1d(
+            x=self.q_proj(hidden_states),
+            cache=conv_state_q,
+            output_final_state=use_cache,
+            cu_seqlens=cu_seqlens,
+        )
+        k, conv_state_k = self.k_conv1d(
+            x=self.k_proj(hidden_states),
+            cache=conv_state_k,
+            output_final_state=use_cache,
+            cu_seqlens=cu_seqlens,
+        )
+        v, conv_state_v = self.v_conv1d(
+            x=self.v_proj(hidden_states),
+            cache=conv_state_v,
+            output_final_state=use_cache,
+            cu_seqlens=cu_seqlens,
+        )
+
+        if self.no_kda_lora:
+            g = self.f_proj(hidden_states)
+        else:
+            g = self.f_b_proj(self.f_a_proj(hidden_states))
+
+        beta = self.b_proj(hidden_states).float().sigmoid()
+
+        q, k = map(lambda x: rearrange(x, '... (h d) -> ... h d', d=self.head_k_dim), (q, k))
+        v = rearrange(v, '... (h d) -> ... h d', d=self.head_dim)
+        g = rearrange(g, '... (h d) -> ... h d', d=self.head_dim)
+
+        if mode == 'chunk':
+            o, recurrent_state = chunk_kda(
+                q=q,
+                k=k,
+                v=v,
+                g=g,
+                beta=beta,
+                A_log=self.A_log,
+                dt_bias=self.dt_bias,
+                initial_state=recurrent_state,
+                output_final_state=True,
+                use_qk_l2norm_in_kernel=True,
+                use_gate_in_kernel=True,
+                safe_gate=self.safe_gate,
+                lower_bound=self.lower_bound,
+                cu_seqlens=cu_seqlens,
+            )
+        else:
+            o, recurrent_state = fused_recurrent_kda(
+                q=q,
+                k=k,
+                v=v,
+                g=g,
+                beta=beta,
+                A_log=self.A_log,
+                dt_bias=self.dt_bias,
+                initial_state=recurrent_state,
+                output_final_state=True,
+                use_qk_l2norm_in_kernel=True,
+                use_gate_in_kernel=True,
+                lower_bound=self.lower_bound,
+                cu_seqlens=cu_seqlens,
+            )
+
+        if use_cache and past_key_value is not None and isinstance(past_key_value, Cache):
+            target_device = None
+            for cache in past_key_value.layers:
+                if cache.keys is not None:
+                    target_device = cache.keys.device
+                    break
+            if target_device is None:
+                target_device = recurrent_state.device
+
+            # move to target device
+            if recurrent_state.device != target_device:
+                recurrent_state = recurrent_state.to(target_device)
+
+            past_key_value.layers[self.layer_idx].keys = recurrent_state
+            past_key_value.layers[self.layer_idx].values = (conv_state_q, conv_state_k, conv_state_v)
+
+        if self.no_kda_lora:
+            g = self.g_proj(hidden_states)
+        else:
+            g = self.g_b_proj(self.g_a_proj(hidden_states))
+        g = rearrange(g, '... (h d) -> ... h d', d=self.head_dim)
+        o = self.o_norm(o, g)
+
+        o = rearrange(o, 'b t h d -> b t (h d)')
+        o = self.o_proj(o)
+        if attention_mask is not None:
+            o = pad_input(o.squeeze(0), indices, batch_size, q_len)
+
+        return o, None, past_key_value
+
+
+class BailingMoeV3MTPLayer(nn.Module):
+    def __init__(self, config: BailingMoeV3Config, layer_idx: int):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.input_layernorm = BailingMoeV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.enorm = BailingMoeV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)
+        self.post_attention_layernorm = BailingMoeV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.attention = BailingMoeV3MultiLatentAttention(config=config, layer_idx=layer_idx)
+        self.mlp = BailingMoeV3SparseMoeBlock(config)
+
+        self.hnorm = BailingMoeV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.final_layernorm = BailingMoeV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_embeds,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        input_embeds = self.enorm(input_embeds)
+        hidden_states = self.hnorm(hidden_states)
+        hidden_states = self.eh_proj(torch.cat([input_embeds, hidden_states], dim=-1))
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.attention(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            position_embeddings=position_embeddings,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        if isinstance(hidden_states, tuple):
+            hidden_states, router_logits = hidden_states
+        else:
+            router_logits = None
+        hidden_states = residual + hidden_states.to(residual.device)
+        hidden_states = self.final_layernorm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+class BailingMoeV3DecoderLayer(nn.Module):
+    def __init__(self, config: BailingMoeV3Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.layer_idx = layer_idx
+        self.attention_layer_type = (
+            "attention"
+            if (layer_idx + 1) % config.layer_group_size == 0
+            or layer_idx >= config.num_hidden_layers // config.layer_group_size * config.layer_group_size
+            else "linear_attention"
+        )
+
+        if self.attention_layer_type == "attention":
+            self.attention = BailingMoeV3MultiLatentAttention(config=config, layer_idx=layer_idx)
+        else:
+            self.attention = BailingMoeV3KimiDeltaAttention(config=config, layer_idx=layer_idx)
+
+        self.mlp = (
+            BailingMoeV3SparseMoeBlock(config)
+            if (config.num_experts is not None and layer_idx >= config.first_k_dense_replace)
+            else BailingMoeV3MLP(config=config, intermediate_size=config.intermediate_size)
+        )
+        self.input_layernorm = BailingMoeV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = BailingMoeV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = False,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*):
+                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
+                query_sequence_length, key_sequence_length)` if default attention is used.
+            position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+                config.n_positions - 1]`.
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*):
+                cached past key and value projection states
+            output_attentions (`bool`, *optional*):
+                Whether to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_router_logits (`bool`, *optional*):
+                Whether or not to return the logits of all the routers. They are useful for computing the router loss,
+                and should not be returned during inference.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        if self.attention_layer_type == "attention":
+            hidden_states, self_attn_weights, present_key_value = self.attention(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_value,
+                use_cache=use_cache,
+                cache_position=cache_position,  #
+                position_embeddings=position_embeddings,  #
+                **kwargs,
+            )
+        else:
+            batch_size, seq_len = hidden_states.shape[0], hidden_states.shape[1]
+            device = hidden_states.device
+
+            if attention_mask is None:
+                # if attention_mask is None, create a full mask
+                attention_mask = torch.ones((batch_size, seq_len), dtype=torch.int32, device=device)
+            elif attention_mask.dim() == 4 and attention_mask.shape[1] == 1:
+                attention_mask = attention_mask[:, 0, -1, :].to(torch.int32)
+                attention_mask = (attention_mask > -1e4).to(torch.int32)
+            elif attention_mask.dim() == 2:
+                attention_mask = attention_mask.to(torch.int32)
+            else:
+                raise ValueError(f"Unsupported mask dimension: {attention_mask.shape}")
+
+            hidden_states, self_attn_weights, present_key_value = self.attention(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                past_key_value=past_key_value,
+                position_ids=position_ids,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+            )
+
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        if isinstance(hidden_states, tuple):
+            hidden_states, router_logits = hidden_states
+        else:
+            router_logits = None
+        hidden_states = residual + hidden_states.to(residual.device)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+BAILINGMOEV3_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+    Parameters:
+        config ([`BailingMoeV3Config`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare BailingMoeV3 Model outputting raw hidden-states without any specific head on top.",
+    BAILINGMOEV3_START_DOCSTRING,
+)
+class BailingMoeV3PreTrainedModel(PreTrainedModel):
+    config_class = BailingMoeV3Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["BailingMoeV3DecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+BAILINGMOEV3_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+            [What are attention masks?](../glossary#attention-mask)
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare BailingMoeV3 Model outputting raw hidden-states without any specific head on top.",
+    BAILINGMOEV3_START_DOCSTRING,
+)
+class BailingMoeV3Model(BailingMoeV3PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`BailingMoeV3DecoderLayer`]
+    Args:
+        config: BailingMoeV3Config
+    """
+
+    def __init__(self, config: BailingMoeV3Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.num_nextn_predict_layers = config.num_nextn_predict_layers
+
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = []
+        for layer_idx in range(config.num_hidden_layers + config.num_nextn_predict_layers):
+            layer_cls = BailingMoeV3DecoderLayer if layer_idx < config.num_hidden_layers else BailingMoeV3MTPLayer
+            self.layers.append(layer_cls(config, layer_idx))
+
+        self.layers = nn.ModuleList(self.layers)
+
+        self._use_sdpa = config._attn_implementation == "sdpa"
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self.norm = BailingMoeV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = BailingMoeV3RotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.word_embeddings = value
+
+    @add_start_docstrings_to_model_forward(BAILINGMOEV3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoeV3ModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`transformers."
+                )
+                use_cache = False
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        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.Tensor = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        softmax_attention_layer_id = self.config.layer_group_size - 1
+        past_seen_tokens = (
+            past_key_values.get_seq_length(layer_idx=softmax_attention_layer_id) if past_key_values is not None else 0
+        )
+
+        if position_ids is None:
+            position_ids = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+            position_ids = position_ids.unsqueeze(0)
+
+        if self._use_flash_attention_2:
+            # 2d mask is passed through the layers
+            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        elif self._use_sdpa and not output_attentions:
+            # output_attentions=True can not be supported when using SDPA, and we fall back on
+            # the manual implementation that requires a 4D causal mask in all cases.
+            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_seen_tokens,
+            )
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask, (batch_size, seq_length), inputs_embeds, past_seen_tokens
+            )
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_logits = () if output_router_logits else None
+        next_decoder_cache = None
+        layers = self.layers[: -self.num_nextn_predict_layers] if self.num_nextn_predict_layers > 0 else self.layers
+        mtp_layers = self.layers[-self.num_nextn_predict_layers :] if self.num_nextn_predict_layers > 0 else None
+
+        # tptest miss causal_mask = create_causal_mask(
+
+        for decoder_layer in layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    past_key_values,
+                    cache_position,
+                    output_attentions,
+                    output_router_logits,
+                    use_cache,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    cache_position=cache_position,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    position_embeddings=position_embeddings,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            if output_router_logits and layer_outputs[-1] is not None:
+                all_router_logits += (layer_outputs[-1],)
+
+        hidden_states = self.norm(hidden_states)
+        main_hidden_states = hidden_states
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (main_hidden_states,)
+
+        mtp_hidden_states = None
+
+        if mtp_layers:
+            for decoder_layer in mtp_layers:
+                input_ids, _ = roll_tensor(input_ids, shifts=-1, dims=-1)
+                inputs_embeds = self.word_embeddings(input_ids)
+
+                if self.gradient_checkpointing and self.training:
+                    layer_outputs = self._gradient_checkpointing_func(
+                        decoder_layer.__call__,
+                        inputs_embeds,
+                        hidden_states,
+                        attention_mask,
+                        position_ids,
+                        past_key_values,
+                        output_attentions,
+                        output_router_logits,
+                        use_cache,
+                        position_embeddings,
+                    )
+                else:
+                    layer_outputs = decoder_layer(
+                        inputs_embeds,
+                        hidden_states,
+                        attention_mask=attention_mask,
+                        position_ids=position_ids,
+                        past_key_value=past_key_values,
+                        output_attentions=output_attentions,
+                        output_router_logits=output_router_logits,
+                        use_cache=use_cache,
+                        position_embeddings=position_embeddings,
+                    )
+                if mtp_hidden_states is None:
+                    mtp_hidden_states = []
+                hidden_states = layer_outputs[0]
+                mtp_hidden_states.append(hidden_states)
+
+                if output_hidden_states:
+                    all_hidden_states += (hidden_states,)
+
+                if use_cache:
+                    next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+                if output_attentions:
+                    all_self_attns += (layer_outputs[1],)
+
+                if output_router_logits and layer_outputs[-1] is not None:
+                    all_router_logits += (layer_outputs[-1],)
+
+        next_cache = None
+        if use_cache:
+            next_cache = next_decoder_cache
+        if not return_dict:
+            return tuple(
+                v
+                for v in [main_hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
+                if v is not None
+            )
+        return MoeV3ModelOutputWithPast(
+            last_hidden_state=main_hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            mtp_hidden_states=mtp_hidden_states,
+            attentions=all_self_attns,
+            router_logits=all_router_logits,
+        )
+
+
+class BailingMoeV3ForCausalLM(BailingMoeV3PreTrainedModel, GenerationMixin):
+    # transformers 5.x expects a dict (4.x used a list); this model sets
+    # tie_word_embeddings=false so nothing is tied -> empty mapping.
+    _tied_weights_keys = {}
+
+    def __init__(self, config: BailingMoeV3Config):
+        super().__init__(config)
+        self.model = BailingMoeV3Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.num_nextn_predict_layers = config.num_nextn_predict_layers
+        self.mtp_loss_scaling_factor = config.mtp_loss_scaling_factor
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.word_embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(BAILINGMOEV3_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=MoEV3CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoEV3CausalLMOutputWithPast]:
+        r"""
+        Args:
+            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]`.
+        Returns:
+        Example:
+        ```python
+        >>> from transformers import AutoTokenizer
+        >>> model = BailingMoeV3ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            output_router_logits=output_router_logits,
+            return_dict=return_dict,
+            **kwargs,
+        )
+
+        loss = None
+        all_mtp_loss = None
+        aux_loss = None
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.config.vocab_size, **kwargs)
+
+        all_mtp_logits = None
+        if self.num_nextn_predict_layers > 0:
+            mtp_hidden_states = outputs.mtp_hidden_states
+            shift_labels_mtp = None
+            for i in range(self.num_nextn_predict_layers):
+                mtp_hidden_states = mtp_hidden_states[i]
+                mtp_logits = self.lm_head(mtp_hidden_states).float()
+                if all_mtp_logits is None:
+                    all_mtp_logits = []
+                all_mtp_logits.append(mtp_logits)
+                if labels is not None:
+                    if shift_labels_mtp is None:
+                        shift_labels_mtp = labels.clone()
+                    shift_labels_mtp, _ = roll_tensor(shift_labels_mtp, shifts=-1, dims=-1, fill_value=-100)
+                    mtp_logits_ = mtp_logits.view(-1, self.config.vocab_size)
+                    mtp_loss = self.loss_function(
+                        mtp_logits_, shift_labels_mtp.to(mtp_logits_.device).view(-1), self.config.vocab_size, **kwargs
+                    )
+                    if loss is not None:
+                        loss += self.mtp_loss_scaling_factor * mtp_loss
+                    else:
+                        loss = self.mtp_loss_scaling_factor * mtp_loss
+
+                    if all_mtp_loss is None:
+                        all_mtp_loss = []
+                    all_mtp_loss.append(mtp_loss)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            if output_router_logits:
+                output = (aux_loss,) + output
+            return (loss,) + output if loss is not None else output
+
+        return MoEV3CausalLMOutputWithPast(
+            loss=loss,
+            mtp_loss=all_mtp_loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            mtp_logits=all_mtp_logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98279ab5817f4e892add29cc6184c941461a3ff8daf6254f5ef0f0d318998317
+size 12205899

tokenizer_config.json

@@ -0,0 +1,18 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<|startoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "[CLS]",
+  "eos_token": "<|role_end|>",
+  "fast_tokenizer": true,
+  "gmask_token": "[gMASK]",
+  "is_local": true,
+  "local_files_only": false,
+  "merges_file": null,
+  "model_max_length": 1000000000000000019884624838656,
+  "model_specific_special_tokens": {
+    "gmask_token": "[gMASK]"
+  },
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "TokenizersBackend"
+}