Upload folder using huggingface_hub

README.md

@@ -0,0 +1,75 @@
+---
+license: other
+license_name: tencent-youtu
+tags:
+- mlx
+- apple-silicon
+- tencent
+- youtu
+- reasoning
+- mla
+- 4-bit
+base_model: tencent/Youtu-LLM-2B
+library_name: mlx-lm
+pipeline_tag: text-generation
+---
+
+# Youtu-LLM-2B 4-bit MLX
+
+MLX-optimized 4-bit quantized version of [tencent/Youtu-LLM-2B](https://huggingface.co/tencent/Youtu-LLM-2B) for Apple Silicon.
+
+## Quick Start
+
+```bash
+pip install mlx-lm
+
+mlx_lm.generate \
+  --model mlx-community/Youtu-LLM-2B-4bit \
+  --prompt "Hello, what can you do?" \
+  --max-tokens 100
+```
+
+## Model Details
+
+- **Base Model:** tencent/Youtu-LLM-2B
+- **Parameters:** 1.96B
+- **Quantization:** 4-bit (4.5 bits/weight)
+- **Context:** 128K tokens
+- **Architecture:** Dense MLA (Multi-head Latent Attention)
+- **Framework:** MLX (Apple Silicon optimized)
+
+## Performance
+
+| Metric | Value |
+|--------|-------|
+| Size | 1.2GB |
+| Speed | ~209 tokens/sec |
+| Peak Memory | ~1.4GB |
+
+## Features
+
+- **Reasoning Mode:** Uses `<think>` tags for Chain of Thought
+- **128K Context:** Long document understanding
+- **Agentic:** Strong on SWE-Bench, GAIA benchmarks
+- **Edge-friendly:** Runs on any Apple Silicon Mac
+
+## Benchmarks (vs Qwen3-4B)
+
+| Benchmark | Youtu-LLM-2B | Qwen3-4B |
+|-----------|--------------|----------|
+| HumanEval | **95.9%** | 95.4% |
+| SWE-Bench | **17.7%** | 5.7% |
+| GAIA | **33.9%** | 25.5% |
+
+## Other Quantizations
+
+- [Full precision](https://huggingface.co/mlx-community/Youtu-LLM-2B) (4.4GB)
+- [4-bit](https://huggingface.co/mlx-community/Youtu-LLM-2B-4bit) (1.2GB)
+
+## Technical Note
+
+Converted using deepseek_v2 architecture mapping (compatible MLA implementation).
+
+## License
+
+See [original model license](https://huggingface.co/tencent/Youtu-LLM-2B/blob/main/LICENSE.txt).

chat_template.jinja

@@ -0,0 +1 @@
+{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_tool_message=false, first_tool_index=messages|length, is_output_first=true, system_prompt='', is_first_sp=true, is_last_user=false) %}{% for message in messages %}{% if message['role'] == 'system' %}{% if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{% else %}{% set ns.system_prompt = ns.system_prompt + '\n\n' + message['content'] %}{% endif %}{% endif %}{% if not ns.is_tool_message and (message['role'] == 'tool' or (message['role'] == 'user' and message['content'].startswith('<tool_response>') and message['content'].endswith('</tool_response>'))) %}{% set ns.is_tool_message = true %}{% set ns.first_tool_index = loop.index0 %}{% endif %}{% endfor %}{% if tools is defined and tools is not none %}{% set tool_ns = namespace(text='<|begin_of_tool_description|>Tool calling capabilities.\nYou may call one or more functions to assist with the user query. You have the following functions available:', return_text='For tool call returns, you MUST use the following format:\n<tool_call>{\"name\": \"function-name\", \"arguments\": {\"param1\": \"value1\", \"param2\": \"value2\"}}</tool_call>\n<|end_of_tool_description|>') %}{% for tool in tools %}{% set tool_ns.text = tool_ns.text + '\n```json\n' + (tool | tojson) + '\n```' %}{% endfor %}{% set tool_ns.text = tool_ns.text + '\n' + tool_ns.return_text %}{% if ns.system_prompt == '' %}{% set ns.system_prompt = tool_ns.text %}{% else %}{% set ns.system_prompt = ns.system_prompt + '\n\n' + tool_ns.text %}{% endif %}{% endif %}{{ bos_token }}{{ ns.system_prompt }}{% for message in messages %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{% set ns.is_tool = false %}{% set ns.is_first = false %}{% set ns.is_last_user = true %}{{ '<|User|>' + content }}{% endif %}{% if message['role'] == 'assistant' %}{% if '</think>' in content and not loop.last and loop.index0 < (ns.first_tool_index - 1) %}{% set content = content.rsplit('</think>', 1)[-1].lstrip('\n') %}{% endif %}{% if '<think>' not in content and '</think>' not in content and loop.last %}{% set content = '<think>\n\n</think>\n\n' + content %}{% endif %}{% endif %}{% if message['role'] == 'assistant' and message['tool_calls'] is defined and message['tool_calls'] is not none %}{% set ns.is_last_user = false %}{{ '<|Assistant|>' }}{% if content is not none %}{{ content }}{% endif %}{% set ns.is_first = false %}{% set ns.is_tool = false %}{% set ns.is_output_first = true %}{% for tool in message['tool_calls'] %}{% if tool['function']['arguments'] is string %}{% set tool_call_str = '{\"name\": \"' + tool['function']['name'] + '\", \"arguments\": ' + tool['function']['arguments'] + '}' %}{% else %}{% set tool_call_str = '{\"name\": \"' + tool['function']['name'] + '\", \"arguments\": ' + tool['function']['arguments']|tojson + '}' %}{% endif %}{% if not ns.is_first %}{{ '<tool_call>' + tool_call_str + '</tool_call>' }}{% set ns.is_first = true %}{% else %}{{ '\n' + '<tool_call>' + tool_call_str + '</tool_call>' }}{% endif %}{% endfor %}{{ '<|end_of_text|>' }}{% endif %}{% if message['role'] == 'assistant' and (message['tool_calls'] is not defined or message['tool_calls'] is none)%}{% set ns.is_last_user = false %}{% set ns.is_tool = false %}{% set ns.is_output_first = true %}{{ '<|Assistant|>' + content + '<|end_of_text|>' }}{% endif %}{% if message['role'] == 'tool' %}{% set ns.is_last_user = false %}{% set ns.is_tool = true %}{% if ns.is_output_first %}{{ '<|User|><tool_response>' + content + '</tool_response>' }}{% set ns.is_output_first = false %}{% else %}{{ '\n<tool_response>' + content + '</tool_response>' }}{% endif %}{% endif %}{% endfor %}{% if add_generation_prompt and (ns.is_last_user or ns.is_tool) %}{{ '<|Assistant|>' }}{% if enable_thinking is defined and enable_thinking is false %}{{ '<think>\n\n</think>\n\n' }}{% elif forced_thinking is defined and forced_thinking is true %}{{ '<think>\n' }}{% endif %}{% endif %}

config.json

@@ -0,0 +1,53 @@
+{
+    "architectures": [
+        "YoutuForCausalLM"
+    ],
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "auto_map": {
+        "AutoConfig": "configuration_youtu.YoutuConfig",
+        "AutoModel": "modeling_youtu.YoutuModel",
+        "AutoModelForCausalLM": "modeling_youtu.YoutuForCausalLM"
+    },
+    "bos_token_id": 128000,
+    "embedding_initializer_range": null,
+    "eos_token_id": 128001,
+    "hidden_act": "silu",
+    "hidden_size": 2048,
+    "initializer_range": null,
+    "intermediate_size": 6144,
+    "kv_lora_rank": 512,
+    "max_position_embeddings": 131072,
+    "mlp_bias": false,
+    "model_type": "deepseek_v2",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 32,
+    "num_key_value_heads": 16,
+    "q_lora_rank": 1536,
+    "qk_nope_head_dim": 128,
+    "qk_rope_head_dim": 64,
+    "quantization": {
+        "group_size": 64,
+        "bits": 4,
+        "mode": "affine"
+    },
+    "quantization_config": {
+        "group_size": 64,
+        "bits": 4,
+        "mode": "affine"
+    },
+    "rms_norm_eps": 1e-06,
+    "rope_interleave": true,
+    "rope_scaling": {
+        "type": "yarn",
+        "factor": 1.0,
+        "mscale_all_dim": 0
+    },
+    "rope_theta": 1600000,
+    "tie_word_embeddings": true,
+    "torch_dtype": "bfloat16",
+    "transformers_version": "4.56.0",
+    "use_cache": true,
+    "v_head_dim": 128,
+    "vocab_size": 128256
+}

configuration_youtu.py

@@ -0,0 +1,198 @@
+# coding=utf-8
+# Copyright 2025 Tencent Youtu Lab and the HuggingFace Inc. team. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+
+
+Youtu_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+
+
+class YoutuConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`YoutuModel`]. It is used to instantiate an Youtu
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the Youtu-LLM-2B.
+    e.g. [tencent/Youtu-LLM-2B](https://huggingface.co/tencent/Youtu-LLM-2B)
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 128256):
+            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`YoutuModel`]
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 6144):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*, defaults to 16):
+            In MLA, num_key_value_heads=num_attention_heads.
+        kv_lora_rank (`int`, *optional*, defaults to 512):
+            Rank of the LoRA matrices for key and value projections.
+        q_lora_rank (`int`, *optional*, defaults to 1536):
+            Rank of the LoRA matrices for query projections.
+        qk_rope_head_dim (`int`, *optional*, defaults to 64):
+            Dimension of the query/key heads that use rotary position embeddings.
+        v_head_dim (`int`, *optional*, defaults to 128):
+            Dimension of the value heads.
+        qk_nope_head_dim (`int`, *optional*, defaults to 128):
+            Dimension of the query/key heads that don't use rotary position embeddings.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 131072):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to None):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices, except embedding matrices.
+        embedding_initializer_range (`float`, *optional*, defaults to None):
+            The standard deviation of the truncated_normal_initializer for initializing all embedding matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 128000):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 128001):
+            End of stream token id.
+        tie_word_embeddings (`bool`, *optional*, defaults to `True`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 1600000):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*, defaults to `None`):
+            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
+            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
+            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
+            `max_position_embeddings` to the expected new maximum.
+        rope_interleave (`bool`, *optional*, defaults to `True`):
+            Whether to interleave the rotary position embeddings.
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import YoutuModel, YoutuConfig
+
+    >>> # Initializing a Youtu-LLM-2B style configuration
+    >>> configuration = YoutuConfig()
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "youtu_llm"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "layers.*.mlp.gate_proj": "local_colwise",
+        "layers.*.mlp.up_proj": "local_colwise",
+        "layers.*.mlp.down_proj": "local_rowwise",
+        "layers.*.mlp": "gather",  # This is the only moment where results are gathered
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=128256,
+        hidden_size=2048,
+        intermediate_size=6144,
+        num_hidden_layers=32,
+        num_attention_heads=16,
+        num_key_value_heads=16,
+        kv_lora_rank=512,
+        q_lora_rank=1536,
+        qk_rope_head_dim=64,
+        v_head_dim=128,
+        qk_nope_head_dim=128,
+        hidden_act="silu",
+        max_position_embeddings=131072,
+        initializer_range=None,
+        embedding_initializer_range=None,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=128000,
+        eos_token_id=128001,
+        tie_word_embeddings=True,
+        rope_theta=1600000,
+        rope_scaling=None,
+        rope_interleave=True,
+        attention_bias=False,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        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.head_dim = qk_rope_head_dim
+        self.rope_interleave = rope_interleave
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.mlp_bias = False
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        # if initializer_range is None, set it to 2.0 / (5.0 * self.hidden_size) ** 0.5
+        self.initializer_range = (2.0 / (5.0 * self.hidden_size)) ** 0.5 if initializer_range is None else initializer_range
+        # if embedding_initializer_range is None, set it to 2.0 * self.initializer_range
+        self.embedding_initializer_range = self.initializer_range * 2.0 if embedding_initializer_range is None else embedding_initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, copy it it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+
+        if self.rope_scaling is not None:
+            for key in ["beta_fast", "beta_slow", "factor"]:
+                if key in self.rope_scaling:
+                    self.rope_scaling[key] = float(self.rope_scaling[key])
+        
+        rope_config_validation(self)
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["YoutuConfig"]

generation_config.json

@@ -0,0 +1,11 @@
+{
+    "_from_model_config": true,
+    "bos_token_id": 128000,
+    "eos_token_id": 128001,
+    "do_sample": true,
+    "temperature": 1.0,
+    "top_k": 20,
+    "top_p": 0.95,
+    "transformers_version": "4.56.0",
+    "use_cache": false
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf243e7e3a01ad5934078fa0e8195ed27b17d3f90b8d54c0cbb8fa0ee1a013a6
+size 1251517638

model.safetensors.index.json

@@ -0,0 +1,911 @@
+{
+    "metadata": {
+        "total_size": 1251414016,
+        "total_parameters": 2224228352
+    },
+    "weight_map": {
+        "lm_head.biases": "model.safetensors",
+        "lm_head.scales": "model.safetensors",
+        "lm_head.weight": "model.safetensors",
+        "model.embed_tokens.biases": "model.safetensors",
+        "model.embed_tokens.scales": "model.safetensors",
+        "model.embed_tokens.weight": "model.safetensors",
+        "model.layers.0.input_layernorm.weight": "model.safetensors",
+        "model.layers.0.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.0.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.0.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.0.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.0.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.0.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.0.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.0.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.0.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.0.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.0.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.0.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.0.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.0.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.0.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.0.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.0.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.0.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.0.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.0.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.0.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.0.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.0.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.0.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.0.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.0.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.0.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.1.input_layernorm.weight": "model.safetensors",
+        "model.layers.1.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.1.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.1.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.1.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.1.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.1.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.1.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.1.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.1.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.1.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.1.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.1.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.1.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.1.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.1.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.1.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.1.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.1.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.1.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.1.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.1.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.1.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.1.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.1.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.1.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.1.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.1.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.10.input_layernorm.weight": "model.safetensors",
+        "model.layers.10.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.10.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.10.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.10.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.10.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.10.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.10.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.10.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.10.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.10.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.10.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.10.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.10.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.10.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.10.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.10.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.10.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.10.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.10.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.10.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.10.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.10.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.10.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.10.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.10.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.10.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.10.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.11.input_layernorm.weight": "model.safetensors",
+        "model.layers.11.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.11.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.11.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.11.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.11.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.11.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.11.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.11.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.11.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.11.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.11.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.11.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.11.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.11.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.11.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.11.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.11.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.11.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.11.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.11.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.11.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.11.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.11.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.11.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.11.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.11.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.11.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.12.input_layernorm.weight": "model.safetensors",
+        "model.layers.12.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.12.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.12.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.12.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.12.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.12.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.12.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.12.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.12.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.12.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.12.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.12.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.12.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.12.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.12.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.12.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.12.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.12.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.12.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.12.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.12.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.12.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.12.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.12.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.12.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.12.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.12.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.13.input_layernorm.weight": "model.safetensors",
+        "model.layers.13.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.13.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.13.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.13.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.13.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.13.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.13.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.13.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.13.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.13.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.13.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.13.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.13.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.13.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.13.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.13.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.13.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.13.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.13.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.13.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.13.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.13.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.13.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.13.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.13.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.13.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.13.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.14.input_layernorm.weight": "model.safetensors",
+        "model.layers.14.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.14.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.14.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.14.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.14.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.14.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.14.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.14.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.14.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.14.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.14.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.14.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.14.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.14.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.14.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.14.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.14.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.14.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.14.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.14.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.14.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.14.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.14.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.14.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.14.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.14.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.14.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.15.input_layernorm.weight": "model.safetensors",
+        "model.layers.15.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.15.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.15.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.15.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.15.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.15.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.15.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.15.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.15.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.15.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.15.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.15.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.15.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.15.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.15.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.15.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.15.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.15.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.15.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.15.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.15.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.15.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.15.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.15.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.15.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.15.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.15.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.16.input_layernorm.weight": "model.safetensors",
+        "model.layers.16.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.16.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.16.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.16.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.16.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.16.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.16.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.16.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.16.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.16.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.16.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.16.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.16.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.16.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.16.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.16.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.16.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.16.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.16.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.16.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.16.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.16.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.16.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.16.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.16.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.16.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.16.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.17.input_layernorm.weight": "model.safetensors",
+        "model.layers.17.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.17.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.17.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.17.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.17.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.17.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.17.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.17.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.17.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.17.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.17.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.17.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.17.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.17.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.17.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.17.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.17.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.17.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.17.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.17.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.17.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.17.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.17.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.17.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.17.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.17.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.17.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.18.input_layernorm.weight": "model.safetensors",
+        "model.layers.18.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.18.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.18.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.18.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.18.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.18.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.18.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.18.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.18.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.18.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.18.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.18.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.18.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.18.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.18.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.18.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.18.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.18.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.18.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.18.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.18.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.18.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.18.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.18.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.18.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.18.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.18.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.19.input_layernorm.weight": "model.safetensors",
+        "model.layers.19.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.19.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.19.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.19.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.19.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.19.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.19.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.19.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.19.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.19.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.19.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.19.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.19.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.19.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.19.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.19.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.19.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.19.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.19.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.19.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.19.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.19.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.19.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.19.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.19.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.19.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.19.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.2.input_layernorm.weight": "model.safetensors",
+        "model.layers.2.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.2.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.2.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.2.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.2.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.2.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.2.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.2.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.2.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.2.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.2.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.2.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.2.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.2.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.2.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.2.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.2.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.2.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.2.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.2.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.2.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.2.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.2.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.2.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.2.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.2.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.2.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.20.input_layernorm.weight": "model.safetensors",
+        "model.layers.20.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.20.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.20.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.20.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.20.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.20.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.20.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.20.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.20.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.20.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.20.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.20.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.20.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.20.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.20.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.20.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.20.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.20.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.20.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.20.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.20.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.20.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.20.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.20.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.20.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.20.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.20.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.21.input_layernorm.weight": "model.safetensors",
+        "model.layers.21.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.21.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.21.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.21.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.21.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.21.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.21.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.21.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.21.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.21.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.21.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.21.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.21.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.21.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.21.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.21.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.21.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.21.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.21.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.21.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.21.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.21.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.21.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.21.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.21.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.21.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.21.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.22.input_layernorm.weight": "model.safetensors",
+        "model.layers.22.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.22.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.22.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.22.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.22.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.22.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.22.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.22.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.22.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.22.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.22.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.22.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.22.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.22.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.22.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.22.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.22.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.22.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.22.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.22.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.22.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.22.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.22.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.22.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.22.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.22.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.22.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.23.input_layernorm.weight": "model.safetensors",
+        "model.layers.23.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.23.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.23.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.23.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.23.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.23.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.23.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.23.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.23.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.23.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.23.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.23.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.23.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.23.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.23.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.23.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.23.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.23.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.23.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.23.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.23.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.23.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.23.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.23.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.23.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.23.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.23.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.24.input_layernorm.weight": "model.safetensors",
+        "model.layers.24.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.24.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.24.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.24.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.24.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.24.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.24.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.24.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.24.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.24.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.24.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.24.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.24.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.24.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.24.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.24.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.24.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.24.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.24.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.24.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.24.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.24.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.24.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.24.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.24.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.24.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.24.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.25.input_layernorm.weight": "model.safetensors",
+        "model.layers.25.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.25.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.25.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.25.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.25.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.25.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.25.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.25.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.25.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.25.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.25.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.25.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.25.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.25.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.25.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.25.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.25.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.25.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.25.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.25.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.25.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.25.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.25.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.25.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.25.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.25.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.25.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.26.input_layernorm.weight": "model.safetensors",
+        "model.layers.26.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.26.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.26.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.26.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.26.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.26.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.26.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.26.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.26.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.26.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.26.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.26.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.26.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.26.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.26.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.26.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.26.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.26.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.26.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.26.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.26.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.26.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.26.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.26.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.26.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.26.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.26.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.27.input_layernorm.weight": "model.safetensors",
+        "model.layers.27.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.27.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.27.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.27.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.27.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.27.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.27.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.27.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.27.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.27.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.27.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.27.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.27.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.27.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.27.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.27.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.27.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.27.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.27.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.27.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.27.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.27.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.27.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.27.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.27.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.27.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.27.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.28.input_layernorm.weight": "model.safetensors",
+        "model.layers.28.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.28.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.28.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.28.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.28.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.28.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.28.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.28.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.28.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.28.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.28.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.28.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.28.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.28.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.28.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.28.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.28.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.28.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.28.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.28.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.28.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.28.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.28.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.28.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.28.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.28.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.28.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.29.input_layernorm.weight": "model.safetensors",
+        "model.layers.29.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.29.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.29.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.29.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.29.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.29.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.29.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.29.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.29.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.29.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.29.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.29.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.29.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.29.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.29.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.29.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.29.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.29.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.29.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.29.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.29.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.29.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.29.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.29.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.29.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.29.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.29.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.3.input_layernorm.weight": "model.safetensors",
+        "model.layers.3.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.3.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.3.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.3.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.3.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.3.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.3.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.3.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.3.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.3.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.3.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.3.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.3.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.3.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.3.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.3.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.3.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.3.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.3.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.3.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.3.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.3.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.3.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.3.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.3.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.3.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.3.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.30.input_layernorm.weight": "model.safetensors",
+        "model.layers.30.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.30.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.30.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.30.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.30.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.30.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.30.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.30.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.30.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.30.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.30.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.30.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.30.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.30.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.30.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.30.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.30.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.30.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.30.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.30.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.30.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.30.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.30.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.30.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.30.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.30.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.30.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.31.input_layernorm.weight": "model.safetensors",
+        "model.layers.31.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.31.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.31.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.31.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.31.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.31.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.31.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.31.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.31.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.31.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.31.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.31.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.31.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.31.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.31.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.31.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.31.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.31.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.31.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.31.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.31.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.31.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.31.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.31.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.31.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.31.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.31.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.4.input_layernorm.weight": "model.safetensors",
+        "model.layers.4.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.4.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.4.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.4.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.4.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.4.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.4.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.4.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.4.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.4.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.4.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.4.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.4.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.4.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.4.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.4.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.4.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.4.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.4.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.4.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.4.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.4.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.4.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.4.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.4.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.4.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.4.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.5.input_layernorm.weight": "model.safetensors",
+        "model.layers.5.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.5.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.5.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.5.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.5.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.5.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.5.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.5.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.5.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.5.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.5.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.5.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.5.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.5.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.5.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.5.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.5.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.5.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.5.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.5.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.5.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.5.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.5.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.5.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.5.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.5.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.5.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.6.input_layernorm.weight": "model.safetensors",
+        "model.layers.6.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.6.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.6.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.6.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.6.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.6.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.6.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.6.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.6.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.6.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.6.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.6.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.6.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.6.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.6.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.6.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.6.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.6.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.6.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.6.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.6.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.6.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.6.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.6.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.6.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.6.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.6.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.7.input_layernorm.weight": "model.safetensors",
+        "model.layers.7.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.7.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.7.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.7.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.7.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.7.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.7.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.7.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.7.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.7.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.7.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.7.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.7.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.7.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.7.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.7.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.7.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.7.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.7.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.7.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.7.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.7.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.7.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.7.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.7.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.7.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.7.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.8.input_layernorm.weight": "model.safetensors",
+        "model.layers.8.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.8.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.8.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.8.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.8.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.8.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.8.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.8.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.8.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.8.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.8.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.8.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.8.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.8.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.8.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.8.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.8.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.8.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.8.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.8.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.8.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.8.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.8.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.8.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.8.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.8.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.8.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.layers.9.input_layernorm.weight": "model.safetensors",
+        "model.layers.9.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.9.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.9.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.9.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.9.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.9.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.9.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.9.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.9.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.9.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.9.self_attn.kv_a_layernorm.weight": "model.safetensors",
+        "model.layers.9.self_attn.kv_a_proj_with_mqa.biases": "model.safetensors",
+        "model.layers.9.self_attn.kv_a_proj_with_mqa.scales": "model.safetensors",
+        "model.layers.9.self_attn.kv_a_proj_with_mqa.weight": "model.safetensors",
+        "model.layers.9.self_attn.kv_b_proj.biases": "model.safetensors",
+        "model.layers.9.self_attn.kv_b_proj.scales": "model.safetensors",
+        "model.layers.9.self_attn.kv_b_proj.weight": "model.safetensors",
+        "model.layers.9.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.9.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.9.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.9.self_attn.q_a_layernorm.weight": "model.safetensors",
+        "model.layers.9.self_attn.q_a_proj.biases": "model.safetensors",
+        "model.layers.9.self_attn.q_a_proj.scales": "model.safetensors",
+        "model.layers.9.self_attn.q_a_proj.weight": "model.safetensors",
+        "model.layers.9.self_attn.q_b_proj.biases": "model.safetensors",
+        "model.layers.9.self_attn.q_b_proj.scales": "model.safetensors",
+        "model.layers.9.self_attn.q_b_proj.weight": "model.safetensors",
+        "model.norm.weight": "model.safetensors"
+    }
+}

modeling_youtu.py

@@ -0,0 +1,610 @@
+# coding=utf-8
+# Copyright 2025 Tencent Youtu lab, DeepSeek-AI 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.
+import math
+from typing import Callable, Optional, Union
+
+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.generation import GenerationMixin
+from transformers.integrations import use_kernel_forward_from_hub
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import check_model_inputs
+from .configuration_youtu import YoutuConfig
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class YoutuRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        YoutuRMSNorm 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)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class YoutuRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: YoutuConfig, 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 = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        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)
+
+
+class YoutuMLP(nn.Module):
+    def __init__(self, config, hidden_size=None, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
+        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
+        self.mlp_bias = config.mlp_bias
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=self.mlp_bias)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=self.mlp_bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=self.mlp_bias)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+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)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, 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.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        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 YoutuMLAttention(nn.Module):
+    """Multi-latent attention from 'DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model' paper"""
+
+    def __init__(self, config: YoutuConfig, 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.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.attention_bias)
+            self.q_a_layernorm = YoutuRMSNorm(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.attention_bias,
+        )
+        self.kv_a_layernorm = YoutuRMSNorm(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,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            config.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        self.scaling = self.qk_head_dim ** (-0.5)
+        if self.config.rope_scaling 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
+        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:
+            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
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        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]
+
+        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class YoutuDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: YoutuConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = YoutuMLAttention(config=config, layer_idx=layer_idx)
+        self.mlp = YoutuMLP(config)
+        self.input_layernorm = YoutuRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = YoutuRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+@auto_docstring
+class YoutuPreTrainedModel(PreTrainedModel):
+    config: YoutuConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["YoutuDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _can_compile_fullgraph = False
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": YoutuDecoderLayer,
+        "attentions": YoutuMLAttention,
+    }
+
+    def init_weights(self):
+        """
+        If needed prunes and maybe initializes weights. If using a custom `PreTrainedModel`, you need to implement any
+        initialization logic in `_init_weights`.
+        """
+        # Prune heads if needed
+        if self.config.pruned_heads:
+            self.prune_heads(self.config.pruned_heads)
+
+        if "-init" in self.name_or_path:
+            # Initialize weights
+            self.apply(self._initialize_weights)
+
+            # Adjust weights of o_proj in Attention and down_proj in MLP
+            for name, module in self.named_modules():
+                if "o_proj" in name or "down_proj" in name:
+                    # For the output projection, we reinitialize the weights
+                    scaled_std = self.config.initializer_range * (1.0 / self.config.num_hidden_layers) ** 0.5
+                    module.weight.data.normal_(mean=0.0, std=scaled_std)
+
+            # Tie weights should be skipped when not initializing all weights
+            # since from_pretrained(...) calls tie weights anyways
+            self.tie_weights()
+
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        std = self.config.initializer_range
+        embedding_std = self.config.embedding_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=embedding_std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+@auto_docstring
+class YoutuModel(YoutuPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [""]
+
+    def __init__(self, config: YoutuConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [YoutuDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = YoutuRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = YoutuRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds: torch.Tensor = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache(config=self.config)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position: torch.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)
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            input_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class YoutuForCausalLM(YoutuPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = YoutuModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        Example:
+
+        ```python
+        >>> from transformers import YoutuTokenizer, YoutuForCausalLM
+
+        >>> model = YoutuForCausalLM.from_pretrained("tencent/Youtu-LLM-2B")
+        >>> tokenizer = YoutuTokenizer.from_pretrained("tencent/Youtu-LLM-2B")
+
+        >>> 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]
+        ```"""
+        outputs: BaseModelOutputWithPast = 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,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = ["YoutuPreTrainedModel", "YoutuModel", "YoutuForCausalLM"]

tokenizer_config.json

@@ -0,0 +1,17 @@
+{
+  "additional_special_tokens": null,
+  "backend": "tokenizers",
+  "bos_token": "<|begin_of_text|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|end_of_text|>",
+  "is_local": true,
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 131072,
+  "model_specific_special_tokens": {},
+  "pad_token": "<|end_of_text|>",
+  "tokenizer_class": "TokenizersBackend",
+  "truncation_side": "left"
+}