Upload folder using huggingface_hub

__init__.py

@@ -0,0 +1,26 @@
+"""IQuestLoopCoder model package."""
+
+from .configuration_iquestloopcoder import IQuestLoopCoderConfig
+from .modeling_iquestloopcoder import (
+    IQuestLoopCoderPreTrainedModel,
+    IQuestLoopCoderModel,
+    IQuestLoopCoderForCausalLM,
+    IQuestLoopCoderCache,
+)
+from .tokenization_iquestcoder import IQuestCoderTokenizer
+
+try:
+    from .tokenization_iquestcoder import IQuestCoderTokenizerFast
+except ImportError:
+    IQuestCoderTokenizerFast = None
+
+__all__ = [
+    "IQuestLoopCoderConfig",
+    "IQuestLoopCoderPreTrainedModel",
+    "IQuestLoopCoderModel",
+    "IQuestLoopCoderForCausalLM",
+    "IQuestLoopCoderCache",
+    "IQuestCoderTokenizer",
+    "IQuestCoderTokenizerFast",
+]
+

config.json

@@ -0,0 +1,34 @@
+{
+  "_name_or_path": "iquestloopcoder",
+  "architectures": [
+    "IQuestLoopCoderForCausalLM"
+  ],
+  "model_type": "iquestloopcoder",
+  "vocab_size": 76800,
+  "hidden_size": 5120,
+  "intermediate_size": 27648,
+  "num_hidden_layers": 80,
+  "eos_token_id": [2, 75864, 75869],
+  "num_attention_heads": 40,
+  "num_key_value_heads": 8,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "max_position_embeddings": 131072,
+  "initializer_range": 0.02,
+  "rms_norm_eps": 1e-05,
+  "use_cache": true,
+  "tie_word_embeddings": false,
+  "rope_theta": 500000,
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "mlp_bias": false,
+  "loop_num": 2,
+  "loop_window_size": 64,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.40.0",
+  "auto_map": {
+    "AutoConfig": "configuration_iquestloopcoder.IQuestLoopCoderConfig",
+    "AutoModel": "modeling_iquestloopcoder.IQuestLoopCoderModel",
+    "AutoModelForCausalLM": "modeling_iquestloopcoder.IQuestLoopCoderForCausalLM"
+  }
+}

configuration_iquestloopcoder.py

@@ -0,0 +1,132 @@
+# Copyright 2024 IQuestLoopCoder Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+"""IQuestLoopCoder model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class IQuestLoopCoderConfig(PretrainedConfig):
+    r"""
+    Configuration class for IQuestLoopCoder model.
+    
+    IQuestLoopCoder extends the standard LLaMA architecture with a loop mechanism:
+    - Loop 1: Standard attention, stores K1, V1
+    - Loop 2+: Mixed attention with gated combination of global (K1,V1) and local (K2,V2) KV
+    
+    The gate is computed as: gate = sigmoid(W @ Q + bias)
+    Mixed output = gate * Attention(Q, K1, V1) + (1 - gate) * SlidingWindowAttention(Q, K2, V2)
+    
+    Args:
+        vocab_size (`int`, *optional*, defaults to 76800):
+            Vocabulary size of the model.
+        hidden_size (`int`, *optional*, defaults to 5120):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 27648):
+            Dimension of the MLP representations (FFN hidden size).
+        num_hidden_layers (`int`, *optional*, defaults to 80):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 40):
+            Number of attention heads for each attention layer.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            Number of key-value heads (for GQA). If None, defaults to num_attention_heads.
+        head_dim (`int`, *optional*, defaults to 128):
+            Dimension of each attention head (hidden_size // num_attention_heads).
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            Activation function in the MLP.
+        max_position_embeddings (`int`, *optional*, defaults to 8192):
+            Maximum sequence length.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            Standard deviation for weight initialization.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-5):
+            Epsilon for RMS normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether to use past key/values for generation.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie input and output embeddings.
+        rope_theta (`float`, *optional*, defaults to 500000.0):
+            Base value for rotary position embeddings.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in attention layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            Dropout ratio for attention weights.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in MLP layers.
+            
+        # Loop-specific parameters
+        loop_num (`int`, *optional*, defaults to 2):
+            Number of loops through the decoder.
+        loop_window_size (`int`, *optional*, defaults to 64):
+            Window size for sliding window attention in Loop 2+.
+    """
+    
+    model_type = "iquestloopcoder"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=76800,
+        hidden_size=5120,
+        intermediate_size=27648,
+        num_hidden_layers=80,
+        num_attention_heads=40,
+        num_key_value_heads=8,
+        head_dim=128,
+        hidden_act="silu",
+        max_position_embeddings=8192,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        rope_theta=500000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        # Loop-specific parameters
+        loop_num=2,
+        loop_window_size=64,
+        **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.head_dim = head_dim
+        
+        # GQA support
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        
+        self.hidden_act = hidden_act
+        self.initializer_range = 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
+        self.mlp_bias = mlp_bias
+        
+        # Loop-specific
+        self.loop_num = loop_num
+        self.loop_window_size = loop_window_size
+
+        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,
+        )
+

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": [2, 75864, 75869],
+  "transformers_version": "4.55.4"
+}

model-00001-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9433151153a8a83cc7aa4c15338e61fb73022b9377099596858181ba5efc7ce
+size 5359151696

model-00002-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00cf8bdb3506d972f5d528a9b070cbf66662109cfbea43fddcd15595cbe282de
+size 5274455744

model-00003-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d75d4294504e0eed0095f0528adb9ed74617305fa87d5f915a68d6edf5cfcb5
+size 5159101664

model-00004-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d86956a47df06e404ac8c00a2996eb180b67bb03ca4aa464c971a3f643a726d0
+size 5284951976

model-00005-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e667fa53b09c758774232aa81d232aad08179c3aa46892647538e21c0e332cc
+size 5159101664

model-00006-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fd7b5d45765e95e62c94508bc757d8d63e4eb29d646ac29ca672170c64f88a9
+size 5159101680

model-00007-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6b756b0bdff8aa89599d49b43f6765b53c3b6e2a5b811255e100d538e8f093d
+size 5284951968

model-00008-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78794dd0071ae57746030f5dd1ba2d8cba2546d44992ebb065454183525067b5
+size 5159101680

model-00009-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b80829c307ef680e282aae5153a7c78494664e6268e96fbffaaf1d714925e33f
+size 5159101664

model-00010-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccd839c88a7efca21e2cb18f7ea2764d1ba5deaff22259e0f315d6bc6b20a426
+size 5284951976

model-00011-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a3bca95b8e7cc54c8ff72182046031b5c3ea500ac6453fb99d00a84f0b3ed32
+size 5159101664

model-00012-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5eea564372bb8b8fe83908d608c11af4bd22e54a76b9a386737cdeb54c77dce7
+size 5159101680

model-00013-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffb8e154ba4f93956ace20501496bae0c9b9283ac2ddb9c41054a68750f0dfc6
+size 5284951968

model-00014-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee0dc913f924a572e61548035c151c772341c6a5dcc7d210a033f5d2f87ce71c
+size 5159101680

model-00015-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aab0983411682f912be4c40bc32d9c1d58b098a12424b15e0afe8d871898a72f
+size 5159101672

model-00016-of-00016.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83cd92cdee77b61c58076ade54f843bbc017ed60fddad00fb220ace6805481f1
+size 1384163096

model.safetensors.index.json

@@ -0,0 +1,890 @@
+{
+  "metadata": {
+    "total_size": 79589392640
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00001-of-00016.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.0.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.0.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.1.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.1.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.10.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.10.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.11.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.11.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.12.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.12.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.13.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.13.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.14.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.14.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.15.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.15.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.16.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.16.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.17.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.17.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.18.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.18.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.19.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.19.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.2.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.2.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.20.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.20.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.21.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.21.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.22.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.22.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.23.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.23.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.24.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.24.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.25.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.25.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.26.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.26.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.27.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.27.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.28.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.28.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.29.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.29.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.3.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.3.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.30.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.30.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.31.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.31.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.32.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.32.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.33.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.33.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.34.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.34.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.35.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.35.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.36.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.36.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.37.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.37.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.38.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.38.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.39.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.39.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.4.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.4.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.40.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.40.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.41.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.41.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.42.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.42.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.43.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.43.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.44.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.44.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.45.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.45.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.46.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.46.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.47.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.47.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.48.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.48.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.49.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.49.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.5.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.5.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.50.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.50.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.51.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.51.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.52.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.52.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.53.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.53.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.54.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.54.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.55.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.55.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.56.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.56.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.57.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.57.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.58.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.58.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.59.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.59.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.6.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.6.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.60.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.60.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.61.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.61.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.62.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.62.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.63.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.63.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.64.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.64.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.65.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.65.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.66.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.66.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.67.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.67.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.68.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.68.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.69.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.69.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.7.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.7.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.70.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.70.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.71.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.71.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.72.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.72.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.73.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.73.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.74.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.74.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.75.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.75.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.76.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.76.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.77.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.77.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.78.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.78.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.79.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.79.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.8.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.8.weight": "model-00001-of-00016.safetensors",
+    "model.gate_projections.9.bias": "model-00001-of-00016.safetensors",
+    "model.gate_projections.9.weight": "model-00001-of-00016.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00016.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00016.safetensors",
+    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00016.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.1.mlp.up_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00016.safetensors",
+    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00001-of-00016.safetensors",
+    "model.layers.10.mlp.down_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.10.mlp.gate_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.10.mlp.up_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00001-of-00016.safetensors",
+    "model.layers.10.self_attn.k_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.10.self_attn.v_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00001-of-00016.safetensors",
+    "model.layers.11.mlp.down_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.11.mlp.gate_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.11.mlp.up_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00001-of-00016.safetensors",
+    "model.layers.11.self_attn.k_proj.weight": "model-00001-of-00016.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.11.self_attn.q_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.11.self_attn.v_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.12.mlp.down_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.12.mlp.gate_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.12.mlp.up_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.12.self_attn.k_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.12.self_attn.q_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.12.self_attn.v_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.13.mlp.down_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.13.mlp.up_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.14.mlp.down_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.14.mlp.up_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.15.mlp.down_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.15.mlp.gate_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.15.mlp.up_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.15.self_attn.k_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.15.self_attn.v_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.16.mlp.down_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.16.mlp.gate_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.16.mlp.up_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.16.self_attn.k_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.16.self_attn.q_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.16.self_attn.v_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00002-of-00016.safetensors",
+    "model.layers.17.mlp.down_proj.weight": "model-00002-of-00016.safetensors",
+    "model.layers.17.mlp.gate_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.17.mlp.up_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.17.self_attn.k_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.17.self_attn.q_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.17.self_attn.v_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.18.mlp.down_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.18.mlp.gate_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.18.mlp.up_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.18.self_attn.k_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.18.self_attn.q_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.18.self_attn.v_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.19.mlp.down_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.19.mlp.gate_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.19.mlp.up_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.19.self_attn.k_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.19.self_attn.q_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.19.self_attn.v_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.2.mlp.down_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.2.mlp.gate_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.2.mlp.up_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.2.self_attn.k_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.2.self_attn.q_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.2.self_attn.v_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.20.mlp.down_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.20.mlp.gate_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.20.mlp.up_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.20.self_attn.k_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.20.self_attn.q_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.20.self_attn.v_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00003-of-00016.safetensors",
+    "model.layers.21.mlp.down_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.21.mlp.gate_proj.weight": "model-00003-of-00016.safetensors",
+    "model.layers.21.mlp.up_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.21.self_attn.k_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.21.self_attn.q_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.21.self_attn.v_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.22.mlp.down_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.22.mlp.gate_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.22.mlp.up_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.22.self_attn.k_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.22.self_attn.q_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.22.self_attn.v_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.23.mlp.down_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.23.mlp.gate_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.23.mlp.up_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.23.self_attn.k_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.23.self_attn.q_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.23.self_attn.v_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.24.mlp.down_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.24.mlp.gate_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.24.mlp.up_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.24.self_attn.k_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.24.self_attn.q_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.24.self_attn.v_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.25.mlp.down_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.25.mlp.gate_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.25.mlp.up_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.25.self_attn.k_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.25.self_attn.q_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.25.self_attn.v_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.26.mlp.down_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.26.mlp.gate_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.26.mlp.up_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.26.self_attn.k_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.26.self_attn.q_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.26.self_attn.v_proj.weight": "model-00004-of-00016.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00004-of-00016.safetensors",
+    "model.layers.27.mlp.down_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.27.mlp.gate_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.27.mlp.up_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.27.self_attn.k_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.27.self_attn.q_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.27.self_attn.v_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.28.input_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.28.mlp.down_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.28.mlp.gate_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.28.mlp.up_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.28.post_attention_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.28.self_attn.k_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.28.self_attn.o_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.28.self_attn.q_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.28.self_attn.v_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.29.input_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.29.mlp.down_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.29.mlp.gate_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.29.mlp.up_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.29.post_attention_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.29.self_attn.k_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.29.self_attn.o_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.29.self_attn.q_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.29.self_attn.v_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.3.mlp.down_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.3.mlp.gate_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.3.mlp.up_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.3.self_attn.k_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.3.self_attn.q_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.3.self_attn.v_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.30.input_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.30.mlp.down_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.30.mlp.gate_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.30.mlp.up_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.30.post_attention_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.30.self_attn.k_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.30.self_attn.o_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.30.self_attn.q_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.30.self_attn.v_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.31.input_layernorm.weight": "model-00005-of-00016.safetensors",
+    "model.layers.31.mlp.down_proj.weight": "model-00005-of-00016.safetensors",
+    "model.layers.31.mlp.gate_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.31.mlp.up_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.31.post_attention_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.31.self_attn.k_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.31.self_attn.o_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.31.self_attn.q_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.31.self_attn.v_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.32.input_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.32.mlp.down_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.32.mlp.gate_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.32.mlp.up_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.32.post_attention_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.32.self_attn.k_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.32.self_attn.o_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.32.self_attn.q_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.32.self_attn.v_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.33.input_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.33.mlp.down_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.33.mlp.gate_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.33.mlp.up_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.33.post_attention_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.33.self_attn.k_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.33.self_attn.o_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.33.self_attn.q_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.33.self_attn.v_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.34.input_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.34.mlp.down_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.34.mlp.gate_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.34.mlp.up_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.34.post_attention_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.34.self_attn.k_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.34.self_attn.o_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.34.self_attn.q_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.34.self_attn.v_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.35.input_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.35.mlp.down_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.35.mlp.gate_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.35.mlp.up_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.35.post_attention_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.35.self_attn.k_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.35.self_attn.o_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.35.self_attn.q_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.35.self_attn.v_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.36.input_layernorm.weight": "model-00006-of-00016.safetensors",
+    "model.layers.36.mlp.down_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.36.mlp.gate_proj.weight": "model-00006-of-00016.safetensors",
+    "model.layers.36.mlp.up_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.36.post_attention_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.36.self_attn.k_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.36.self_attn.o_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.36.self_attn.q_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.36.self_attn.v_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.37.input_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.37.mlp.down_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.37.mlp.gate_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.37.mlp.up_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.37.post_attention_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.37.self_attn.k_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.37.self_attn.o_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.37.self_attn.q_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.37.self_attn.v_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.38.input_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.38.mlp.down_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.38.mlp.gate_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.38.mlp.up_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.38.post_attention_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.38.self_attn.k_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.38.self_attn.o_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.38.self_attn.q_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.38.self_attn.v_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.39.input_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.39.mlp.down_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.39.mlp.gate_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.39.mlp.up_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.39.post_attention_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.39.self_attn.k_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.39.self_attn.o_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.39.self_attn.q_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.39.self_attn.v_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.4.mlp.down_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.4.mlp.gate_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.4.mlp.up_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.4.self_attn.k_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.4.self_attn.q_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.4.self_attn.v_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.40.input_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.40.mlp.down_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.40.mlp.gate_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.40.mlp.up_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.40.post_attention_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.40.self_attn.k_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.40.self_attn.o_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.40.self_attn.q_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.40.self_attn.v_proj.weight": "model-00007-of-00016.safetensors",
+    "model.layers.41.input_layernorm.weight": "model-00007-of-00016.safetensors",
+    "model.layers.41.mlp.down_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.41.mlp.gate_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.41.mlp.up_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.41.post_attention_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.41.self_attn.k_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.41.self_attn.o_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.41.self_attn.q_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.41.self_attn.v_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.42.input_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.42.mlp.down_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.42.mlp.gate_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.42.mlp.up_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.42.post_attention_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.42.self_attn.k_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.42.self_attn.o_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.42.self_attn.q_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.42.self_attn.v_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.43.input_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.43.mlp.down_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.43.mlp.gate_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.43.mlp.up_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.43.post_attention_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.43.self_attn.k_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.43.self_attn.o_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.43.self_attn.q_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.43.self_attn.v_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.44.input_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.44.mlp.down_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.44.mlp.gate_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.44.mlp.up_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.44.post_attention_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.44.self_attn.k_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.44.self_attn.o_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.44.self_attn.q_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.44.self_attn.v_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.45.input_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.45.mlp.down_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.45.mlp.gate_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.45.mlp.up_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.45.post_attention_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.45.self_attn.k_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.45.self_attn.o_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.45.self_attn.q_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.45.self_attn.v_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.46.input_layernorm.weight": "model-00008-of-00016.safetensors",
+    "model.layers.46.mlp.down_proj.weight": "model-00008-of-00016.safetensors",
+    "model.layers.46.mlp.gate_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.46.mlp.up_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.46.post_attention_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.46.self_attn.k_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.46.self_attn.o_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.46.self_attn.q_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.46.self_attn.v_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.47.input_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.47.mlp.down_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.47.mlp.gate_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.47.mlp.up_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.47.post_attention_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.47.self_attn.k_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.47.self_attn.o_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.47.self_attn.q_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.47.self_attn.v_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.48.input_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.48.mlp.down_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.48.mlp.gate_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.48.mlp.up_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.48.post_attention_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.48.self_attn.k_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.48.self_attn.o_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.48.self_attn.q_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.48.self_attn.v_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.49.input_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.49.mlp.down_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.49.mlp.gate_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.49.mlp.up_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.49.post_attention_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.49.self_attn.k_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.49.self_attn.o_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.49.self_attn.q_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.49.self_attn.v_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.5.mlp.down_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.5.mlp.gate_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.5.mlp.up_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.5.self_attn.k_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.5.self_attn.q_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.5.self_attn.v_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.50.input_layernorm.weight": "model-00009-of-00016.safetensors",
+    "model.layers.50.mlp.down_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.50.mlp.gate_proj.weight": "model-00009-of-00016.safetensors",
+    "model.layers.50.mlp.up_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.50.post_attention_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.50.self_attn.k_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.50.self_attn.o_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.50.self_attn.q_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.50.self_attn.v_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.51.input_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.51.mlp.down_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.51.mlp.gate_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.51.mlp.up_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.51.post_attention_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.51.self_attn.k_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.51.self_attn.o_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.51.self_attn.q_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.51.self_attn.v_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.52.input_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.52.mlp.down_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.52.mlp.gate_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.52.mlp.up_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.52.post_attention_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.52.self_attn.k_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.52.self_attn.o_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.52.self_attn.q_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.52.self_attn.v_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.53.input_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.53.mlp.down_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.53.mlp.gate_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.53.mlp.up_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.53.post_attention_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.53.self_attn.k_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.53.self_attn.o_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.53.self_attn.q_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.53.self_attn.v_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.54.input_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.54.mlp.down_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.54.mlp.gate_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.54.mlp.up_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.54.post_attention_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.54.self_attn.k_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.54.self_attn.o_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.54.self_attn.q_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.54.self_attn.v_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.55.input_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.55.mlp.down_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.55.mlp.gate_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.55.mlp.up_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.55.post_attention_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.55.self_attn.k_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.55.self_attn.o_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.55.self_attn.q_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.55.self_attn.v_proj.weight": "model-00010-of-00016.safetensors",
+    "model.layers.56.input_layernorm.weight": "model-00010-of-00016.safetensors",
+    "model.layers.56.mlp.down_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.56.mlp.gate_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.56.mlp.up_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.56.post_attention_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.56.self_attn.k_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.56.self_attn.o_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.56.self_attn.q_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.56.self_attn.v_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.57.input_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.57.mlp.down_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.57.mlp.gate_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.57.mlp.up_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.57.post_attention_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.57.self_attn.k_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.57.self_attn.o_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.57.self_attn.q_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.57.self_attn.v_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.58.input_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.58.mlp.down_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.58.mlp.gate_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.58.mlp.up_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.58.post_attention_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.58.self_attn.k_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.58.self_attn.o_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.58.self_attn.q_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.58.self_attn.v_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.59.input_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.59.mlp.down_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.59.mlp.gate_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.59.mlp.up_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.59.post_attention_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.59.self_attn.k_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.59.self_attn.o_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.59.self_attn.q_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.59.self_attn.v_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.6.mlp.down_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.6.mlp.gate_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.6.mlp.up_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.6.self_attn.k_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.6.self_attn.q_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.6.self_attn.v_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.60.input_layernorm.weight": "model-00011-of-00016.safetensors",
+    "model.layers.60.mlp.down_proj.weight": "model-00011-of-00016.safetensors",
+    "model.layers.60.mlp.gate_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.60.mlp.up_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.60.post_attention_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.60.self_attn.k_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.60.self_attn.o_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.60.self_attn.q_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.60.self_attn.v_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.61.input_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.61.mlp.down_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.61.mlp.gate_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.61.mlp.up_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.61.post_attention_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.61.self_attn.k_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.61.self_attn.o_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.61.self_attn.q_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.61.self_attn.v_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.62.input_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.62.mlp.down_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.62.mlp.gate_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.62.mlp.up_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.62.post_attention_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.62.self_attn.k_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.62.self_attn.o_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.62.self_attn.q_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.62.self_attn.v_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.63.input_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.63.mlp.down_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.63.mlp.gate_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.63.mlp.up_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.63.post_attention_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.63.self_attn.k_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.63.self_attn.o_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.63.self_attn.q_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.63.self_attn.v_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.64.input_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.64.mlp.down_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.64.mlp.gate_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.64.mlp.up_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.64.post_attention_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.64.self_attn.k_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.64.self_attn.o_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.64.self_attn.q_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.64.self_attn.v_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.65.input_layernorm.weight": "model-00012-of-00016.safetensors",
+    "model.layers.65.mlp.down_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.65.mlp.gate_proj.weight": "model-00012-of-00016.safetensors",
+    "model.layers.65.mlp.up_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.65.post_attention_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.65.self_attn.k_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.65.self_attn.o_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.65.self_attn.q_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.65.self_attn.v_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.66.input_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.66.mlp.down_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.66.mlp.gate_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.66.mlp.up_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.66.post_attention_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.66.self_attn.k_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.66.self_attn.o_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.66.self_attn.q_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.66.self_attn.v_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.67.input_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.67.mlp.down_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.67.mlp.gate_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.67.mlp.up_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.67.post_attention_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.67.self_attn.k_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.67.self_attn.o_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.67.self_attn.q_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.67.self_attn.v_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.68.input_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.68.mlp.down_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.68.mlp.gate_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.68.mlp.up_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.68.post_attention_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.68.self_attn.k_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.68.self_attn.o_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.68.self_attn.q_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.68.self_attn.v_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.69.input_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.69.mlp.down_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.69.mlp.gate_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.69.mlp.up_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.69.post_attention_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.69.self_attn.k_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.69.self_attn.o_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.69.self_attn.q_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.69.self_attn.v_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.7.mlp.down_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.7.mlp.gate_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.7.mlp.up_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.7.self_attn.k_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.7.self_attn.q_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.7.self_attn.v_proj.weight": "model-00013-of-00016.safetensors",
+    "model.layers.70.input_layernorm.weight": "model-00013-of-00016.safetensors",
+    "model.layers.70.mlp.down_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.70.mlp.gate_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.70.mlp.up_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.70.post_attention_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.70.self_attn.k_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.70.self_attn.o_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.70.self_attn.q_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.70.self_attn.v_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.71.input_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.71.mlp.down_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.71.mlp.gate_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.71.mlp.up_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.71.post_attention_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.71.self_attn.k_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.71.self_attn.o_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.71.self_attn.q_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.71.self_attn.v_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.72.input_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.72.mlp.down_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.72.mlp.gate_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.72.mlp.up_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.72.post_attention_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.72.self_attn.k_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.72.self_attn.o_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.72.self_attn.q_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.72.self_attn.v_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.73.input_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.73.mlp.down_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.73.mlp.gate_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.73.mlp.up_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.73.post_attention_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.73.self_attn.k_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.73.self_attn.o_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.73.self_attn.q_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.73.self_attn.v_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.74.input_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.74.mlp.down_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.74.mlp.gate_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.74.mlp.up_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.74.post_attention_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.74.self_attn.k_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.74.self_attn.o_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.74.self_attn.q_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.74.self_attn.v_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.75.input_layernorm.weight": "model-00014-of-00016.safetensors",
+    "model.layers.75.mlp.down_proj.weight": "model-00014-of-00016.safetensors",
+    "model.layers.75.mlp.gate_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.75.mlp.up_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.75.post_attention_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.75.self_attn.k_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.75.self_attn.o_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.75.self_attn.q_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.75.self_attn.v_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.76.input_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.76.mlp.down_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.76.mlp.gate_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.76.mlp.up_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.76.post_attention_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.76.self_attn.k_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.76.self_attn.o_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.76.self_attn.q_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.76.self_attn.v_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.77.input_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.77.mlp.down_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.77.mlp.gate_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.77.mlp.up_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.77.post_attention_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.77.self_attn.k_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.77.self_attn.o_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.77.self_attn.q_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.77.self_attn.v_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.78.input_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.78.mlp.down_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.78.mlp.gate_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.78.mlp.up_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.78.post_attention_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.78.self_attn.k_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.78.self_attn.o_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.78.self_attn.q_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.78.self_attn.v_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.79.input_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.79.mlp.down_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.79.mlp.gate_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.79.mlp.up_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.79.post_attention_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.79.self_attn.k_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.79.self_attn.o_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.79.self_attn.q_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.79.self_attn.v_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00015-of-00016.safetensors",
+    "model.layers.8.mlp.down_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.8.mlp.gate_proj.weight": "model-00015-of-00016.safetensors",
+    "model.layers.8.mlp.up_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00016-of-00016.safetensors",
+    "model.layers.8.self_attn.k_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.8.self_attn.q_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.8.self_attn.v_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00016-of-00016.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.9.mlp.gate_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.9.mlp.up_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00016-of-00016.safetensors",
+    "model.layers.9.self_attn.k_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.9.self_attn.q_proj.weight": "model-00016-of-00016.safetensors",
+    "model.layers.9.self_attn.v_proj.weight": "model-00016-of-00016.safetensors",
+    "model.norm.weight": "model-00016-of-00016.safetensors"
+  }
+}

modeling_iquestloopcoder.py

@@ -0,0 +1,1411 @@
+# Copyright 2024 IQuestLoopCoder Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+"""
+IQuestLoopCoder Model Implementation for HuggingFace.
+
+Loop model passes hidden states through the decoder multiple times:
+- Loop 1: Standard attention, stores K1, V1 for each layer
+- Loop 2+: Mixed attention with gated combination of:
+    - A: Full attention with Loop1's KV (global context)
+    - B: Sliding window attention with Loop2's KV (local, high-precision context)
+    - Gate g = sigmoid(linear(Q)), per-head
+    - Output = g * A + (1 - g) * B
+"""
+
+import math
+from typing import Any, List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, StaticCache
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.generation.utils import GenerationMixin
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+
+from .configuration_iquestloopcoder import IQuestLoopCoderConfig
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "IQuestLoopCoderConfig"
+
+
+class IQuestLoopCoderCache(Cache):
+    """Cache implementation for IQuestLoopCoder that manages shared and local KV caches.
+    
+    - shared_key_cache/shared_value_cache: Stores KV from Loop 1 (global context)
+    - local_key_cache/local_value_cache: Stores KV from Loop 2+ (local window, only window_size tokens)
+    """
+    
+    def __init__(self, window_size: int, num_layers: int):
+        # We intentionally don't call super().__init__ because the parent assumes static cache sizes.
+        self.window_size = window_size
+        self.num_layers = num_layers
+        
+        # Shared cache: stores Loop 1 KV (global context)
+        self.shared_key_cache: List[Optional[torch.Tensor]] = [None] * num_layers
+        self.shared_value_cache: List[Optional[torch.Tensor]] = [None] * num_layers
+        
+        # Local cache: stores Loop 2+ KV (sliding window, only window_size tokens)
+        self.local_key_cache: List[Optional[torch.Tensor]] = [None] * num_layers
+        self.local_value_cache: List[Optional[torch.Tensor]] = [None] * num_layers
+        
+        self.layers: List[Any] = []  # attribute expected by HF Cache utilities
+        self._seen_tokens = 0
+    
+    def update_shared(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[dict] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Update shared cache (Loop 1 KV)."""
+        if layer_idx < 0 or layer_idx >= self.num_layers:
+            raise ValueError(f"layer_idx must be in [0, {self.num_layers}), got {layer_idx}")
+        
+        cached_key = self.shared_key_cache[layer_idx]
+        cached_value = self.shared_value_cache[layer_idx]
+        
+        if cached_key is None:
+            self.shared_key_cache[layer_idx] = key_states
+            self.shared_value_cache[layer_idx] = value_states
+        else:
+            if (
+                key_states.shape[0] != cached_key.shape[0]
+                or key_states.shape[1] != cached_key.shape[1]
+                or key_states.shape[3] != cached_key.shape[3]
+            ):
+                raise ValueError(
+                    "Cached and incoming key/value tensors must match on batch, head, and head_dim dimensions."
+                )
+            assert cached_value is not None
+            self.shared_key_cache[layer_idx] = torch.cat([cached_key, key_states], dim=2)
+            self.shared_value_cache[layer_idx] = torch.cat([cached_value, value_states], dim=2)
+        
+        result_key = self.shared_key_cache[layer_idx]
+        result_value = self.shared_value_cache[layer_idx]
+        assert result_key is not None and result_value is not None
+        
+        # Track sequence length
+        self._seen_tokens = result_key.shape[2]
+        return result_key, result_value
+    
+    def update_local(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[dict] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Update local cache (Loop 2+ KV) with sliding window management.
+        
+        If the cache is full (window_size tokens), remove the oldest token and add the new one.
+        """
+        if layer_idx < 0 or layer_idx >= self.num_layers:
+            raise ValueError(f"layer_idx must be in [0, {self.num_layers}), got {layer_idx}")
+        
+        cached_key = self.local_key_cache[layer_idx]
+        cached_value = self.local_value_cache[layer_idx]
+        
+        if cached_key is None:
+            # First token in local cache
+            self.local_key_cache[layer_idx] = key_states
+            self.local_value_cache[layer_idx] = value_states
+        else:
+            if (
+                key_states.shape[0] != cached_key.shape[0]
+                or key_states.shape[1] != cached_key.shape[1]
+                or key_states.shape[3] != cached_key.shape[3]
+            ):
+                raise ValueError(
+                    "Cached and incoming key/value tensors must match on batch, head, and head_dim dimensions."
+                )
+            assert cached_value is not None
+            
+            # Check if we need to remove the oldest token
+            current_len = cached_key.shape[2]
+            if current_len >= self.window_size:
+                # Remove the first token (oldest) and add the new one
+                self.local_key_cache[layer_idx] = torch.cat([cached_key[:, :, 1:, :], key_states], dim=2)
+                self.local_value_cache[layer_idx] = torch.cat([cached_value[:, :, 1:, :], value_states], dim=2)
+            else:
+                # Just append
+                self.local_key_cache[layer_idx] = torch.cat([cached_key, key_states], dim=2)
+                self.local_value_cache[layer_idx] = torch.cat([cached_value, value_states], dim=2)
+        
+        result_key = self.local_key_cache[layer_idx]
+        result_value = self.local_value_cache[layer_idx]
+        assert result_key is not None and result_value is not None
+        
+        return result_key, result_value
+    
+    def get_shared(self, layer_idx: int) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+        """Get shared cache for a layer."""
+        if layer_idx < 0 or layer_idx >= self.num_layers:
+            return None, None
+        return self.shared_key_cache[layer_idx], self.shared_value_cache[layer_idx]
+    
+    def get_local(self, layer_idx: int) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+        """Get local cache for a layer."""
+        if layer_idx < 0 or layer_idx >= self.num_layers:
+            return None, None
+        return self.local_key_cache[layer_idx], self.local_value_cache[layer_idx]
+    
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[dict] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Default update method (for compatibility, updates shared cache)."""
+        return self.update_shared(key_states, value_states, layer_idx, cache_kwargs)
+    
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Get sequence length from shared cache."""
+        if layer_idx is None:
+            layer_idx = 0
+        if layer_idx < 0 or layer_idx >= len(self.shared_key_cache):
+            return 0
+        cached = self.shared_key_cache[layer_idx]
+        if cached is None:
+            return 0
+        return cached.shape[2]
+    
+    def get_max_length(self) -> Optional[int]:
+        return None
+    
+    def get_usable_length(
+        self, new_seq_length: int, layer_idx: Optional[int] = 0
+    ) -> int:
+        return self.get_seq_length(layer_idx)
+    
+    def reorder_cache(self, beam_idx: torch.LongTensor) -> None:
+        """Reorder cache for beam search."""
+        for layer_idx in range(self.num_layers):
+            if self.shared_key_cache[layer_idx] is not None:
+                device = self.shared_key_cache[layer_idx].device
+                self.shared_key_cache[layer_idx] = self.shared_key_cache[layer_idx].index_select(0, beam_idx.to(device))
+                self.shared_value_cache[layer_idx] = self.shared_value_cache[layer_idx].index_select(0, beam_idx.to(device))
+            
+            if self.local_key_cache[layer_idx] is not None:
+                device = self.local_key_cache[layer_idx].device
+                self.local_key_cache[layer_idx] = self.local_key_cache[layer_idx].index_select(0, beam_idx.to(device))
+                self.local_value_cache[layer_idx] = self.local_value_cache[layer_idx].index_select(0, beam_idx.to(device))
+    
+    @property
+    def is_compileable(self) -> bool:
+        return False
+    
+    def clear(self) -> None:
+        """Clear all caches."""
+        logger.debug("Clearing IQuestLoopCoderCache")
+        self.shared_key_cache = [None] * self.num_layers
+        self.shared_value_cache = [None] * self.num_layers
+        self.local_key_cache = [None] * self.num_layers
+        self.local_value_cache = [None] * self.num_layers
+        self._seen_tokens = 0
+
+
+class IQuestLoopCoderRMSNorm(nn.Module):
+    """RMS Normalization layer."""
+    
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class IQuestLoopCoderRotaryEmbedding(nn.Module):
+    """Rotary Position Embedding (RoPE)."""
+    
+    def __init__(self, dim, max_position_embeddings=8192, base=500000.0, device=None, scaling_factor=1.0):
+        super().__init__()
+        self.scaling_factor = scaling_factor
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.max_seq_len_cached = max_position_embeddings
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        # x: [batch_size, num_heads, seq_len, head_dim]
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        
+        device_type = x.device.type
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+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."""
+    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:
+    """Expand KV heads to match query heads for GQA."""
+    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)
+
+
+class IQuestLoopCoderMLP(nn.Module):
+    """MLP with SwiGLU activation."""
+    
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class LoopGateProjection(nn.Module):
+    """Gate projection for mixed attention in Loop 2+.
+    
+    Computes: g = sigmoid(linear(Q)) for each head independently.
+    This gate determines how much to use Loop1's KV (global) vs current loop's KV (local).
+    """
+    
+    def __init__(self, num_heads: int, head_dim: int):
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+        # Each head has its own gate: Linear(head_dim -> 1) per head
+        # Implemented as [num_heads, head_dim] weight + [num_heads] bias
+        self.weight = nn.Parameter(torch.zeros(num_heads, head_dim))
+        self.bias = nn.Parameter(torch.zeros(num_heads))
+    
+    def forward(self, query: torch.Tensor) -> torch.Tensor:
+        """Compute gate values from query tensor.
+        
+        Args:
+            query: [batch, num_heads, seq_len, head_dim]
+            
+        Returns:
+            gate: [batch, num_heads, seq_len, 1]
+        """
+        # query: [batch, num_heads, seq_len, head_dim]
+        # weight: [num_heads, head_dim]
+        # For each head h: gate_h = query[:, h, :, :] @ weight[h, :].T + bias[h]
+        # Using einsum: gate = einsum('bhsd,hd->bhs', query, weight) + bias
+        gate_logits = torch.einsum('bhsd,hd->bhs', query, self.weight)  # [batch, num_heads, seq_len]
+        gate_logits = gate_logits + self.bias[None, :, None]  # broadcast bias
+        gate = torch.sigmoid(gate_logits)
+        return gate.unsqueeze(-1)  # [batch, num_heads, seq_len, 1]
+
+
+class IQuestLoopCoderAttention(nn.Module):
+    """Multi-head attention with GQA support."""
+    
+    def __init__(self, config: IQuestLoopCoderConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = config.head_dim
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.attention_dropout = config.attention_dropout
+        
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
+        
+        self.rotary_emb = IQuestLoopCoderRotaryEmbedding(
+            self.head_dim,
+            max_position_embeddings=self.max_position_embeddings,
+            base=self.rope_theta,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # Repeat KV for GQA
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        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_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights if output_attentions else None, past_key_value
+    
+    def forward_with_external_kv(
+        self,
+        hidden_states: torch.Tensor,
+        external_key: torch.Tensor,
+        external_value: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        sliding_window: Optional[int] = None,
+    ) -> torch.Tensor:
+        """Forward pass using external K, V (for Loop 2+ mixed attention).
+        
+        Args:
+            hidden_states: Input for computing Q
+            external_key: Pre-computed K (already with RoPE applied)
+            external_value: Pre-computed V
+            attention_mask: Causal attention mask
+            position_ids: Position IDs
+            sliding_window: If set, apply sliding window attention
+            
+        Returns:
+            Attention output [batch, seq_len, num_heads, head_dim]
+        """
+        bsz, q_len, _ = hidden_states.size()
+        
+        # Compute Q from current hidden states
+        query_states = self.q_proj(hidden_states)
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        
+        # Apply RoPE to Q
+        cos, sin = self.rotary_emb(query_states, position_ids)
+        query_states = (query_states * cos.unsqueeze(1)) + (rotate_half(query_states) * sin.unsqueeze(1))
+        
+        # Use external K, V (already have RoPE for K)
+        key_states = external_key
+        value_states = external_value
+        
+        # Repeat KV for GQA
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        
+        # Compute attention
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+        
+        # Apply attention mask (causal)
+        if attention_mask is not None:
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+        
+        # Apply sliding window mask if needed
+        if sliding_window is not None and q_len > sliding_window:
+            # Create sliding window mask
+            # For each position i, can only attend to [i-window+1, i]
+            seq_len = key_states.shape[2]
+            row_idx = torch.arange(q_len, device=query_states.device).unsqueeze(1)
+            col_idx = torch.arange(seq_len, device=query_states.device).unsqueeze(0)
+            window_mask = (col_idx > row_idx) | (col_idx < row_idx - sliding_window + 1)
+            window_mask = window_mask.unsqueeze(0).unsqueeze(0)  # [1, 1, q_len, seq_len]
+            attn_weights = attn_weights.masked_fill(window_mask, float('-inf'))
+        
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+        
+        # Don't apply o_proj here - return raw attention output
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        return attn_output  # [batch, seq_len, num_heads, head_dim]
+    
+    def get_qkv(
+        self,
+        hidden_states: torch.Tensor,
+        position_ids: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Get Q, K, V tensors with RoPE applied.
+        
+        Returns:
+            query: [batch, num_heads, seq_len, head_dim]
+            key: [batch, num_kv_heads, seq_len, head_dim]
+            value: [batch, num_kv_heads, seq_len, head_dim]
+        """
+        bsz, q_len, _ = hidden_states.size()
+        
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+        
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+        
+        return query_states, key_states, value_states
+    
+    def forward_decode_loop1(
+        self,
+        hidden_states: torch.Tensor,
+        past_shared_key: Optional[torch.Tensor],
+        past_shared_value: Optional[torch.Tensor],
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Forward pass for Loop 1 in decode stage.
+        
+        Args:
+            hidden_states: Current hidden states [batch, 1, hidden_size]
+            past_shared_key: Past shared keys from cache [batch, num_kv_heads, past_len, head_dim]
+            past_shared_value: Past shared values from cache [batch, num_kv_heads, past_len, head_dim]
+            attention_mask: Causal attention mask
+            position_ids: Position IDs
+            cache_position: Cache position
+            
+        Returns:
+            output: Attention output [batch, 1, hidden_size]
+            k1: Current key [batch, num_kv_heads, 1, head_dim] (only current token)
+            v1: Current value [batch, num_kv_heads, 1, head_dim] (only current token)
+        """
+        bsz, q_len, _ = hidden_states.size()
+        
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+        
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        
+        cos, sin = self.rotary_emb(value_states, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+        
+        # Store current token's k1, v1 for return (before concatenation)
+        k1_current = key_states  # [batch, num_kv_heads, 1, head_dim]
+        v1_current = value_states  # [batch, num_kv_heads, 1, head_dim]
+        
+        # Concatenate with past shared KV cache for attention computation
+        if past_shared_key is not None and past_shared_value is not None:
+            key_states = torch.cat([past_shared_key, key_states], dim=2)
+            value_states = torch.cat([past_shared_value, value_states], dim=2)
+        
+        # Repeat KV for GQA
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+        
+        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_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+        
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+        attn_output = self.o_proj(attn_output)
+        
+        return attn_output, k1_current, v1_current
+    
+    def forward_decode_loop2(
+        self,
+        hidden_states: torch.Tensor,
+        k1: torch.Tensor,
+        v1: torch.Tensor,
+        past_shared_key: Optional[torch.Tensor],
+        past_shared_value: Optional[torch.Tensor],
+        past_local_key: Optional[torch.Tensor],
+        past_local_value: Optional[torch.Tensor],
+        gate_proj: LoopGateProjection,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        loop_window_size: int = 64,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Forward pass for Loop 2 in decode stage with mixed attention.
+        
+        Args:
+            hidden_states: Current hidden states [batch, 1, hidden_size]
+            k1: Key from Loop 1 (current token) [batch, num_kv_heads, 1, head_dim]
+            v1: Value from Loop 1 (current token) [batch, num_kv_heads, 1, head_dim]
+            past_shared_key: Past shared keys from cache [batch, num_kv_heads, past_len, head_dim]
+            past_shared_value: Past shared values from cache [batch, num_kv_heads, past_len, head_dim]
+            past_local_key: Past local keys from cache [batch, num_kv_heads, window_len, head_dim]
+            past_local_value: Past local values from cache [batch, num_kv_heads, window_len, head_dim]
+            gate_proj: Gate projection module
+            attention_mask: Causal attention mask
+            position_ids: Position IDs
+            loop_window_size: Window size for sliding window attention
+            
+        Returns:
+            output: Attention output [batch, 1, hidden_size]
+            k2: Current key [batch, num_kv_heads, 1, head_dim]
+            v2: Current value [batch, num_kv_heads, 1, head_dim]
+        """
+        bsz, q_len, _ = hidden_states.size()
+        
+        # Get Q2, K2, V2 for current loop
+        q2, k2, v2 = self.get_qkv(hidden_states, position_ids)
+        
+        # Compute gate: g = sigmoid(linear(Q2))
+        gate = gate_proj(q2)  # [batch, num_heads, 1, 1]
+        
+        # For attention A: concatenate past shared KV with current k1, v1 (full global context)
+        if past_shared_key is not None and past_shared_value is not None:
+            k1_full = torch.cat([past_shared_key, k1], dim=2)
+            v1_full = torch.cat([past_shared_value, v1], dim=2)
+        else:
+            k1_full = k1
+            v1_full = v1
+        
+        # For attention B: concatenate past local KV with current k2, v2 (sliding window)
+        if past_local_key is not None and past_local_value is not None:
+            k2_full = torch.cat([past_local_key, k2], dim=2)
+            v2_full = torch.cat([past_local_value, v2], dim=2)
+        else:
+            k2_full = k2
+            v2_full = v2
+        
+        # Repeat KV for GQA
+        k1_expanded = repeat_kv(k1_full, self.num_key_value_groups)
+        v1_expanded = repeat_kv(v1_full, self.num_key_value_groups)
+        k2_expanded = repeat_kv(k2_full, self.num_key_value_groups)
+        v2_expanded = repeat_kv(v2_full, self.num_key_value_groups)
+        
+        # Attention A: Q2 @ K1_full, V1_full (global, full sequence)
+        head_dim = q2.shape[-1]
+        attn_weights_A = torch.matmul(q2, k1_expanded.transpose(2, 3)) / math.sqrt(head_dim)
+        if attention_mask is not None:
+            causal_mask = attention_mask[:, :, :, : k1_expanded.shape[-2]]
+            attn_weights_A = attn_weights_A + causal_mask
+        attn_weights_A = nn.functional.softmax(attn_weights_A, dim=-1, dtype=torch.float32).to(q2.dtype)
+        attn_A = torch.matmul(attn_weights_A, v1_expanded)
+        
+        # Attention B: Q2 @ K2_full, V2_full (local sliding window)
+        attn_weights_B = torch.matmul(q2, k2_expanded.transpose(2, 3)) / math.sqrt(head_dim)
+        if attention_mask is not None:
+            causal_mask = attention_mask[:, :, :, : k2_expanded.shape[-2]]
+            attn_weights_B = attn_weights_B + causal_mask
+        
+        # Apply sliding window mask
+        q_len_attn = q2.shape[2]
+        k_len_attn = k2_expanded.shape[2]
+        if q_len_attn <= loop_window_size:
+            # If sequence fits in window, use standard attention
+            attn_weights_B = nn.functional.softmax(attn_weights_B, dim=-1, dtype=torch.float32).to(q2.dtype)
+        else:
+            # Apply sliding window mask
+            row_idx = torch.arange(q_len_attn, device=q2.device).unsqueeze(1)
+            col_idx = torch.arange(k_len_attn, device=q2.device).unsqueeze(0)
+            window_mask = (col_idx > row_idx) | (col_idx < row_idx - loop_window_size + 1)
+            window_mask = window_mask.unsqueeze(0).unsqueeze(0)
+            attn_weights_B = attn_weights_B.masked_fill(window_mask, float('-inf'))
+            attn_weights_B = nn.functional.softmax(attn_weights_B, dim=-1, dtype=torch.float32).to(q2.dtype)
+        attn_B = torch.matmul(attn_weights_B, v2_expanded)
+        
+        # Mixed attention: gate * A + (1 - gate) * B
+        mixed_attn = gate * attn_A + (1 - gate) * attn_B
+        
+        # Reshape and apply output projection
+        bsz, num_heads, seq_len, head_dim = mixed_attn.shape
+        mixed_attn = mixed_attn.transpose(1, 2).contiguous().reshape(bsz, seq_len, -1)
+        attn_output = self.o_proj(mixed_attn)
+        
+        return attn_output, k2, v2
+
+
+class IQuestLoopCoderDecoderLayer(nn.Module):
+    """Transformer decoder layer."""
+    
+    def __init__(self, config: IQuestLoopCoderConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = IQuestLoopCoderAttention(config=config, layer_idx=layer_idx)
+        self.mlp = IQuestLoopCoderMLP(config)
+        self.input_layernorm = IQuestLoopCoderRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = IQuestLoopCoderRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+        if use_cache:
+            outputs += (present_key_value,)
+        return outputs
+    
+    def forward_loop2_mixed(
+        self,
+        hidden_states: torch.Tensor,
+        k1: torch.Tensor,
+        v1: torch.Tensor,
+        gate_proj: LoopGateProjection,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        loop_window_size: int = 64,
+    ) -> Tuple[torch.Tensor, float]:
+        """Forward pass for Loop 2+ with mixed attention.
+        
+        Args:
+            hidden_states: Current hidden states
+            k1: Key from Loop 1 [batch, num_kv_heads, seq_len, head_dim]
+            v1: Value from Loop 1 [batch, num_kv_heads, seq_len, head_dim]
+            gate_proj: Gate projection module for this layer
+            attention_mask: Causal attention mask
+            position_ids: Position IDs
+            loop_window_size: Window size for sliding window attention
+            
+        Returns:
+            output hidden states, gate mean value
+        """
+        residual = hidden_states
+        hidden_states_normed = self.input_layernorm(hidden_states)
+        
+        # Get Q2, K2, V2 for current loop
+        q2, k2, v2 = self.self_attn.get_qkv(hidden_states_normed, position_ids)
+        
+        # Compute gate: g = sigmoid(linear(Q2))
+        # q2: [batch, num_heads, seq_len, head_dim]
+        gate = gate_proj(q2)  # [batch, num_heads, seq_len, 1]
+        gate_mean = gate.detach().mean().item()
+        
+        # Repeat K1, V1 for GQA
+        k1_expanded = repeat_kv(k1, self.self_attn.num_key_value_groups)
+        v1_expanded = repeat_kv(v1, self.self_attn.num_key_value_groups)
+        k2_expanded = repeat_kv(k2, self.self_attn.num_key_value_groups)
+        v2_expanded = repeat_kv(v2, self.self_attn.num_key_value_groups)
+        
+        # Attention A: Q2 @ K1, V1 (global, full sequence)
+        attn_A = self._compute_attention(q2, k1_expanded, v1_expanded, attention_mask)
+        
+        # Attention B: Q2 @ K2, V2 (local sliding window)
+        attn_B = self._compute_attention_with_window(q2, k2_expanded, v2_expanded, attention_mask, loop_window_size)
+        
+        # Mixed attention: gate * A + (1 - gate) * B
+        # attn_A, attn_B: [batch, num_heads, seq_len, head_dim]
+        mixed_attn = gate * attn_A + (1 - gate) * attn_B
+        
+        # Reshape and apply output projection
+        bsz, num_heads, seq_len, head_dim = mixed_attn.shape
+        mixed_attn = mixed_attn.transpose(1, 2).contiguous().reshape(bsz, seq_len, -1)
+        hidden_states = self.self_attn.o_proj(mixed_attn)
+        
+        hidden_states = residual + hidden_states
+        
+        # MLP
+        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, gate_mean
+    
+    def _compute_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+    ) -> torch.Tensor:
+        """Standard attention computation."""
+        head_dim = query.shape[-1]
+        attn_weights = torch.matmul(query, key.transpose(2, 3)) / math.sqrt(head_dim)
+        
+        if attention_mask is not None:
+            causal_mask = attention_mask[:, :, :, : key.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+        
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+        attn_output = torch.matmul(attn_weights, value)
+        return attn_output
+    
+    def _compute_attention_with_window(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+        window_size: int,
+    ) -> torch.Tensor:
+        """Attention with sliding window."""
+        q_len = query.shape[2]
+        k_len = key.shape[2]
+        head_dim = query.shape[-1]
+        
+        # If sequence fits in window, use standard attention
+        if q_len <= window_size:
+            return self._compute_attention(query, key, value, attention_mask)
+        
+        attn_weights = torch.matmul(query, key.transpose(2, 3)) / math.sqrt(head_dim)
+        
+        # Apply causal mask
+        if attention_mask is not None:
+            causal_mask = attention_mask[:, :, :, : key.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+        
+        # Apply sliding window mask
+        row_idx = torch.arange(q_len, device=query.device).unsqueeze(1)
+        col_idx = torch.arange(k_len, device=query.device).unsqueeze(0)
+        # Can only attend to positions in [i - window_size + 1, i]
+        window_mask = (col_idx > row_idx) | (col_idx < row_idx - window_size + 1)
+        window_mask = window_mask.unsqueeze(0).unsqueeze(0)
+        attn_weights = attn_weights.masked_fill(window_mask, float('-inf'))
+        
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+        attn_output = torch.matmul(attn_weights, value)
+        return attn_output
+
+
+class IQuestLoopCoderPreTrainedModel(PreTrainedModel):
+    """Base class for IQuestLoopCoder models."""
+    config_class = IQuestLoopCoderConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["IQuestLoopCoderDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_cache_class = True
+    _supports_static_cache = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+class IQuestLoopCoderModel(IQuestLoopCoderPreTrainedModel):
+    """IQuestLoopCoder Transformer decoder model."""
+    
+    def __init__(self, config: IQuestLoopCoderConfig):
+        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([
+            IQuestLoopCoderDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)
+        ])
+        self.norm = IQuestLoopCoderRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        
+        # Gate projections for Loop 2+ (one per layer)
+        self.gate_projections = nn.ModuleList([
+            LoopGateProjection(config.num_attention_heads, config.head_dim)
+            for _ in range(config.num_hidden_layers)
+        ])
+        
+        # Loop configuration
+        self.loop_num = config.loop_num
+        self.loop_window_size = config.loop_window_size
+        
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: 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,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        seq_length = inputs_embeds.shape[1]
+        
+        # Determine which forward path to use:
+        # 1. If past_key_values exists and seq_length == 1: autoregressive generation step
+        #    -> Use standard attention with KV cache (no loop needed for single token)
+        # 2. Otherwise (prefill or training): use loop mechanism
+        
+        is_generation_step = past_key_values is not None and seq_length == 1
+        # import pdb; pdb.set_trace()
+        
+        if is_generation_step:
+            # Autoregressive generation: single token, use KV cache
+            return self._forward_with_cache(
+                inputs_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                cache_position=cache_position,
+            )
+        
+        # Prefill or training: use loop mechanism
+        return self._forward_loop(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            use_cache=use_cache,
+            cache_position=cache_position,
+        )
+    
+    def _forward_loop(
+        self,
+        inputs_embeds: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+        position_ids: Optional[torch.LongTensor],
+        output_attentions: bool,
+        output_hidden_states: bool,
+        return_dict: bool,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        """Forward with loop mechanism (for training and prefill).
+        
+        This implements the Loop mechanism:
+        - Loop 1: Standard attention, stores K1, V1 for each layer
+        - Loop 2+: Mixed attention with gated combination of global (K1,V1) and local (K2,V2)
+        """
+        batch_size, seq_length, _ = inputs_embeds.shape
+        
+        if position_ids is None:
+            device = inputs_embeds.device
+            position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0)
+        
+        if cache_position is None:
+            cache_position = torch.arange(seq_length, device=inputs_embeds.device)
+        
+        # Create causal mask
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, None, output_attentions)
+        
+        hidden_states = inputs_embeds
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        
+        # For KV cache during prefill - use IQuestLoopCoderCache
+        # In prefill, past_key_values should be None, so we create a new cache
+        if use_cache:
+            next_decoder_cache = IQuestLoopCoderCache(self.loop_window_size, len(self.layers))
+        else:
+            next_decoder_cache = None
+        
+        # ============ Loop 1: Standard forward, store K1, V1 in shared cache ============
+        for layer_idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            
+            # Get K1, V1 before standard forward (from original hidden_states, after layernorm)
+            hidden_states_normed = decoder_layer.input_layernorm(hidden_states)
+            q1, k1, v1 = decoder_layer.self_attn.get_qkv(hidden_states_normed, position_ids)
+            
+            # Store K1, V1 in shared cache
+            if use_cache:
+                next_decoder_cache.update_shared(k1, v1, layer_idx)
+            
+            # Standard forward
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                past_key_value=None,
+                output_attentions=output_attentions,
+                use_cache=False,
+            )
+            hidden_states = layer_outputs[0]
+            
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+        
+        # ============ Loop 2 to loop_num: Mixed attention, store in local cache ============
+        for loop_idx in range(2, self.loop_num + 1):
+            for layer_idx, decoder_layer in enumerate(self.layers):
+                # Get K1, V1 from shared cache
+                k1, v1 = next_decoder_cache.get_shared(layer_idx) if use_cache else (None, None)
+                if k1 is None or v1 is None:
+                    # Fallback: compute K1, V1 if not in cache (shouldn't happen in prefill)
+                    hidden_states_normed = decoder_layer.input_layernorm(hidden_states)
+                    _, k1, v1 = decoder_layer.self_attn.get_qkv(hidden_states_normed, position_ids)
+                
+                gate_proj = self.gate_projections[layer_idx]
+                
+                hidden_states, gate_mean = decoder_layer.forward_loop2_mixed(
+                    hidden_states,
+                    k1=k1,
+                    v1=v1,
+                    gate_proj=gate_proj,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    loop_window_size=self.loop_window_size,
+                )
+                
+                # Store Loop 2+ KV in local cache (only for loop_idx == 2)
+                if use_cache and loop_idx == 2:
+                    hidden_states_normed = decoder_layer.input_layernorm(hidden_states)
+                    _, k2, v2 = decoder_layer.self_attn.get_qkv(hidden_states_normed, position_ids)
+                    next_decoder_cache.update_local(k2, v2, layer_idx)
+        
+        hidden_states = self.norm(hidden_states)
+        
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+        
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_decoder_cache, all_hidden_states, all_self_attns] if v is not None)
+        
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+    
+    def _forward_with_cache(
+        self,
+        inputs_embeds: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+        position_ids: Optional[torch.LongTensor],
+        past_key_values: Optional[Cache],
+        use_cache: bool,
+        output_attentions: bool,
+        output_hidden_states: bool,
+        return_dict: bool,
+        cache_position: Optional[torch.LongTensor],
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        """Forward with KV cache using loop mechanism (for inference generation).
+        
+        Loop 1: Standard attention, uses shared KV cache (previous tokens + current token)
+        Loop 2+: Mixed attention, uses local KV cache (sliding window)
+        """
+        batch_size, seq_length, _ = inputs_embeds.shape
+        
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(past_seen_tokens, past_seen_tokens + seq_length, device=inputs_embeds.device)
+        
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+        
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions)
+        
+        # Ensure we're using IQuestLoopCoderCache
+        if use_cache:
+            if not isinstance(past_key_values, IQuestLoopCoderCache):
+                # Convert to IQuestLoopCoderCache if needed
+                next_decoder_cache = IQuestLoopCoderCache(self.loop_window_size, len(self.layers))
+                # Copy existing cache if possible
+                if past_key_values is not None:
+                    for layer_idx in range(len(self.layers)):
+                        try:
+                            past_k = past_key_values.key_cache[layer_idx] if hasattr(past_key_values, 'key_cache') else None
+                            past_v = past_key_values.value_cache[layer_idx] if hasattr(past_key_values, 'value_cache') else None
+                            if past_k is not None and past_v is not None:
+                                next_decoder_cache.update_shared(past_k, past_v, layer_idx)
+                        except:
+                            pass
+            else:
+                next_decoder_cache = past_key_values
+        else:
+            next_decoder_cache = None
+        
+        hidden_states = inputs_embeds
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        
+        # ============ Loop 1: Standard attention, store in shared cache ============
+        for layer_idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            
+            # Get past shared KV cache
+            past_shared_key, past_shared_value = None, None
+            if next_decoder_cache is not None:
+                past_shared_key, past_shared_value = next_decoder_cache.get_shared(layer_idx)
+            
+            # Forward Loop 1
+            attn_output, k1, v1 = decoder_layer.self_attn.forward_decode_loop1(
+                hidden_states=decoder_layer.input_layernorm(hidden_states),
+                past_shared_key=past_shared_key,
+                past_shared_value=past_shared_value,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                cache_position=cache_position,
+            )
+            
+            # Update shared cache with current token's Loop 1 KV
+            if use_cache:
+                next_decoder_cache.update_shared(k1, v1, layer_idx)
+            
+            hidden_states = hidden_states + attn_output
+            
+            # MLP
+            residual = hidden_states
+            hidden_states = decoder_layer.post_attention_layernorm(hidden_states)
+            hidden_states = decoder_layer.mlp(hidden_states)
+            hidden_states = residual + hidden_states
+            
+            if output_attentions:
+                all_self_attns += (None,)  # We don't return attention weights in decode loop
+        
+        # ============ Loop 2 to loop_num: Mixed attention, store in local cache ============
+        # Store k1, v1 from Loop 1 for use in Loop 2+
+        loop1_kv = []
+        for layer_idx in range(len(self.layers)):
+            if next_decoder_cache is not None:
+                k1_full, v1_full = next_decoder_cache.get_shared(layer_idx)
+                if k1_full is not None and v1_full is not None:
+                    # Get only the last token (current token)
+                    loop1_kv.append((k1_full[:, :, -1:, :], v1_full[:, :, -1:, :], k1_full, v1_full))
+                else:
+                    loop1_kv.append((None, None, None, None))
+            else:
+                loop1_kv.append((None, None, None, None))
+        
+        for loop_idx in range(2, self.loop_num + 1):
+            for layer_idx, decoder_layer in enumerate(self.layers):
+                # Get k1, v1 (current token's Loop 1 KV) and full shared cache
+                k1_current, v1_current, k1_full, v1_full = loop1_kv[layer_idx]
+                if k1_current is None or v1_current is None:
+                    continue
+                
+                # Get past local KV cache
+                past_local_key, past_local_value = None, None
+                if next_decoder_cache is not None:
+                    past_local_key, past_local_value = next_decoder_cache.get_local(layer_idx)
+                
+                gate_proj = self.gate_projections[layer_idx]
+                
+                # Forward Loop 2+
+                attn_output, k2, v2 = decoder_layer.self_attn.forward_decode_loop2(
+                    hidden_states=decoder_layer.input_layernorm(hidden_states),
+                    k1=k1_current,
+                    v1=v1_current,
+                    past_shared_key=k1_full[:, :, :-1, :] if k1_full is not None and k1_full.shape[2] > 1 else None,
+                    past_shared_value=v1_full[:, :, :-1, :] if v1_full is not None and v1_full.shape[2] > 1 else None,
+                    past_local_key=past_local_key,
+                    past_local_value=past_local_value,
+                    gate_proj=gate_proj,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    loop_window_size=self.loop_window_size,
+                )
+                
+                # Update local cache with current token's Loop 2+ KV
+                if use_cache and loop_idx == 2:
+                    next_decoder_cache.update_local(k2, v2, layer_idx)
+                
+                hidden_states = hidden_states + attn_output
+                
+                # MLP
+                residual = hidden_states
+                hidden_states = decoder_layer.post_attention_layernorm(hidden_states)
+                hidden_states = decoder_layer.mlp(hidden_states)
+                hidden_states = residual + hidden_states
+        
+        hidden_states = self.norm(hidden_states)
+        
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+        
+        next_cache = next_decoder_cache if use_cache else None
+        
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+    
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        """Create causal attention mask."""
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        
+        # Determine target length for attention
+        if past_key_values is not None:
+            # For DynamicCache: use get_seq_length() to get cached length
+            # target_length = cached_length + current_sequence_length
+            past_length = past_key_values.get_seq_length()
+            target_length = past_length + sequence_length
+        elif attention_mask is not None:
+            target_length = attention_mask.shape[-1]
+        else:
+            target_length = sequence_length
+        
+        # Create causal mask
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            # For prefill: standard causal mask
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        
+        # Adjust for cache position (for generation steps after prefill)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
+        
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()
+            mask_length = attention_mask.shape[-1]
+            if mask_length <= target_length:
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(padding_mask, min_dtype)
+        
+        return causal_mask
+
+
+class IQuestLoopCoderForCausalLM(IQuestLoopCoderPreTrainedModel, GenerationMixin):
+    """IQuestLoopCoder model with a causal language modeling head."""
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = IQuestLoopCoderModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: 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,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = nn.CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        use_cache=True,
+        **kwargs,
+    ):
+        past_length = 0
+        if past_key_values is not None:
+            past_length = past_key_values.get_seq_length()
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+
+        if cache_position is None:
+            cache_position = torch.arange(past_length, past_length + input_ids.shape[1], device=input_ids.device)
+        elif use_cache:
+            cache_position = cache_position[-input_ids.shape[1]:]
+
+        position_ids = cache_position.unsqueeze(0)
+
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids.contiguous()}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+

tokenization_iquestcoder.py

@@ -0,0 +1,552 @@
+"""Tokenization classes for IQuestCoder."""
+
+import os
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import sentencepiece as spm
+
+from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {},
+    "tokenizer_file": {},
+}
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}
+
+
+
+class IQuestCoderTokenizer(PreTrainedTokenizer):
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        unk_token="<unk>",
+        bos_token="<s>",
+        eos_token="</s>",
+        pad_token=None,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        add_bos_token=True,
+        add_eos_token=False,
+        clean_up_tokenization_spaces=False,
+        add_prefix_space=False,
+        legacy=None,
+        use_default_system_prompt=False,
+        chat_template=None,
+        **kwargs,
+    ):
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        
+        # Legacy behavior handling
+        if legacy is None:
+            logger.warning_once(
+                f"You are using the default legacy behaviour of the {self.__class__.__name__}. This is"
+                " expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you."
+                " If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it"
+                " means, and thoroughly read the reason why this was added as explained in"
+                " https://github.com/huggingface/transformers/pull/24565"
+            )
+            legacy = True
+        
+        self.legacy = legacy
+        self.vocab_file = vocab_file
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+        self.add_prefix_space = add_prefix_space
+        self.use_default_system_prompt = use_default_system_prompt
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+        
+
+        
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            add_bos_token=add_bos_token,
+            add_eos_token=add_eos_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            add_prefix_space=add_prefix_space,
+            legacy=legacy,
+            use_default_system_prompt=use_default_system_prompt,
+            chat_template=chat_template,
+            **kwargs,
+        )
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab_file)
+
+    @property
+    def vocab_size(self) -> int:
+        """Returns the vocabulary size."""
+        return self.sp_model.get_piece_size()
+
+    def get_vocab(self) -> Dict[str, int]:
+        """Returns the vocabulary as a dictionary of token to index."""
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text: str) -> List[str]:
+        """
+        Tokenize a string.
+        
+        Args:
+            text (`str`): The text to tokenize.
+            
+        Returns:
+            `List[str]`: The list of tokens.
+        """
+        if self.add_prefix_space:
+            text = " " + text
+            
+        if self.legacy:
+            return self.sp_model.encode(text, out_type=str)
+        
+        # Non-legacy behavior: handle special tokens properly
+        return self.sp_model.encode(text, out_type=str)
+
+    def _convert_token_to_id(self, token: str) -> int:
+        """Converts a token (str) to an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index: int) -> str:
+        """Converts an index (integer) to a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """
+        Converts a sequence of tokens (strings) to a single string.
+        
+        This method handles special tokens separately to ensure they are not
+        decoded using the SentencePiece model.
+        
+        Args:
+            tokens (`List[str]`): The list of tokens to convert.
+            
+        Returns:
+            `str`: The decoded string.
+        """
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for i, token in enumerate(tokens):
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special and i != 0:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+        
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+            filename_prefix (`str`, *optional*):
+                An optional prefix to add to the named of the saved files.
+                
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    def build_inputs_with_special_tokens(
+        self, 
+        token_ids_0: List[int], 
+        token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequences for sequence classification tasks by concatenating
+        and adding special tokens.
+        
+        An IQuestCoder sequence has the following format:
+        
+        - single sequence: `<s> X </s>` (if add_eos_token is True) or `<s> X` (default)
+        - pair of sequences: `<s> A </s> <s> B </s>` (if add_eos_token is True) or `<s> A <s> B` (default)
+        
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+                
+        Returns:
+            `List[int]`: List of input IDs with the appropriate special tokens.
+        """
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = bos_token_id + token_ids_0 + eos_token_id
+
+        if token_ids_1 is not None:
+            output = output + bos_token_id + token_ids_1 + eos_token_id
+
+        return output
+
+    def get_special_tokens_mask(
+        self, 
+        token_ids_0: List[int], 
+        token_ids_1: Optional[List[int]] = None, 
+        already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+        
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+                
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        bos_token_id = [1] if self.add_bos_token else []
+        eos_token_id = [1] if self.add_eos_token else []
+
+        if token_ids_1 is None:
+            return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
+        return (
+            bos_token_id
+            + ([0] * len(token_ids_0))
+            + eos_token_id
+            + bos_token_id
+            + ([0] * len(token_ids_1))
+            + eos_token_id
+        )
+
+    def create_token_type_ids_from_sequences(
+        self, 
+        token_ids_0: List[int], 
+        token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task.
+        
+        An IQuestCoder sequence pair mask has the following format:
+        
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+        
+        If `token_ids_1` is `None`, this method only returns the first portion of the mask (0s).
+        
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+                
+        Returns:
+            `List[int]`: List of token type IDs according to the given sequence(s).
+        """
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
+
+        if token_ids_1 is not None:
+            output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
+
+        return output
+
+    @property
+    def default_chat_template(self) -> str:
+        """
+        Returns the default chat template for IQuestCoder.
+        
+        This template formats conversations with system, user, and assistant roles.
+        """
+        return DEFAULT_CHAT_TEMPLATE
+
+    def apply_chat_template(
+        self,
+        conversation: Union[List[Dict[str, str]], "Conversation"],
+        chat_template: Optional[str] = None,
+        add_generation_prompt: bool = False,
+        tokenize: bool = True,
+        padding: bool = False,
+        truncation: bool = False,
+        max_length: Optional[int] = None,
+        return_tensors: Optional[str] = None,
+        return_dict: bool = False,
+        **tokenizer_kwargs,
+    ):
+        """
+        Apply a chat template to format a conversation.
+        
+        Args:
+            conversation (`List[Dict[str, str]]` or `Conversation`):
+                A list of dicts with "role" and "content" keys, representing the conversation history.
+            chat_template (`str`, *optional*):
+                A Jinja template to use for formatting. If not provided, the tokenizer's default will be used.
+            add_generation_prompt (`bool`, *optional*, defaults to `False`):
+                Whether to add a generation prompt at the end for the assistant to continue.
+            tokenize (`bool`, *optional*, defaults to `True`):
+                Whether to tokenize the output. If `False`, returns a string.
+            padding (`bool`, *optional*, defaults to `False`):
+                Whether to pad sequences.
+            truncation (`bool`, *optional*, defaults to `False`):
+                Whether to truncate sequences.
+            max_length (`int`, *optional*):
+                Maximum length of the output.
+            return_tensors (`str`, *optional*):
+                The type of tensors to return ("pt", "tf", "np", or None).
+            return_dict (`bool`, *optional*, defaults to `False`):
+                Whether to return a dictionary with additional information.
+            **tokenizer_kwargs:
+                Additional keyword arguments passed to the tokenizer.
+                
+        Returns:
+            `Union[str, List[int], BatchEncoding]`: The formatted (and optionally tokenized) conversation.
+            
+        Example:
+            ```python
+            >>> tokenizer = IQuestCoderTokenizer.from_pretrained("path/to/model")
+            >>> conversation = [
+            ...     {"role": "system", "content": "You are a helpful assistant."},
+            ...     {"role": "user", "content": "Hello!"},
+            ...     {"role": "assistant", "content": "Hi there! How can I help you today?"},
+            ...     {"role": "user", "content": "What's the weather like?"},
+            ... ]
+            >>> tokenizer.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
+            '<|system|>\\nYou are a helpful assistant.\\n</|system|><|user|>\\nHello!\\n</|user|>...'
+            ```
+        """
+        # Use parent class implementation with our template
+        return super().apply_chat_template(
+            conversation,
+            chat_template=chat_template,
+            add_generation_prompt=add_generation_prompt,
+            tokenize=tokenize,
+            padding=padding,
+            truncation=truncation,
+            max_length=max_length,
+            return_tensors=return_tensors,
+            return_dict=return_dict,
+            **tokenizer_kwargs,
+        )
+
+
+# Try to import and create Fast tokenizer version
+try:
+    from transformers import PreTrainedTokenizerFast
+    from tokenizers import Tokenizer, decoders, models, normalizers, pre_tokenizers, processors
+    
+    class IQuestCoderTokenizerFast(PreTrainedTokenizerFast):
+        """
+        Construct a "fast" IQuestCoder tokenizer (backed by HuggingFace's *tokenizers* library).
+        
+        This is a fast implementation of [`IQuestCoderTokenizer`] using the 🤗 Tokenizers library.
+        
+        Args:
+            vocab_file (`str`, *optional*):
+                Path to the vocabulary file (SentencePiece model).
+            tokenizer_file (`str`, *optional*):
+                Path to a tokenizer JSON file.
+            unk_token (`str`, *optional*, defaults to `"<unk>"`):
+                The unknown token.
+            bos_token (`str`, *optional*, defaults to `"<s>"`):
+                The beginning of sequence token.
+            eos_token (`str`, *optional*, defaults to `"</s>"`):
+                The end of sequence token.
+            pad_token (`str`, *optional*):
+                The token used for padding.
+            add_bos_token (`bool`, *optional*, defaults to `True`):
+                Whether to add a BOS token at the start of sequences.
+            add_eos_token (`bool`, *optional*, defaults to `False`):
+                Whether to add an EOS token at the end of sequences.
+            add_prefix_space (`bool`, *optional*, defaults to `False`):
+                Whether to add an initial space to the input.
+            use_default_system_prompt (`bool`, *optional*, defaults to `False`):
+                Whether to use the default system prompt.
+            chat_template (`str`, *optional*):
+                A Jinja template for formatting conversations.
+                
+        Example:
+            ```python
+            >>> from tokenization_iquestcoder import IQuestCoderTokenizerFast
+            
+            >>> tokenizer = IQuestCoderTokenizerFast.from_pretrained("path/to/model")
+            >>> tokenizer.encode("Hello, world!")
+            [1, 15043, 29892, 3186, 29991]
+            ```
+        """
+        
+        vocab_files_names = VOCAB_FILES_NAMES
+        pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+        max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+        model_input_names = ["input_ids", "attention_mask"]
+        slow_tokenizer_class = IQuestCoderTokenizer
+        
+        def __init__(
+            self,
+            vocab_file=None,
+            tokenizer_file=None,
+            unk_token="<unk>",
+            bos_token="<s>",
+            eos_token="</s>",
+            pad_token=None,
+            add_bos_token=True,
+            add_eos_token=False,
+            add_prefix_space=False,
+            use_default_system_prompt=False,
+            chat_template=None,
+            **kwargs,
+        ):
+            self.add_bos_token = add_bos_token
+            self.add_eos_token = add_eos_token
+            self.add_prefix_space = add_prefix_space
+            self.use_default_system_prompt = use_default_system_prompt
+            
+            if chat_template is None:
+                chat_template = DEFAULT_CHAT_TEMPLATE
+            
+            super().__init__(
+                vocab_file=vocab_file,
+                tokenizer_file=tokenizer_file,
+                unk_token=unk_token,
+                bos_token=bos_token,
+                eos_token=eos_token,
+                pad_token=pad_token,
+                add_bos_token=add_bos_token,
+                add_eos_token=add_eos_token,
+                add_prefix_space=add_prefix_space,
+                use_default_system_prompt=use_default_system_prompt,
+                chat_template=chat_template,
+                **kwargs,
+            )
+        
+        @property
+        def can_save_slow_tokenizer(self) -> bool:
+            return os.path.isfile(self.vocab_file) if self.vocab_file else False
+        
+        @property
+        def default_chat_template(self) -> str:
+            """Returns the default chat template."""
+            return DEFAULT_CHAT_TEMPLATE
+        
+        def build_inputs_with_special_tokens(
+            self, 
+            token_ids_0: List[int], 
+            token_ids_1: Optional[List[int]] = None
+        ) -> List[int]:
+            """Build model inputs with special tokens."""
+            bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+            eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+            output = bos_token_id + token_ids_0 + eos_token_id
+
+            if token_ids_1 is not None:
+                output = output + bos_token_id + token_ids_1 + eos_token_id
+
+            return output
+        
+        def get_special_tokens_mask(
+            self, 
+            token_ids_0: List[int], 
+            token_ids_1: Optional[List[int]] = None, 
+            already_has_special_tokens: bool = False
+        ) -> List[int]:
+            """Retrieve special tokens mask."""
+            if already_has_special_tokens:
+                return super().get_special_tokens_mask(
+                    token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+                )
+
+            bos_token_id = [1] if self.add_bos_token else []
+            eos_token_id = [1] if self.add_eos_token else []
+
+            if token_ids_1 is None:
+                return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
+            return (
+                bos_token_id
+                + ([0] * len(token_ids_0))
+                + eos_token_id
+                + bos_token_id
+                + ([0] * len(token_ids_1))
+                + eos_token_id
+            )
+        
+        def create_token_type_ids_from_sequences(
+            self, 
+            token_ids_0: List[int], 
+            token_ids_1: Optional[List[int]] = None
+        ) -> List[int]:
+            """Create token type IDs from sequences."""
+            bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+            eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+            output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
+
+            if token_ids_1 is not None:
+                output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
+
+            return output
+
+except ImportError:
+    # tokenizers library not available, Fast tokenizer not supported
+    IQuestCoderTokenizerFast = None
+    logger.info(
+        "The `tokenizers` library is not installed. "
+        "IQuestCoderTokenizerFast will not be available. "
+        "Install it with `pip install tokenizers`."
+    )
+

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d3be68e090a927f31e0e378d7599b15c206dd47e4a73933775a746cc9c1cd91
+size 1345108

tokenizer_config.json

@@ -0,0 +1,242 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": true,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": true,
+      "special": true
+    },
+    "75858": {
+      "content": "<CLS>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75859": {
+      "content": "<SEP>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75860": {
+      "content": "<EOD>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75861": {
+      "content": "<MASK>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75862": {
+      "content": "<PAD>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75863": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75864": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75865": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75866": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75867": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75868": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75869": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75870": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75871": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "75872": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "75873": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "75874": {
+      "content": "<tools>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "75875": {
+      "content": "</tools>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "75876": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "75877": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "75878": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "75879": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|CLS|>",
+    "<|SEP|>",
+    "<|EOD|>",
+    "<|MASK|>",
+    "<|PAD|>",
+    "<|fim_prefix|>",
+    "<|fim_middle|>",
+    "<|fim_suffix|>",
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|fim_pad|>",
+    "<|endoftext|>",
+    "<|repo_name|>",
+    "<|file_sep|>",
+    "<think>",
+    "</think>"
+  ],
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_iquestcoder.IQuestCoderTokenizer",
+      null
+    ]
+  },
+  "bos_token": "<s>",
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0].role == 'system' %}\n        {{- messages[0].content + '\\n\\n' }}\n    {%- else %}\n        {{- 'You are LoopCoder, a helpful assistant developed by IQuest.' }}\n    {%- endif %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are LoopCoder, a helpful assistant developed by IQuest.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set content = message.content %}\n        {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- if tool_call.function %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>\\n{\"name\": \"' }}\n                {{- tool_call.name }}\n                {{- '\", \"arguments\": ' }}\n                {%- if tool_call.arguments is string %}\n                    {{- tool_call.arguments }}\n                {%- else %}\n                    {{- tool_call.arguments | tojson }}\n                {%- endif %}\n                {{- '}\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "split_special_tokens": false,
+  "tokenizer_class": "IQuestCoderTokenizer",
+  "unk_token": "<unk>",
+  "use_fast": false
+}