Upload folder using huggingface_hub

.gitattributes

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

chat_template.jinja

@@ -0,0 +1,86 @@
+[gMASK]<sop>
+{%- if tools -%}
+<|system|>
+# Tools
+
+You may call one or more functions to assist with the user query.
+
+You are provided with function signatures within <tools></tools> XML tags:
+<tools>
+{% for tool in tools %}
+{{ tool | tojson(ensure_ascii=False) }}
+{% endfor %}
+</tools>
+
+For each function call, output the function name and arguments within the following XML format:
+<tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call>{%- endif -%}
+{%- macro visible_text(content) -%}
+    {%- if content is string -%}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping -%}
+        {%- for item in content -%}
+            {%- if item is mapping and item.type == 'text' -%}
+                {{- item.text }}
+            {%- elif item is string -%}
+                {{- item }}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- else -%}
+        {{- content }}
+    {%- endif -%}
+{%- endmacro -%}
+{%- set ns = namespace(last_user_index=-1) %}
+{%- for m in messages %}
+    {%- if m.role == 'user' %}
+        {% set ns.last_user_index = loop.index0 -%}
+    {%- endif %}
+{%- endfor %}
+{% for m in messages %}
+{%- if m.role == 'user' -%}<|user|>{{ visible_text(m.content) }}
+{%- elif m.role == 'assistant' -%}
+<|assistant|>
+{%- set reasoning_content = '' %}
+{%- set content = visible_text(m.content) %}
+{%- if m.reasoning_content is string %}
+    {%- set reasoning_content = m.reasoning_content %}
+{%- else %}
+    {%- if '</think>' in content %}
+        {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+        {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+    {%- endif %}
+{%- endif %}
+{%- if ((clear_thinking is defined and not clear_thinking) or loop.index0 > ns.last_user_index) and reasoning_content -%}
+{{ '<think>' + reasoning_content.strip() +  '</think>'}}
+{%- else -%}
+{{ '</think>' }}
+{%- endif -%}
+{%- if content.strip() -%}
+{{ content.strip() }}
+{%- endif -%}
+{% if m.tool_calls %}
+{% for tc in m.tool_calls %}
+{%- if tc.function %}
+    {%- set tc = tc.function %}
+{%- endif %}
+{{- '<tool_call>' + tc.name -}}
+{% set _args = tc.arguments %}{% for k, v in _args.items() %}<arg_key>{{ k }}</arg_key><arg_value>{{ v | tojson(ensure_ascii=False) if v is not string else v }}</arg_value>{% endfor %}</tool_call>{% endfor %}
+{% endif %}
+{%- elif m.role == 'tool' -%}
+{%- if m.content is string -%}
+{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+    {{- '<|observation|>' }}
+{%- endif %}
+{{- '<tool_response>' }}
+{{- m.content }}
+{{- '</tool_response>' }}
+{%- else -%}
+<|observation|>{% for tr in m.content %}
+<tool_response>{{ tr.output if tr.output is defined else tr }}</tool_response>{% endfor -%}
+{% endif -%}
+{%- elif m.role == 'system' -%}
+<|system|>{{ visible_text(m.content) }}
+{%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    <|assistant|>{{- '</think>' if (enable_thinking is defined and not enable_thinking) else '<think>' -}}
+{%- endif -%}

config.json

@@ -0,0 +1,68 @@
+{
+  "architectures": [
+    "RWKV07IMoEForCausalLM"
+  ],
+  "auto_map": {
+      "AutoConfig": "configuration_rwkv07i.RWKV07IConfig",
+      "AutoModelForCausalLM": "modeling_rwkv07i.RWKV07IMoEForCausalLM"
+  },
+  "description": "Prime-RWKV TICA (Tiny Infused Causal Attention)",
+  "base_model": "GLM4.7-Flash",
+  "model_revision": "alpha",
+  "transformer_layers": [-1], 
+  
+  "rwkv_layers": [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46],
+  "tiny_attention_layers": [21, 23, 25, 27, 28, 30, 33, 37, 38, 41, 43],
+  "rwkv_architecture": "hxa07i",
+  "enable_qk_norm": true,
+  "nope_in_transformer": true,
+  "nope_in_rwkv": false,
+  "tiny_head_dim": 128,
+  "tiny_n_heads": 4,
+  "tiny_kv_heads": 2,
+  "lora_rank_decay": 512,
+  "lora_rank_iclr":256,
+  "lora_rank_value_residual_mix":-1,
+  "lora_rank_key_residual_mix":-1,
+  "lora_rank_gate":384,
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "pad_token_id": 154820,
+  "eos_token_id": [
+    154820,
+    154827,
+    154829
+  ],
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "intermediate_size": 10240,
+  "max_position_embeddings": 202752,
+  "model_type": "rwkv07i_moe",
+  "moe_intermediate_size": 1536,
+  "topk_method": "noaux_tc",
+  "norm_topk_prob": true,
+  "num_attention_heads": 20,
+  "n_group": 1,
+  "topk_group": 1,
+  "n_routed_experts": 64,
+  "n_shared_experts": 1,
+  "routed_scaling_factor": 1.8,
+  "num_experts_per_tok": 4,
+  "first_k_dense_replace": 1,
+  "num_hidden_layers": 47,
+  "num_key_value_heads": 20,
+  "num_nextn_predict_layers": 1,
+  "partial_rotary_factor": 1.0,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "tie_word_embeddings": false,
+  "dtype": "bfloat16",
+  "transformers_version": "5.0.0",
+  "q_lora_rank": 768,
+  "kv_lora_rank": 512,
+  "qk_nope_head_dim": 192,
+  "qk_rope_head_dim": 64,
+  "v_head_dim": 256,
+  "vocab_size": 154880
+}

configuration_rwkv07i.py

@@ -0,0 +1,324 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""RWKV07DQwen3 model configuration"""
+
+#Never gonna give you up
+
+from transformers.configuration_utils import PretrainedConfig, layer_type_validation
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+from transformers.configuration_utils import PretrainedConfig, layer_type_validation
+#from transformers.modeling_rope_utils import RopeParameters
+from typing import Optional, TypedDict
+#from transformers.modeling_rope_utils import RopeParameters
+class RopeParameters(TypedDict):
+    """
+    Args:
+        rope_theta (`float`):
+            The base period of the RoPE embeddings.
+        rope_type (`str`, *optional*, defaults to "default"):
+            The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+            'llama3'], with 'default' being the original RoPE implementation.
+        factor (`float`, *optional*):
+            Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+            most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+            original maximum pre-trained length.
+        original_max_position_embeddings (`int`, *optional*):
+            Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+            pretraining.
+        attention_factor (`float`, *optional*):
+            Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+            computation. If unspecified, it defaults to value recommended by the implementation, using the
+            `factor` field to infer the suggested value.
+        beta_fast (`float`, *optional*):
+            Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+            ramp function. If unspecified, it defaults to 32.
+        beta_slow (`float`, *optional*):
+            Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+            ramp function. If unspecified, it defaults to 1.
+        short_factor (`list[float]`, *optional*):
+            Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+            `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+            size divided by the number of attention heads divided by 2
+        long_factor (`list[float]`, *optional*):
+            Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+            `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+            size divided by the number of attention heads divided by 2
+        low_freq_factor (`float`, *optional*):
+            Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+        high_freq_factor (`float`, *optional*):
+            Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+    """
+
+    rope_theta: float
+    rope_type: Optional[str]
+    factor: Optional[float]
+    original_max_position_embeddings: Optional[int]
+    attention_factor: Optional[float]
+    beta_fast: Optional[float]
+    beta_slow: Optional[float]
+    short_factor: Optional[list[float]]
+    long_factor: Optional[list[float]]
+    low_freq_factor: Optional[float]
+    high_freq_factor: Optional[float]
+logger = logging.get_logger(__name__)
+
+
+class RWKV07IConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`RWKV07BModel`]. It is used to instantiate a
+    RWKV079Qwen3 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen3-7B-beta [Qwen/Qwen3-7B-beta](https://huggingface.co/Qwen/Qwen3-7B-beta).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the RWKV079Qwen3 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`RWKV07BModel`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+        lora_rank_decay (`int`, *optional*):
+            The rank of the lora used to generate decay.
+        lora_rank_iclr (`int`, *optional*):
+            The rank of the lora used to generate the in-context learning rate.
+        lora_rank_value_residual_mix (`int`, *optional*):
+            The rank of the lora used to generate the value residual mix amount.
+        lora_rank_value_gate (`int`, *optional*):
+            The rank of the lora used to generate the gate.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import RWKV07BModel, RWKV079Qwen3Config
+
+    >>> # Initializing a RWKV079Qwen3 style configuration
+    >>> configuration = RWKV079Qwen3Config()
+
+    >>> # Initializing a model from the RWKV079Qwen3-7B style configuration
+    >>> model = RWKV07BModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "rwkv07i_moe"
+
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.experts.gate_up_proj": "local_rowwise",
+        "layers.*.mlp.experts.down_proj": "local_rowwise",
+        "layers.*.mlp.experts": "gather",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+    attribute_map = {
+        "num_local_experts": "n_routed_experts",
+    }
+
+    def __init__(
+        self,
+        lora_rank_tokenshift=None,
+        lora_rank_decay=None,
+        lora_rank_iclr=None,
+        lora_rank_value_residual_mix=None,
+        lora_rank_value_key_mix=None,
+        lora_rank_gate=None,
+
+        vocab_size: int | None = 154880,
+        hidden_size: int | None = 2048,
+        intermediate_size: int | None = 10240,
+        moe_intermediate_size: int | None = 1536,
+        num_hidden_layers: int | None = 47,
+        num_attention_heads: int | None = 20,
+        num_key_value_heads: int | None = 20,
+        n_shared_experts: int | None = 1,
+        n_routed_experts: int | None = 64,
+        routed_scaling_factor: float | None = 1.8,
+        kv_lora_rank: int | None = 512,
+        q_lora_rank: int | None = 768,
+        qk_rope_head_dim: int | None = 64,
+        v_head_dim: int | None = 256,
+        qk_nope_head_dim: int | None = 192,
+        n_group: int | None = 1,
+        topk_group: int | None = 1,
+        num_experts_per_tok: int | None = 4,
+        norm_topk_prob: bool | None = True,
+        hidden_act: str | None = "silu",
+        max_position_embeddings: int | None = 202752,
+        initializer_range: float | None = 0.02,
+        rms_norm_eps: int | None = 1e-5,
+        use_cache: bool | None = True,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = 0,
+        eos_token_id: int | None = 1,
+        pretraining_tp: int | None = 1,
+        tie_word_embeddings: bool | None = False,
+        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
+        rope_interleave: bool | None = True,
+        mlp_layer_types=None,
+        attention_bias: bool | None = False,
+        attention_dropout: float | None = 0.0,
+        
+        **kwargs,
+    ):
+
+
+        self.num_key_value_heads = num_key_value_heads
+        self.lora_rank_tokenshift = lora_rank_tokenshift
+        self.lora_rank_decay = lora_rank_decay
+        self.lora_rank_iclr = lora_rank_iclr
+        self.lora_rank_value_residual_mix = lora_rank_value_residual_mix
+        self.lora_rank_gate = lora_rank_gate
+
+        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
+
+        # Default to MoE from the second layer and on
+        self.mlp_layer_types = mlp_layer_types
+        if self.mlp_layer_types is None:
+            self.mlp_layer_types = ["dense"] + ["sparse"] * (self.num_hidden_layers - 1)
+        layer_type_validation(self.mlp_layer_types, self.num_hidden_layers, attention=False)
+
+        self.layer_types = None
+        self.sliding_window = None
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention"
+                if self.sliding_window is not None and i >= self.max_window_layers
+                else "full_attention"
+                for i in range(self.num_hidden_layers)
+            ]
+
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
+        self.head_dim = qk_rope_head_dim
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.norm_topk_prob = norm_topk_prob
+        self.rope_interleave = rope_interleave
+        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.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.rope_parameters = rope_parameters
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.tie_word_embeddings = tie_word_embeddings
+
+        super().__init__(**kwargs)
+
+__all__ = ["RWKV07IConfig"]

modeling_rwkv07i.py

@@ -0,0 +1,1528 @@
+"""
+PyTorch RWKV07I model.
+base code from SmerkyG @ recursal.ai, featherless.ai
+hxa07i implementation RWKV07I Few-Head-Gated-Attention
+
+"""
+
+import math
+import inspect
+from typing import List, Optional, Tuple, Union, Dict, Any
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers import initialization as init
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, CacheLayerMixin
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_kernel_forward_from_hub
+from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
+from transformers.integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernel_func_from_hub
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+
+from transformers.modeling_layers import (
+    GenericForQuestionAnswering,
+    GenericForSequenceClassification,
+    GenericForTokenClassification,
+    GradientCheckpointingLayer,
+)
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple
+from transformers.utils.generic import check_model_inputs
+from transformers.utils.generic import check_model_inputs, is_flash_attention_requested, maybe_autocast
+from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, is_grouped_mm_available
+
+from .configuration_rwkv07i import RWKV07IConfig
+
+import torch
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple
+
+
+@dataclass
+class LayerCache:
+    """
+    Cache storage for a single layer.
+    
+    Supports two independent cache types per layer:
+      - rwkv_state / shift_state:  overwrite semantics (copy_)
+      - attn_key_cache / attn_value_cache: append semantics (cat along seq dim)
+    
+    A pure RWKV layer uses only rwkv_state/shift_state.
+    A pure Attention layer uses only attn_key/value_cache.
+    A hybrid RWKV+small-Attention layer uses all four.
+    """
+    rwkv_state: Optional[torch.Tensor] = None
+    shift_state: Optional[torch.Tensor] = None
+    attn_key_cache: Optional[torch.Tensor] = None
+    attn_value_cache: Optional[torch.Tensor] = None
+
+    @property
+    def has_rwkv_state(self) -> bool:
+        return self.rwkv_state is not None
+
+    @property
+    def has_kv_cache(self) -> bool:
+        return self.attn_key_cache is not None
+
+    @property
+    def kv_seq_length(self) -> int:
+        """Current sequence length stored in the KV cache."""
+        if self.attn_key_cache is not None:
+            return self.attn_key_cache.size(-2)
+        return 0
+
+    def reset_kv(self):
+        """Clear KV cache (e.g. for sliding window truncation)."""
+        self.attn_key_cache = None
+        self.attn_value_cache = None
+
+    def reset_rwkv(self):
+        """Clear RWKV state."""
+        self.rwkv_state = None
+        self.shift_state = None
+
+    def reset(self):
+        """Clear all cached data for this layer."""
+        self.reset_kv()
+        self.reset_rwkv()
+
+
+class RWKV07IState:
+    """
+    Cache manager for RWKV-Attention hybrid models.
+
+    Each layer gets an independent LayerCache that can hold:
+      - RWKV recurrent state  (overwrite on update)
+      - Token shift state     (overwrite on update)
+      - Attention K cache     (append on update)
+      - Attention V cache     (append on update)
+
+    Usage in model forward:
+        # Pure RWKV layer
+        cache.update_rwkv_state(layer_idx, new_state, new_shift, token_count=seq_len)
+
+        # Pure Attention layer
+        cache.update_kv_cache(layer_idx, key, value, token_count=seq_len)
+
+        # Hybrid RWKV layer (RWKV + small Attention)
+        cache.update_rwkv_state(layer_idx, new_state, new_shift, token_count=seq_len)
+        cache.update_kv_cache(layer_idx, key, value)  # token_count=0, already counted
+    """
+
+    def __init__(self) -> None:
+        self._seen_tokens: int = 0
+        self._layers: List[LayerCache] = []
+        # RoPE cache (shared across layers)
+        self.sin: List[torch.Tensor] = []
+        self.cos: List[torch.Tensor] = []
+        self.cumulative_scores: List[torch.Tensor] = []
+
+    # ------------------------------------------------------------------ #
+    #  Internal helpers
+    # ------------------------------------------------------------------ #
+
+    def _ensure_layer(self, layer_idx: int) -> LayerCache:
+        """Ensure LayerCache exists for the given index, padding with empties if needed."""
+        while layer_idx >= len(self._layers):
+            self._layers.append(LayerCache())
+        return self._layers[layer_idx]
+
+    # ------------------------------------------------------------------ #
+    #  Core update methods (separated by semantics & timing)
+    # ------------------------------------------------------------------ #
+
+    @torch.no_grad()
+    def update_rwkv_state(
+        self,
+        layer_idx: int,
+        rwkv_state: torch.Tensor,
+        shift_state: torch.Tensor,
+        token_count: int = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Update RWKV recurrent state — **overwrite** semantics.
+
+        Called AFTER the RWKV state recurrence computation.
+
+        Args:
+            layer_idx:   Index of the layer.
+            rwkv_state:  New recurrent state tensor.
+            shift_state: New token-shift state tensor.
+            token_count: Number of new tokens processed.
+                         Pass seq_len for the first cache-counting call per step,
+                         pass 0 if another update already counted this step.
+
+        Returns:
+            (rwkv_state, shift_state) stored in cache.
+        """
+        if layer_idx == 0:
+            self._seen_tokens += token_count
+
+        cache = self._ensure_layer(layer_idx)
+
+        if cache.rwkv_state is None:
+            # First call — store directly
+            cache.rwkv_state = rwkv_state
+            cache.shift_state = shift_state
+        else:
+            # Subsequent calls — overwrite in-place
+            cache.rwkv_state.copy_(rwkv_state)
+            cache.shift_state.copy_(shift_state)
+
+        return cache.rwkv_state, cache.shift_state
+
+    @torch.no_grad()
+    def update_kv_cache(
+        self,
+        layer_idx: int,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        token_count: int = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Update Attention KV cache — **append** semantics.
+
+        Called AFTER K, V projection.
+
+        Args:
+            layer_idx:   Index of the layer.
+            key:         New key tensor,   shape (..., new_seq_len, head_dim).
+            value:       New value tensor, shape (..., new_seq_len, head_dim).
+            token_count: Number of new tokens processed.
+                         Pass seq_len for the first cache-counting call per step,
+                         pass 0 if another update already counted this step
+                         (e.g. hybrid layer where update_rwkv_state was called first).
+
+        Returns:
+            (full_key_cache, full_value_cache) after concatenation.
+        """
+        if layer_idx == 0:
+            self._seen_tokens += token_count
+
+        cache = self._ensure_layer(layer_idx)
+
+        if cache.attn_key_cache is None:
+            # First call — store directly
+            cache.attn_key_cache = key
+            cache.attn_value_cache = value
+        else:
+            # Subsequent calls — append along sequence dimension
+            cache.attn_key_cache = torch.cat(
+                [cache.attn_key_cache, key], dim=-2
+            )
+            cache.attn_value_cache = torch.cat(
+                [cache.attn_value_cache, value], dim=-2
+            )
+
+        return cache.attn_key_cache, cache.attn_value_cache
+
+    @torch.no_grad()
+    def update(
+        self,
+        kv_state: torch.Tensor,
+        shift_state: torch.Tensor,
+        layer_idx: int,
+        token_count: int = 0,
+        is_attention_layer: bool = True,
+        cache_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Legacy unified update for backward compatibility.
+
+        Dispatches to update_kv_cache or update_rwkv_state based on
+        is_attention_layer flag.
+
+        New code should prefer the explicit update_rwkv_state / update_kv_cache
+        methods directly.
+        """
+        if is_attention_layer:
+            return self.update_kv_cache(
+                layer_idx=layer_idx,
+                key=kv_state,
+                value=shift_state,
+                token_count=token_count,
+            )
+        else:
+            return self.update_rwkv_state(
+                layer_idx=layer_idx,
+                rwkv_state=kv_state,
+                shift_state=shift_state,
+                token_count=token_count,
+            )
+
+    # ------------------------------------------------------------------ #
+    #  Accessors
+    # ------------------------------------------------------------------ #
+
+    def get_layer_cache(self, layer_idx: int) -> LayerCache:
+        """Get the full LayerCache for a given layer."""
+        if layer_idx < len(self._layers):
+            return self._layers[layer_idx]
+        raise KeyError(
+            f"Cache only has {len(self._layers)} layers, "
+            f"attempted to access layer with index {layer_idx}"
+        )
+
+    def get_rwkv_state(self, layer_idx: int) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+        """Get (rwkv_state, shift_state) for a layer. Returns (None, None) if not set."""
+        if layer_idx < len(self._layers):
+            c = self._layers[layer_idx]
+            return c.rwkv_state, c.shift_state
+        return None, None
+
+    def get_kv_cache(self, layer_idx: int) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+        """Get (key_cache, value_cache) for a layer. Returns (None, None) if not set."""
+        if layer_idx < len(self._layers):
+            c = self._layers[layer_idx]
+            return c.attn_key_cache, c.attn_value_cache
+        return None, None
+
+    # ------------------------------------------------------------------ #
+    #  Backward-compatible dict-like interface
+    # ------------------------------------------------------------------ #
+
+    def __getitem__(self, layer_idx: int) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Backward-compatible indexing: cache[layer_idx] returns
+        (layer_kv_states, layer_shift_states).
+
+        For Attention layers this returns (key_cache, value_cache).
+        For RWKV layers this returns (rwkv_state, shift_state).
+        For hybrid layers, this returns KV cache if present, else RWKV state.
+
+        Prefer get_layer_cache / get_rwkv_state / get_kv_cache for new code.
+        """
+        if layer_idx >= len(self._layers):
+            raise KeyError(
+                f"Cache only has {len(self._layers)} layers, "
+                f"attempted to access layer with index {layer_idx}"
+            )
+        c = self._layers[layer_idx]
+        # Priority: KV cache (for seq length queries etc.), fallback to RWKV state
+        if c.has_kv_cache:
+            return c.attn_key_cache, c.attn_value_cache
+        return c.rwkv_state, c.shift_state
+
+    def __iter__(self):
+        for layer_idx in range(len(self._layers)):
+            yield self[layer_idx]
+
+    def __len__(self) -> int:
+        return len(self._layers)
+
+    # ------------------------------------------------------------------ #
+    #  Sequence / shape queries
+    # ------------------------------------------------------------------ #
+
+    def get_seq_length(self, layer_idx: int = 0) -> int:
+        """Returns the total number of tokens seen so far."""
+        return self._seen_tokens
+
+    def get_kv_seq_length(self, layer_idx: int = 0) -> int:
+        """Returns the sequence length stored in KV cache for a specific layer."""
+        if layer_idx < len(self._layers):
+            return self._layers[layer_idx].kv_seq_length
+        return 0
+
+    def get_usable_length(self, new_seq_length: int, layer_idx: Optional[int] = 0) -> int:
+        """Given the sequence length of new inputs, returns the usable length.
+        Linear attention / RWKV layers have no maximum length constraint."""
+        return new_seq_length
+
+    def get_max_cache_shape(self) -> Optional[int]:
+        return None
+
+    def get_max_length(self) -> Optional[int]:
+        return None
+
+    def get_mask_sizes(
+        self, cache_position: torch.Tensor, layer_idx: int
+    ) -> Tuple[int, int]:
+        """Return (kv_length, kv_offset) used to generate the attention mask."""
+        kv_offset = 0
+        query_length = cache_position.shape[0]
+        past_seen_tokens = self.get_seq_length()
+        kv_length = query_length + past_seen_tokens
+        return kv_length, kv_offset
+
+    # ------------------------------------------------------------------ #
+    #  Housekeeping
+    # ------------------------------------------------------------------ #
+
+    def crop(self, max_length: int):
+        """Crop KV caches to max_length. RWKV state is unaffected."""
+        for cache in self._layers:
+            if cache.has_kv_cache and cache.kv_seq_length > max_length:
+                cache.attn_key_cache = cache.attn_key_cache[..., -max_length:, :]
+                cache.attn_value_cache = cache.attn_value_cache[..., -max_length:, :]
+
+    def reset_layer(self, layer_idx: int):
+        """Clear all cached data for a specific layer."""
+        if layer_idx < len(self._layers):
+            self._layers[layer_idx].reset()
+
+    def reset(self):
+        """Clear the entire cache."""
+        self._layers.clear()
+        self._seen_tokens = 0
+        self.sin.clear()
+        self.cos.clear()
+        self.cumulative_scores.clear()
+
+    def reorder_cache(self, beam_idx: torch.LongTensor):
+        """
+        Reorder cache for beam search.
+        KV caches can be reordered; RWKV states are reordered along batch dim.
+        """
+        for cache in self._layers:
+            if cache.has_kv_cache:
+                cache.attn_key_cache = cache.attn_key_cache.index_select(0, beam_idx)
+                cache.attn_value_cache = cache.attn_value_cache.index_select(0, beam_idx)
+            if cache.has_rwkv_state:
+                cache.rwkv_state = cache.rwkv_state.index_select(0, beam_idx)
+                cache.shift_state = cache.shift_state.index_select(0, beam_idx)
+
+    @property
+    def is_compileable(self) -> bool:
+        return True
+
+    # ------------------------------------------------------------------ #
+    #  Debug / inspection
+    # ------------------------------------------------------------------ #
+
+    def summary(self) -> str:
+        """Human-readable summary of cache contents."""
+        lines = [f"HybridCache: {len(self._layers)} layers, {self._seen_tokens} tokens seen"]
+        for i, cache in enumerate(self._layers):
+            parts = []
+            if cache.has_rwkv_state:
+                parts.append(f"rwkv_state={list(cache.rwkv_state.shape)}")
+                parts.append(f"shift_state={list(cache.shift_state.shape)}")
+            if cache.has_kv_cache:
+                parts.append(f"key_cache={list(cache.attn_key_cache.shape)}")
+                parts.append(f"value_cache={list(cache.attn_value_cache.shape)}")
+            if not parts:
+                parts.append("empty")
+            lines.append(f"  layer {i:3d}: {', '.join(parts)}")
+        return "\n".join(lines)
+
+try:
+    from fla.ops.rwkv7.chunk import chunk_rwkv7
+    from fla.ops.rwkv7.fused_recurrent import fused_recurrent_rwkv7
+except ImportError:
+    print("Required module is not installed. Please install it using the following commands:")
+    print("pip install --no-use-pep517 flash-linear-attention")
+    print("Additionally, ensure you have at least version 2.2.0 of Triton installed:")
+    print("pip install triton>=2.2.0")
+
+
+
+def apply_rotary_pos_emb_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    r"""
+    TODO let's just use the original freqcis computation to not have the view
+    transpose + reshape! This is not optimized!
+    Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    b, h, s, d = q.shape
+    q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    b, h, s, d = k.shape
+    k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def yarn_get_mscale(scale=1, mscale=1):
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+# def is_layer_attention(config, layer_id):
+#     return layer_id >= config.first_attention_layer and layer_id < config.first_post_attention_layer and  (layer_id > min(config.num_hidden_layers, config.last_striping_layer) or (min(config.num_hidden_layers-1, config.last_striping_layer) - layer_id) % config.attention_striping == 0)
+
+def is_layer_attention(config, layer_id):
+    return layer_id in config.transformer_layers
+
+def repeat_kv_rwkv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    Repeat KV heads along the head dimension (GQA).
+    Input:  (B, T, H_kv, D)
+    Output: (B, T, H_kv * n_rep, D)
+    """
+    B, T, H_kv, D = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    # Expand head dim
+    hidden_states = hidden_states[:, :, :, None, :]  # (B, T, H_kv, 1, D)
+    hidden_states = hidden_states.expand(B, T, H_kv, n_rep, D)  # (B, T, H_kv, n_rep, D)
+    return hidden_states.reshape(B, T, H_kv * n_rep, D).contiguous()
+
+
+
+class Glm4MoeLiteRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: RWKV07IConfig, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_type = self.config.rope_parameters["rope_type"]
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: RWKV07IConfig | None = None,
+        device: Optional["torch.device"] = None,
+        seq_len: int | None = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Computes the inverse frequencies according to the original RoPE implementation
+        Args:
+            config ([`~transformers.PreTrainedConfig`]):
+                The model configuration.
+            device (`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (`int`, *optional*):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+        """
+        base = config.rope_parameters["rope_theta"]
+        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
+        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        dim = int(head_dim * partial_rotary_factor)
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        # Compute the inverse frequencies
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+class Glm4MoeLiteMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class Glm4MoeLiteTopkRouter(nn.Module):
+    def __init__(self, config: RWKV07IConfig):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.norm_topk_prob = config.norm_topk_prob
+
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size)))
+        self.register_buffer("e_score_correction_bias", torch.zeros((self.n_routed_experts), dtype=torch.float32))
+
+    def forward(self, hidden_states):
+        hidden_states = hidden_states.view(-1, self.config.hidden_size)
+        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
+        return router_logits
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class Glm4MoeLiteRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-5):
+        """
+        Glm4MoeLiteRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+@use_experts_implementation
+class Glm4MoeLiteNaiveMoe(nn.Module):
+    """Collection of expert weights stored as 3D tensors."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.num_experts = config.num_local_experts
+        self.hidden_dim = config.hidden_size
+        self.intermediate_dim = config.moe_intermediate_size
+        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
+        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        top_k_index: torch.Tensor,
+        top_k_weights: torch.Tensor,
+    ) -> torch.Tensor:
+        final_hidden_states = torch.zeros_like(hidden_states)
+        with torch.no_grad():
+            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
+            expert_mask = expert_mask.permute(2, 1, 0)
+            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
+
+        for expert_idx in expert_hit:
+            expert_idx = expert_idx[0]
+            if expert_idx == self.num_experts:
+                continue
+            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
+            current_state = hidden_states[token_idx]
+            gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
+            current_hidden_states = self.act_fn(gate) * up
+            current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
+            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
+            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
+
+        return final_hidden_states
+
+
+class Glm4MoeLiteMoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.experts = Glm4MoeLiteNaiveMoe(config)
+        self.gate = Glm4MoeLiteTopkRouter(config)
+        self.shared_experts = Glm4MoeLiteMLP(
+            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
+        )
+        self.n_routed_experts = config.n_routed_experts
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.norm_topk_prob = config.norm_topk_prob
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.top_k = config.num_experts_per_tok
+
+    def route_tokens_to_experts(self, router_logits):
+        router_logits = router_logits.sigmoid()
+        router_logits_for_choice = router_logits + self.gate.e_score_correction_bias
+        group_scores = (
+            router_logits_for_choice.view(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .topk(2, dim=-1)[0]
+            .sum(dim=-1)
+        )
+        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(-1, self.n_group, self.n_routed_experts // self.n_group)
+            .reshape(-1, self.n_routed_experts)
+        )
+        scores_for_choice = router_logits_for_choice.masked_fill(~score_mask.bool(), 0.0)
+        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
+        topk_weights = router_logits.gather(1, topk_indices)
+        if self.norm_topk_prob:
+            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weights /= denominator
+        topk_weights = topk_weights * self.routed_scaling_factor
+        return topk_indices, topk_weights
+
+    def forward(self, hidden_states):
+        residuals = hidden_states
+        orig_shape = hidden_states.shape
+        router_logits = self.gate(hidden_states)
+        topk_indices, topk_weights = self.route_tokens_to_experts(router_logits)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape)
+        hidden_states = hidden_states + self.shared_experts(residuals)
+        return hidden_states
+
+
+
+def rms_norm(hidden_states, eps = 1e-6):
+    #print('ugyuugyu')
+    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 + eps)
+    return hidden_states.to(input_dtype)
+
+def generate_rotary_embedding(max_seqlen:int, dim:int, theta:float = 10000.0, scale:float = 1):
+    #inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float).to(device) / dim))
+
+    angular_velocity = theta ** -(torch.arange(0, dim, 2, dtype=torch.float) / dim) / scale # frequencies from 1.0 ... 1/theta
+    angles = torch.outer(torch.arange(max_seqlen), angular_velocity)
+    # Different from paper, but it uses a different permutation in order to obtain the same calculation
+    emb = torch.cat((angles, angles), dim=-1)
+    return torch.stack([emb.cos(), emb.sin()], dim=0)
+    #return torch.polar(torch.ones_like(angles), angles)
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb_single(x, cos, sin, unsqueeze_dim=1):
+    return (x * cos.unsqueeze(unsqueeze_dim)) + (rotate_half(x) * sin.unsqueeze(unsqueeze_dim))
+
+from typing import Callable, Optional, Tuple, Union
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+ 
+
+
+      
+    
+
+class RWKV07I_Attention(nn.Module):
+    # this supports
+    # Prime-RWKV mode(FHGA Assist)
+    # Effecient-RWKV mode
+
+    def __init__(self, config: RWKV07IConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.attention_dropout = config.attention_dropout
+        self.num_heads = config.num_attention_heads
+
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.qk_head_dim = config.qk_head_dim
+
+
+        self.is_causal = True
+        if self.q_lora_rank is None:
+            self.receptance_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
+        else:
+            self.receptance_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)
+            self.receptance_a_layernorm = Glm4MoeLiteRMSNorm(config.q_lora_rank,eps=config.rms_norm_eps)
+            self.receptance_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)
+
+        self.keyvalue_a_proj_with_mqa = nn.Linear(
+            config.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.keyvalue_a_layernorm = Glm4MoeLiteRMSNorm(self.kv_lora_rank,eps=config.rms_norm_eps)
+        self.keyvalue_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        self.output_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            config.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        self.scaling = self.qk_head_dim ** (-0.5)
+ 
+
+        lora_rank_decay = config.lora_rank_decay
+        lora_rank_iclr = config.lora_rank_iclr
+        lora_rank_gate = config.lora_rank_gate
+
+        H = self.num_heads * 2
+        N = self.qk_head_dim // 2
+
+        self.w0 = nn.Parameter(torch.empty(1,1,H*N))
+        self.w1 = nn.Parameter(torch.empty(config.hidden_size, lora_rank_decay))
+        self.w2 = nn.Parameter(torch.empty(lora_rank_decay, H*N))
+
+        self.a0 = nn.Parameter(torch.empty(1,1,H*N))
+        self.a1 = nn.Parameter(torch.empty(config.hidden_size, lora_rank_iclr))
+        self.a2 = nn.Parameter(torch.empty(lora_rank_iclr, H*N))
+
+        self.g1 = nn.Parameter(torch.empty(config.hidden_size, lora_rank_gate))
+        self.g2 = nn.Parameter(torch.empty(lora_rank_gate, H*N))
+
+        # ---- Tiny Attention (optional per layer) ----
+        self.TinyAttention = layer_idx in config.tiny_attention_layers
+        if self.TinyAttention:
+            #print(f"[Layer {layer_idx}] tiny Attention enabled")
+            self.tiny_n_heads = config.tiny_n_heads#4
+            self.tiny_head_dim = config.tiny_head_dim#128
+            self.tiny_kv_heads = config.tiny_kv_heads#2
+            self.tiny_kv_groups = self.tiny_n_heads // self.tiny_kv_heads
+
+            self.tiny_q_proj = nn.Linear(config.hidden_size, self.tiny_n_heads * self.tiny_head_dim, bias=False)
+            self.tiny_k_proj = nn.Linear(config.hidden_size, self.tiny_kv_heads * self.tiny_head_dim, bias=False)
+            self.tiny_v_proj = nn.Linear(config.hidden_size, self.tiny_kv_heads * self.tiny_head_dim, bias=False)
+            self.tiny_o_proj = nn.Linear(self.tiny_n_heads * self.tiny_head_dim, config.hidden_size, bias=False)
+
+            self.tiny_q_norm = Glm4MoeLiteRMSNorm(self.tiny_head_dim, eps=config.rms_norm_eps)
+            self.tiny_k_norm = Glm4MoeLiteRMSNorm(self.tiny_head_dim, eps=config.rms_norm_eps)
+
+            # near-zero init so tiny-attention starts as small perturbation
+            # nn.init.zeros_(self.tiny_o_proj.weight)
+            # nn.init.xavier_uniform_(self.tiny_q_proj.weight)
+            # nn.init.xavier_uniform_(self.tiny_k_proj.weight)
+            # nn.init.xavier_uniform_(self.tiny_v_proj.weight)
+
+            # LoRA gate for mixing
+            D_tiny_GATE = 128
+            self.tiny_g1 = nn.Parameter(torch.zeros(config.hidden_size, D_tiny_GATE))
+            self.tiny_g2 = nn.Parameter(torch.randn(D_tiny_GATE, self.tiny_n_heads * self.tiny_head_dim) * 0.01)
+
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: bool = True,
+        v_first: Optional[torch.Tensor] = None, 
+        k_first: Optional[torch.Tensor] = None, 
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+        batch_size, seq_length = hidden_states.shape[:-1]
+        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)
+        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)
+
+        self.head_dim = self.qk_head_dim
+
+        output_shift_state = hidden_states[:, -1:].detach().clone()
+
+        x = hidden_states
+
+        B, T, C = hidden_states.shape
+
+        #RWKV Block uses 128
+        H = self.num_heads * 2
+        N = self.head_dim // 2
+
+        # ============================================================ #
+        #  1. Retrieve cached RWKV state (if any)
+        # ============================================================ #
+
+        input_vk_state = None
+        input_shift_state = None
+
+        if use_cache and past_key_values is not None and len(past_key_values) > self.layer_idx:
+            input_vk_state, input_shift_state = past_key_values.get_rwkv_state(self.layer_idx)
+
+        if input_vk_state is None:
+            input_vk_state = torch.zeros(
+                B, H, N, N,
+                dtype=torch.bfloat16, device=hidden_states.device
+            )
+        if input_shift_state is None:
+            input_shift_state = torch.zeros_like(hidden_states[:, -1:])
+
+        xr = xw = xk = xa = xg = xe = x
+
+        if self.q_lora_rank is None:
+            q_states = self.receptance_proj(hidden_states)
+        else:
+            q_states = self.receptance_b_proj(self.receptance_a_layernorm(self.receptance_a_proj(xr)))
+        q_states = q_states.view(query_shape).transpose(1, 2)
+        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+
+        compressed_kv = self.keyvalue_a_proj_with_mqa(xk)
+        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+
+        k_pass = self.keyvalue_b_proj(self.keyvalue_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
+        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)
+
+        cos, sin = position_embeddings
+        if self.config.nope_in_rwkv == False:
+            if self.config.rope_interleave:  # support using interleaved weights for efficiency
+                q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
+            else:
+                q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
+        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)
+
+        query_states = torch.cat((q_pass, q_rot), dim=-1)
+        key_states = torch.cat((k_pass, k_rot), dim=-1)
+
+        query_states = query_states.transpose(1, 2).contiguous()
+        key_states = key_states.transpose(1, 2).contiguous()
+        value_states = value_states.transpose(1, 2).contiguous()
+
+        
+        q_len = T
+
+        r = query_states.reshape(B,T,-1) 
+        k = key_states.reshape(B,T,-1) 
+        v = value_states.reshape(B,T,-1) 
+ 
+        log_neglog_w = -F.softplus(-(self.w0 + torch.tanh(xw @ self.w1) @ self.w2)) -0.5
+        a = torch.sigmoid(self.a0 + (xa @ self.a1) @ self.a2)
+        g = torch.sigmoid(xg @ self.g1) @ self.g2
+
+
+
+        kk = (k).view(B,T,H,-1).float()
+        kk = (kk / (torch.norm(kk, dim=-1, keepdim=True) + 1e-12)).view(B,T,-1).to(k.dtype)
+
+        w = (-log_neglog_w.float().exp()).exp()        
+        k = k * (1.0 - w + a).to(dtype=torch.bfloat16)
+
+        aa = -kk
+        bb = kk * a
+        w = -log_neglog_w.float().exp()
+
+        r_,w_,k_,v_,aa_,bb_ = [i.view(B,T,H,N) for i in [r,w,k,v,aa,bb]]
+
+        if attention_mask is not None:
+            if attention_mask is not None:
+                if attention_mask.ndim == 2:
+                    # [B, S]
+                    mask = attention_mask[:, -T:]             # [B, T]
+                    v_ = v_ * mask[:, :, None, None]            # → [B, T, 1, 1] に拡張して掛け算
+                elif attention_mask.ndim == 4:
+                    # [B, 1, L, S]
+                    mask = attention_mask[:, 0, -1, -T:]      # [B, T]
+                    v_ = v_ * mask[:, :, None, None]            # 同上
+
+        x, output_vk_state = fused_recurrent_rwkv7(r_, w_, k_, v_, aa_, bb_, scale=1.0, initial_state=input_vk_state, output_final_state=True, head_first=False)
+
+        x = x.view(B,T,-1) * (float(N) ** -0.5)
+
+        if past_key_values is not None:
+            past_key_values.update_rwkv_state(
+                layer_idx=self.layer_idx,
+                rwkv_state=output_vk_state,
+                shift_state=output_shift_state,
+                token_count=T,  # count tokens here
+            )
+        
+        tiny_out = None
+        if self.TinyAttention:
+            # --- Q projection (always from current hidden_states) ---
+            mq = self.tiny_q_norm(
+                self.tiny_q_proj(hidden_states).view(B, T, self.tiny_n_heads, self.tiny_head_dim)
+            ).transpose(1, 2)  # (B, n_heads, T, head_dim)
+
+            # --- K, V projection (current step) ---
+            mk_new = self.tiny_k_norm(
+                self.tiny_k_proj(hidden_states).view(B, T, self.tiny_kv_heads, self.tiny_head_dim)
+            ).transpose(1, 2)  # (B, kv_heads, T, head_dim)
+
+            mv_new = self.tiny_v_proj(hidden_states).view(
+                B, T, self.tiny_kv_heads, self.tiny_head_dim
+            ).transpose(1, 2)  # (B, kv_heads, T, head_dim)
+
+            # ---- Update KV cache (append, AFTER projection) ----
+            if past_key_values is not None:
+                mk_full, mv_full = past_key_values.update_kv_cache(
+                    layer_idx=self.layer_idx,
+                    key=mk_new,
+                    value=mv_new,
+                    token_count=0,  # already counted in update_rwkv_state
+                )
+            else:
+                mk_full, mv_full = mk_new, mv_new
+
+            # GQA expand: (B, kv_heads, S, D) → (B, n_heads, S, D)
+            S = mk_full.size(2)  # full sequence length including past
+            mk_expanded = mk_full[:, :, None, :, :].expand(
+                B, self.tiny_kv_heads, self.tiny_kv_groups, S, self.tiny_head_dim
+            ).reshape(B, self.tiny_n_heads, S, self.tiny_head_dim)
+
+            mv_expanded = mv_full[:, :, None, :, :].expand(
+                B, self.tiny_kv_heads, self.tiny_kv_groups, S, self.tiny_head_dim
+            ).reshape(B, self.tiny_n_heads, S, self.tiny_head_dim)
+
+            # SDPA — use is_causal only during prefill (T > 1)
+            # During generation (T == 1), no causal mask needed
+            # When past KV exists and T > 1, need explicit causal mask
+            if T == 1:
+                # Single-token generation: attend to all cached positions
+                tiny_out = F.scaled_dot_product_attention(
+                    mq, mk_expanded, mv_expanded,
+                    is_causal=False,
+                    dropout_p=self.attention_dropout if self.training else 0.0,
+                )
+            elif past_key_values is None or S == T:
+                # Prefill (no past cache, or first forward): standard causal
+                tiny_out = F.scaled_dot_product_attention(
+                    mq, mk_expanded, mv_expanded,
+                    is_causal=True,
+                    dropout_p=self.attention_dropout if self.training else 0.0,
+                )
+            else:
+                # Chunked prefill with existing cache:
+                # build explicit causal mask allowing full attention to past
+                # Q positions: [past_len, past_len + T)
+                # K positions: [0, past_len + T)
+                past_len = S - T
+                causal_mask = torch.ones(T, S, dtype=torch.bool, device=hidden_states.device).tril(diagonal=past_len)
+                # SDPA expects (B, n_heads, T, S) or broadcastable
+                attn_mask = causal_mask.unsqueeze(0).unsqueeze(0)  # (1, 1, T, S)
+                tiny_out = F.scaled_dot_product_attention(
+                    mq, mk_expanded, mv_expanded,
+                    attn_mask=attn_mask,
+                    dropout_p=self.attention_dropout if self.training else 0.0,
+                )
+
+            # (B, n_heads, T, head_dim) → (B, T, n_heads * head_dim)
+            tiny_out = tiny_out.transpose(1, 2).contiguous().view(B, T, -1)
+
+            # Gated mixing via LoRA
+            tiny_gate = torch.sigmoid(hidden_states @ self.tiny_g1) @ self.tiny_g2
+            tiny_out = self.tiny_o_proj(tiny_out * tiny_gate)
+
+        if tiny_out is not None:
+            x = self.output_proj(x * g) + tiny_out
+        else:
+            x = self.output_proj(x * g)
+
+        return x, v_first, k_first
+    
+
+
+class RWKV07IDecoderLayer(nn.Module):
+    def __init__(self, config: RWKV07IConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.layer_idx = layer_idx
+
+
+        if layer_idx in config.tiny_attention_layers:
+            print(f'layer {layer_idx} : Prime-RWKV')
+        else:
+            print(f'layer {layer_idx} : Effecient-RWKV')
+
+        att_fn = RWKV07I_Attention
+        
+        self.self_attn = att_fn(config, layer_idx)
+
+        #Qwen Variant
+        # if (layer_idx not in config.mlp_only_layers) and (
+        #     config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
+        # ):
+        #     self.mlp = Qwen3MoeSparseMoeBlock(config)
+        # else:
+        #     self.mlp = Qwen3MLP(config)
+
+        #GLM4-Lite variant
+        if config.mlp_layer_types[layer_idx] == "sparse":
+            self.mlp = Glm4MoeLiteMoE(config)
+            print(f'Sparse MoE Mode = {layer_idx}')
+        else:
+            self.mlp = Glm4MoeLiteMLP(config)
+            print(f'Dense MoE Mode = {layer_idx}')
+
+        self.input_layernorm = Glm4MoeLiteRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Glm4MoeLiteRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.attention_type = config.layer_types[layer_idx]
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        frozen_residual: torch.Tensor,
+        v_first: Optional[torch.Tensor],
+        k_first: Optional[torch.Tensor],
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = True,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, sequence_length)` where padding elements are indicated by 0.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_router_logits (`bool`, *optional*):
+                Whether or not to return the logits of all the routers. They are useful for computing the router loss,
+                and should not be returned during inference.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_values (`Cache`, *optional*): cached past key and value projection states
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
+                with `head_dim` being the embedding dimension of each attention head.
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, v_first, k_first = self.self_attn(
+            hidden_states=hidden_states,
+            frozen_residual=frozen_residual,
+            v_first=v_first,
+            k_first=k_first,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+            #is_causal=True,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        # For the MoE layers, we need to unpack
+        if isinstance(hidden_states, tuple):
+            hidden_states, _ = hidden_states
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, v_first,k_first,)
+
+        return outputs
+  
+
+@auto_docstring
+class RWKV07IPreTrainedModel(PreTrainedModel):
+    config: RWKV07IConfig
+    config_class = RWKV07IConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = False
+    _no_split_modules = ["RWKV07IDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+
+    _can_compile_fullgraph = (
+        is_grouped_mm_available()
+    )  # https://huggingface.co/docs/transformers/experts_interface#torchcompile
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": RWKV07IDecoderLayer,
+        #"attentions": Glm4MoeLiteAttention,
+    }
+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]
+
+    # @torch.no_grad()
+    # def _init_weights(self, module):
+    #     super()._init_weights(module)
+    #     if isinstance(module, Glm4MoeLiteTopkRouter):
+    #         init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+    #         init.zeros_(module.e_score_correction_bias)
+    #     elif isinstance(module, Glm4MoeLiteNaiveMoe):
+    #         init.normal_(module.gate_up_proj, mean=0.0, std=self.config.initializer_range)
+    #         init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range)
+
+    # 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_()
+
+
+
+@auto_docstring
+class RWKV07IModel(RWKV07IPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen3DecoderLayer`]
+
+    Args:
+        config: RWKV07EConfig
+    """
+
+    def __init__(self, config: RWKV07IConfig):
+        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(
+            [RWKV07IDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Glm4MoeLiteRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Glm4MoeLiteRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+        self.has_sliding_layers = "sliding_attention" in self.config.layer_types
+        
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> 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
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and not isinstance(past_key_values, RWKV07IState):
+            past_key_values = RWKV07IState()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        # It may already have been prepared by e.g. `generate`
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            # Prepare mask arguments
+            mask_kwargs = {
+                "config": self.config,
+                "input_embeds": inputs_embeds,
+                "attention_mask": attention_mask,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "position_ids": position_ids,
+            }
+            # Create the masks
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(**mask_kwargs),
+            }
+            # The sliding window alternating layers are not always activated depending on the config
+            if self.has_sliding_layers:
+                causal_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        #if self.config.use_rope:
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        # else:
+        #     position_embeddings = None
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+        v_first = None
+        k_first = None
+        frozen_residual = None
+        
+        for decoder_layer in self.layers:
+            if not is_layer_attention(self.config, decoder_layer.layer_idx):
+                frozen_residual = hidden_states#rms_norm(hidden_states)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            attention_mask = causal_mask_mapping[decoder_layer.attention_type]
+            if attention_mask is not None and attention_mask.ndim == 1:
+                attention_mask = None
+            #attention_mask = None
+
+            layer_outputs = decoder_layer(
+                hidden_states,
+                frozen_residual=frozen_residual,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                v_first=v_first,
+                k_first=k_first
+            )
+
+            hidden_states = layer_outputs[0]
+            v_first = layer_outputs[1]
+            k_first = layer_outputs[2]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[2],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        #if return_legacy_cache:
+        #    next_cache = next_cache.to_legacy_cache()
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+class RWKV07IMoEForCausalLM(RWKV07IPreTrainedModel, GenerationMixin):
+
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = RWKV07IModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = True,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **loss_kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            num_logits_to_keep (`int`, *optional*):
+                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, RWKV07EQwen3ForCausalLM
+
+        >>> model = RWKV07EQwen3ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        # # run the prefill only up to the last token, then run one more for the actual result
+        # # we do this so that called code doesn't have to handle the dichotomy specially and can just check for L==1
+        # for i in range(2):
+        #     all_but_one = max(1, input_ids.size(-1)-1)
+        #     iid = input_ids[..., i*all_but_one:(i+1)*all_but_one]
+        #     if iid.size(-1) == 0: 
+        #         continue
+        #     pids = position_ids
+        #     if pids is not None:
+        #         pids = position_ids[..., i*all_but_one:(i+1)*all_but_one]
+        #     cp = cache_position
+        #     if cp is not None:
+        #         cp = cache_position[..., i*all_but_one:(i+1)*all_but_one]
+        #     rv = self.forward_inner(iid, attention_mask=attention_mask, position_ids=pids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, cache_position=cp, num_logits_to_keep=num_logits_to_keep, **loss_kwargs)
+        #     past_key_values = rv.past_key_values
+    #     return rv
+
+    # def forward_inner(
+    #     self,
+    #     input_ids: torch.LongTensor = None,
+    #     attention_mask: Optional[torch.Tensor] = None,
+    #     position_ids: Optional[torch.LongTensor] = None,
+    #     past_key_values: Optional[List[torch.FloatTensor]] = None,
+    #     inputs_embeds: Optional[torch.FloatTensor] = None,
+    #     labels: Optional[torch.LongTensor] = None,
+    #     use_cache: Optional[bool] = None,
+    #     output_attentions: Optional[bool] = None,
+    #     output_hidden_states: Optional[bool] = None,
+    #     cache_position: Optional[torch.LongTensor] = None,
+    #     num_logits_to_keep: int = 0,
+    #     **loss_kwargs,
+    # ) -> 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
+        )
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.vocab_size, **loss_kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+@auto_docstring
+class RWKV07IQwen3ForSequenceClassification(RWKV07IPreTrainedModel):
+    pass
+
+@auto_docstring
+class RWKV07IQwen3ForTokenClassification(RWKV07IPreTrainedModel):
+    pass
+
+@auto_docstring
+class RWKV07IQwen3ForQuestionAnswering(RWKV07IPreTrainedModel):
+    base_model_prefix = "transformer"  # For BC, where `transformer` was used instead of `model`

requirements.txt

@@ -0,0 +1,4 @@
+transformers==5.0.0
+flash-linear-attention
+fastapi
+uvicorn

teacher.txt

@@ -0,0 +1,79 @@
+hf (pretrained=/workspace/llm/GLM-4.7-Flash,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 32
+|                 Tasks                 |Version|Filter|n-shot|Metric|   |Value |   |Stderr|
+|---------------------------------------|------:|------|-----:|------|---|-----:|---|-----:|
+|mmlu                                   |      2|none  |      |acc   |↑  |0.7072|±  |0.0035|
+| - humanities                          |      2|none  |      |acc   |↑  |0.6113|±  |0.0063|
+|  - formal_logic                       |      1|none  |     0|acc   |↑  |0.5873|±  |0.0440|
+|  - high_school_european_history       |      1|none  |     0|acc   |↑  |0.7939|±  |0.0316|
+|  - high_school_us_history             |      1|none  |     0|acc   |↑  |0.8725|±  |0.0234|
+|  - high_school_world_history          |      1|none  |     0|acc   |↑  |0.8692|±  |0.0219|
+|  - international_law                  |      1|none  |     0|acc   |↑  |0.8512|±  |0.0325|
+|  - jurisprudence                      |      1|none  |     0|acc   |↑  |0.7870|±  |0.0396|
+|  - logical_fallacies                  |      1|none  |     0|acc   |↑  |0.8160|±  |0.0304|
+|  - moral_disputes                     |      1|none  |     0|acc   |↑  |0.7630|±  |0.0229|
+|  - moral_scenarios                    |      1|none  |     0|acc   |↑  |0.2380|±  |0.0142|
+|  - philosophy                         |      1|none  |     0|acc   |↑  |0.8006|±  |0.0227|
+|  - prehistory                         |      1|none  |     0|acc   |↑  |0.8210|±  |0.0213|
+|  - professional_law                   |      1|none  |     0|acc   |↑  |0.5385|±  |0.0127|
+|  - world_religions                    |      1|none  |     0|acc   |↑  |0.8655|±  |0.0262|
+| - other                               |      2|none  |      |acc   |↑  |0.7757|±  |0.0071|
+|  - business_ethics                    |      1|none  |     0|acc   |↑  |0.7300|±  |0.0446|
+|  - clinical_knowledge                 |      1|none  |     0|acc   |↑  |0.7585|±  |0.0263|
+|  - college_medicine                   |      1|none  |     0|acc   |↑  |0.7514|±  |0.0330|
+|  - global_facts                       |      1|none  |     0|acc   |↑  |0.3700|±  |0.0485|
+|  - human_aging                        |      1|none  |     0|acc   |↑  |0.7444|±  |0.0293|
+|  - management                         |      1|none  |     0|acc   |↑  |0.8641|±  |0.0339|
+|  - marketing                          |      1|none  |     0|acc   |↑  |0.9017|±  |0.0195|
+|  - medical_genetics                   |      1|none  |     0|acc   |↑  |0.8600|±  |0.0349|
+|  - miscellaneous                      |      1|none  |     0|acc   |↑  |0.8710|±  |0.0120|
+|  - nutrition                          |      1|none  |     0|acc   |↑  |0.8072|±  |0.0226|
+|  - professional_accounting            |      1|none  |     0|acc   |↑  |0.5780|±  |0.0295|
+|  - professional_medicine              |      1|none  |     0|acc   |↑  |0.8493|±  |0.0217|
+|  - virology                           |      1|none  |     0|acc   |↑  |0.5663|±  |0.0386|
+| - social sciences                     |      2|none  |      |acc   |↑  |0.8203|±  |0.0068|
+|  - econometrics                       |      1|none  |     0|acc   |↑  |0.6228|±  |0.0456|
+|  - high_school_geography              |      1|none  |     0|acc   |↑  |0.8333|±  |0.0266|
+|  - high_school_government_and_politics|      1|none  |     0|acc   |↑  |0.9223|±  |0.0193|
+|  - high_school_macroeconomics         |      1|none  |     0|acc   |↑  |0.7641|±  |0.0215|
+|  - high_school_microeconomics         |      1|none  |     0|acc   |↑  |0.8782|±  |0.0212|
+|  - high_school_psychology             |      1|none  |     0|acc   |↑  |0.9064|±  |0.0125|
+|  - human_sexuality                    |      1|none  |     0|acc   |↑  |0.8550|±  |0.0309|
+|  - professional_psychology            |      1|none  |     0|acc   |↑  |0.7778|±  |0.0168|
+|  - public_relations                   |      1|none  |     0|acc   |↑  |0.7091|±  |0.0435|
+|  - security_studies                   |      1|none  |     0|acc   |↑  |0.7510|±  |0.0277|
+|  - sociology                          |      1|none  |     0|acc   |↑  |0.8507|±  |0.0252|
+|  - us_foreign_policy                  |      1|none  |     0|acc   |↑  |0.8800|±  |0.0327|
+| - stem                                |      2|none  |      |acc   |↑  |0.6727|±  |0.0080|
+|  - abstract_algebra                   |      1|none  |     0|acc   |↑  |0.5400|±  |0.0501|
+|  - anatomy                            |      1|none  |     0|acc   |↑  |0.7037|±  |0.0394|
+|  - astronomy                          |      1|none  |     0|acc   |↑  |0.8289|±  |0.0306|
+|  - college_biology                    |      1|none  |     0|acc   |↑  |0.8264|±  |0.0317|
+|  - college_chemistry                  |      1|none  |     0|acc   |↑  |0.5600|±  |0.0499|
+|  - college_computer_science           |      1|none  |     0|acc   |↑  |0.6400|±  |0.0482|
+|  - college_mathematics                |      1|none  |     0|acc   |↑  |0.4700|±  |0.0502|
+|  - college_physics                    |      1|none  |     0|acc   |↑  |0.5490|±  |0.0495|
+|  - computer_security                  |      1|none  |     0|acc   |↑  |0.7900|±  |0.0409|
+|  - conceptual_physics                 |      1|none  |     0|acc   |↑  |0.8255|±  |0.0248|
+|  - electrical_engineering             |      1|none  |     0|acc   |↑  |0.7310|±  |0.0370|
+|  - elementary_mathematics             |      1|none  |     0|acc   |↑  |0.6270|±  |0.0249|
+|  - high_school_biology                |      1|none  |     0|acc   |↑  |0.8677|±  |0.0193|
+|  - high_school_chemistry              |      1|none  |     0|acc   |↑  |0.6946|±  |0.0324|
+|  - high_school_computer_science       |      1|none  |     0|acc   |↑  |0.8300|±  |0.0378|
+|  - high_school_mathematics            |      1|none  |     0|acc   |↑  |0.4148|±  |0.0300|
+|  - high_school_physics                |      1|none  |     0|acc   |↑  |0.5894|±  |0.0402|
+|  - high_school_statistics             |      1|none  |     0|acc   |↑  |0.6111|±  |0.0332|
+|  - machine_learning                   |      1|none  |     0|acc   |↑  |0.5536|±  |0.0472|
+
+|      Groups      |Version|Filter|n-shot|Metric|   |Value |   |Stderr|
+|------------------|------:|------|------|------|---|-----:|---|-----:|
+|mmlu              |      2|none  |      |acc   |↑  |0.7072|±  |0.0035|
+| - humanities     |      2|none  |      |acc   |↑  |0.6113|±  |0.0063|
+| - other          |      2|none  |      |acc   |↑  |0.7757|±  |0.0071|
+| - social sciences|      2|none  |      |acc   |↑  |0.8203|±  |0.0068|
+| - stem           |      2|none  |      |acc   |↑  |0.6727|±  |0.0080|
+
+hf (pretrained=/workspace/llm/GLM-4.7-Flash,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 32
+|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
+|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
+|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8264|±  |0.0104|
+|     |       |strict-match    |     5|exact_match|↑  |0.8271|±  |0.0104|

test.py

@@ -0,0 +1,47 @@
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+#quantization_config = BitsAndBytesConfig(load_in_8bit=True)
+
+MODEL_PATH = "/workspace/output/glm4_7_30b/hf_temp_07i/"
+#MODEL_PATH = "/workspace/llm/GLM-4.7-Flash/"
+
+messages = [{"role": "user", "content": "who is rick astley?"}]
+tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH
+    ,torch_dtype="auto",
+    device_map="auto",
+    trust_remote_code=True,
+    #quantization_config=quantization_config
+)
+inputs = tokenizer.apply_chat_template(
+    messages,
+    tokenize=True,
+    add_generation_prompt=True,
+    return_dict=True,
+    enable_thinking=False,
+    
+    return_tensors="pt",
+)
+
+print(type(tokenizer))
+print("chat_template is None?", tokenizer.chat_template is None)
+print("chat_template head:\n", (tokenizer.chat_template or "")[:400])
+
+print(inputs)
+
+
+print('---------------------------')
+print(tokenizer.decode(inputs['input_ids']))
+#exit()
+model = AutoModelForCausalLM.from_pretrained(
+    pretrained_model_name_or_path=MODEL_PATH,
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    trust_remote_code=True,
+)
+inputs = inputs.to(model.device)
+generated_ids = model.generate(**inputs, max_new_tokens=256,use_cache=True, do_sample=True)
+output_text = tokenizer.decode(generated_ids[0][inputs.input_ids.shape[1]:])
+
+print('--------------------------------------------------------------------------------------')
+print(output_text)
+ 

test_client_api.py

@@ -0,0 +1,234 @@
+import requests
+import json
+import time
+import sys
+
+BASE_URL = "http://localhost:8000/v1"
+MODEL_NAME = "RWKV-GLM-4.7-Flash-Preview-v0.1"
+
+# ==========================================================
+# Utility
+# ==========================================================
+def print_section(title):
+    print("\n" + "=" * 60)
+    print(title)
+    print("=" * 60)
+
+
+def safe_json(resp):
+    try:
+        return resp.json()
+    except:
+        print("❌ JSON decode failed")
+        print(resp.text)
+        sys.exit(1)
+
+
+# ==========================================================
+# 1️⃣ Models API
+# ==========================================================
+def test_models():
+    print_section("TEST: /v1/models")
+
+    resp = requests.get(f"{BASE_URL}/models")
+    assert resp.status_code == 200, "Models API failed"
+
+    data = safe_json(resp)
+
+    assert "data" in data, "No model list returned"
+    assert len(data["data"]) > 0, "Empty model list"
+
+    print("✅ Models endpoint OK")
+    print("Available models:", [m["id"] for m in data["data"]])
+
+
+# ==========================================================
+# 2️⃣ Non-stream basic
+# ==========================================================
+def test_basic_completion():
+    print_section("TEST: Basic Non-Streaming Completion")
+
+    payload = {
+        "model": MODEL_NAME,
+        "messages": [{"role": "user", "content": "Say hello."}],
+        "max_tokens": 30,
+        "stream": False
+    }
+
+    resp = requests.post(
+        f"{BASE_URL}/chat/completions",
+        headers={"Content-Type": "application/json"},
+        data=json.dumps(payload)
+    )
+
+    assert resp.status_code == 200, "Completion failed"
+
+    data = safe_json(resp)
+
+    assert "choices" in data, "No choices returned"
+    assert "usage" in data, "No usage returned"
+
+    print("Assistant:", data["choices"][0]["message"]["content"])
+    print("Usage:", data["usage"])
+    print("✅ Basic completion OK")
+
+
+# ==========================================================
+# 3️⃣ Streaming
+# ==========================================================
+def test_streaming():
+    print_section("TEST: Streaming Completion")
+
+    payload = {
+        "model": MODEL_NAME,
+        "messages": [{"role": "user", "content": "Count from 1 to 5."}],
+        "max_tokens": 50,
+        "stream": True
+    }
+
+    full_text = ""
+
+    with requests.post(
+        f"{BASE_URL}/chat/completions",
+        headers={"Content-Type": "application/json"},
+        data=json.dumps(payload),
+        stream=True
+    ) as resp:
+
+        assert resp.status_code == 200, "Streaming failed"
+
+        for line in resp.iter_lines():
+            if line:
+                decoded = line.decode("utf-8")
+
+                if decoded.startswith("data: "):
+                    content = decoded[len("data: "):]
+
+                    if content == "[DONE]":
+                        break
+
+                    chunk = json.loads(content)
+                    delta = chunk["choices"][0]["delta"]
+
+                    if "content" in delta:
+                        print(delta["content"], end="", flush=True)
+                        full_text += delta["content"]
+
+    print("\n\n✅ Streaming OK")
+    assert len(full_text) > 0, "Streaming returned empty"
+
+
+# ==========================================================
+# 4️⃣ Sampling Variations
+# ==========================================================
+def test_sampling_variations():
+    print_section("TEST: Sampling Variations")
+
+    base_payload = {
+        "model": MODEL_NAME,
+        "messages": [{"role": "user", "content": "Write a creative sentence about AI."}],
+        "max_tokens": 50,
+        "stream": False
+    }
+
+    configs = [
+        {"temperature": 0.0},
+        {"temperature": 0.7},
+        {"top_p": 0.8},
+        {"top_k": 20},
+        {"repetition_penalty": 1.2},
+        {"presence_penalty": 0.5},
+        {"frequency_penalty": 0.5}
+    ]
+
+    for cfg in configs:
+        payload = base_payload.copy()
+        payload.update(cfg)
+
+        resp = requests.post(
+            f"{BASE_URL}/chat/completions",
+            headers={"Content-Type": "application/json"},
+            data=json.dumps(payload)
+        )
+
+        assert resp.status_code == 200, f"Sampling failed: {cfg}"
+
+        data = safe_json(resp)
+
+        text = data["choices"][0]["message"]["content"]
+
+        print(f"\nConfig: {cfg}")
+        print("Output:", text[:120], "...")
+
+    print("\n✅ Sampling parameter variations OK")
+
+
+# ==========================================================
+# 5️⃣ Deterministic Check (temperature=0)
+# ==========================================================
+def test_deterministic():
+    print_section("TEST: Deterministic Mode (temperature=0)")
+
+    payload = {
+        "model": MODEL_NAME,
+        "messages": [{"role": "user", "content": "Define gravity in one sentence."}],
+        "temperature": 0.0,
+        "max_tokens": 50,
+        "stream": False
+    }
+
+    resp1 = requests.post(f"{BASE_URL}/chat/completions",
+                          headers={"Content-Type": "application/json"},
+                          data=json.dumps(payload))
+    resp2 = requests.post(f"{BASE_URL}/chat/completions",
+                          headers={"Content-Type": "application/json"},
+                          data=json.dumps(payload))
+
+    out1 = safe_json(resp1)["choices"][0]["message"]["content"]
+    out2 = safe_json(resp2)["choices"][0]["message"]["content"]
+
+    print("Run1:", out1)
+    print("Run2:", out2)
+
+    assert out1 == out2, "❌ Deterministic mode not deterministic"
+    print("✅ Deterministic check OK")
+
+
+# ==========================================================
+# 6️⃣ Error Handling
+# ==========================================================
+def test_error_handling():
+    print_section("TEST: Error Handling")
+
+    payload = {
+        "model": MODEL_NAME,
+        # missing messages intentionally
+    }
+
+    resp = requests.post(
+        f"{BASE_URL}/chat/completions",
+        headers={"Content-Type": "application/json"},
+        data=json.dumps(payload)
+    )
+
+    if resp.status_code != 200:
+        print("✅ Server correctly handled bad request")
+    else:
+        print("⚠️ Warning: server did not reject bad request")
+
+
+# ==========================================================
+# Main
+# ==========================================================
+if __name__ == "__main__":
+    start = time.time()
+
+    test_models()
+    test_basic_completion()
+    test_streaming()
+    test_sampling_variations()
+    test_deterministic()
+    test_error_handling()
+
+    print_section("ALL TESTS PASSED")
+    print(f"Total time: {round(time.time() - start, 2)} sec")

test_openai_api.py

@@ -0,0 +1,242 @@
+import time
+import json
+import uuid
+import torch
+from threading import Thread, Event
+from fastapi import FastAPI, Request
+from fastapi.responses import StreamingResponse
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    TextIteratorStreamer,
+    LogitsProcessor,
+    LogitsProcessorList,
+    StoppingCriteria,
+    StoppingCriteriaList,
+)
+
+# ==========================================================
+# 設定
+# ==========================================================
+MODEL_ID = "/workspace/output/glm4_7_30b/hf_temp_07i"
+VIEW_NAME = "RWKV-GLM-4.7-Flash"
+HOST = "0.0.0.0"
+PORT = 8000
+
+# ==========================================================
+# モデルロード
+# ==========================================================
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID,
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    trust_remote_code=True,
+)
+
+app = FastAPI()
+
+
+# ==========================================================
+# Logits Processors
+# ==========================================================
+class PresencePenaltyProcessor(LogitsProcessor):
+    def __init__(self, penalty):
+        self.penalty = penalty
+
+    def __call__(self, input_ids, scores):
+        for batch_idx in range(input_ids.shape[0]):
+            unique_tokens = torch.unique(input_ids[batch_idx])
+            scores[batch_idx, unique_tokens] -= self.penalty
+        return scores
+
+
+class FrequencyPenaltyProcessor(LogitsProcessor):
+    def __init__(self, penalty):
+        self.penalty = penalty
+
+    def __call__(self, input_ids, scores):
+        for batch_idx in range(input_ids.shape[0]):
+            token_counts = torch.bincount(
+                input_ids[batch_idx], minlength=scores.shape[-1]
+            )
+            scores[batch_idx] -= token_counts * self.penalty
+        return scores
+
+
+# ==========================================================
+# Cancellable Stopping Criteria
+# ==========================================================
+class CancelledStoppingCriteria(StoppingCriteria):
+    """threading.Event がセットされたら生成を打ち切る"""
+
+    def __init__(self, stop_event: Event):
+        self.stop_event = stop_event
+
+    def __call__(self, input_ids, scores, **kwargs):
+        return self.stop_event.is_set()
+
+
+# ==========================================================
+# Models Endpoint
+# ==========================================================
+@app.get("/v1/models")
+async def list_models():
+    return {
+        "object": "list",
+        "data": [
+            {
+                "id": VIEW_NAME,
+                "object": "model",
+                "created": int(time.time()),
+                "owned_by": "local",
+            }
+        ],
+    }
+
+
+# ==========================================================
+# Chat Completions Endpoint
+# ==========================================================
+@app.post("/v1/chat/completions")
+async def chat_completions(request: Request):
+    body = await request.json()
+
+    model_name = body.get("model", MODEL_ID)
+    messages = body["messages"]
+    stream = body.get("stream", False)
+
+    temperature = body.get("temperature", 1.0)
+    top_p = body.get("top_p", 1.0)
+    top_k = body.get("top_k", 50)
+    repetition_penalty = body.get("repetition_penalty", 1.0)
+    presence_penalty = body.get("presence_penalty", 0.0)
+    frequency_penalty = body.get("frequency_penalty", 0.0)
+    max_tokens = body.get("max_tokens", 2048)
+
+    prompt = tokenizer.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True#,enable_thinking=False
+    )
+    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+
+    processors = LogitsProcessorList()
+    if presence_penalty > 0:
+        processors.append(PresencePenaltyProcessor(presence_penalty))
+    if frequency_penalty > 0:
+        processors.append(FrequencyPenaltyProcessor(frequency_penalty))
+
+    generate_kwargs = dict(
+        **inputs,
+        max_new_tokens=max_tokens,
+        temperature=temperature,
+        top_p=top_p,
+        top_k=top_k,
+        repetition_penalty=repetition_penalty,
+        logits_processor=processors,
+        do_sample=temperature > 0,
+        use_cache=True,  # 生成時に明示的に有効化
+    )
+
+    # ================= Non-stream =================
+    if not stream:
+        outputs = model.generate(**generate_kwargs)
+        completion_tokens = outputs.shape[1] - inputs["input_ids"].shape[1]
+        generated_text = tokenizer.decode(
+            outputs[0][inputs["input_ids"].shape[1] :], skip_special_tokens=False
+        )
+        return {
+            "id": f"chatcmpl-{uuid.uuid4().hex}",
+            "object": "chat.completion",
+            "created": int(time.time()),
+            "model": model_name,
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {"role": "assistant", "content": generated_text},
+                    "finish_reason": "stop",
+                }
+            ],
+            "usage": {
+                "prompt_tokens": inputs["input_ids"].shape[1],
+                "completion_tokens": completion_tokens,
+                "total_tokens": inputs["input_ids"].shape[1] + completion_tokens,
+            },
+        }
+
+    # ================= Streaming =================
+    stop_event = Event()
+
+    stopping_criteria = StoppingCriteriaList(
+        [CancelledStoppingCriteria(stop_event)]
+    )
+
+    streamer = TextIteratorStreamer(
+        tokenizer, skip_prompt=True, skip_special_tokens=True
+    )
+
+    generation_kwargs = dict(
+        **generate_kwargs,
+        streamer=streamer,
+        stopping_criteria=stopping_criteria,
+    )
+
+    thread = Thread(target=model.generate, kwargs=generation_kwargs)
+    thread.start()
+
+    async def event_generator():
+        completion_id = f"chatcmpl-{uuid.uuid4().hex}"
+        firsttime = "<think>"
+        cancelled = False
+
+        try:
+            for new_text in streamer:
+                if await request.is_disconnected():
+                    stop_event.set()
+                    cancelled = True
+                    break
+
+                chunk = {
+                    "id": completion_id,
+                    "object": "chat.completion.chunk",
+                    "created": int(time.time()),
+                    "model": model_name,
+                    "choices": [
+                        {
+                            "index": 0,
+                            "delta": {"content": firsttime + new_text},
+                            "finish_reason": None,
+                        }
+                    ],
+                }
+                firsttime = ""
+                yield f"data: {json.dumps(chunk)}\n\n"
+
+            if not cancelled:
+                yield "data: [DONE]\n\n"
+
+        except Exception:
+            stop_event.set()
+            cancelled = True
+        finally:
+            if cancelled:
+                for _ in streamer:
+                    pass
+            thread.join(timeout=10)
+
+    return StreamingResponse(
+        event_generator(), media_type="text/event-stream"
+    )
+
+
+# ==========================================================
+# Python実行時に自動起動
+# ==========================================================
+if __name__ == "__main__":
+    import uvicorn
+
+    uvicorn.run(
+        "test_openai_api:app",
+        host=HOST,
+        port=PORT,
+        reload=False,
+    )

tokenizer_config.json

@@ -0,0 +1,321 @@
+{
+  "added_tokens_decoder": {
+    "154820": {
+      "content": "<|endoftext|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154821": {
+      "content": "[MASK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154822": {
+      "content": "[gMASK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154823": {
+      "content": "[sMASK]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154824": {
+      "content": "<sop>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154825": {
+      "content": "<eop>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154826": {
+      "content": "<|system|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154827": {
+      "content": "<|user|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154828": {
+      "content": "<|assistant|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154829": {
+      "content": "<|observation|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154830": {
+      "content": "<|begin_of_image|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154831": {
+      "content": "<|end_of_image|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154832": {
+      "content": "<|begin_of_video|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154833": {
+      "content": "<|end_of_video|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154834": {
+      "content": "<|begin_of_audio|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154835": {
+      "content": "<|end_of_audio|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154836": {
+      "content": "<|begin_of_transcription|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154837": {
+      "content": "<|end_of_transcription|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "154838": {
+      "content": "<|code_prefix|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154839": {
+      "content": "<|code_middle|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154840": {
+      "content": "<|code_suffix|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154841": {
+      "content": "<think>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154842": {
+      "content": "</think>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154843": {
+      "content": "<tool_call>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154844": {
+      "content": "</tool_call>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154845": {
+      "content": "<tool_response>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154846": {
+      "content": "</tool_response>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154847": {
+      "content": "<arg_key>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154848": {
+      "content": "</arg_key>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154849": {
+      "content": "<arg_value>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154850": {
+      "content": "</arg_value>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154851": {
+      "content": "/nothink",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154852": {
+      "content": "<|begin_of_box|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154853": {
+      "content": "<|end_of_box|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154854": {
+      "content": "<|image|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    },
+    "154855": {
+      "content": "<|video|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|endoftext|>",
+    "[MASK]",
+    "[gMASK]",
+    "[sMASK]",
+    "<sop>",
+    "<eop>",
+    "<|system|>",
+    "<|user|>",
+    "<|assistant|>",
+    "<|observation|>",
+    "<|begin_of_image|>",
+    "<|end_of_image|>",
+    "<|begin_of_video|>",
+    "<|end_of_video|>",
+    "<|begin_of_audio|>",
+    "<|end_of_audio|>",
+    "<|begin_of_transcription|>",
+    "<|end_of_transcription|>"
+  ],
+  "clean_up_tokenization_spaces": false,
+  "do_lower_case": false,
+  "eos_token": "<|endoftext|>",
+  "extra_special_tokens": {},
+  "model_max_length": 128000,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "left",
+  "remove_space": false,
+  "tokenizer_class": "PreTrainedTokenizer"
+}