Upload folder using huggingface_hub

YuLan-Mini-Nanbeige-Distill/config.json

@@ -0,0 +1,66 @@
+{
+  "architectures": [
+    "Qwen3NextForCausalLM"
+  ],
+  "attention_bias": true,
+  "attention_dropout": 0.0,
+  "attn_output_gate": false,
+  "auto_map": {
+    "AutoConfig": "configuration_qwen3_next.Qwen3NextConfig",
+    "AutoModel": "modeling_qwen3_next.Qwen3NextForCausalLM",
+    "AutoModelForCausalLM": "modeling_qwen3_next.Qwen3NextForCausalLM"
+  },
+  "bos_token_id": 1,
+  "decoder_sparse_step": 1,
+  "dtype": "float32",
+  "enable_qk_norm": false,
+  "eos_token_id": 2,
+  "full_attention_interval": 0,
+  "head_dim": 64,
+  "hidden_act": "silu",
+  "hidden_size": 1920,
+  "initializer_range": 0.02,
+  "intermediate_size": 4800,
+  "layer_types": ["linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "full_attention", "linear_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "full_attention", "linear_attention", "full_attention", "full_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention", "linear_attention"],
+  "linear_conv_kernel_dim": 4,
+  "linear_key_head_dim": 64,
+  "linear_num_key_heads": 8,
+  "linear_num_value_heads": 32,
+  "linear_value_head_dim": 64,
+  "max_position_embeddings": 32768,
+  "mlp_only_layers": [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55],
+  "num_experts_per_tok": 2,
+  "num_experts": 0,
+  "model_type": "qwen3_next",
+  "moe_intermediate_size": 0,
+  "norm_topk_prob": true,
+  "num_attention_heads": 30,
+  "num_hidden_layers": 56,
+  "num_key_value_heads": 6,
+  "output_router_logits": false,
+  "partial_rotary_factor": 1.0,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 490000,
+  "router_aux_loss_coef": 0.001,
+  "router_bias": false,
+  "moe_router_score_function": "softmax",
+  "shared_expert_intermediate_size": 0,
+  "use_shared_expert_gate": true,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.1",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "ffn_token_shift": null,
+  "ffn_intermediate_token_shift": null,
+  "attn_token_shift": null,
+  "attn_q_token_shift": null,
+  "attn_k_token_shift": null,
+  "attn_v_token_shift": null,
+  "token_shift_conv_size": 4,
+  "token_shift_conv_init": "default",
+  "attn_position_embedding_type": "rope",
+  "rnn_position_embedding_type": "nope",
+  "attn_logits_scaling": null,
+  "vocab_size": 99000
+}

YuLan-Mini-Nanbeige-Distill/configuration_qwen3_next.py

@@ -0,0 +1,329 @@
+# coding=utf-8
+# Copyright 2025 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.
+"""Qwen3-Next model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig, layer_type_validation
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen3NextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen3NextModel`]. It is used to instantiate a
+    Qwen3-Next 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-Next-80B-A3B-Instruct [Qwen/Qwen3-Next-80B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct).
+
+    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 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids`.
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 5632):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 48):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 2):
+            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`.
+        hidden_act (`str`, *optional*, defaults to `"silu"`):
+            The non-linear activation function 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
+        partial_rotary_factor (`float`, *optional*, defaults to 0.25):
+            Percentage of the query and keys which will have rotary embedding.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        head_dim (`int`, *optional*, defaults to 256):
+            Projection weights dimension in multi-head attention.
+        linear_conv_kernel_dim (`int`, *optional*, defaults to 4):
+            Kernel size of the convolution used in linear attention layers.
+        linear_key_head_dim (`int`, *optional*, defaults to 128):
+            Dimension of each key head in linear attention.
+        linear_value_head_dim (`int`, *optional*, defaults to 128):
+            Dimension of each value head in linear attention.
+        linear_num_key_heads (`int`, *optional*, defaults to 16):
+            Number of key heads used in linear attention layers.
+        linear_num_value_heads (`int`, *optional*, defaults to 32):
+            Number of value heads used in linear attention layers.
+        decoder_sparse_step (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer.
+        moe_intermediate_size (`int`, *optional*, defaults to 512):
+            Intermediate size of the routed expert.
+        shared_expert_intermediate_size (`int`, *optional*, defaults to 512):
+            Intermediate size of the shared expert.
+        num_experts_per_tok (`int`, *optional*, defaults to 10):
+            Number of selected experts.
+        num_experts (`int`, *optional*, defaults to 512):
+            Number of routed experts.
+        norm_topk_prob (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the topk probabilities.
+        output_router_logits (`bool`, *optional*, defaults to `False`):
+            Whether or not the router logits should be returned by the model. Enabling this will also
+            allow the model to output the auxiliary loss, including load balancing loss and router z-loss.
+        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
+            The aux loss factor for the total loss.
+        mlp_only_layers (`list[int]`, *optional*, defaults to `[]`):
+            Indicate which layers use Qwen3NextMLP rather than Qwen3NextSparseMoeBlock
+            The list contains layer index, from 0 to num_layers-1 if we have num_layers layers
+            If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity.
+        layer_types (`list[str]`, *optional*):
+            Types of each layer (attention or linear).
+        enable_qk_norm (`bool`, *optional*, defaults to `False`):
+            Whether to apply L2 normalization to the query and key embeddings.
+        router_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in the router logits.
+        moe_router_score_function (`str`, *optional*, defaults to `"softmax"`):
+            The score function used in the MoE router.
+        ffn_token_shift (`str`, *optional*): Token shift before FFN/MoE. `None`, `"cat"`, or `"conv"`.
+        ffn_intermediate_token_shift (`str`, *optional*): Token shift in MLP before down_proj. `None`, `"cat"`, or `"conv"`.
+        attn_token_shift (`str`, *optional*): Token shift before attention. `None`, `"cat"`, or `"conv"`.
+        attn_q_token_shift (`str`, *optional*): Token shift on query after projection. `None`, `"cat"`, or `"conv"`.
+        attn_k_token_shift (`str`, *optional*): Token shift on key after projection. `None`, `"cat"`, or `"conv"`.
+        attn_v_token_shift (`str`, *optional*): Token shift on value after projection. `None`, `"cat"`, or `"conv"`.
+        token_shift_conv_size (`int`, *optional*, defaults to 4): Kernel size for token-shift Conv1d.
+        token_shift_conv_init (`str`, *optional*, defaults to `"default"`): Init for token-shift Conv1d; `"identity"` for causal identity.
+    ```python
+    >>> from transformers import Qwen3NextModel, Qwen3NextConfig
+
+    >>> # Initializing a Qwen3Next style configuration
+    >>> configuration =  Qwen3NextConfig()
+
+    >>> # Initializing a model from the Qwen3-Next-80B-A3B style configuration
+    >>> model = Qwen3NextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = "qwen3_next"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.experts.*.gate_proj": "colwise",
+        "layers.*.mlp.experts.*.up_proj": "colwise",
+        "layers.*.mlp.experts.*.down_proj": "rowwise",
+        "layers.*.mlp.shared_experts.gate_proj": "colwise",
+        "layers.*.mlp.shared_experts.up_proj": "colwise",
+        "layers.*.mlp.shared_experts.down_proj": "rowwise",
+        "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"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=2048,
+        intermediate_size=5632,
+        num_hidden_layers=48,
+        num_attention_heads=16,
+        num_key_value_heads=2,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        partial_rotary_factor=0.25,
+        attention_bias=False,
+        attention_dropout=0.0,
+        head_dim=256,
+        linear_conv_kernel_dim=4,
+        linear_key_head_dim=128,
+        linear_value_head_dim=128,
+        linear_num_key_heads=16,
+        linear_num_value_heads=32,
+        decoder_sparse_step=1,
+        moe_intermediate_size=512,
+        shared_expert_intermediate_size=512,
+        num_experts_per_tok=10,
+        num_experts=512,
+        norm_topk_prob=True,
+        output_router_logits=False,
+        router_aux_loss_coef=0.001,
+        mlp_only_layers=[],
+        layer_types=None,
+        enable_qk_norm=False, # @o2iginal
+        router_bias=False, # @o2iginal
+        attn_output_gate=False, # @o2iginal
+        moe_router_score_function="softmax", # @xcx
+        # Cannon layer / token shifting (align with Megatron) @o2iginal
+        ffn_token_shift=None,
+        ffn_intermediate_token_shift=None,
+        attn_token_shift=None,
+        attn_q_token_shift=None,
+        attn_k_token_shift=None,
+        attn_v_token_shift=None,
+        token_shift_conv_size=4,
+        token_shift_conv_init="default",
+        # Separate RoPE for attention vs linear/RNN (GDN): "rope" or "nope"
+        attn_position_embedding_type="rope",
+        rnn_position_embedding_type="nope",
+        # Optional logits scaling for length extrapolation (attention only): None, float, or "log" / "log <a>"
+        attn_logits_scaling=None,
+        **kwargs,
+    ):
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+        assert attn_position_embedding_type in ("rope", "nope"), (
+            f"attn_position_embedding_type must be 'rope' or 'nope', got {attn_position_embedding_type}"
+        )
+        assert rnn_position_embedding_type in ("rope", "nope"), (
+            f"rnn_position_embedding_type must be 'rope' or 'nope', got {rnn_position_embedding_type}"
+        )
+        self.attn_position_embedding_type = attn_position_embedding_type
+        self.rnn_position_embedding_type = rnn_position_embedding_type
+        self.attn_logits_scaling = attn_logits_scaling
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.partial_rotary_factor = partial_rotary_factor
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.head_dim = head_dim
+        rope_config_validation(self)
+
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            interval_pattern = kwargs.get("full_attention_interval", 4)
+            self.layer_types = [
+                "linear_attention" if bool((i + 1) % interval_pattern) else "full_attention"
+                for i in range(self.num_hidden_layers)
+            ]
+        layer_type_validation(self.layer_types)
+
+        # linear attention part
+        self.linear_conv_kernel_dim = linear_conv_kernel_dim
+        self.linear_key_head_dim = linear_key_head_dim
+        self.linear_value_head_dim = linear_value_head_dim
+        self.linear_num_key_heads = linear_num_key_heads
+        self.linear_num_value_heads = linear_num_value_heads
+
+        # MoE arguments
+        self.decoder_sparse_step = decoder_sparse_step
+        self.moe_intermediate_size = moe_intermediate_size
+        self.shared_expert_intermediate_size = shared_expert_intermediate_size
+        self.num_experts_per_tok = num_experts_per_tok
+        self.num_experts = num_experts
+        self.norm_topk_prob = norm_topk_prob
+        self.output_router_logits = output_router_logits
+        self.router_aux_loss_coef = router_aux_loss_coef
+        self.mlp_only_layers = mlp_only_layers
+        self.enable_qk_norm = enable_qk_norm
+        self.router_bias = router_bias
+        self.attn_output_gate = attn_output_gate
+        self.moe_router_score_function = moe_router_score_function
+        self.enable_qk_norm = enable_qk_norm
+        self.router_bias = router_bias
+        self.attn_output_gate = attn_output_gate
+        self.moe_router_score_function = moe_router_score_function
+
+        # Token shifting (cannon layer): None | "cat" | "conv"
+        self.ffn_token_shift = ffn_token_shift
+        self.ffn_intermediate_token_shift = ffn_intermediate_token_shift
+        self.attn_token_shift = attn_token_shift
+        self.attn_q_token_shift = attn_q_token_shift
+        self.attn_k_token_shift = attn_k_token_shift
+        self.attn_v_token_shift = attn_v_token_shift
+        self.token_shift_conv_size = token_shift_conv_size
+        self.token_shift_conv_init = token_shift_conv_init
+
+__all__ = ["Qwen3NextConfig"]

YuLan-Mini-Nanbeige-Distill/hf2mcore.log

@@ -0,0 +1,1291 @@
+torchrun --nproc_per_node 1 --nnodes 1 --node_rank 0 --master_addr localhost --master_port 41921 /workspace/lvzhihao/PostTrain/YuLan-Pretrain/scripts/distributed_checkpoints_convertor/impl/convert.py --tokenizer-type HuggingFaceTokenizer --tokenizer-model /tmp/tmp.FZZhIF5Vmh --hf-dir /tmp/tmp.FZZhIF5Vmh --mcore2hf --use-gpu --bf16 --normalization RMSNorm --swiglu --disable-bias-linear --seq-length 1 --max-position-embeddings 490000 --attention-backend auto --position-embedding-type rope --kv-channels 64 --group-query-attention --add-qkv-bias --num-layers 56 --hidden-size 1920 --ffn-hidden-size 4800 --num-attention-heads 30 --untie-embeddings-and-output-weights --rotary-base 490000 --rotary-percent 1.00 --num-query-groups 6 --normalization RMSNorm --norm-epsilon 1e-6 --linear-attention-type gated_delta_net --linear-attention-freq [1,1,1,1,1,1,1,1,1,1,1,1,0,1,1,1,1,1,1,1,0,0,1,0,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0,1,0,0,1,1,1,1,1,1] --linear-conv-kernel-dim 4 --linear-key-head-dim 64 --linear-value-head-dim 64 --linear-num-key-heads 8 --linear-num-value-heads 32 --micro-batch-size 1 --global-batch-size 1024 --train-iters 500000 --weight-decay 0.1 --adam-beta1 0.9 --adam-beta2 0.95 --init-method-std 0.006 --clip-grad 1.0 --lr 2.0e-5 --lr-decay-style cosine --min-lr 6.0e-6 --lr-warmup-fraction .001 --lr-decay-iters 430000 --bf16 --tensor-model-parallel-size 1 --pipeline-model-parallel-size 1 --expert-tensor-parallel-size 1 --expert-model-parallel-size 1 --log-interval 100 --save-interval 10000 --eval-interval 1000 --eval-iters 10 --model-type GPT --load-dir /capacity/userdata/vc0e4b0o65t5/lvzhihao/PostTrain/YuLan-Pretrain/outputs/yulan_mini_sft/run_sl16384_tp1_pp1_cp2/checkpoint/yulan-gdn-sft-1b-sl16384-lr1e-5-gbs64-mb1-tp1-pp1-cp2 --save-dir /capacity/userdata/vc0e4b0o65t5/lvzhihao/PostTrain/YuLan-Pretrain/outputs/yulan_mini_sft/run_sl16384_tp1_pp1_cp2/checkpoint/yulan-gdn-sft-1b-sl16384-lr1e-5-gbs64-mb1-tp1-pp1-cp2/iter_2340-hf --dist-ckpt-optim-fully-reshardable --skip-train --use-cpu-initialization --padded-vocab-size 99000 --no-load-optim --no-load-rng --logging-level 1 --attention-backend auto --synchronizer mcore_gdn_moe --pretrain-script mcore_gdn_moe.model_provider --debug --max-shard-size 20GB
+W0416 17:58:28.424000 71892 .venv/lib/python3.11/site-packages/torch/utils/cpp_extension.py:2425] TORCH_CUDA_ARCH_LIST is not set, all archs for visible cards are included for compilation. 
+W0416 17:58:28.424000 71892 .venv/lib/python3.11/site-packages/torch/utils/cpp_extension.py:2425] If this is not desired, please set os.environ['TORCH_CUDA_ARCH_LIST'] to specific architectures.
+fused_indices_to_multihot has reached end of life. Please migrate to a non-experimental function.
+/workspace/lvzhihao/PostTrain/YuLan-Pretrain/.venv/lib/python3.11/site-packages/modelopt/torch/utils/import_utils.py:32: UserWarning: Failed to import diffusers plugin due to: ImportError('Requires Flash-Attention version >=2.7.1,<=2.8.2 but got 2.8.3.'). You may ignore this warning if you do not need this plugin.
+  warnings.warn(
+INFO 04-16 17:58:48 [__init__.py:216] Automatically detected platform cuda.
+/workspace/lvzhihao/PostTrain/YuLan-Pretrain/.venv/lib/python3.11/site-packages/modelopt/torch/__init__.py:36: UserWarning: transformers version 4.57.1 is incompatible with nvidia-modelopt and may cause issues. Please install recommended version with `pip install nvidia-modelopt[hf]` if working with HF models.
+  _warnings.warn(
+Warning: Pai-Megatron-Patch arguments not available, some arguments may not be recognized
+using world size: 1, data-parallel size: 1, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1
+Number of virtual stages per pipeline stage: None
+accumulate and all-reduce gradients in fp32 for bfloat16 data type.
+using torch.bfloat16 for parameters ...
+------------------------ arguments ------------------------
+  account_for_embedding_in_pipeline_split ......... False
+  account_for_loss_in_pipeline_split .............. False
+  accumulate_allreduce_grads_in_fp32 .............. True
+  activation_func_clamp_value ..................... None
+  adam_beta1 ...................................... 0.9
+  adam_beta2 ...................................... 0.95
+  adam_eps ........................................ 1e-08
+  adamw_lr_mup_scaler ............................. False
+  add_bias_linear ................................. False
+  add_position_embedding .......................... True
+  add_qkv_bias .................................... True
+  adlr_autoresume ................................. False
+  adlr_autoresume_interval ........................ 1000
+  align_grad_reduce ............................... True
+  align_param_gather .............................. False
+  allow_ambiguous_pad_tokens ...................... False
+  app_tag_run_name ................................ None
+  app_tag_run_version ............................. 0.0.0
+  apply_layernorm_1p .............................. False
+  apply_query_key_layer_scaling ................... False
+  apply_residual_connection_post_layernorm ........ False
+  apply_rope_fusion ............................... True
+  async_save ...................................... None
+  async_tensor_model_parallel_allreduce ........... True
+  attention_backend ............................... AttnBackend.auto
+  attention_dropout ............................... 0.1
+  attention_output_gate ........................... False
+  attention_softmax_in_fp32 ....................... False
+  attn_k_token_shift .............................. None
+  attn_output_gate ................................ None
+  attn_output_gate_rand_init ...................... False
+  attn_q_token_shift .............................. None
+  attn_token_shift ................................ None
+  attn_v_token_shift .............................. None
+  auto_detect_ckpt_format ......................... False
+  auto_generate_cu_seqlens ........................ False
+  auto_model ...................................... AutoModelForCausalLM
+  barrier_with_L1_time ............................ True
+  benchmark_eval .................................. False
+  benchmark_global_batch .......................... None
+  benchmark_interval .............................. None
+  benchmark_micro_batch ........................... None
+  benchmark_sequence_length ....................... None
+  benchmark_tasks ................................. None
+  bert_binary_head ................................ True
+  bert_embedder_type .............................. megatron
+  bert_load ....................................... None
+  bf16 ............................................ True
+  bias_dropout_fusion ............................. True
+  bias_gelu_fusion ................................ False
+  bias_swiglu_fusion .............................. True
+  biencoder_projection_dim ........................ 0
+  biencoder_shared_query_context_model ............ False
+  block_data_path ................................. None
+  cache_mla_latents ............................... False
+  calc_ft_timeouts ................................ False
+  calculate_per_token_loss ........................ False
+  check_for_large_grads ........................... False
+  check_for_nan_in_loss_and_grad .................. True
+  check_for_spiky_loss ............................ False
+  check_weight_hash_across_dp_replicas_interval ... None
+  ckpt_assume_constant_structure .................. False
+  ckpt_convert_format ............................. None
+  ckpt_convert_save ............................... None
+  ckpt_convert_update_legacy_dist_opt_format ...... False
+  ckpt_format ..................................... torch_dist
+  ckpt_fully_parallel_load ........................ False
+  ckpt_fully_parallel_save ........................ True
+  ckpt_fully_parallel_save_deprecated ............. False
+  ckpt_step ....................................... None
+  classes_fraction ................................ 1.0
+  clip_grad ....................................... 1.0
+  clone_scatter_output_in_embedding ............... True
+  config_logger_dir ............................... 
+  consumed_train_samples .......................... 0
+  consumed_valid_samples .......................... 0
+  context_parallel_size ........................... 1
+  cp_comm_type .................................... ['p2p']
+  create_attention_mask_in_dataloader ............. True
+  cross_entropy_fusion_impl ....................... native
+  cross_entropy_loss_fusion ....................... False
+  cuda_graph_impl ................................. none
+  cuda_graph_scope ................................ []
+  cuda_graph_warmup_steps ......................... 3
+  data_args_path .................................. None
+  data_cache_path ................................. None
+  data_parallel_random_init ....................... False
+  data_parallel_sharding_strategy ................. no_shard
+  data_parallel_size .............................. 1
+  data_path ....................................... None
+  data_per_class_fraction ......................... 1.0
+  data_sharding ................................... True
+  dataloader_type ................................. single
+  ddp_average_in_collective ....................... False
+  ddp_bucket_size ................................. None
+  ddp_num_buckets ................................. None
+  ddp_pad_buckets_for_high_nccl_busbw ............. False
+  debug ........................................... True
+  decode_only_cuda_graphs ......................... False
+  decoder_first_pipeline_num_layers ............... None
+  decoder_last_pipeline_num_layers ................ None
+  decoder_num_layers .............................. None
+  decoder_seq_length .............................. None
+  decoupled_lr .................................... None
+  decoupled_min_lr ................................ None
+  decrease_batch_size_if_needed ................... False
+  defer_embedding_wgrad_compute ................... False
+  delay_wgrad_compute ............................. False
+  deprecated_use_mcore_models ..................... False
+  deterministic_mode .............................. False
+  dino_bottleneck_size ............................ 256
+  dino_freeze_last_layer .......................... 1
+  dino_head_hidden_size ........................... 2048
+  dino_local_crops_number ......................... 10
+  dino_local_img_size ............................. 96
+  dino_norm_last_layer ............................ False
+  dino_teacher_temp ............................... 0.07
+  dino_warmup_teacher_temp ........................ 0.04
+  dino_warmup_teacher_temp_epochs ................. 30
+  disable_attn_output_gate ........................ False
+  disable_bf16_reduced_precision_matmul ........... False
+  disable_chunked_prefill ......................... False
+  disable_explicit_attention_mask ................. False
+  disable_mamba_mem_eff_path ...................... False
+  disable_straggler_on_startup .................... False
+  disable_symmetric_registration .................. False
+  dist_ckpt_format_deprecated ..................... None
+  dist_ckpt_optim_fully_reshardable ............... True
+  dist_ckpt_save_pre_mcore_014 .................... False
+  dist_ckpt_strictness ............................ assume_ok_unexpected
+  distrib_optim_fully_reshardable_mem_efficient ... False
+  distribute_saved_activations .................... False
+  distributed_backend ............................. nccl
+  distributed_timeout_minutes ..................... 10
+  distributed_timeout_seconds_after_init .......... None
+  document_packing_algorithm ...................... random
+  dryrun .......................................... False
+  dump_param_to_param_group_map ................... None
+  emb_deviation_loss_coeff ........................ 0
+  emb_deviation_type .............................. None
+  embedding_init_method_std ....................... None
+  embedding_path .................................. None
+  empty_unused_memory_level ....................... 0
+  enable_cuda_graph ............................... False
+  enable_debug_logging ............................ False
+  enable_experimental ............................. False
+  enable_ft_package ............................... False
+  enable_full_sharding_in_hsdp .................... False
+  enable_gloo_process_groups ...................... True
+  enable_msc ...................................... True
+  enable_one_logger ............................... True
+  encoder_num_layers .............................. 56
+  encoder_seq_length .............................. 1
+  end_weight_decay ................................ 0.1
+  eod_mask_loss ................................... False
+  error_injection_rate ............................ 0
+  error_injection_type ............................ transient_error
+  eval_interval ................................... 1000
+  eval_iters ...................................... 10
+  evidence_data_path .............................. None
+  exit_duration_in_mins ........................... None
+  exit_interval ................................... None
+  exit_on_missing_checkpoint ...................... False
+  exit_signal_handler ............................. False
+  exp_avg_dtype ................................... torch.float32
+  exp_avg_sq_dtype ................................ torch.float32
+  expert_model_parallel_size ...................... 1
+  expert_tensor_parallel_size ..................... 1
+  external_cuda_graph ............................. False
+  ffn_hidden_size ................................. 4800
+  ffn_intermediate_token_shift .................... None
+  ffn_token_shift ................................. None
+  fine_grained_activation_offloading .............. False
+  finetune ........................................ False
+  first_last_layers_bf16 .......................... False
+  flash_decode .................................... False
+  fp16 ............................................ False
+  fp16_lm_cross_entropy ........................... False
+  fp32_residual_connection ........................ False
+  fp4 ............................................. None
+  fp4_param ....................................... False
+  fp4_recipe ...................................... nvfp4
+  fp8 ............................................. None
+  fp8_amax_compute_algo ........................... most_recent
+  fp8_amax_history_len ............................ 1
+  fp8_interval .................................... 1
+  fp8_margin ...................................... 0
+  fp8_param_gather ................................ False
+  fp8_recipe ...................................... delayed
+  fp8_wgrad ....................................... True
+  freeze_layernorm_weight ......................... False
+  freeze_non_mamba ................................ False
+  fsdp_double_buffer .............................. False
+  full_validation ................................. False
+  gdn_cp_impl ..................................... cp2hp
+  geglu ........................................... False
+  global_batch_size ............................... 1024
+  glu_linear_offset ............................... 0.0
+  grad_reduce_in_bf16 ............................. False
+  gradient_accumulation_fusion .................... True
+  gradient_reduce_div_fusion ...................... True
+  group_query_attention ........................... True
+  grpo_clamp_eps_lower ............................ 0.01
+  grpo_clamp_eps_upper ............................ 0.01
+  grpo_default_temperature ........................ 1.0
+  grpo_default_top_p .............................. 0
+  grpo_entropy_term_weight ........................ 0.0
+  grpo_filter_groups_with_same_reward ............. False
+  grpo_group_size ................................. 2
+  grpo_iterations ................................. 2
+  grpo_kl_beta .................................... 0.001
+  grpo_prompts_per_step ........................... 32
+  head_lr_mult .................................... 1.0
+  heterogeneous_layers_config_encoded_json ........ None
+  heterogeneous_layers_config_path ................ None
+  hf_dir .......................................... /tmp/tmp.FZZhIF5Vmh
+  hidden_dropout .................................. 0.1
+  hidden_size ..................................... 1920
+  hierarchical_context_parallel_sizes ............. None
+  high_priority_stream_groups ..................... []
+  hybrid_attention_ratio .......................... 0.0
+  hybrid_context_parallel ......................... False
+  hybrid_mlp_ratio ................................ 0.0
+  hybrid_override_pattern ......................... None
+  hysteresis ...................................... 2
+  ict_head_size ................................... None
+  ict_load ........................................ None
+  img_h ........................................... 224
+  img_w ........................................... 224
+  increase_log_level_interval ..................... 1000
+  increase_log_level_iters ........................ 5
+  indexer_batch_size .............................. 128
+  indexer_log_interval ............................ 1000
+  inference_batch_times_seqlen_threshold .......... -1
+  inference_dynamic_batching ...................... False
+  inference_dynamic_batching_block_size ........... 256
+  inference_dynamic_batching_buffer_guaranteed_fraction  0.2
+  inference_dynamic_batching_buffer_overflow_factor  None
+  inference_dynamic_batching_buffer_size_gb ....... 40.0
+  inference_dynamic_batching_max_requests_override  None
+  inference_dynamic_batching_max_tokens_override .. None
+  inference_dynamic_batching_num_cuda_graphs ...... 16
+  inference_dynamic_batching_track_paused_request_events  False
+  inference_dynamic_batching_unified_memory_level . 0
+  inference_max_batch_size ........................ 8
+  inference_max_seq_length ........................ 2560
+  inference_rng_tracker ........................... False
+  init_method_std ................................. 0.006
+  init_method_xavier_uniform ...................... False
+  init_model_with_meta_device ..................... False
+  initial_loss_scale .............................. 4294967296
+  inprocess_active_world_size ..................... 1
+  inprocess_barrier_timeout ....................... 120
+  inprocess_completion_timeout .................... 120
+  inprocess_empty_cuda_cache ...................... False
+  inprocess_granularity ........................... node
+  inprocess_hard_timeout .......................... 90
+  inprocess_heartbeat_interval .................... 30
+  inprocess_heartbeat_timeout ..................... 60
+  inprocess_last_call_wait ........................ 1
+  inprocess_max_iterations ........................ None
+  inprocess_monitor_process_interval .............. 1.0
+  inprocess_monitor_thread_interval ............... 1.0
+  inprocess_progress_watchdog_interval ............ 1.0
+  inprocess_restart ............................... False
+  inprocess_soft_timeout .......................... 60
+  inprocess_termination_grace_time ................ 1
+  is_hybrid_model ................................. False
+  iter_per_epoch .................................. 1250
+  iterations_to_skip .............................. []
+  keep_fp8_transpose_cache ........................ False
+  kitchen_config_file ............................. None
+  kitchen_recipe_number ........................... None
+  kv_channels ..................................... 64
+  kv_lora_rank .................................... 32
+  langrl_env_config ............................... None
+  langrl_external_server .......................... False
+  langrl_inference_server_conversation_template ... None
+  langrl_inference_server_type .................... inplace_megatron
+  lazy_mpu_init ................................... None
+  legacy_tokenizer ................................ False
+  linear_attention_freq ........................... [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1]
+  linear_attention_type ........................... gated_delta_net
+  linear_conv_kernel_dim .......................... 4
+  linear_key_head_dim ............................. 64
+  linear_num_key_heads ............................ 8
+  linear_num_value_heads .......................... 32
+  linear_value_head_dim ........................... 64
+  load ............................................ None
+  load_complemental_dataset ....................... None
+  load_dir ........................................ /capacity/userdata/vc0e4b0o65t5/lvzhihao/PostTrain/YuLan-Pretrain/outputs/yulan_mini_sft/run_sl16384_tp1_pp1_cp2/checkpoint/yulan-gdn-sft-1b-sl16384-lr1e-5-gbs64-mb1-tp1-pp1-cp2
+  load_main_params_from_ckpt ...................... None
+  local_rank ...................................... 0
+  log_energy ...................................... False
+  log_hidden_states ............................... []
+  log_interval .................................... 100
+  log_loss_scale_to_tensorboard ................... True
+  log_memory_to_tensorboard ....................... False
+  log_num_zeros_in_grad ........................... False
+  log_params ...................................... []
+  log_params_norm ................................. False
+  log_per_module_grad_rms ......................... False
+  log_per_module_update_rms ....................... False
+  log_progress .................................... False
+  log_straggler ................................... False
+  log_throughput .................................. False
+  log_timers_to_tensorboard ....................... False
+  log_validation_ppl_to_tensorboard ............... False
+  log_world_size_to_tensorboard ................... False
+  logging_level ................................... 1
+  loss_scale ...................................... None
+  loss_scale_window ............................... 1000
+  lr .............................................. 2e-05
+  lr_decay_iters .................................. 430000
+  lr_decay_samples ................................ None
+  lr_decay_style .................................. cosine
+  lr_warmup_fraction .............................. 0.001
+  lr_warmup_init .................................. 0.0
+  lr_warmup_iters ................................. 0
+  lr_warmup_samples ............................... 0
+  lr_wsd_decay_iters .............................. None
+  lr_wsd_decay_samples ............................ None
+  lr_wsd_decay_style .............................. exponential
+  main_grads_dtype ................................ torch.float32
+  main_params_dtype ............................... torch.float32
+  make_vocab_size_divisible_by .................... 128
+  mamba_disable_cp ................................ False
+  mamba_expand .................................... 2
+  mamba_head_dim .................................. 64
+  mamba_num_groups ................................ 8
+  mamba_num_heads ................................. None
+  mamba_state_dim ................................. 128
+  manual_gc ....................................... False
+  manual_gc_eval .................................. True
+  manual_gc_interval .............................. 0
+  mask_factor ..................................... 1.0
+  mask_prob ....................................... 0.15
+  mask_type ....................................... random
+  masked_softmax_fusion ........................... True
+  max_position_embeddings ......................... 490000
+  max_seqlen_per_cp_rank .......................... None
+  max_shard_size .................................. 20GB
+  max_tokens_to_oom ............................... 12000
+  mcore2hf ........................................ True
+  memory_snapshot_path ............................ None
+  merge_file ...................................... None
+  micro_batch_size ................................ 1
+  microbatch_group_size_per_vp_stage .............. None
+  mid_level_dataset_surplus ....................... 0.005
+  min_loss_scale .................................. 1.0
+  min_lr .......................................... 6e-06
+  min_offloaded_tensor_size ....................... 1048576
+  mlp_chunks_for_prefill .......................... 1
+  mmap_bin_files .................................. True
+  mock_data ....................................... False
+  model_type ...................................... GPT
+  moe_apply_probs_on_input ........................ False
+  moe_aux_loss_coeff .............................. 0.0
+  moe_deepep_num_sms .............................. 20
+  moe_enable_deepep ............................... False
+  moe_expert_capacity_factor ...................... None
+  moe_extended_tp ................................. False
+  moe_ffn_hidden_size ............................. None
+  moe_flex_dispatcher_backend ..................... deepep
+  moe_grouped_gemm ................................ False
+  moe_hybridep_num_sms ............................ 16
+  moe_input_jitter_eps ............................ None
+  moe_layer_freq .................................. 1
+  moe_layer_recompute ............................. False
+  moe_pad_expert_input_to_capacity ................ False
+  moe_pad_experts_for_cuda_graph_inference ........ False
+  moe_per_layer_logging ........................... False
+  moe_permute_fusion .............................. False
+  moe_router_bias_update_method ................... sign
+  moe_router_bias_update_rate ..................... 0.001
+  moe_router_dtype ................................ None
+  moe_router_enable_expert_bias ................... False
+  moe_router_force_load_balancing ................. False
+  moe_router_fusion ............................... False
+  moe_router_group_topk ........................... None
+  moe_router_load_balancing_type .................. aux_loss
+  moe_router_num_groups ........................... None
+  moe_router_padding_for_fp8 ...................... False
+  moe_router_padding_for_quantization ............. False
+  moe_router_pre_softmax .......................... False
+  moe_router_score_function ....................... softmax
+  moe_router_topk ................................. 2
+  moe_router_topk_scaling_factor .................. None
+  moe_shared_expert_gate .......................... False
+  moe_shared_expert_intermediate_size ............. None
+  moe_shared_expert_overlap ....................... False
+  moe_token_dispatcher_type ....................... allgather
+  moe_token_drop_policy ........................... probs
+  moe_upcycling_granularity ....................... 1
+  moe_use_legacy_grouped_gemm ..................... False
+  moe_use_upcycling ............................... False
+  moe_z_loss_coeff ................................ None
+  mrope_section ................................... None
+  mscale .......................................... 1.0
+  mscale_all_dim .................................. 0.0
+  mtp_linear_attention_type ....................... None
+  mtp_loss_scaling_factor ......................... 0.1
+  mtp_num_layers .................................. None
+  multi_latent_attention .......................... False
+  multiple_validation_sets ........................ False
+  muon_ball_momentum .............................. 0.9
+  muon_ball_msign_steps ........................... 5
+  muon_ball_power_iteration_steps ................. 10
+  muon_ball_qkv_split_mode ........................ component
+  muon_ball_radius_mode ........................... spectral_mup
+  muon_ball_retract_alpha ......................... 0.05
+  muon_ball_retract_mode .......................... hard
+  muon_ball_scale_mode ............................ spectral_mup
+  muon_ball_split_fc1 ............................. True
+  muon_ball_split_moe_experts ..................... True
+  muon_ball_split_qkv ............................. True
+  muon_ball_use_nesterov .......................... True
+  muon_extra_scale_factor ......................... 1.0
+  muon_fp32_matmul_prec ........................... medium
+  muon_momentum ................................... 0.9
+  muon_num_ns_steps ............................... 5
+  muon_qkv_split_mode ............................. component
+  muon_scale_mode ................................. spectral_mup
+  muon_scale_vectorized_mode ...................... full
+  muon_split_fc1 .................................. True
+  muon_split_moe_experts .......................... True
+  muon_split_qkv .................................. True
+  muon_tp_mode .................................... blockwise
+  muon_use_nesterov ............................... False
+  muon_vectorize .................................. []
+  muon_vectorize_attn_dim ......................... hidden_size
+  nccl_all_reduce_for_prefill ..................... False
+  nccl_communicator_config_path ................... None
+  nccl_ub ......................................... False
+  no_load_optim ................................... True
+  no_load_rng ..................................... True
+  no_load_scheduler ............................... None
+  no_persist_layer_norm ........................... False
+  no_rope_freq .................................... None
+  no_save_optim ................................... None
+  no_save_rng ..................................... None
+  no_save_step_one ................................ None
+  no_weight_decay_cond_type ....................... None
+  non_persistent_ckpt_type ........................ None
+  non_persistent_global_ckpt_dir .................. None
+  non_persistent_local_ckpt_algo .................. fully_parallel
+  non_persistent_local_ckpt_dir ................... None
+  non_persistent_save_interval .................... None
+  norm_epsilon .................................... 1e-06
+  normalization ................................... RMSNorm
+  num_attention_heads ............................. 30
+  num_channels .................................... 3
+  num_classes ..................................... 1000
+  num_dataset_builder_threads ..................... 1
+  num_distributed_optimizer_instances ............. 1
+  num_experts ..................................... None
+  num_hf_saver .................................... None
+  num_layers ...................................... 56
+  num_layers_at_end_in_bf16 ....................... 1
+  num_layers_at_start_in_bf16 ..................... 1
+  num_layers_per_virtual_pipeline_stage ........... None
+  num_query_groups ................................ 6
+  num_virtual_stages_per_pipeline_rank ............ None
+  num_workers ..................................... 2
+  object_storage_cache_path ....................... None
+  offload_modules ................................. []
+  one_logger_async ................................ False
+  one_logger_project .............................. megatron-lm
+  one_logger_run_name ............................. None
+  onnx_safe ....................................... None
+  openai_gelu ..................................... False
+  optimizer ....................................... adam
+  optimizer_cpu_offload ........................... False
+  optimizer_offload_fraction ...................... 1.0
+  output_bert_embeddings .......................... False
+  overlap_cpu_optimizer_d2h_h2d ................... False
+  overlap_grad_reduce ............................. False
+  overlap_moe_expert_parallel_comm ................ False
+  overlap_p2p_comm ................................ False
+  overlap_p2p_comm_warmup_flush ................... False
+  overlap_param_gather ............................ False
+  overlap_param_gather_with_optimizer_step ........ False
+  override_hf_eod_token_id ........................ None
+  override_opt_param_scheduler .................... False
+  padded_vocab_size ............................... 99000
+  params_dtype .................................... torch.bfloat16
+  patch_dim ....................................... 16
+  per_split_data_args_path ........................ None
+  perform_initialization .......................... True
+  perform_rl_step ................................. False
+  pin_cpu_grads ................................... True
+  pin_cpu_params .................................. True
+  pipeline_model_parallel_comm_backend ............ None
+  pipeline_model_parallel_layout .................. None
+  pipeline_model_parallel_size .................... 1
+  position_embedding_type ......................... rope
+  pretrain_script ................................. mcore_gdn_moe.model_provider
+  pretrained_checkpoint ........................... None
+  profile ......................................... False
+  profile_ranks ................................... [0]
+  profile_step_end ................................ 12
+  profile_step_start .............................. 10
+  q_lora_rank ..................................... None
+  qk_head_dim ..................................... 128
+  qk_l2_norm ...................................... False
+  qk_layernorm .................................... False
+  qk_pos_emb_head_dim ............................. 64
+  query_in_block_prob ............................. 0.1
+  quick_geglu ..................................... False
+  rampup_batch_size ............................... None
+  rank ............................................ 0
+  recompute_granularity ........................... None
+  recompute_method ................................ None
+  recompute_modules ............................... None
+  recompute_num_layers ............................ None
+  record_memory_history ........................... False
+  relative_attention_max_distance ................. 128
+  relative_attention_num_buckets .................. 32
+  reparam_checkpoint .............................. None
+  reparam_fallback_value .......................... None
+  reparam_keys .................................... None
+  replication ..................................... False
+  replication_factor .............................. 2
+  replication_jump ................................ None
+  rerun_mode ...................................... validate_results
+  reset_attention_mask ............................ False
+  reset_iteration_on_load ......................... False
+  reset_iteration_one_to_zero ..................... False
+  reset_position_ids .............................. False
+  reset_scheduler_steps_on_load ................... False
+  result_rejected_tracker_filename ................ None
+  retriever_report_topk_accuracies ................ []
+  retriever_score_scaling ......................... False
+  retriever_seq_length ............................ 256
+  retro_add_retriever ............................. False
+  retro_attention_gate ............................ 1
+  retro_cyclic_train_iters ........................ None
+  retro_encoder_attention_dropout ................. 0.1
+  retro_encoder_hidden_dropout .................... 0.1
+  retro_encoder_layers ............................ 2
+  retro_num_neighbors ............................. 2
+  retro_num_retrieved_chunks ...................... 2
+  retro_project_dir ............................... None
+  retro_verify_neighbor_count ..................... True
+  reuse_grad_buf_for_mxfp8_param_ag ............... False
+  reweight_loss_by_sample ......................... False
+  rl_calculate_intra_group_similarity ............. False
+  rl_importance_sampling_truncation_coef .......... None
+  rl_inference_logprobs_is_correction ............. False
+  rl_offload_kv_cache_during_training ............. False
+  rl_offload_optimizer_during_inference ........... False
+  rl_partial_rollouts ............................. False
+  rl_prompts_per_eval ............................. 32
+  rl_remove_kv_cache_during_training .............. False
+  rl_reset_cuda_graphs ............................ False
+  rl_sequence_packing_algo ........................ fifo
+  rl_sequence_packing_bin_size .................... 8192
+  rl_use_sequence_packing ......................... False
+  rope_scaling_factor ............................. 8.0
+  rope_type ....................................... None
+  rotary_base ..................................... 490000
+  rotary_interleaved .............................. False
+  rotary_percent .................................. 1.0
+  rotary_scaling_factor ........................... 1.0
+  rotary_seq_len_interpolation_factor ............. None
+  run_workload_inspector_server ................... False
+  sample_rate ..................................... 1.0
+  save ............................................ None
+  save_after_load ................................. False
+  save_dir ........................................ /capacity/userdata/vc0e4b0o65t5/lvzhihao/PostTrain/YuLan-Pretrain/outputs/yulan_mini_sft/run_sl16384_tp1_pp1_cp2/checkpoint/yulan-gdn-sft-1b-sl16384-lr1e-5-gbs64-mb1-tp1-pp1-cp2/iter_2340-hf
+  save_interval ................................... 10000
+  save_retain_interval ............................ None
+  scatter_gather_tensors_in_pipeline .............. True
+  seed ............................................ 1234
+  seq_length ...................................... 1
+  sequence_parallel ............................... False
+  sft ............................................. False
+  sft_tokenizer_prompt_format ..................... nemotron-h-aligned
+  sgd_momentum .................................... 0.9
+  sharp_enabled_group ............................. None
+  short_seq_prob .................................. 0.1
+  skip_train ...................................... True
+  skipped_train_samples ........................... 0
+  softmax_type .................................... vanilla
+  spec ............................................ None
+  spectral_ball_momentum .......................... 0.9
+  spectral_ball_msign_steps ....................... 8
+  spectral_ball_power_iteration_steps ............. 20
+  spectral_ball_qkv_split_mode .................... component
+  spectral_ball_radius_mode ....................... spectral_mup
+  spectral_ball_retract_alpha ..................... 0.05
+  spectral_ball_retract_mode ...................... hard
+  spectral_ball_scale_mode ........................ spectral_mup
+  spectral_ball_solver ............................ bisection
+  spectral_ball_solver_max_iterations ............. 20
+  spectral_ball_solver_tolerance_f ................ 1e-08
+  spectral_ball_split_fc1 ......................... True
+  spectral_ball_split_moe_experts ................. True
+  spectral_ball_split_qkv ......................... True
+  spectral_ball_use_nesterov ...................... True
+  spectral_mup_init ............................... False
+  split ........................................... None
+  split_expert_init ............................... True
+  split_fc1_init .................................. True
+  split_qkv_init .................................. True
+  split_qkv_init_mode ............................. group
+  sqreglu ......................................... False
+  squared_relu .................................... False
+  start_samples ................................... None
+  start_weight_decay .............................. 0.1
+  straggler_ctrlr_port ............................ 65535
+  straggler_minmax_count .......................... 1
+  strict_fsdp_dtensor_load ........................ True
+  suggested_communication_unit_size ............... None
+  swanlab_exp_name ................................ 
+  swanlab_project ................................. 
+  swanlab_save_dir ................................ 
+  swanlab_workspace ............................... 
+  swiglu .......................................... True
+  swin_backbone_type .............................. tiny
+  symmetric_ar_type ............................... None
+  synchronizer .................................... mcore_gdn_moe
+  target_ckpt_format .............................. torch_dist
+  te_rng_tracker .................................. False
+  tensor_model_parallel_size ...................... 1
+  tensorboard_dir ................................. None
+  tensorboard_log_interval ........................ 1
+  tensorboard_queue_size .......................... 1000
+  test_data_path .................................. None
+  test_mode ....................................... False
+  tiktoken_num_special_tokens ..................... 1000
+  tiktoken_pattern ................................ None
+  tiktoken_special_tokens ......................... None
+  timing_log_level ................................ 0
+  timing_log_option ............................... minmax
+  titles_data_path ................................ None
+  token_shift_conv_init ........................... default
+  token_shift_conv_size ........................... 4
+  tokenizer_metadata .............................. None
+  tokenizer_model ................................. /tmp/tmp.FZZhIF5Vmh
+  tokenizer_type .................................. HuggingFaceTokenizer
+  torch_fsdp2_reshard_after_forward ............... True
+  tp_comm_bootstrap_backend ....................... nccl
+  tp_comm_bulk_dgrad .............................. True
+  tp_comm_bulk_wgrad .............................. True
+  tp_comm_overlap ................................. False
+  tp_comm_overlap_ag .............................. True
+  tp_comm_overlap_cfg ............................. None
+  tp_comm_overlap_rs .............................. True
+  tp_comm_overlap_rs_dgrad ........................ False
+  tp_comm_split_ag ................................ True
+  tp_comm_split_rs ................................ True
+  train_data_path ................................. None
+  train_iters ..................................... 500000
+  train_samples ................................... None
+  train_sync_interval ............................. None
+  transformer_impl ................................ transformer_engine
+  transformer_pipeline_model_parallel_size ........ 1
+  trust_remote_code ............................... False
+  untie_embeddings_and_output_weights ............. True
+  use_checkpoint_args ............................. False
+  use_checkpoint_opt_param_scheduler .............. False
+  use_cpu_initialization .......................... True
+  use_dist_ckpt ................................... True
+  use_dist_ckpt_deprecated ........................ False
+  use_distributed_optimizer ....................... False
+  use_flash_attn .................................. False
+  use_fused_weighted_squared_relu ................. False
+  use_gpu ......................................... True
+  use_legacy_models ............................... False
+  use_megatron_fsdp ............................... False
+  use_mp_args_from_checkpoint_args ................ False
+  use_one_sent_docs ............................... False
+  use_persistent_ckpt_worker ...................... False
+  use_precision_aware_optimizer ................... False
+  use_pytorch_profiler ............................ False
+  use_ring_exchange_p2p ........................... False
+  use_rope_scaling ................................ False
+  use_rotary_position_embeddings .................. False
+  use_sharp ....................................... False
+  use_te_activation_func .......................... False
+  use_tokenizer_model_from_checkpoint_args ........ True
+  use_torch_fsdp2 ................................. False
+  use_torch_optimizer_for_cpu_offload ............. False
+  use_tp_pp_dp_mapping ............................ False
+  v_head_dim ...................................... 128
+  valid_data_path ................................. None
+  variable_seq_lengths ............................ False
+  virtual_pipeline_model_parallel_size ............ None
+  vision_backbone_type ............................ vit
+  vision_pretraining .............................. False
+  vision_pretraining_type ......................... classify
+  vocab_extra_ids ................................. 0
+  vocab_file ...................................... None
+  vocab_size ...................................... None
+  wandb_entity .................................... 
+  wandb_exp_name .................................. 
+  wandb_project ................................... 
+  wandb_save_dir .................................. 
+  weight_decay .................................... 0.1
+  weight_decay_incr_style ......................... constant
+  wgrad_deferral_limit ............................ 0
+  window_attn_skip_freq ........................... None
+  window_size ..................................... None
+  word_embedding_dropout_prob ..................... 0.0
+  world_size ...................................... 1
+  yaml_cfg ........................................ None
+-------------------- end of arguments ---------------------
+INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1024
+> building HuggingFaceTokenizer tokenizer ...
+You are using the default legacy behaviour of the <class 'transformers.models.llama.tokenization_llama_fast.LlamaTokenizerFast'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565 - if you loaded a llama tokenizer from a GGUF file you can ignore this message.
+WARNING: one_logger package is required to enable e2e metrics tracking. please go to https://confluence.nvidia.com/display/MLWFO/Package+Repositories for details to install it
+INFO:megatron.training.initialize:Setting logging level to 1
+WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
+> initializing torch distributed ...
+[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+> initialized tensor model parallel with size 1
+> initialized pipeline model parallel with size 1
+> setting random seeds to 1234 ...
+> compiling dataset index builder ...
+make: Entering directory '/capacity/userdata/vc0e4b0o65t5/lvzhihao/PostTrain/YuLan-Pretrain/megatron/core/datasets'
+make: Nothing to be done for 'default'.
+make: Leaving directory '/capacity/userdata/vc0e4b0o65t5/lvzhihao/PostTrain/YuLan-Pretrain/megatron/core/datasets'
+>>> done with dataset index builder. Compilation time: 0.063 seconds
+WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
+> compiling and loading fused kernels ...
+/workspace/lvzhihao/PostTrain/YuLan-Pretrain/.venv/lib/python3.11/site-packages/torch/distributed/distributed_c10d.py:4807: UserWarning: No device id is provided via `init_process_group` or `barrier `. Using the current device set by the user. 
+  warnings.warn(  # warn only once
+[rank0]:[W416 17:58:49.023645313 ProcessGroupNCCL.cpp:5023] [PG ID 0 PG GUID 0 Rank 0]  using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
+>>> done with compiling and loading fused kernels. Compilation time: 0.506 seconds
+WORLD_SIZE: 1, RANK: 0, LOCAL_RANK: 0
+building GPT model ...
+`torch_dtype` is deprecated! Use `dtype` instead!
+`torch_dtype` is deprecated! Use `dtype` instead!
+INFO:transformers_modules.tmp_dot_FZZhIF5Vmh.modeling_qwen3_next:[Qwen3Next custom] attn_position_embedding_type=rope, rnn_position_embedding_type=nope, attn_logits_scaling=None
+Qwen3NextForCausalLM(
+  (model): Qwen3NextModel(
+    (embed_tokens): Embedding(99000, 1920)
+    (layers): ModuleList(
+      (0-11): 12 x Qwen3NextDecoderLayer(
+        (linear_attn): Qwen3NextGatedDeltaNet(
+          (act): SiLUActivation()
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (in_proj_qkvz): Linear(in_features=1920, out_features=5120, bias=False)
+          (in_proj_ba): Linear(in_features=1920, out_features=64, bias=False)
+          (norm): FusedRMSNormGated(64, eps=1e-06, activation=silu)
+          (out_proj): Linear(in_features=2048, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (12): Qwen3NextDecoderLayer(
+        (self_attn): Qwen3NextAttention(
+          (q_proj): Linear(in_features=1920, out_features=1920, bias=True)
+          (k_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (v_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (o_proj): Linear(in_features=1920, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (13-19): 7 x Qwen3NextDecoderLayer(
+        (linear_attn): Qwen3NextGatedDeltaNet(
+          (act): SiLUActivation()
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (in_proj_qkvz): Linear(in_features=1920, out_features=5120, bias=False)
+          (in_proj_ba): Linear(in_features=1920, out_features=64, bias=False)
+          (norm): FusedRMSNormGated(64, eps=1e-06, activation=silu)
+          (out_proj): Linear(in_features=2048, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (20-21): 2 x Qwen3NextDecoderLayer(
+        (self_attn): Qwen3NextAttention(
+          (q_proj): Linear(in_features=1920, out_features=1920, bias=True)
+          (k_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (v_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (o_proj): Linear(in_features=1920, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (22): Qwen3NextDecoderLayer(
+        (linear_attn): Qwen3NextGatedDeltaNet(
+          (act): SiLUActivation()
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (in_proj_qkvz): Linear(in_features=1920, out_features=5120, bias=False)
+          (in_proj_ba): Linear(in_features=1920, out_features=64, bias=False)
+          (norm): FusedRMSNormGated(64, eps=1e-06, activation=silu)
+          (out_proj): Linear(in_features=2048, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (23): Qwen3NextDecoderLayer(
+        (self_attn): Qwen3NextAttention(
+          (q_proj): Linear(in_features=1920, out_features=1920, bias=True)
+          (k_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (v_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (o_proj): Linear(in_features=1920, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (24-45): 22 x Qwen3NextDecoderLayer(
+        (linear_attn): Qwen3NextGatedDeltaNet(
+          (act): SiLUActivation()
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (in_proj_qkvz): Linear(in_features=1920, out_features=5120, bias=False)
+          (in_proj_ba): Linear(in_features=1920, out_features=64, bias=False)
+          (norm): FusedRMSNormGated(64, eps=1e-06, activation=silu)
+          (out_proj): Linear(in_features=2048, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (46): Qwen3NextDecoderLayer(
+        (self_attn): Qwen3NextAttention(
+          (q_proj): Linear(in_features=1920, out_features=1920, bias=True)
+          (k_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (v_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (o_proj): Linear(in_features=1920, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (47): Qwen3NextDecoderLayer(
+        (linear_attn): Qwen3NextGatedDeltaNet(
+          (act): SiLUActivation()
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (in_proj_qkvz): Linear(in_features=1920, out_features=5120, bias=False)
+          (in_proj_ba): Linear(in_features=1920, out_features=64, bias=False)
+          (norm): FusedRMSNormGated(64, eps=1e-06, activation=silu)
+          (out_proj): Linear(in_features=2048, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (48-49): 2 x Qwen3NextDecoderLayer(
+        (self_attn): Qwen3NextAttention(
+          (q_proj): Linear(in_features=1920, out_features=1920, bias=True)
+          (k_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (v_proj): Linear(in_features=1920, out_features=384, bias=True)
+          (o_proj): Linear(in_features=1920, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+      (50-55): 6 x Qwen3NextDecoderLayer(
+        (linear_attn): Qwen3NextGatedDeltaNet(
+          (act): SiLUActivation()
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (in_proj_qkvz): Linear(in_features=1920, out_features=5120, bias=False)
+          (in_proj_ba): Linear(in_features=1920, out_features=64, bias=False)
+          (norm): FusedRMSNormGated(64, eps=1e-06, activation=silu)
+          (out_proj): Linear(in_features=2048, out_features=1920, bias=False)
+        )
+        (mlp): Qwen3NextMLP(
+          (gate_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (up_proj): Linear(in_features=1920, out_features=4800, bias=False)
+          (down_proj): Linear(in_features=4800, out_features=1920, bias=False)
+          (act_fn): SiLUActivation()
+        )
+        (input_layernorm): LlamaRMSNorm()
+        (post_attention_layernorm): LlamaRMSNorm()
+      )
+    )
+    (norm): LlamaRMSNorm()
+    (rotary_emb): Qwen3NextRotaryEmbedding()
+  )
+  (lm_head): Linear(in_features=1920, out_features=99000, bias=False)
+)
+GPTModel(
+  (embedding): LanguageModelEmbedding(
+    (word_embeddings): VocabParallelEmbedding()
+    (embedding_dropout): Dropout(p=0.1, inplace=False)
+  )
+  (rotary_pos_emb): RotaryEmbedding()
+  (decoder): TransformerBlock(
+    (layers): ModuleList(
+      (0-11): 12 x TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): GatedDeltaNet(
+          (in_proj): TELayerNormColumnParallelLinear(in_features=1920, out_features=5184, bias=False, TP=1)
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (out_norm): RMSNorm()
+          (out_proj): TERowParallelLinear(in_features=2048, out_features=1920, bias=False, TP=1)
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (12): TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): SelfAttention(
+          (core_attention): TEDotProductAttention(
+            (flash_attention): FlashAttention()
+            (fused_attention): FusedAttention()
+            (unfused_attention): UnfusedDotProductAttention(
+              (scale_mask_softmax): FusedScaleMaskSoftmax()
+              (attention_dropout): Dropout(p=0.1, inplace=False)
+            )
+          )
+          (linear_proj): TERowParallelLinear(in_features=1920, out_features=1920, bias=False, TP=1)
+          (linear_qkv): TELayerNormColumnParallelLinear(in_features=1920, out_features=2688, bias=True, TP=1)
+          (q_layernorm): IdentityOp()
+          (k_layernorm): IdentityOp()
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (13-19): 7 x TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): GatedDeltaNet(
+          (in_proj): TELayerNormColumnParallelLinear(in_features=1920, out_features=5184, bias=False, TP=1)
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (out_norm): RMSNorm()
+          (out_proj): TERowParallelLinear(in_features=2048, out_features=1920, bias=False, TP=1)
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (20-21): 2 x TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): SelfAttention(
+          (core_attention): TEDotProductAttention(
+            (flash_attention): FlashAttention()
+            (fused_attention): FusedAttention()
+            (unfused_attention): UnfusedDotProductAttention(
+              (scale_mask_softmax): FusedScaleMaskSoftmax()
+              (attention_dropout): Dropout(p=0.1, inplace=False)
+            )
+          )
+          (linear_proj): TERowParallelLinear(in_features=1920, out_features=1920, bias=False, TP=1)
+          (linear_qkv): TELayerNormColumnParallelLinear(in_features=1920, out_features=2688, bias=True, TP=1)
+          (q_layernorm): IdentityOp()
+          (k_layernorm): IdentityOp()
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (22): TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): GatedDeltaNet(
+          (in_proj): TELayerNormColumnParallelLinear(in_features=1920, out_features=5184, bias=False, TP=1)
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (out_norm): RMSNorm()
+          (out_proj): TERowParallelLinear(in_features=2048, out_features=1920, bias=False, TP=1)
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (23): TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): SelfAttention(
+          (core_attention): TEDotProductAttention(
+            (flash_attention): FlashAttention()
+            (fused_attention): FusedAttention()
+            (unfused_attention): UnfusedDotProductAttention(
+              (scale_mask_softmax): FusedScaleMaskSoftmax()
+              (attention_dropout): Dropout(p=0.1, inplace=False)
+            )
+          )
+          (linear_proj): TERowParallelLinear(in_features=1920, out_features=1920, bias=False, TP=1)
+          (linear_qkv): TELayerNormColumnParallelLinear(in_features=1920, out_features=2688, bias=True, TP=1)
+          (q_layernorm): IdentityOp()
+          (k_layernorm): IdentityOp()
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (24-45): 22 x TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): GatedDeltaNet(
+          (in_proj): TELayerNormColumnParallelLinear(in_features=1920, out_features=5184, bias=False, TP=1)
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (out_norm): RMSNorm()
+          (out_proj): TERowParallelLinear(in_features=2048, out_features=1920, bias=False, TP=1)
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (46): TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): SelfAttention(
+          (core_attention): TEDotProductAttention(
+            (flash_attention): FlashAttention()
+            (fused_attention): FusedAttention()
+            (unfused_attention): UnfusedDotProductAttention(
+              (scale_mask_softmax): FusedScaleMaskSoftmax()
+              (attention_dropout): Dropout(p=0.1, inplace=False)
+            )
+          )
+          (linear_proj): TERowParallelLinear(in_features=1920, out_features=1920, bias=False, TP=1)
+          (linear_qkv): TELayerNormColumnParallelLinear(in_features=1920, out_features=2688, bias=True, TP=1)
+          (q_layernorm): IdentityOp()
+          (k_layernorm): IdentityOp()
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (47): TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): GatedDeltaNet(
+          (in_proj): TELayerNormColumnParallelLinear(in_features=1920, out_features=5184, bias=False, TP=1)
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (out_norm): RMSNorm()
+          (out_proj): TERowParallelLinear(in_features=2048, out_features=1920, bias=False, TP=1)
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (48-49): 2 x TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): SelfAttention(
+          (core_attention): TEDotProductAttention(
+            (flash_attention): FlashAttention()
+            (fused_attention): FusedAttention()
+            (unfused_attention): UnfusedDotProductAttention(
+              (scale_mask_softmax): FusedScaleMaskSoftmax()
+              (attention_dropout): Dropout(p=0.1, inplace=False)
+            )
+          )
+          (linear_proj): TERowParallelLinear(in_features=1920, out_features=1920, bias=False, TP=1)
+          (linear_qkv): TELayerNormColumnParallelLinear(in_features=1920, out_features=2688, bias=True, TP=1)
+          (q_layernorm): IdentityOp()
+          (k_layernorm): IdentityOp()
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+      (50-55): 6 x TransformerLayer(
+        (input_layernorm): IdentityOp()
+        (self_attention): GatedDeltaNet(
+          (in_proj): TELayerNormColumnParallelLinear(in_features=1920, out_features=5184, bias=False, TP=1)
+          (conv1d): Conv1d(3072, 3072, kernel_size=(4,), stride=(1,), padding=(3,), groups=3072, bias=False)
+          (out_norm): RMSNorm()
+          (out_proj): TERowParallelLinear(in_features=2048, out_features=1920, bias=False, TP=1)
+        )
+        (pre_cross_attn_layernorm): IdentityOp()
+        (cross_attention): IdentityOp()
+        (cross_attn_bda): IdentityFuncOp()
+        (pre_mlp_layernorm): IdentityOp()
+        (mlp): MLP(
+          (linear_fc1): TELayerNormColumnParallelLinear(in_features=1920, out_features=9600, bias=False, TP=1)
+          (linear_fc2): TERowParallelLinear(in_features=4800, out_features=1920, bias=False, TP=1)
+        )
+      )
+    )
+    (final_layernorm): RMSNorm()
+  )
+  (output_layer): ColumnParallelLinear(in_features=1920, out_features=99000, bias=False, TP=1)
+)
+/workspace/lvzhihao/PostTrain/YuLan-Pretrain/megatron/core/dist_checkpointing/strategies/common.py:89: UserWarning: Environment variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD detected, since the`weights_only` argument was not explicitly passed to `torch.load`, forcing weights_only=False.
+  return torch.load(load_path, map_location='cpu')
+sharded_state_dict metadata loaded from the checkpoint: {'singleton_local_shards': True, 'distrib_optim_sharding_type': 'fully_reshardable', 'distrib_optim_fully_reshardable_mem_efficient': False, 'chained_optim_avoid_prefix': True}
+Job sharding has changed: Rerun state will be ignored
+ loading distributed checkpoint from /capacity/userdata/vc0e4b0o65t5/lvzhihao/PostTrain/YuLan-Pretrain/outputs/yulan_mini_sft/run_sl16384_tp1_pp1_cp2/checkpoint/yulan-gdn-sft-1b-sl16384-lr1e-5-gbs64-mb1-tp1-pp1-cp2 at iteration 2340
+/workspace/lvzhihao/PostTrain/YuLan-Pretrain/megatron/core/dist_checkpointing/strategies/torch.py:956: FutureWarning: `load_state_dict` is deprecated and will be removed in future versions. Please use `load` instead.
+  checkpoint.load_state_dict(
+ checkpoint version 3.0
+  successfully loaded checkpoint from /capacity/userdata/vc0e4b0o65t5/lvzhihao/PostTrain/YuLan-Pretrain/outputs/yulan_mini_sft/run_sl16384_tp1_pp1_cp2/checkpoint/yulan-gdn-sft-1b-sl16384-lr1e-5-gbs64-mb1-tp1-pp1-cp2 [ t 1/1, p 1/1 ] at iteration 2340
+INFO:root:Converting layer 0 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 1 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 2 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 3 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 4 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 5 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 6 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 7 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 8 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 9 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 10 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 11 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 12 is_gdn=False is_not_moe=True
+INFO:root:[DEBUG] Layer 12: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: attention_output_gate=False, linear_layer=TELayerNormColumnParallelLinear
+INFO:root:Converting layer 13 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 14 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 15 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 16 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 17 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 18 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 19 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 20 is_gdn=False is_not_moe=True
+INFO:root:[DEBUG] Layer 20: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: attention_output_gate=False, linear_layer=TELayerNormColumnParallelLinear
+INFO:root:Converting layer 21 is_gdn=False is_not_moe=True
+INFO:root:[DEBUG] Layer 21: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: attention_output_gate=False, linear_layer=TELayerNormColumnParallelLinear
+INFO:root:Converting layer 22 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 23 is_gdn=False is_not_moe=True
+INFO:root:[DEBUG] Layer 23: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: attention_output_gate=False, linear_layer=TELayerNormColumnParallelLinear
+INFO:root:Converting layer 24 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 25 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 26 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 27 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 28 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 29 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 30 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 31 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 32 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 33 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 34 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 35 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 36 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 37 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 38 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 39 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 40 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 41 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 42 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 43 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 44 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 45 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 46 is_gdn=False is_not_moe=True
+INFO:root:[DEBUG] Layer 46: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: attention_output_gate=False, linear_layer=TELayerNormColumnParallelLinear
+INFO:root:Converting layer 47 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 48 is_gdn=False is_not_moe=True
+INFO:root:[DEBUG] Layer 48: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: attention_output_gate=False, linear_layer=TELayerNormColumnParallelLinear
+INFO:root:Converting layer 49 is_gdn=False is_not_moe=True
+INFO:root:[DEBUG] Layer 49: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: args.attention_output_gate=False
+INFO:root:[DEBUG] set_gated_selfattn_state: attention_output_gate=False, linear_layer=TELayerNormColumnParallelLinear
+INFO:root:Converting layer 50 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 51 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 52 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 53 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 54 is_gdn=True is_not_moe=True
+INFO:root:Converting layer 55 is_gdn=True is_not_moe=True
+DEBUG:root:[RANK 0] 0 send op & 0 recv op.
+INFO:root:[Iters 0 RANK 0] starts synchronizing parameters with other ranks...
+INFO:root:[Iters 0 RANK 0] finishes synchronizing
+[Iters 0 RANK 0] model.safetensors is saved.
+DEBUG:root:[Iters 0 RANK 0] joined
+Conversion finished in 25.8255398273468 seconds.
+DEBUG:filelock:Attempting to acquire lock 140691634004240 on /root/.triton/autotune/Fp16Matmul_2d_kernel.pickle.lock
+DEBUG:filelock:Lock 140691634004240 acquired on /root/.triton/autotune/Fp16Matmul_2d_kernel.pickle.lock
+DEBUG:filelock:Attempting to release lock 140691634004240 on /root/.triton/autotune/Fp16Matmul_2d_kernel.pickle.lock
+DEBUG:filelock:Lock 140691634004240 released on /root/.triton/autotune/Fp16Matmul_2d_kernel.pickle.lock
+DEBUG:filelock:Attempting to acquire lock 140691634004560 on /root/.triton/autotune/Fp16Matmul_4d_kernel.pickle.lock
+DEBUG:filelock:Lock 140691634004560 acquired on /root/.triton/autotune/Fp16Matmul_4d_kernel.pickle.lock
+DEBUG:filelock:Attempting to release lock 140691634004560 on /root/.triton/autotune/Fp16Matmul_4d_kernel.pickle.lock
+DEBUG:filelock:Lock 140691634004560 released on /root/.triton/autotune/Fp16Matmul_4d_kernel.pickle.lock
+[rank0]:[W416 17:59:16.552855020 ProcessGroupNCCL.cpp:1538] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

YuLan-Mini-Nanbeige-Distill/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44938572364bf8fd69d242c80835739972d73d311ef23ba535bef86de5934087
+size 5343303296

YuLan-Mini-Nanbeige-Distill/modeling_qwen3_next.py

@@ -0,0 +1,1561 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/qwen3_next/modular_qwen3_next.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_qwen3_next.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 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.
+
+import math
+from typing import Any, Callable, Optional, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache
+from transformers.generation import GenerationMixin
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_layers import (
+    GradientCheckpointingLayer,
+)
+from transformers.modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
+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, logging
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import OutputRecorder, check_model_inputs
+from transformers.utils.import_utils import (
+    is_causal_conv1d_available,
+    is_flash_linear_attention_available,
+)
+try:
+    from configuration_qwen3_next import Qwen3NextConfig
+except ImportError:
+    from .configuration_qwen3_next import Qwen3NextConfig
+
+if is_causal_conv1d_available():
+    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
+else:
+    causal_conv1d_update, causal_conv1d_fn = None, None
+
+if is_flash_linear_attention_available():
+    from fla.modules import FusedRMSNormGated
+    from fla.ops.gated_delta_rule import chunk_gated_delta_rule, fused_recurrent_gated_delta_rule
+else:
+    chunk_gated_delta_rule, fused_recurrent_gated_delta_rule = None, None
+    FusedRMSNormGated = None
+
+logger = logging.get_logger(__name__)
+
+# ANSI colors for console (custom features log)
+_CG = "\033[92m"   # green
+_CY = "\033[93m"   # yellow
+_CC = "\033[96m"   # cyan
+_CR = "\033[0m"    # reset
+
+
+def _log_custom_features(config: Qwen3NextConfig) -> None:
+    """Log attn/rnn RoPE and attn_logits_scaling settings (once per model load)."""
+    attn_pe = getattr(config, "attn_position_embedding_type", "rope")
+    rnn_pe = getattr(config, "rnn_position_embedding_type", "nope")
+    logits_scaling = getattr(config, "attn_logits_scaling", None)
+    msg = (
+        f"{_CG}[Qwen3Next custom]{_CR} "
+        f"{_CC}attn_position_embedding_type{_CR}={_CY}{attn_pe}{_CR}, "
+        f"{_CC}rnn_position_embedding_type{_CR}={_CY}{rnn_pe}{_CR}, "
+        f"{_CC}attn_logits_scaling{_CR}={_CY}{logits_scaling}{_CR}"
+    )
+    logger.info(msg)
+
+
+class Qwen3NextRMSNormGated(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6, **kwargs):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states, gate=None):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        # Norm before gate
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        hidden_states = self.weight * hidden_states.to(input_dtype)
+        hidden_states = hidden_states * F.silu(gate.to(torch.float32))
+
+        return hidden_states.to(input_dtype)
+
+
+class Qwen3NextDynamicCache:
+    """
+    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the linear attention
+    cache (which has a constant shape regardless of seq_len).
+
+    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
+    and `ssm_states` for gated deltanet cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
+    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
+    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
+    For linear attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
+    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
+    and `recurrent_states` represents the recurrent state and has a shape of `(batch_size, d_inner, d_state)`.
+    """
+
+    is_compileable = False
+
+    def __init__(self, config: Qwen3NextConfig):
+        super().__init__()
+        self.layer_types = config.layer_types
+        self.transformer_layers = [
+            i for i in range(config.num_hidden_layers) if self.layer_types[i] == "full_attention"
+        ]
+        self.last_linear_layer = len(self.layer_types) - 1 - self.layer_types[::-1].index("linear_attention")
+
+        # Initialize everything to None -> will be lazy initialized to allow multi-gpu (device_map) inference
+        self.conv_states = [None for _ in range(config.num_hidden_layers)]
+        self.recurrent_states = [None for _ in range(config.num_hidden_layers)]
+        self.key_cache = [None for _ in range(config.num_hidden_layers)]
+        self.value_cache = [None for _ in range(config.num_hidden_layers)]
+
+    def __len__(self):
+        return len(self.layer_types)
+
+    def __getitem__(self, layer_idx: int) -> tuple[torch.Tensor, torch.Tensor]:
+        return self.key_cache[layer_idx], self.value_cache[layer_idx]
+
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[dict[str, Any]] = None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if self.key_cache[layer_idx] is None:
+            self.key_cache[layer_idx] = key_states
+            self.value_cache[layer_idx] = value_states
+        else:
+            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
+            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)
+
+        return self.key_cache[layer_idx], self.value_cache[layer_idx]
+
+    def reorder_cache(self, beam_idx: torch.LongTensor):
+        """Reorders the cache for beam search, given the selected beam indices."""
+        for layer_idx in range(len(self.key_cache)):
+            if self.key_cache[layer_idx] is not None:
+                device = self.key_cache[layer_idx].device
+                beam_idx = beam_idx.to(device)
+                self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx)
+                self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx)
+
+            if self.conv_states[layer_idx] is not None:
+                device = self.conv_states[layer_idx].device
+                beam_idx = beam_idx.to(device)
+                self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx)
+                self.recurrent_states[layer_idx] = self.recurrent_states[layer_idx].index_select(0, beam_idx)
+
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        # take any layer that contains cache and not empty tensor
+        layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
+        if len(self.key_cache) <= layer_idx or self.key_cache[layer_idx] is None:
+            return 0
+        return self.key_cache[layer_idx].shape[-2]
+
+    def get_mask_sizes(self, cache_position: torch.Tensor, layer_idx: int) -> tuple[int, int]:
+        """
+        Return a tuple (kv_length, kv_offset) corresponding to the length and offset that will be returned for
+        the given layer at `layer_idx`.
+        The masks are then prepared according to the given lengths (kv_length, kv_offset) and patterns for each layer.
+        """
+        kv_offset = 0
+        query_length = cache_position.shape[0]
+        past_seen_tokens = self.get_seq_length(layer_idx)
+        kv_length = query_length + past_seen_tokens
+        return kv_length, kv_offset
+
+    @property
+    def has_previous_state(self):
+        """We have a previous state if the last linear (conv) layer was already updated."""
+        return self.conv_states[self.last_linear_layer] is not None
+
+
+class Qwen3NextRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: Qwen3NextConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Qwen3NextRMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.zeros(dim))
+
+    def _norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        output = self._norm(x.float())
+        # Llama does x.to(float16) * w whilst Qwen3Next is (x * w).to(float16)
+        # See https://github.com/huggingface/transformers/pull/29402
+        output = output * (1.0 + self.weight.float())
+        return output.type_as(x)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+
+class LlamaRMSNorm(nn.Module):  # Copy from Llama
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+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)
+
+
+# Adapted from transformers.models.glm.modular_glm.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Removes the interleaving of cos and sin from GLM
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    # Keep half or full tensor for later concatenation
+    rotary_dim = cos.shape[-1]
+    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
+    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
+
+    # Apply rotary embeddings on the first half or full tensor
+    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
+    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)
+
+    # Concatenate back to full shape
+    q_embed = torch.cat([q_embed, q_pass], dim=-1)
+    k_embed = torch.cat([k_embed, k_pass], dim=-1)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class Qwen3NextAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Qwen3NextConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+        self.attn_output_gate = config.attn_output_gate
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim * (1 + self.attn_output_gate), bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=False
+        )
+        self.enable_qk_norm = config.enable_qk_norm
+        if self.enable_qk_norm:
+            self.q_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
+            self.k_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
+        else:
+            self.q_norm = None
+            self.k_norm = None
+
+        # Separate RoPE for attention: "rope" or "nope" (no RoPE)
+        self.attn_position_embedding_type = getattr(config, "attn_position_embedding_type", "rope")
+        # Optional logits scaling for length extrapolation: None, float, or "log" / "log <a>"
+        self.attn_logits_scaling = getattr(config, "attn_logits_scaling", None)
+
+        # Token shift on Q/K/V after projection (cannon layer, conv mode); per-head conv (head_dim, 1, kernel_size)
+        kernel_size = getattr(config, "token_shift_conv_size", 4)
+        self.attn_q_token_shift = getattr(config, "attn_q_token_shift", None)
+        self.attn_k_token_shift = getattr(config, "attn_k_token_shift", None)
+        self.attn_v_token_shift = getattr(config, "attn_v_token_shift", None)
+        if self.attn_q_token_shift == "conv":
+            self.q_token_shift_conv = nn.Conv1d(
+                self.head_dim,
+                self.head_dim,
+                kernel_size=kernel_size,
+                padding=0,
+                groups=self.head_dim,
+                bias=False,
+            )
+        else:
+            self.q_token_shift_conv = None
+        if self.attn_k_token_shift == "conv":
+            self.k_token_shift_conv = nn.Conv1d(
+                self.head_dim,
+                self.head_dim,
+                kernel_size=kernel_size,
+                padding=0,
+                groups=self.head_dim,
+                bias=False,
+            )
+        else:
+            self.k_token_shift_conv = None
+        if self.attn_v_token_shift == "conv":
+            self.v_token_shift_conv = nn.Conv1d(
+                self.head_dim,
+                self.head_dim,
+                kernel_size=kernel_size,
+                padding=0,
+                groups=self.head_dim,
+                bias=False,
+            )
+        else:
+            self.v_token_shift_conv = None
+
+    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        if self.attn_output_gate:
+            query_states, gate = torch.chunk(
+                self.q_proj(hidden_states).view(*input_shape, -1, self.head_dim * 2), 2, dim=-1
+            )
+            gate = gate.reshape(*input_shape, -1)
+        else:
+            query_states = self.q_proj(hidden_states).view(*input_shape, -1, self.head_dim)
+            gate = None
+
+        if self.enable_qk_norm:
+            query_states = self.q_norm(query_states.view(hidden_shape)).transpose(1, 2)
+            key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        else:
+            query_states = query_states.view(hidden_shape).transpose(1, 2)
+            key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        # Token shift on Q/K/V (cannon layer, conv mode): [batch, heads, seq, head_dim] -> [batch*heads, head_dim, seq]
+        if self.attn_q_token_shift == "conv" and self.q_token_shift_conv is not None:
+            b, nh, sq, hn = query_states.shape
+            q_bcl = query_states.reshape(b * nh, hn, sq)
+            q_bcl = apply_causal_depthwise_conv1d_bcl(q_bcl, self.q_token_shift_conv.weight)
+            query_states = q_bcl.reshape(b, nh, hn, sq).permute(0, 1, 3, 2).contiguous()
+        if self.attn_k_token_shift == "conv" and self.k_token_shift_conv is not None:
+            b, ng, sq, hn = key_states.shape
+            k_bcl = key_states.reshape(b * ng, hn, sq)
+            k_bcl = apply_causal_depthwise_conv1d_bcl(k_bcl, self.k_token_shift_conv.weight)
+            key_states = k_bcl.reshape(b, ng, hn, sq).permute(0, 1, 3, 2).contiguous()
+        if self.attn_v_token_shift == "conv" and self.v_token_shift_conv is not None:
+            b, ng, sq, hn = value_states.shape
+            v_bcl = value_states.reshape(b * ng, hn, sq)
+            v_bcl = apply_causal_depthwise_conv1d_bcl(v_bcl, self.v_token_shift_conv.weight)
+            value_states = v_bcl.reshape(b, ng, hn, sq).permute(0, 1, 3, 2).contiguous()
+
+        cos, sin = position_embeddings
+        if self.attn_position_embedding_type == "rope":
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+        # when "nope", do not apply RoPE
+
+        if past_key_values is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # Optional logits scaling for length extrapolation (before attention, same as FLA)
+        if self.attn_logits_scaling is not None:
+            if isinstance(self.attn_logits_scaling, (int, float)):
+                query_states = query_states * float(self.attn_logits_scaling)
+            elif isinstance(self.attn_logits_scaling, str):
+                position_ids = kwargs.get("position_ids")
+                if position_ids is None:
+                    batch_size, num_heads, seq_len, _ = query_states.shape
+                    position_ids = torch.arange(
+                        seq_len, device=query_states.device, dtype=torch.long
+                    ).unsqueeze(0).expand(batch_size, -1)
+                    logger.warning_once(
+                        "attn_logits_scaling uses position-dependent scaling but position_ids was not passed; "
+                        "using arange(0, seq_len). Pass position_ids for correct behavior with padding."
+                    )
+                parts = self.attn_logits_scaling.split()
+                a = float(parts[1]) if len(parts) > 1 else 362.0
+                position_ids_f = position_ids.to(device=query_states.device, dtype=torch.float32)
+                scale = (torch.log(position_ids_f + a) / math.log(a)).to(query_states.dtype)
+                # query_states: (B, H, T, D); scale: (B, T) -> broadcast
+                query_states = query_states * scale.unsqueeze(1).unsqueeze(-1)
+            else:
+                raise TypeError(
+                    f"attn_logits_scaling must be float, str or None, got {type(self.attn_logits_scaling)}"
+                )
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        if self.attn_output_gate:
+            attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+            attn_output = attn_output * torch.sigmoid(gate)
+        else:
+            attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+def apply_mask_to_padding_states(hidden_states, attention_mask):
+    """
+    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
+    """
+    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
+        dtype = hidden_states.dtype
+        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
+
+    return hidden_states
+
+
+is_fast_path_available = all(
+    (causal_conv1d_fn, causal_conv1d_update, chunk_gated_delta_rule, fused_recurrent_gated_delta_rule)
+)
+
+
+def torch_causal_conv1d_update(
+    hidden_states,
+    conv_state,
+    weight,
+    bias=None,
+    activation=None,
+):
+    _, hidden_size, seq_len = hidden_states.shape
+    state_len = conv_state.shape[-1]
+
+    hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
+    conv_state.copy_(hidden_states_new[:, :, -state_len:])
+    out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size)
+    out = F.silu(out[:, :, -seq_len:])
+    out = out.to(hidden_states.dtype)
+    return out
+
+
+def apply_causal_depthwise_conv1d_bcl(
+    x_bcl: torch.Tensor,
+    weight_c1w: torch.Tensor,
+) -> torch.Tensor:
+    """Apply causal depthwise Conv1d for token shifting. Aligns with Megatron cannon layer.
+
+    x_bcl: [B, C, L], weight_c1w: [C, 1, W]. Output: [B, C, L].
+    """
+    if causal_conv1d_fn is not None:
+        return causal_conv1d_fn(
+            x=x_bcl,
+            weight=weight_c1w.squeeze(1),
+            bias=None,
+            activation=None,
+        )
+    w = weight_c1w.shape[-1]
+    x_pad = F.pad(x_bcl, (w - 1, 0))
+    y = F.conv1d(
+        x_pad,
+        weight_c1w,
+        bias=None,
+        stride=1,
+        padding=0,
+        groups=weight_c1w.shape[0],
+    )
+    return y[..., : x_bcl.shape[-1]]
+
+
+def l2norm(x: torch.FloatTensor, dim: int = -1, eps: float = 1e-6):
+    """This function is intended to align with the l2norm implementation in the FLA library."""
+    inv_norm = 1 / torch.sqrt((x * x).sum(dim=dim, keepdim=True) + eps)
+    return x * inv_norm
+
+
+def torch_chunk_gated_delta_rule(
+    query,
+    key,
+    value,
+    g,
+    beta,
+    chunk_size=64,
+    initial_state=None,
+    output_final_state=False,
+    use_qk_l2norm_in_kernel=False,
+):
+    initial_dtype = query.dtype
+    if use_qk_l2norm_in_kernel:
+        query = l2norm(query, dim=-1, eps=1e-6)
+        key = l2norm(key, dim=-1, eps=1e-6)
+    query, key, value, beta, g = [
+        x.transpose(1, 2).contiguous().to(torch.float32) for x in (query, key, value, beta, g)
+    ]
+
+    batch_size, sequence_length, num_heads, k_head_dim = key.shape
+    v_head_dim = value.shape[-1]
+    pad_size = (chunk_size - num_heads % chunk_size) % chunk_size
+    query = F.pad(query, (0, 0, 0, pad_size))
+    key = F.pad(key, (0, 0, 0, pad_size))
+    value = F.pad(value, (0, 0, 0, pad_size))
+    beta = F.pad(beta, (0, pad_size))
+    g = F.pad(g, (0, pad_size))
+    tot_heads = num_heads + pad_size
+    scale = 1 / (query.shape[-1] ** 0.5)
+    query = query * scale
+
+    v_beta = value * beta.unsqueeze(-1)
+    k_beta = key * beta.unsqueeze(-1)
+    # reshape to chunks
+    query, key, value, k_beta, v_beta = [
+        x.reshape(x.shape[0], x.shape[1], -1, chunk_size, x.shape[-1]) for x in (query, key, value, k_beta, v_beta)
+    ]
+    g = g.reshape(g.shape[0], g.shape[1], -1, chunk_size)
+    mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=0)
+
+    # chunk decay
+    g = g.cumsum(dim=-1)
+    decay_mask = ((g.unsqueeze(-1) - g.unsqueeze(-2)).tril().exp().float()).tril()
+    attn = -((k_beta @ key.transpose(-1, -2)) * decay_mask).masked_fill(mask, 0)
+    for i in range(1, chunk_size):
+        row = attn[..., i, :i].clone()
+        sub = attn[..., :i, :i].clone()
+        attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)
+    attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
+    value = attn @ v_beta
+    k_cumdecay = attn @ (k_beta * g.exp().unsqueeze(-1))
+    last_recurrent_state = (
+        torch.zeros(batch_size, sequence_length, k_head_dim, v_head_dim).to(value)
+        if initial_state is None
+        else initial_state.to(value)
+    )
+    core_attn_out = torch.zeros_like(value)
+    mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=1)
+
+    # for each chunk
+    for i in range(0, tot_heads // chunk_size):
+        q_i, k_i, v_i = query[:, :, i], key[:, :, i], value[:, :, i]
+        attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
+        v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
+        v_new = v_i - v_prime
+        attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
+        core_attn_out[:, :, i] = attn_inter + attn @ v_new
+        last_recurrent_state = (
+            last_recurrent_state * g[:, :, i, -1, None, None].exp()
+            + (k_i * (g[:, :, i, -1, None] - g[:, :, i]).exp()[..., None]).transpose(-1, -2) @ v_new
+        )
+
+    if not output_final_state:
+        last_recurrent_state = None
+    core_attn_out = core_attn_out.reshape(core_attn_out.shape[0], core_attn_out.shape[1], -1, core_attn_out.shape[-1])
+    core_attn_out = core_attn_out[:, :, :num_heads]
+    core_attn_out = core_attn_out.transpose(1, 2).contiguous().to(initial_dtype)
+    return core_attn_out, last_recurrent_state
+
+
+def torch_recurrent_gated_delta_rule(
+    query, key, value, g, beta, initial_state, output_final_state, use_qk_l2norm_in_kernel=False
+):
+    initial_dtype = query.dtype
+    if use_qk_l2norm_in_kernel:
+        query = l2norm(query, dim=-1, eps=1e-6)
+        key = l2norm(key, dim=-1, eps=1e-6)
+    query, key, value, beta, g = [
+        x.transpose(1, 2).contiguous().to(torch.float32) for x in (query, key, value, beta, g)
+    ]
+
+    batch_size, sequence_length, num_heads, k_head_dim = key.shape
+    v_head_dim = value.shape[-1]
+    scale = 1 / (query.shape[-1] ** 0.5)
+    query = query * scale
+
+    core_attn_out = torch.zeros(batch_size, sequence_length, num_heads, v_head_dim).to(value)
+    last_recurrent_state = (
+        torch.zeros(batch_size, sequence_length, k_head_dim, v_head_dim).to(value)
+        if initial_state is None
+        else initial_state.to(value)
+    )
+
+    for i in range(num_heads):
+        q_t = query[:, :, i]
+        k_t = key[:, :, i]
+        v_t = value[:, :, i]
+        g_t = g[:, :, i].exp().unsqueeze(-1).unsqueeze(-1)
+        beta_t = beta[:, :, i].unsqueeze(-1)
+
+        last_recurrent_state = last_recurrent_state * g_t
+        kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
+        delta = (v_t - kv_mem) * beta_t
+        last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta.unsqueeze(-2)
+        core_attn_out[:, :, i] = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
+
+    if not output_final_state:
+        last_recurrent_state = None
+    core_attn_out = core_attn_out.transpose(1, 2).contiguous().to(initial_dtype)
+    return core_attn_out, last_recurrent_state
+
+
+class Qwen3NextGatedDeltaNet(nn.Module):
+    def __init__(self, config: Qwen3NextConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.num_v_heads = config.linear_num_value_heads
+        self.num_k_heads = config.linear_num_key_heads
+        self.head_k_dim = config.linear_key_head_dim
+        self.head_v_dim = config.linear_value_head_dim
+        self.key_dim = self.head_k_dim * self.num_k_heads
+        self.value_dim = self.head_v_dim * self.num_v_heads
+
+        self.conv_kernel_size = config.linear_conv_kernel_dim
+        self.layer_idx = layer_idx
+
+        # Optional RoPE for linear/RNN path: "rope" or "nope" (same as FLA rnn_position_embedding_type)
+        self.rnn_position_embedding_type = getattr(config, "rnn_position_embedding_type", "nope")
+        rope_theta = getattr(config, "rope_theta", 10000.0)
+        if self.rnn_position_embedding_type == "rope":
+            inv_freq = 1.0 / (
+                rope_theta ** (torch.arange(0, self.head_k_dim, 2, dtype=torch.float32) / self.head_k_dim)
+            )
+            self.register_buffer("_inv_freq", inv_freq, persistent=False)
+        else:
+            self._inv_freq = None
+        self.activation = config.hidden_act
+        self.act = ACT2FN[config.hidden_act]
+        self.layer_norm_epsilon = config.rms_norm_eps
+
+        # QKV
+        self.conv_dim = self.key_dim * 2 + self.value_dim
+        self.conv1d = nn.Conv1d(
+            in_channels=self.conv_dim,
+            out_channels=self.conv_dim,
+            bias=False,
+            kernel_size=self.conv_kernel_size,
+            groups=self.conv_dim,
+            padding=self.conv_kernel_size - 1,
+        )
+
+        # projection of the input hidden states
+        projection_size_qkvz = self.key_dim * 2 + self.value_dim * 2
+        projection_size_ba = self.num_v_heads * 2
+        self.in_proj_qkvz = nn.Linear(self.hidden_size, projection_size_qkvz, bias=False)
+        self.in_proj_ba = nn.Linear(self.hidden_size, projection_size_ba, bias=False)
+
+        # time step projection (discretization)
+        # instantiate once and copy inv_dt in init_weights of PretrainedModel
+        self.dt_bias = nn.Parameter(torch.ones(self.num_v_heads))
+
+        A = torch.empty(self.num_v_heads).uniform_(0, 16)
+        self.A_log = nn.Parameter(torch.log(A))
+
+        # self.norm = FusedRMSNormGated(self.head_v_dim, eps=self.layer_norm_epsilon)
+        self.norm = (
+            Qwen3NextRMSNormGated(self.head_v_dim, eps=self.layer_norm_epsilon)
+            if FusedRMSNormGated is None
+            else FusedRMSNormGated(
+                self.head_v_dim,
+                eps=self.layer_norm_epsilon,
+                activation=self.activation,
+                device=torch.cuda.current_device(),
+                dtype=config.dtype if config.dtype is not None else torch.get_current_dtype(),
+            )
+        )
+
+        self.out_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)
+
+        self.causal_conv1d_fn = causal_conv1d_fn
+        self.causal_conv1d_update = causal_conv1d_update or torch_causal_conv1d_update
+        self.chunk_gated_delta_rule = chunk_gated_delta_rule or torch_chunk_gated_delta_rule
+        self.recurrent_gated_delta_rule = fused_recurrent_gated_delta_rule or torch_recurrent_gated_delta_rule
+
+        if not is_fast_path_available:
+            logger.warning_once(
+                "The fast path is not available because one of the required library is not installed. Falling back to "
+                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and"
+                " https://github.com/Dao-AILab/causal-conv1d"
+            )
+
+    def fix_query_key_value_ordering(self, mixed_qkvz, mixed_ba):
+        """
+        Derives `query`, `key` and `value` tensors from `mixed_qkvz` and `mixed_ba`.
+        """
+        # (b, s, d_model) -> (b, s, num_k_heads, 2 * head_k_dim + 2 * head_v_dim * num_v_heads // num_k_heads)
+        new_tensor_shape_qkvz = mixed_qkvz.size()[:-1] + (
+            self.num_k_heads,
+            2 * self.head_k_dim + 2 * self.head_v_dim * self.num_v_heads // self.num_k_heads,
+        )
+        new_tensor_shape_ba = mixed_ba.size()[:-1] + (self.num_k_heads, 2 * self.num_v_heads // self.num_k_heads)
+
+        mixed_qkvz = mixed_qkvz.view(*new_tensor_shape_qkvz)
+        mixed_ba = mixed_ba.view(*new_tensor_shape_ba)
+        split_arg_list_qkvz = [
+            self.head_k_dim,
+            self.head_k_dim,
+            (self.num_v_heads // self.num_k_heads * self.head_v_dim),
+            (self.num_v_heads // self.num_k_heads * self.head_v_dim),
+        ]
+        split_arg_list_ba = [self.num_v_heads // self.num_k_heads, self.num_v_heads // self.num_k_heads]
+        query, key, value, z = torch.split(mixed_qkvz, split_arg_list_qkvz, dim=3)
+        b, a = torch.split(mixed_ba, split_arg_list_ba, dim=3)
+        # [b, sq, ng, np/ng * hn] -> [b, sq, np, hn]
+        value = value.reshape(value.size(0), value.size(1), -1, self.head_v_dim)
+        z = z.reshape(z.size(0), z.size(1), -1, self.head_v_dim)
+        b = b.reshape(b.size(0), b.size(1), self.num_v_heads)
+        a = a.reshape(a.size(0), a.size(1), self.num_v_heads)
+        return query, key, value, z, b, a
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cache_params: Optional[Qwen3NextDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+    ):
+        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
+
+        # Set up dimensions for reshapes later
+        batch_size, seq_len, _ = hidden_states.shape
+
+        use_precomputed_states = (
+            cache_params is not None
+            and cache_params.has_previous_state
+            and seq_len == 1
+            and cache_position is not None
+        )
+
+        # getting projected states from cache if it exists
+        if cache_params is not None:
+            conv_state = cache_params.conv_states[self.layer_idx]
+            recurrent_state = cache_params.recurrent_states[self.layer_idx]
+
+        projected_states_qkvz = self.in_proj_qkvz(hidden_states)
+        projected_states_ba = self.in_proj_ba(hidden_states)
+        query, key, value, z, b, a = self.fix_query_key_value_ordering(projected_states_qkvz, projected_states_ba)
+        query, key, value = (x.reshape(x.shape[0], x.shape[1], -1) for x in (query, key, value))
+
+        mixed_qkv = torch.cat((query, key, value), dim=-1)
+        mixed_qkv = mixed_qkv.transpose(1, 2)
+
+        if use_precomputed_states:
+            # 2. Convolution sequence transformation
+            # NOTE: the conv state is updated in `causal_conv1d_update`
+            mixed_qkv = self.causal_conv1d_update(
+                mixed_qkv,
+                conv_state,
+                self.conv1d.weight.squeeze(1),
+                self.conv1d.bias,
+                self.activation,
+            )
+        else:
+            if cache_params is not None:
+                conv_state = F.pad(mixed_qkv, (self.conv_kernel_size - mixed_qkv.shape[-1], 0))
+                cache_params.conv_states[self.layer_idx] = conv_state
+            if self.causal_conv1d_fn is not None:
+                mixed_qkv = self.causal_conv1d_fn(
+                    x=mixed_qkv,
+                    weight=self.conv1d.weight.squeeze(1),
+                    bias=self.conv1d.bias,
+                    activation=self.activation,
+                    seq_idx=None,
+                )
+            else:
+                mixed_qkv = F.silu(self.conv1d(mixed_qkv)[:, :, :seq_len])
+
+        mixed_qkv = mixed_qkv.transpose(1, 2)
+        query, key, value = torch.split(
+            mixed_qkv,
+            [
+                self.key_dim,
+                self.key_dim,
+                self.value_dim,
+            ],
+            dim=-1,
+        )
+        query = query.reshape(query.shape[0], query.shape[1], -1, self.head_k_dim)
+        key = key.reshape(key.shape[0], key.shape[1], -1, self.head_k_dim)
+        value = value.reshape(value.shape[0], value.shape[1], -1, self.head_v_dim)
+
+        # Optional RoPE for linear/RNN (same as FLA GatedDeltaNet when rnn_position_embedding_type=="rope")
+        if self._inv_freq is not None and position_ids is not None:
+            # query, key: (batch, seq, num_k_heads, head_k_dim); need cos, sin (batch, seq, head_k_dim)
+            inv_freq = self._inv_freq.to(query.device)
+            freqs = position_ids[:, :, None].float() * inv_freq[None, None, :]
+            emb = torch.cat([freqs, freqs], dim=-1)
+            cos = emb.cos().to(query.dtype)
+            sin = emb.sin().to(query.dtype)
+            query, key = apply_rotary_pos_emb(query, key, cos, sin, unsqueeze_dim=2)
+
+        beta = b.sigmoid()
+        # If the model is loaded in fp16, without the .float() here, A might be -inf
+        g = -self.A_log.float().exp() * F.softplus(a.float() + self.dt_bias)
+        if self.num_v_heads // self.num_k_heads > 1:
+            query = query.repeat_interleave(self.num_v_heads // self.num_k_heads, dim=2)
+            key = key.repeat_interleave(self.num_v_heads // self.num_k_heads, dim=2)
+
+        if not use_precomputed_states:
+            core_attn_out, last_recurrent_state = self.chunk_gated_delta_rule(
+                query,
+                key,
+                value,
+                g=g,
+                beta=beta,
+                initial_state=None,
+                output_final_state=cache_params is not None,
+                use_qk_l2norm_in_kernel=True,
+            )
+
+        else:
+            core_attn_out, last_recurrent_state = self.recurrent_gated_delta_rule(
+                query,
+                key,
+                value,
+                g=g,
+                beta=beta,
+                initial_state=recurrent_state,
+                output_final_state=cache_params is not None,
+                use_qk_l2norm_in_kernel=True,
+            )
+        # Update cache
+        if cache_params is not None:
+            cache_params.recurrent_states[self.layer_idx] = last_recurrent_state
+
+        z_shape_og = z.shape
+        # reshape input data into 2D tensor
+        # core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1])
+        # z = z.reshape(-1, z.shape[-1])
+        core_attn_out = core_attn_out
+        z = z
+        core_attn_out = self.norm(core_attn_out, z)
+
+        core_attn_out = core_attn_out.reshape(z_shape_og)
+        core_attn_out = core_attn_out.reshape(core_attn_out.shape[0], core_attn_out.shape[1], -1)
+
+        output = self.out_proj(core_attn_out)
+        return output
+
+
+class Qwen3NextMLP(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size if intermediate_size is not None else config.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]
+
+        # Token shifting (cannon layer): conv mode adds depthwise Conv1d
+        self.ffn_token_shift = getattr(config, "ffn_token_shift", None)
+        self.ffn_intermediate_token_shift = getattr(config, "ffn_intermediate_token_shift", None)
+        kernel_size = getattr(config, "token_shift_conv_size", 4)
+        if self.ffn_token_shift == "conv":
+            self.token_shift_conv = nn.Conv1d(
+                self.hidden_size,
+                self.hidden_size,
+                kernel_size=kernel_size,
+                padding=0,
+                groups=self.hidden_size,
+                bias=False,
+            )
+        else:
+            self.token_shift_conv = None
+        if self.ffn_intermediate_token_shift == "conv":
+            self.intermediate_token_shift_conv = nn.Conv1d(
+                self.intermediate_size,
+                self.intermediate_size,
+                kernel_size=kernel_size,
+                padding=0,
+                groups=self.intermediate_size,
+                bias=False,
+            )
+        else:
+            self.intermediate_token_shift_conv = None
+
+    def forward(self, x, per_token_scale=None):
+        # Token shift at MLP entry (conv mode; cat mode is stateless, not implemented here for loading conv checkpoints)
+        if self.ffn_token_shift == "conv" and self.token_shift_conv is not None:
+            # x: [batch, seq, hidden] -> [batch, hidden, seq]
+            x_bcl = x.transpose(1, 2).contiguous()
+            x_bcl = apply_causal_depthwise_conv1d_bcl(x_bcl, self.token_shift_conv.weight)
+            x = x_bcl.transpose(1, 2).contiguous()
+
+        # Compute gate and up projections
+        gate = self.gate_proj(x)
+        up = self.up_proj(x)
+        # Apply activation: act_fn(gate) * up
+        intermediate = self.act_fn(gate) * up
+        # Apply per_token_scale if provided (to align with Megatron's behavior)
+        if per_token_scale is not None:
+            intermediate = intermediate * per_token_scale.unsqueeze(-1)
+
+        # Intermediate token shift before down_proj (conv mode)
+        if self.ffn_intermediate_token_shift == "conv" and self.intermediate_token_shift_conv is not None:
+            inter_bcl = intermediate.transpose(1, 2).contiguous()
+            inter_bcl = apply_causal_depthwise_conv1d_bcl(inter_bcl, self.intermediate_token_shift_conv.weight)
+            intermediate = inter_bcl.transpose(1, 2).contiguous()
+
+        # Apply down projection
+        down_proj = self.down_proj(intermediate)
+        return down_proj
+
+
+class Qwen3NextSparseMoeBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.norm_topk_prob = config.norm_topk_prob
+        self.score_func = config.moe_router_score_function
+
+        # Token shift at MoE entry (cannon layer, reuses ffn_token_shift)
+        self.ffn_token_shift = getattr(config, "ffn_token_shift", None)
+        kernel_size = getattr(config, "token_shift_conv_size", 4)
+        if self.ffn_token_shift == "conv":
+            self.token_shift_conv = nn.Conv1d(
+                config.hidden_size,
+                config.hidden_size,
+                kernel_size=kernel_size,
+                padding=0,
+                groups=config.hidden_size,
+                bias=False,
+            )
+        else:
+            self.token_shift_conv = None
+
+        # gating
+        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=config.router_bias)
+        self.experts = nn.ModuleList(
+            [Qwen3NextMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
+        )
+
+        if config.shared_expert_intermediate_size > 0:
+            self.shared_expert = Qwen3NextMLP(config, intermediate_size=config.shared_expert_intermediate_size)
+            self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)
+        else:
+            self.shared_expert = None
+            self.shared_expert_gate = None
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+
+        # Token shift at MoE entry (cannon layer, conv mode)
+        if self.ffn_token_shift == "conv" and self.token_shift_conv is not None:
+            h_bcl = hidden_states.transpose(1, 2).contiguous()
+            h_bcl = apply_causal_depthwise_conv1d_bcl(h_bcl, self.token_shift_conv.weight)
+            hidden_states = h_bcl.transpose(1, 2).contiguous()
+
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        if self.score_func == "sigmoid":
+            routing_weights = torch.sigmoid(router_logits.to(torch.float32))
+        elif self.score_func == "softmax":
+            routing_weights = F.softmax(router_logits.to(torch.float32), dim=-1)
+        else:
+            raise NotImplementedError(f"Unknown score function {self.score_func}")
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        if self.norm_topk_prob:
+            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
+        for expert_idx in expert_hit:
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx].squeeze(0))
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert.
+            # To align with Megatron: apply routing_weights after activation, before down_proj
+            # (instead of after the full expert output)
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_routing_weights = routing_weights[top_x, idx]  # [M] where M is num tokens for this expert
+            # Pass routing_weights as per_token_scale to apply it after activation, before down_proj
+            current_hidden_states = expert_layer(current_state, per_token_scale=current_routing_weights)
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+
+        if self.shared_expert is not None:
+            shared_expert_output = self.shared_expert(hidden_states)
+            if self.shared_expert_gate is not None:
+                shared_expert_output = F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_expert_output
+
+            final_hidden_states = final_hidden_states + shared_expert_output
+
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits
+
+
+class Qwen3NextDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: Qwen3NextConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        # token mixer
+        self.layer_type = config.layer_types[layer_idx]
+        if self.layer_type == "linear_attention":
+            self.linear_attn = Qwen3NextGatedDeltaNet(config, layer_idx)
+        elif self.layer_type == "full_attention":
+            self.self_attn = Qwen3NextAttention(config, layer_idx)
+
+        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 = Qwen3NextSparseMoeBlock(config)
+        else:
+            self.mlp = Qwen3NextMLP(config, intermediate_size=config.intermediate_size)
+
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        # Token shift before attention (cannon layer, conv mode)
+        self.attn_token_shift = getattr(config, "attn_token_shift", None)
+        kernel_size = getattr(config, "token_shift_conv_size", 4)
+        if self.attn_token_shift == "conv":
+            self.attn_token_shift_conv = nn.Conv1d(
+                config.hidden_size,
+                config.hidden_size,
+                kernel_size=kernel_size,
+                padding=0,
+                groups=config.hidden_size,
+                bias=False,
+            )
+        else:
+            self.attn_token_shift_conv = None
+
+    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[tuple[torch.Tensor]] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> 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 (`Tuple(torch.FloatTensor)`, *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)
+
+        # Token shift before attention (cannon layer, conv mode)
+        if self.attn_token_shift == "conv" and self.attn_token_shift_conv is not None:
+            # [batch, seq, hidden] -> [batch, hidden, seq]
+            h_bcl = hidden_states.transpose(1, 2).contiguous()
+            h_bcl = apply_causal_depthwise_conv1d_bcl(h_bcl, self.attn_token_shift_conv.weight)
+            hidden_states = h_bcl.transpose(1, 2).contiguous()
+
+        # Token Mixer
+        if self.layer_type == "linear_attention":
+            hidden_states = self.linear_attn(
+                hidden_states=hidden_states,
+                cache_params=past_key_values,
+                cache_position=cache_position,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+            )
+        elif self.layer_type == "full_attention":
+            # Self Attention
+            hidden_states, _ = self.self_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        # For the MoE layers, we need to unpack
+        if isinstance(hidden_states, tuple):
+            hidden_states, _ = hidden_states
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class Qwen3NextPreTrainedModel(PreTrainedModel):
+    config: Qwen3NextConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Qwen3NextDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _keys_to_ignore_on_load_unexpected = [r"^mtp.*"]
+    _can_record_outputs = {
+        "router_logits": OutputRecorder(Qwen3NextSparseMoeBlock, index=1),
+        "hidden_states": Qwen3NextDecoderLayer,
+        "attentions": Qwen3NextAttention,
+    }
+    _is_stateful = True
+
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, Qwen3NextGatedDeltaNet):
+            module.dt_bias.data.fill_(1.0)
+            module.A_log.data.uniform_(0, 16).log_()
+
+
+class Qwen3NextModel(Qwen3NextPreTrainedModel):
+    def __init__(self, config: Qwen3NextConfig):
+        super().__init__(config)
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id)
+        self.layers = nn.ModuleList(
+            [Qwen3NextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Qwen3NextRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+        _log_custom_features(config)
+
+    # @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,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> MoeModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = Qwen3NextDynamicCache(config=self.config)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            input_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+        linear_attn_mask = self._update_linear_attn_mask(attention_mask, cache_position)
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            layer_mask = linear_attn_mask if decoder_layer.layer_type == "linear_attention" else causal_mask
+
+            hidden_states = decoder_layer(
+                hidden_states,
+                position_embeddings=position_embeddings,
+                attention_mask=layer_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+    def _update_linear_attn_mask(self, attention_mask, cache_position):
+        """
+        NOTE: Left-padding is used for linear attention mask.
+        No need for zeroing states when
+            1. Cached forward
+            2. Attending to all inputs
+        """
+        linear_attn_mask = attention_mask
+        if cache_position[0] > 0 or (attention_mask is not None and torch.all(attention_mask == 1)):
+            linear_attn_mask = None
+        return linear_attn_mask
+
+
+def load_balancing_loss_func(
+    gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None],
+    num_experts: Optional[int] = None,
+    top_k=2,
+    attention_mask: Optional[torch.Tensor] = None,
+) -> Union[torch.Tensor, int]:
+    r"""
+    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
+
+    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
+    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
+    experts is too unbalanced.
+
+    Args:
+        gate_logits:
+            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
+            shape [batch_size X sequence_length, num_experts].
+        num_experts:
+            Number of experts
+        top_k:
+            The number of experts to route per-token, can be also interpreted as the `top-k` routing
+            parameter.
+        attention_mask (`torch.Tensor`, *optional*):
+            The attention_mask used in forward function
+            shape [batch_size X sequence_length] if not None.
+
+    Returns:
+        The auxiliary loss.
+    """
+    if gate_logits is None or not isinstance(gate_logits, tuple):
+        return 0
+
+    if isinstance(gate_logits, tuple):
+        compute_device = gate_logits[0].device
+        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
+
+    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
+
+    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+
+    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
+
+    if attention_mask is None:
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.mean(routing_weights, dim=0)
+    else:
+        batch_size, sequence_length = attention_mask.shape
+        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
+        expert_attention_mask = (
+            attention_mask[None, :, :, None, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
+            .reshape(-1, top_k, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
+            expert_attention_mask, dim=0
+        )
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
+        router_per_expert_attention_mask = (
+            attention_mask[None, :, :, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
+            .reshape(-1, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
+            router_per_expert_attention_mask, dim=0
+        )
+
+    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
+    return overall_loss * num_experts
+
+
+@auto_docstring
+class Qwen3NextForCausalLM(Qwen3NextPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Qwen3NextModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.router_aux_loss_coef = config.router_aux_loss_coef
+        self.num_experts = config.num_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Qwen3NextDynamicCache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> MoeCausalLMOutputWithPast:
+        r"""
+        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]`.
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Qwen3NextForCausalLM
+
+        >>> model = Qwen3NextForCausalLM.from_pretrained("Qwen/Qwen3-Next-80B-A3B-Instruct")
+        >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-Next-80B-A3B-Instruct")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs: MoeModelOutputWithPast = 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_router_logits=output_router_logits,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+
+        aux_loss = None
+        if output_router_logits:
+            aux_loss = load_balancing_loss_func(
+                outputs.router_logits,
+                self.num_experts,
+                self.num_experts_per_tok,
+                attention_mask,
+            )
+            if labels is not None:
+                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+
+
+__all__ = [
+    "Qwen3NextForCausalLM",
+    "Qwen3NextModel",
+    "Qwen3NextPreTrainedModel",
+]

YuLan-Mini-Nanbeige-Distill/special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

YuLan-Mini-Nanbeige-Distill/tokenizer_config.json

@@ -0,0 +1,436 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "102": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "103": {
+      "content": "<reasoning_step>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "104": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "105": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "106": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "107": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "108": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "109": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "110": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "111": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "112": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "113": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "114": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "115": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "116": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "117": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "118": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "119": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "120": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "121": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "122": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "123": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "124": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "125": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "126": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "127": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "128": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "1071": {
+      "content": "<|sequence|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1072": {
+      "content": "<|/sequence|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1073": {
+      "content": "<|identity|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1074": {
+      "content": "<|identity|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1075": {
+      "content": "<|tail0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1076": {
+      "content": "<|tail1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1077": {
+      "content": "<|tail2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1078": {
+      "content": "<|tail3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1079": {
+      "content": "<|tail4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1080": {
+      "content": "<|head0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1081": {
+      "content": "<|head1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1082": {
+      "content": "<|head2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1083": {
+      "content": "<|head3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1084": {
+      "content": "<|head4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1085": {
+      "content": "<|chunk_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1086": {
+      "content": "<|/chunk_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1087": {
+      "content": "<|last_chunk_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1088": {
+      "content": "<|/last_chunk_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>",
+    "<|sequence|>",
+    "<|/sequence|>",
+    "<|identity|>",
+    "<|/identity|>",
+    "<|tail0|>",
+    "<|tail1|>",
+    "<|tail2|>",
+    "<|tail3|>",
+    "<|tail4|>",
+    "<|head0|>",
+    "<|head1|>",
+    "<|head2|>",
+    "<|head3|>",
+    "<|head4|>",
+    "<|chunk_id|>",
+    "<|/chunk_id|>",
+    "<|last_chunk_id|>",
+    "<|/last_chunk_id|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0].role == 'system' %}\n        {{- messages[0].content + '\\n\\n' }}\n    {%- endif %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if message.content is string %}\n        {%- set content = message.content %}\n    {%- else %}\n        {%- set content = '' %}\n    {%- endif %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n        {{- '<|im_start|>' + message.role + '\\n' + content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is string %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in content %}\n                {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n                {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- if loop.index0 > ns.last_query_index %}\n            {%- if loop.last or (not loop.last and reasoning_content) %}\n                {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content.strip('\\n') + '\\n</think>\\n\\n' + content.lstrip('\\n') }}\n            {%- else %}\n                {{- '<|im_start|>' + message.role + '\\n' + content }}\n            {%- endif %}\n        {%- else %}\n            {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- endif %}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- if tool_call.function %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>\\n{\"name\": \"' }}\n                {{- tool_call.name }}\n                {{- '\", \"arguments\": ' }}\n                {%- if tool_call.arguments is string %}\n                    {{- tool_call.arguments }}\n                {%- else %}\n                    {{- tool_call.arguments | tojson }}\n                {%- endif %}\n                {{- '}\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n<think>\\n' }}\n{%- endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizerFast",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}