Merge pull request #1210 from kvcache-ai/support-amx-qwen

Support amx and qwen3

ktransformers/models/configuration_qwen2_moe.py

+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Qwen2MoE model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen2MoeConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen2MoeModel`]. It is used to instantiate a
+    Qwen2MoE model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen1.5-MoE-A2.7B" [Qwen/Qwen1.5-MoE-A2.7B"](https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B").
+
+    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 Qwen2MoE model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen2MoeModel`]
+        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 24):
+            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 16):
+            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` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        decoder_sparse_step (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer.
+        moe_intermediate_size (`int`, *optional*, defaults to 1408):
+            Intermediate size of the routed expert.
+        shared_expert_intermediate_size (`int`, *optional*, defaults to 5632):
+            Intermediate size of the shared expert.
+        num_experts_per_tok (`int`, *optional*, defaults to 4):
+            Number of selected experts.
+        num_experts (`int`, *optional*, defaults to 60):
+            Number of routed experts.
+        norm_topk_prob (`bool`, *optional*, defaults to `False`):
+            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. Enabeling 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 Qwen2MoeMLP rather than Qwen2MoeSparseMoeBlock
+            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.
+
+    ```python
+    >>> from transformers import Qwen2MoeModel, Qwen2MoeConfig
+
+    >>> # Initializing a Qwen2MoE style configuration
+    >>> configuration = Qwen2MoeConfig()
+
+    >>> # Initializing a model from the Qwen1.5-MoE-A2.7B" style configuration
+    >>> model = Qwen2MoeModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen2_moe"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=2048,
+        intermediate_size=5632,
+        num_hidden_layers=24,
+        num_attention_heads=16,
+        num_key_value_heads=16,
+        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,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        decoder_sparse_step=1,
+        moe_intermediate_size=1408,
+        shared_expert_intermediate_size=5632,
+        num_experts_per_tok=4,
+        num_experts=60,
+        norm_topk_prob=False,
+        output_router_logits=False,
+        router_aux_loss_coef=0.001,
+        mlp_only_layers=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window if use_sliding_window else None
+        self.max_window_layers = max_window_layers
+
+        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.attention_dropout = attention_dropout
+
+        # 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 = [] if mlp_only_layers is None else mlp_only_layers
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
\ No newline at end of file

ktransformers/models/configuration_qwen3_moe.py

+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Qwen3MoE model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen3MoeConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen3MoeModel`]. It is used to instantiate a
+    Qwen3MoE model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of [Qwen/Qwen3-MoE-15B-A2B](https://huggingface.co/Qwen/Qwen3-15B-A2B).
+    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 Qwen3MoE model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen3MoeModel`]
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 6144):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 4):
+            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` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        decoder_sparse_step (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer.
+        moe_intermediate_size (`int`, *optional*, defaults to 768):
+            Intermediate size of the routed expert.
+        num_experts_per_tok (`int`, *optional*, defaults to 8):
+            Number of selected experts.
+        num_experts (`int`, *optional*, defaults to 128):
+            Number of routed experts.
+        norm_topk_prob (`bool`, *optional*, defaults to `False`):
+            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. Enabeling 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 Qwen3MoeMLP rather than Qwen3MoeSparseMoeBlock
+            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.
+    ```python
+    >>> from transformers import Qwen3MoeModel, Qwen3MoeConfig
+    >>> # Initializing a Qwen3MoE style configuration
+    >>> configuration = Qwen3MoeConfig()
+    >>> # Initializing a model from the Qwen3-15B-A2B" style configuration
+    >>> model = Qwen3MoeModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen3_moe"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `Qwen3Moe`
+    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.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=6144,
+        num_hidden_layers=24,
+        num_attention_heads=32,
+        num_key_value_heads=4,
+        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,
+        attention_bias=False,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        decoder_sparse_step=1,
+        moe_intermediate_size=768,
+        num_experts_per_tok=8,
+        num_experts=128,
+        norm_topk_prob=False,
+        output_router_logits=False,
+        router_aux_loss_coef=0.001,
+        mlp_only_layers=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window if use_sliding_window else None
+        self.max_window_layers = max_window_layers
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        # MoE arguments
+        self.decoder_sparse_step = decoder_sparse_step
+        self.moe_intermediate_size = moe_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 = [] if mlp_only_layers is None else mlp_only_layers
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["Qwen3MoeConfig"]
\ No newline at end of file

ktransformers/models/custom_cache.py

@@ -275,3 +275,59 @@ class KDeepSeekV3Cache(nn.Module):
         
         return page_idx, page_offset
     
+class KGQACache(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        page_size: int = 256,
+        dtype=torch.bfloat16,
+        device=torch.device("cuda:0"),
+        
+    ):
+        super().__init__()
+        self.config = config
+        self.dtype = dtype
+        self.device = device
+        self.page_size = page_size
+        self.k_caches = []
+        self.v_caches = []
+        
+
+    def load(self, inference_context: sched_ext.InferenceContext): 
+        print(self.config.num_hidden_layers)
+        for i in range(self.config.num_hidden_layers):
+            self.k_caches.append(
+                inference_context.k_cache[0][i] 
+            )
+            self.v_caches.append(
+                inference_context.v_cache[0][i]
+            )
+
+
+        self.max_cache_len = self.k_caches[0].shape[0]*self.k_caches[0].shape[1]
+
+
+        
+    def get_page_table(self, cache_position: torch.Tensor, q_indptr: torch.Tensor, kv_indptr: torch.Tensor, kv_indices: torch.Tensor, bsz_tensors: torch.tensor):
+        page_offset = cache_position % self.page_size  
+        page_idx_local = cache_position // self.page_size  
+        query_ids = torch.zeros_like(cache_position)
+        for i in range(len(q_indptr) - 1):
+            start_idx = q_indptr[i]
+            end_idx = q_indptr[i + 1]
+            query_ids[start_idx:end_idx] = i
+        page_idx = torch.zeros_like(page_idx_local)
+        for i in range(bsz_tensors[0]):
+            query_id = query_ids[i]
+            local_block = page_idx_local[i]
+            start_block = kv_indptr[query_id]
+            if local_block < kv_indptr[query_id + 1] - kv_indptr[query_id]:
+                page_idx[i] = kv_indices[start_block + local_block]
+        
+        return page_idx, page_offset
+
+    def get_k_cache(self, layer_idx):
+        return self.k_caches[layer_idx]
+
+    def get_v_cache(self, layer_idx):
+        return self.v_caches[layer_idx]
\ No newline at end of file

ktransformers/models/custom_modeling_qwen2_moe.py

+"""
+Date: 2024-11-06 10:05:11
+LastEditors: djw
+LastEditTime: 2024-11-13 07:50:51
+"""
+
+import math
+from dataclasses import dataclass
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+import math
+from typing import List, Optional, Tuple, Union
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from ktransformers.server.balance_serve.inference.forward_batch import ForwardBatchInput, ForwardBatchOutput
+from ktransformers.models.custom_cache import KGQACache
+from ktransformers.models.modeling_qwen2_moe import Qwen2MoeModel, Qwen2MoePreTrainedModel
+from ktransformers.models.configuration_qwen2_moe import Qwen2MoeConfig
+from ktransformers.operators.flashinfer_batch_prefill_wrapper import flashInferAttn
+
+torch.set_grad_enabled(False)
+torch.set_default_dtype(torch.bfloat16)
+import flashinfer
+
+class KQwen2MoeForCausalLM(Qwen2MoePreTrainedModel):
+
+    cache: KGQACache
+    use_cuda_graph = False
+    def __init__(
+        self,
+        config: Qwen2MoeConfig,
+        cache,
+    ):
+        super().__init__(config)
+        self.model = Qwen2MoeModel(config)
+        self.config = config
+        self.cache = cache
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.attn = [None] * 10
+        
+    def init_wrapper(self, use_cuda_graph, device, max_batch_token, max_batch_size, max_pages, cuda_graph_idx = 0):
+        self.attn[cuda_graph_idx] = flashInferAttn(use_cuda_graph=use_cuda_graph, max_batch_token=max_batch_token, max_batch_size=max_batch_size, max_pages=max_pages, device=device)
+
+
+    def batch_embeddings(self, batch: ForwardBatchInput, device="cuda:0"):
+        features = []
+        for i in range(batch.batch_size):
+            tokens = batch.minibatch.tokens.contiguous()
+            feature = (
+                self.model.embed_tokens(tokens.to(torch.device('cpu')))
+                .to(torch.bfloat16)
+                .to(device=device)
+            )
+            features.append(feature)
+
+        return features
+
+
+    def forward(
+        self,
+        batch: ForwardBatchInput | None = None,
+        features: List[torch.Tensor] | None = None,
+        bsz_tensors: torch.Tensor | None = None,
+        num_tokens_tensors: torch.Tensor | None = None,
+        page_idx: torch.Tensor | None = None,
+        page_offset: torch.Tensor | None = None,
+        cuda_graph_idx: int | None = 0
+    ) -> ForwardBatchOutput:
+        current_stream = torch.cuda.current_stream()
+
+        forward_batch_output = ForwardBatchOutput()
+
+        
+        hidden_states = features[0]
+        self.attn[cuda_graph_idx].calc_batch_indices(hidden_states.shape[0])
+
+        with torch.cuda.stream(current_stream):
+            residual = torch.zeros_like(hidden_states)
+            for i, decode_layer in enumerate(self.model.layers):
+                if self.model.transfer_map is not None and i in self.model.transfer_map:
+                    prev_stream = torch.cuda.current_stream()
+                    cur_device = self.model.transfer_map[i]
+                    if cur_device not in self.model.stream_device_map:
+                        self.model.stream_device_map[cur_device] = torch.cuda.Stream(cur_device)
+                    torch.cuda.set_device(cur_device)
+                    self.model.stream_device_map[cur_device].wait_stream(prev_stream)
+                    torch.cuda.set_stream(self.model.stream_device_map[cur_device])
+                    hidden_states = hidden_states.to(
+                        self.model.transfer_map[i], non_blocking=True
+                    )
+
+                    batch.minibatch.position_ids = (
+                        batch.minibatch.position_ids.to(self.model.transfer_map[i], non_blocking=True)
+                        if batch.minibatch.position_ids is not None
+                        else None
+                    )
+                hidden_states, residual = decode_layer.input_layernorm(hidden_states, num_tokens_tensors, residual)
+                hidden_states = decode_layer.self_attn(hidden_states, self.cache, 
+                                                       position_ids=batch.minibatch.position_ids, 
+                                                       wrapper=self.attn[cuda_graph_idx], bsz_tensors=num_tokens_tensors, 
+                                                       page_idx=page_idx,
+                                                       page_offset=page_offset
+                                                       )
+
+                hidden_states, residual = decode_layer.post_attention_layernorm(hidden_states, num_tokens_tensors, residual)
+                hidden_states = decode_layer.mlp(hidden_states.unsqueeze(0), num_tokens_tensors, cuda_graph_idx)
+                hidden_states = hidden_states.squeeze(0)
+        forward_batch_output = ForwardBatchOutput()
+        with torch.cuda.stream(current_stream):
+            local_logit = self.lm_head(self.model.norm(hidden_states, num_tokens_tensors, residual)[0], num_tokens_tensors)
+            forward_batch_output.logits.append(local_logit)
+
+        return forward_batch_output
+    
+
+               
+    def flash_infer_attn_plan(self, batch: ForwardBatchInput, bsz_tensors, num_tokens_tensors,
+        num_q_heads: int,
+        num_kv_heads: int,
+        head_dim: int,
+        page_size: int,
+        causal: bool,
+        q_data_type: torch.dtype,
+        kv_data_type: torch.dtype,
+        cuda_graph_idx: int = 0
+        ):
+        minibatch = batch.minibatch
+        self.attn[cuda_graph_idx].plan(minibatch.q_indptr, minibatch.kv_indptr, minibatch.kv_indices, 
+                          minibatch.kv_last_page_len, bsz_tensors, num_tokens_tensors,num_q_heads, num_kv_heads, head_dim, page_size, causal=causal, q_data_type=q_data_type, kv_data_type=kv_data_type)
+        
\ No newline at end of file

ktransformers/models/custom_modeling_qwen3_moe.py

+"""
+Date: 2024-11-06 10:05:11
+LastEditors: djw
+LastEditTime: 2024-11-13 07:50:51
+"""
+
+import math
+from dataclasses import dataclass
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+import math
+from typing import List, Optional, Tuple, Union
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from ktransformers.server.balance_serve.inference.forward_batch import ForwardBatchInput, ForwardBatchOutput
+from ktransformers.models.custom_cache import KGQACache
+from ktransformers.models.modeling_qwen3_moe import Qwen3MoeModel, Qwen3MoePreTrainedModel
+from ktransformers.models.configuration_qwen3_moe import Qwen3MoeConfig
+from ktransformers.operators.flashinfer_batch_prefill_wrapper import flashInferAttn
+
+torch.set_grad_enabled(False)
+torch.set_default_dtype(torch.bfloat16)
+import flashinfer
+
+class KQwen3MoeForCausalLM(Qwen3MoePreTrainedModel):
+
+    cache: KGQACache
+    use_cuda_graph = False
+    def __init__(
+        self,
+        config: Qwen3MoeConfig,
+        cache = None,
+    ):
+        super().__init__(config)
+        self.model = Qwen3MoeModel(config)
+        self.config = config
+        self.cache = cache
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.attn = [None] * 10
+        
+    def init_wrapper(self, use_cuda_graph, device, max_batch_token, max_batch_size, max_pages, cuda_graph_idx = 0):
+        self.attn[cuda_graph_idx] = flashInferAttn(use_cuda_graph=use_cuda_graph, max_batch_token=max_batch_token, max_batch_size=max_batch_size, max_pages=max_pages, device=device)
+
+
+    def batch_embeddings(self, batch: ForwardBatchInput, device="cuda:0"):
+        features = []
+        for i in range(batch.batch_size):
+            tokens = batch.minibatch.tokens.contiguous()
+            feature = (
+                self.model.embed_tokens(tokens.to(torch.device('cpu')))
+                .to(torch.bfloat16)
+                .to(device=device)
+            )
+            features.append(feature)
+
+        return features
+
+
+    def forward(
+        self,
+        batch: ForwardBatchInput | None = None,
+        features: List[torch.Tensor] | None = None,
+        bsz_tensors: torch.Tensor | None = None,
+        num_tokens_tensors: torch.Tensor | None = None,
+        page_idx: torch.Tensor | None = None,
+        page_offset: torch.Tensor | None = None,
+        cuda_graph_idx: int | None = 0
+    ) -> ForwardBatchOutput:
+        current_stream = torch.cuda.current_stream()
+
+        forward_batch_output = ForwardBatchOutput()
+
+        
+        hidden_states = features[0]
+        self.attn[cuda_graph_idx].calc_batch_indices(hidden_states.shape[0])
+
+        with torch.cuda.stream(current_stream):
+            residual = torch.zeros_like(hidden_states)
+            for i, decode_layer in enumerate(self.model.layers):
+                if self.model.transfer_map is not None and i in self.model.transfer_map:
+                    prev_stream = torch.cuda.current_stream()
+                    cur_device = self.model.transfer_map[i]
+                    if cur_device not in self.model.stream_device_map:
+                        self.model.stream_device_map[cur_device] = torch.cuda.Stream(cur_device)
+                    torch.cuda.set_device(cur_device)
+                    self.model.stream_device_map[cur_device].wait_stream(prev_stream)
+                    torch.cuda.set_stream(self.model.stream_device_map[cur_device])
+                    hidden_states = hidden_states.to(
+                        self.model.transfer_map[i], non_blocking=True
+                    )
+
+                    batch.minibatch.position_ids = (
+                        batch.minibatch.position_ids.to(self.model.transfer_map[i], non_blocking=True)
+                        if batch.minibatch.position_ids is not None
+                        else None
+                    )
+                hidden_states, residual = decode_layer.input_layernorm(hidden_states, num_tokens_tensors, residual)
+                hidden_states = decode_layer.self_attn(hidden_states, self.cache, 
+                                                       position_ids=batch.minibatch.position_ids, 
+                                                       wrapper=self.attn[cuda_graph_idx], bsz_tensors=num_tokens_tensors, 
+                                                       page_idx=page_idx,
+                                                       page_offset=page_offset
+                                                       )
+
+                hidden_states, residual = decode_layer.post_attention_layernorm(hidden_states, num_tokens_tensors, residual)
+                hidden_states = decode_layer.mlp(hidden_states.unsqueeze(0), num_tokens_tensors, cuda_graph_idx)
+                hidden_states = hidden_states.squeeze(0)
+        forward_batch_output = ForwardBatchOutput()
+        with torch.cuda.stream(current_stream):
+            local_logit = self.lm_head(self.model.norm(hidden_states, num_tokens_tensors, residual)[0], num_tokens_tensors)
+            forward_batch_output.logits.append(local_logit)
+
+        return forward_batch_output
+    
+
+               
+    def flash_infer_attn_plan(self, batch: ForwardBatchInput, bsz_tensors, num_tokens_tensors,
+        num_q_heads: int,
+        num_kv_heads: int,
+        head_dim: int,
+        page_size: int,
+        causal: bool,
+        q_data_type: torch.dtype,
+        kv_data_type: torch.dtype,
+        cuda_graph_idx: int = 0
+        ):
+        minibatch = batch.minibatch
+        self.attn[cuda_graph_idx].plan(minibatch.q_indptr, minibatch.kv_indptr, minibatch.kv_indices, 
+                          minibatch.kv_last_page_len, bsz_tensors, num_tokens_tensors, num_q_heads, num_kv_heads, head_dim, page_size, causal=causal, q_data_type=q_data_type, kv_data_type=kv_data_type)
+        
\ No newline at end of file

ktransformers/models/modeling_qwen3_moe.py

+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/qwen3_moe/modular_qwen3_moe.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_moe.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.
+
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+# from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+# from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.modeling_utils import PreTrainedModel
+# from transformers.processing_utils import Unpack
+from transformers.utils import (
+    # LossKwargs,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+from .configuration_qwen3_moe import Qwen3MoeConfig
+
+from ktransformers.models.modeling_qwen2_moe import Qwen2MoeRotaryEmbedding
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "Qwen/Qwen3-MoE-15B-A2B"
+_CONFIG_FOR_DOC = "Qwen3MoeConfig"
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    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 Qwen3MoeAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Qwen3MoeConfig, 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.num_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, 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=config.attention_bias
+        )
+        self.q_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
+        self.k_norm = Qwen3MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
+
+        self.rotary_emb = Qwen2MoeRotaryEmbedding(
+            self.head_dim,
+            max_position_embeddings=self.max_position_embeddings,
+            base=self.rope_theta,
+        )
+
+        self.sliding_window = config.sliding_window
+        if not (
+            self.config.use_sliding_window
+            and getattr(self.config, "sliding_window", None) is not None
+            and self.layer_idx >= self.config.max_window_layers
+        ):
+            self.sliding_window = None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        # **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value 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_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        # if self.config._attn_implementation != "eager":
+        #     if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+        #         logger.warning_once(
+        #             "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+        #             'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+        #         )
+        #     else:
+        #         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,
+            sliding_window=self.sliding_window,  # diff with Llama
+            # **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Qwen3MoeMLP(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]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class Qwen3MoeSparseMoeBlock(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
+
+        # gating
+        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
+        self.experts = nn.ModuleList(
+            [Qwen3MoeMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        if self.norm_topk_prob:  # only diff with mixtral sparse moe block!
+            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
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+
+            # 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))
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits
+
+
+class Qwen3MoeRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Qwen3MoeRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+        self.hidden_size = hidden_size
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class Qwen3MoeDecoderLayer(nn.Module):
+    def __init__(self, config: Qwen3MoeConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = Qwen3MoeAttention(config, layer_idx)
+        self.mlp = Qwen3MoeMLP(config)
+
+        self.self_attn = Qwen3MoeAttention(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 = Qwen3MoeSparseMoeBlock(config)
+        else:
+            self.mlp = Qwen3MoeMLP(config, intermediate_size=config.intermediate_size)
+
+        self.input_layernorm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        # **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, sequence_length)` where padding elements are indicated by 0.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_router_logits (`bool`, *optional*):
+                Whether or not to return the logits of all the routers. They are useful for computing the router loss,
+                and should not be returned during inference.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`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)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+
+        hidden_states = self.mlp(hidden_states)
+        if isinstance(hidden_states, tuple):
+            hidden_states, router_logits = hidden_states
+        else:
+            router_logits = None
+
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+def _compute_default_rope_parameters(
+    config: Optional[Qwen3MoeConfig] = None,
+    device: Optional["torch.device"] = None,
+    seq_len: Optional[int] = None,
+    **rope_kwargs,
+) -> Tuple["torch.Tensor", float]:
+    """
+    Computes the inverse frequencies according to the original RoPE implementation
+    Args:
+        config ([`~transformers.PretrainedConfig`]):
+            The model configuration.
+        device (`torch.device`):
+            The device to use for initialization of the inverse frequencies.
+        seq_len (`int`, *optional*):
+            The current sequence length. Unused for this type of RoPE.
+        rope_kwargs (`Dict`, *optional*):
+            BC compatibility with the previous RoPE class instantiation, will be removed in v4.45.
+    Returns:
+        Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+        post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+    """
+    if config is not None and len(rope_kwargs) > 0:
+        raise ValueError(
+            "Unexpected arguments: `**rope_kwargs` and `config` are mutually exclusive in "
+            f"`_compute_default_rope_parameters`, got `rope_kwargs`={rope_kwargs} and `config`={config}"
+        )
+    if len(rope_kwargs) > 0:
+        base = rope_kwargs["base"]
+        dim = rope_kwargs["dim"]
+    elif config is not None:
+        base = config.rope_theta
+        partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
+        dim = int(config.head_dim * partial_rotary_factor)
+
+    attention_factor = 1.0  # Unused in this type of RoPE
+
+    # Compute the inverse frequencies
+    inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).float().to(device) / dim))
+    return inv_freq, attention_factor
+
+class Qwen3MoeRotaryEmbedding(nn.Module):
+    def __init__(self, config: Qwen3MoeConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        self.scaling_factor = 1.0
+        self.dim = config.head_dim
+        self.max_position_embeddings = config.max_position_embeddings
+        self.base = config.rope_theta
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
+
+        inv_freq, self.attention_scaling = _compute_default_rope_parameters(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            # This .to() is needed if the model has been moved to a device after being initialized (because
+            # the buffer is automatically moved, but not the original copy)
+            self.original_inv_freq = self.original_inv_freq.to(device)
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+QWEN3_MOE_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Qwen3MoeConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen3Moe Model outputting raw hidden-states without any specific head on top.",
+    QWEN3_MOE_START_DOCSTRING,
+)
+class Qwen3MoePreTrainedModel(PreTrainedModel):
+    config_class = Qwen3MoeConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Qwen3MoeDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = False  # MoE models don't work with torch.compile (`torch.where(condition)` not supported)
+    _supports_attention_backend = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+QWEN3_MOE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen3Moe Model outputting raw hidden-states without any specific head on top.",
+    QWEN3_MOE_START_DOCSTRING,
+)
+class Qwen3MoeModel(Qwen3MoePreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen3MoeDecoderLayer`]
+
+    Args:
+        config: Qwen3MoeConfig
+    """
+
+    def __init__(self, config: Qwen3MoeConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [Qwen3MoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Qwen3MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Qwen3MoeRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN3_MOE_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        # **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.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 = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_logits = () if output_router_logits else None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    output_router_logits,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    # **flash_attn_kwargs,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            if output_router_logits:
+                all_router_logits += (layer_outputs[-1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        output = MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            router_logits=all_router_logits,
+        )
+        return output if return_dict else output.to_tuple()
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool = False,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and past_key_values is not None:
+                is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
+                if is_padding_right:
+                    raise ValueError(
+                        "You are attempting to perform batched generation with padding_side='right'"
+                        " this may lead to unexpected behaviour for Flash Attention version of Qwen3Moe. Make sure to "
+                        " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                    )
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if (
+            self.config._attn_implementation == "sdpa"
+            and not (using_static_cache or using_sliding_window_cache)
+            and not output_attentions
+        ):
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                sliding_window=self.config.sliding_window,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        # SlidingWindowCache or StaticCache
+        if using_sliding_window_cache or using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        # DynamicCache or no cache
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+            config=self.config,
+            past_key_values=past_key_values,
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type in ["cuda", "xpu"]
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        config: Qwen3MoeConfig,
+        past_key_values: Cache,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            device (`torch.device`):
+                The device to place the 4D attention mask on.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+            config (`Qwen3MoeConfig`):
+                The model's configuration class
+            past_key_values (`Cache`):
+                The cache class that is being used currently to generate
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+            )
+            diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            if config.sliding_window is not None:
+                # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
+                # the check is needed to verify is current checkpoint was trained with sliding window or not
+                if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
+                    sliding_attend_mask = torch.arange(target_length, device=device) <= (
+                        cache_position.reshape(-1, 1) - config.sliding_window
+                    )
+                    diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
+            causal_mask *= diagonal_attend_mask
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                if attention_mask.shape[-1] > target_length:
+                    attention_mask = attention_mask[:, :target_length]
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+                    causal_mask.device
+                )
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+        return causal_mask
+
+
+# class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+class KwargsForCausalLM(): ...
+
+
+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://arxiv.org/abs/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
+
+
+class Qwen3MoeForCausalLM(Qwen3MoePreTrainedModel, 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 = Qwen3MoeModel(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()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    @add_start_docstrings_to_model_forward(QWEN3_MOE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        # **kwargs: Unpack[KwargsForCausalLM],
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        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]`.
+
+            logits_to_keep (`int` or `torch.Tensor`, *optional*):
+                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+                This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Qwen3MoeForCausalLM
+
+        >>> model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+        >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            output_router_logits=output_router_logits,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            # **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # 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)
+
+        aux_loss = None
+        if output_router_logits:
+            aux_loss = load_balancing_loss_func(
+                outputs.router_logits if return_dict else outputs[-1],
+                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
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            if output_router_logits:
+                output = (aux_loss,) + output
+            return (loss,) + output if loss is not None else output
+
+        return 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,
+        )
+
+
+@add_start_docstrings(
+    """
+    The Qwen3Moe Model transformer with a sequence classification head on top (linear layer).
+
+    [`Qwen3MoeForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    QWEN3_MOE_START_DOCSTRING,
+)
+class Qwen3MoeForSequenceClassification(Qwen3MoePreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Qwen3MoeModel(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN3_MOE_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            last_non_pad_token = -1
+        elif input_ids is not None:
+            # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
+            non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32)
+            token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32)
+            last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
+        else:
+            last_non_pad_token = -1
+            logger.warning_once(
+                f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+            )
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token]
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config)
+
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The Qwen3Moe Model transformer with a token classification head on top (a linear layer on top of the hidden-states
+    output) e.g. for Named-Entity-Recognition (NER) tasks.
+    """,
+    QWEN3_MOE_START_DOCSTRING,
+)
+class Qwen3MoeForTokenClassification(Qwen3MoePreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Qwen3MoeModel(config)
+        if getattr(config, "classifier_dropout", None) is not None:
+            classifier_dropout = config.classifier_dropout
+        elif getattr(config, "hidden_dropout", None) is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.score = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN3_MOE_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        sequence_output = self.dropout(sequence_output)
+        logits = self.score(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.config)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+The Qwen3Moe Model transformer with a span classification head on top for extractive question-answering tasks like
+SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    QWEN3_MOE_START_DOCSTRING,
+)
+class Qwen3MoeForQuestionAnswering(Qwen3MoePreTrainedModel):
+    base_model_prefix = "transformer"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = Qwen3MoeModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, 2)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.transformer.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.transformer.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN3_MOE_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        loss = None
+        if start_positions is not None and end_positions is not None:
+            loss = self.loss_function(start_logits, end_logits, start_positions, end_positions, **kwargs)
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "Qwen3MoeForCausalLM",
+    "Qwen3MoeForQuestionAnswering",
+    "Qwen3MoeModel",
+    "Qwen3MoePreTrainedModel",
+    "Qwen3MoeForSequenceClassification",
+    "Qwen3MoeForTokenClassification",
+]
\ No newline at end of file

ktransformers/operators/RoPE.py

@@ -411,4 +411,30 @@ class RotaryEmbeddingV4(BaseInjectedModule):
         self.inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
         # self.register_buffer("inv_freq", inv_freq, persistent=False)
         # For BC we register cos and sin cached
-        self.max_seq_len_cached = max_position_embeddings
+        self.max_seq_len_cached = max_position_embeddings
\ No newline at end of file
+
+class KQwen3MoeRotaryEmbedding(BaseInjectedModule, DeepseekV2RotaryEmbedding):
+    def __init__(
+        self,
+        key: str,
+        gguf_loader: GGUFLoader,
+        config: PretrainedConfig,
+        orig_module: nn.Module,
+        #  device: str = "cuda",
+        generate_device: str = "cuda",
+        prefill_device: str = "cuda",
+        **kwargs,
+    ):
+        BaseInjectedModule.__init__(
+            self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs
+        )
+        self.orig_module.__init__(
+            config,
+        )
+        self.generate_device = generate_device
+        self.prefill_device = prefill_device
+
+    def load(self):
+        self.orig_module.__init__(
+            self.orig_module.config
+        )
\ No newline at end of file

ktransformers/operators/attention.py

@@ -762,92 +762,3 @@ class KLlamaAttention(BaseInjectedModule):
             attn_weights = None
 
         return attn_output, attn_weights, past_key_value
-
-class flashinfer_attn(BaseInjectedModule, DeepseekV2Attention):
-    def __init__(self,
-                 key: str,
-                 gguf_loader : GGUFLoader,
-                 config: PretrainedConfig,
-                 orig_module: nn.Module,
-                 prefill_device: str = "cuda",
-                 generate_device: str = "cuda",
-                 chunck_size: int = 1000,
-                 **kwargs):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
-        self.orig_module.__init__(orig_module.config,
-            orig_module.layer_idx)
-        self.chunck_size = chunck_size # TODO, generate chunck_size automatically.
-
-
-    def get_absorbed(self) -> Tuple[torch.Tensor, torch.Tensor]:
-        if not (hasattr(self, 'q_absorb') and hasattr(self, 'out_absorb')):
-            kv_b_proj = self.kv_b_proj.weight.view(self.num_heads, -1, self.kv_lora_rank)
-            q_absorb = kv_b_proj[:, :self.qk_nope_head_dim, :].reshape(-1, self.kv_lora_rank)
-            out_absorb = kv_b_proj[:, self.qk_nope_head_dim:, :].reshape(-1, self.kv_lora_rank)
-            self.q_absorb = nn.Linear(self.kv_lora_rank, self.num_heads * self.qk_nope_head_dim, 
-                                      bias=False, dtype=q_absorb.dtype, device=q_absorb.device)
-            self.q_absorb.weight.data = q_absorb
-            self.out_absorb = nn.Linear(self.kv_lora_rank, self.num_heads * self.v_head_dim, 
-                                        bias=False, dtype=out_absorb.dtype, device=out_absorb.device)
-            self.out_absorb.weight.data = out_absorb
-            #del self.orig_module.kv_b_proj
-        q_absorb = self.q_absorb.weight.view(self.num_heads, self.qk_nope_head_dim, self.kv_lora_rank)
-        out_absorb = self.out_absorb.weight.view(self.num_heads, self.v_head_dim, self.kv_lora_rank)
-        return q_absorb, out_absorb
-    
-
-
-    def forward(self,
-                hidden_states: torch.Tensor,
-                kv_cache: KDeepSeekV3Cache,
-                position_ids: torch.Tensor,
-                wrapper: BatchMLAPagedAttentionWrapper,
-                num_tokens_tensors: torch.Tensor,
-                page_idx: torch.Tensor,
-                page_offset: torch.Tensor,
-                ):
-        q_len, _ = hidden_states.size()
-
-        if self.q_lora_rank is None:
-            q = self.q_proj(hidden_states, num_tokens_tensors)
-        else:
-            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states, num_tokens_tensors), num_tokens_tensors), num_tokens_tensors)
-        q = q.view(q_len, self.num_heads, self.q_head_dim)
-        q_nope, q_pe = torch.split(
-            q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
-        )
-
-        compressed_kv = self.kv_a_proj_with_mqa(hidden_states, num_tokens_tensors)
-        compressed_kv, k_pe = torch.split(
-            compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
-        )
-        compressed_kv = compressed_kv.contiguous()
-        compressed_kv = self.kv_a_layernorm(compressed_kv, num_tokens_tensors)
-        k_pe = k_pe.view(q_len, 1, self.qk_rope_head_dim)
-        compressed_kv = compressed_kv.view(q_len, 1, self.kv_lora_rank)
-        
-        cos, sin = self.rotary_emb(q_pe, position_ids.unsqueeze(0))
-        q_pe, k_pe = apply_rotary_pos_emb(q_pe.unsqueeze(0), k_pe.unsqueeze(0), cos, sin, unsqueeze_dim=2)
-        q_pe = q_pe.squeeze(0)
-        if kv_cache is not None:
-            
-            # page_idx, page_offset = kv_cache.get_page_table(position_ids, q_indptr, kv_indptr, kv_indices)
-            cache_kwargs = {"sin": sin, "cos": cos, "page_idx": page_idx, "page_offset": page_offset}  # Specific to RoPE models
-            compressed_kv_with_k_pe = kv_cache.update(compressed_kv.unsqueeze(0), k_pe, self.layer_idx, page_idx, page_offset, cache_kwargs)
-            compressed_kv = compressed_kv_with_k_pe [:, :, :, :self.kv_lora_rank].view(-1, kv_cache.page_size, self.kv_lora_rank)
-            k_pe = compressed_kv_with_k_pe [:, :, :, self.kv_lora_rank:].view(-1, kv_cache.page_size, self.qk_rope_head_dim)
-            
-        q_absorb, out_absorb = self.get_absorbed()
-        q_nope = q_nope.transpose(0, 1) # q_len is 1, no GPU overhead, same below
-        q_nope = torch.matmul(q_nope, q_absorb) # batched MM
-        q_nope = q_nope.transpose(0, 1)
-        # q_nope.squeeze_(1)
-        # q_pe.squeeze_(1)
-
-        attn_output = wrapper.run(q_nope, q_pe, compressed_kv, k_pe).view(q_len, self.num_heads, self.kv_lora_rank)
-        attn_output = attn_output.transpose(0, 1)
-        attn_output = torch.matmul(attn_output, out_absorb.mT) # [self.num_heads, q_len, self.v_head_dim]
-        attn_output = attn_output.transpose(0, 1)
-        attn_output = attn_output.reshape(q_len, self.num_heads * self.v_head_dim)
-        attn_output = self.o_proj(attn_output, num_tokens_tensors)
-        return attn_output

ktransformers/operators/balance_serve_attention.py

+'''
+Description  :  
+Author       : Boxin Zhang
+Version      : 0.2.5
+Copyright (c) 2024 by KVCache.AI, All Rights Reserved. 
+'''
+import torch
+from torch import nn
+from ktransformers.models.modeling_deepseek import DeepseekV2Attention, apply_rotary_pos_emb
+from ktransformers.models.modeling_qwen2_moe import Qwen2MoeAttention
+from ktransformers.models.modeling_qwen3_moe import Qwen3MoeAttention
+from typing import Optional, Tuple
+from ktransformers.operators.base_operator import BaseInjectedModule
+from ktransformers.util.custom_gguf import GGUFLoader
+import logging
+from transformers.configuration_utils import PretrainedConfig
+from flashinfer import BatchMLAPagedAttentionWrapper
+from ktransformers.operators.flashinfer_batch_prefill_wrapper import flashInferAttn
+from ktransformers.models.custom_cache import KDeepSeekV3Cache, KGQACache
+logger = logging.getLogger("attention")
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+class flashinfer_attn(BaseInjectedModule, DeepseekV2Attention):
+    def __init__(self,
+                 key: str,
+                 gguf_loader : GGUFLoader,
+                 config: PretrainedConfig,
+                 orig_module: nn.Module,
+                 prefill_device: str = "cuda",
+                 generate_device: str = "cuda",
+                 chunck_size: int = 1000,
+                 **kwargs):
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
+        self.orig_module.__init__(orig_module.config,
+            orig_module.layer_idx)
+        self.chunck_size = chunck_size # TODO, generate chunck_size automatically.
+
+
+    def get_absorbed(self) -> Tuple[torch.Tensor, torch.Tensor]:
+        if not (hasattr(self, 'q_absorb') and hasattr(self, 'out_absorb')):
+            kv_b_proj = self.kv_b_proj.weight.view(self.num_heads, -1, self.kv_lora_rank)
+            q_absorb = kv_b_proj[:, :self.qk_nope_head_dim, :].reshape(-1, self.kv_lora_rank)
+            out_absorb = kv_b_proj[:, self.qk_nope_head_dim:, :].reshape(-1, self.kv_lora_rank)
+            self.q_absorb = nn.Linear(self.kv_lora_rank, self.num_heads * self.qk_nope_head_dim, 
+                                      bias=False, dtype=q_absorb.dtype, device=q_absorb.device)
+            self.q_absorb.weight.data = q_absorb
+            self.out_absorb = nn.Linear(self.kv_lora_rank, self.num_heads * self.v_head_dim, 
+                                        bias=False, dtype=out_absorb.dtype, device=out_absorb.device)
+            self.out_absorb.weight.data = out_absorb
+            #del self.orig_module.kv_b_proj
+        q_absorb = self.q_absorb.weight.view(self.num_heads, self.qk_nope_head_dim, self.kv_lora_rank)
+        out_absorb = self.out_absorb.weight.view(self.num_heads, self.v_head_dim, self.kv_lora_rank)
+        return q_absorb, out_absorb
+    
+
+    def forward(self,
+                hidden_states: torch.Tensor,
+                kv_cache: KDeepSeekV3Cache,
+                position_ids: torch.Tensor,
+                wrapper: BatchMLAPagedAttentionWrapper,
+                num_tokens_tensors: torch.Tensor,
+                page_idx: torch.Tensor,
+                page_offset: torch.Tensor,
+                ):
+        q_len, _ = hidden_states.size()
+
+        if self.q_lora_rank is None:
+            q = self.q_proj(hidden_states, num_tokens_tensors)
+        else:
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states, num_tokens_tensors), num_tokens_tensors), num_tokens_tensors)
+        q = q.view(q_len, self.num_heads, self.q_head_dim)
+        q_nope, q_pe = torch.split(
+            q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
+        )
+
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states, num_tokens_tensors)
+        compressed_kv, k_pe = torch.split(
+            compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+        )
+        compressed_kv = compressed_kv.contiguous()
+        compressed_kv = self.kv_a_layernorm(compressed_kv, num_tokens_tensors)
+        k_pe = k_pe.view(q_len, 1, self.qk_rope_head_dim)
+        compressed_kv = compressed_kv.view(q_len, 1, self.kv_lora_rank)
+        
+        cos, sin = self.rotary_emb(q_pe, position_ids.unsqueeze(0))
+        q_pe, k_pe = apply_rotary_pos_emb(q_pe.unsqueeze(0), k_pe.unsqueeze(0), cos, sin, unsqueeze_dim=2)
+        q_pe = q_pe.squeeze(0)
+        if kv_cache is not None:
+            
+            # page_idx, page_offset = kv_cache.get_page_table(position_ids, q_indptr, kv_indptr, kv_indices)
+            cache_kwargs = {"sin": sin, "cos": cos, "page_idx": page_idx, "page_offset": page_offset}  # Specific to RoPE models
+            compressed_kv_with_k_pe = kv_cache.update(compressed_kv.unsqueeze(0), k_pe, self.layer_idx, page_idx, page_offset, cache_kwargs)
+            compressed_kv = compressed_kv_with_k_pe [:, :, :, :self.kv_lora_rank].view(-1, kv_cache.page_size, self.kv_lora_rank)
+            k_pe = compressed_kv_with_k_pe [:, :, :, self.kv_lora_rank:].view(-1, kv_cache.page_size, self.qk_rope_head_dim)
+            
+        q_absorb, out_absorb = self.get_absorbed()
+        q_nope = q_nope.transpose(0, 1) # q_len is 1, no GPU overhead, same below
+        q_nope = torch.matmul(q_nope, q_absorb) # batched MM
+        q_nope = q_nope.transpose(0, 1)
+        # q_nope.squeeze_(1)
+        # q_pe.squeeze_(1)
+
+        attn_output = wrapper.run(q_nope, q_pe, compressed_kv, k_pe).view(q_len, self.num_heads, self.kv_lora_rank)
+        attn_output = attn_output.transpose(0, 1)
+        attn_output = torch.matmul(attn_output, out_absorb.mT) # [self.num_heads, q_len, self.v_head_dim]
+        attn_output = attn_output.transpose(0, 1)
+        attn_output = attn_output.reshape(q_len, self.num_heads * self.v_head_dim)
+        attn_output = self.o_proj(attn_output, num_tokens_tensors)
+        return attn_output
+
+class KQwen2MoeAttention(BaseInjectedModule, Qwen2MoeAttention):
+    def __init__(self,
+                 key: str,
+                 gguf_loader : GGUFLoader,
+                 config: PretrainedConfig,
+                 orig_module: nn.Module,
+                 prefill_device: str = "cuda",
+                 generate_device: str = "cuda",
+                 chunck_size: int = 1000,
+                 **kwargs):
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
+        self.orig_module.__init__(orig_module.config,
+            orig_module.layer_idx)
+        self.chunck_size = chunck_size # TODO, generate chunck_size automatically.
+
+
+    # Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
+    def apply_rotary_pos_emb(self, q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+        """Applies Rotary Position Embedding to the query and key tensors.
+
+        Args:
+            q (`torch.Tensor`): The query tensor.
+            k (`torch.Tensor`): The key tensor.
+            cos (`torch.Tensor`): The cosine part of the rotary embedding.
+            sin (`torch.Tensor`): The sine part of the rotary embedding.
+            position_ids (`torch.Tensor`):
+                Deprecated and unused.
+            unsqueeze_dim (`int`, *optional*, defaults to 1):
+                The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+                sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+                that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+                k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+                cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+                the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+        Returns:
+            `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+        """
+        cos = cos.unsqueeze(unsqueeze_dim)
+        sin = sin.unsqueeze(unsqueeze_dim)
+        q_embed = (q * cos) + (rotate_half(q) * sin)
+        k_embed = (k * cos) + (rotate_half(k) * sin)
+        return q_embed, k_embed
+
+
+    def forward(self,
+                hidden_states: torch.Tensor,
+                kv_cache: KGQACache,
+                position_ids: torch.Tensor,
+                wrapper: flashInferAttn,
+                bsz_tensors: torch.Tensor,
+                page_idx: torch.Tensor,
+                page_offset: torch.Tensor,
+                ):
+        q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states, bsz_tensors)
+        key_states = self.k_proj(hidden_states, bsz_tensors)
+        value_states = self.v_proj(hidden_states, bsz_tensors)
+
+
+        query_states = query_states.view(q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(q_len, self.num_key_value_heads, self.head_dim)
+        value_states = value_states.view(q_len, self.num_key_value_heads, self.head_dim)
+        
+        cos, sin = self.rotary_emb(value_states.unsqueeze(0), position_ids.unsqueeze(0))
+        query_states, key_states = self.apply_rotary_pos_emb(query_states.unsqueeze(0), key_states.unsqueeze(0), cos, sin, unsqueeze_dim=2)
+
+        query_states = query_states.view(q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(
+            q_len, self.num_key_value_heads, self.head_dim
+        )
+        value_states = value_states.view(
+            q_len, self.num_key_value_heads, self.head_dim
+        )
+
+        k_cache = kv_cache.get_k_cache(self.layer_idx)
+        v_cache = kv_cache.get_v_cache(self.layer_idx)
+
+
+        attn_output = wrapper.forward(query_states, k_cache, v_cache, key_states, value_states)
+  
+
+        attn_output = self.o_proj(attn_output.view(q_len, self.num_heads * self.head_dim), bsz_tensors)
+
+        return attn_output
+
+class KQwen3MoeAttention(BaseInjectedModule, Qwen3MoeAttention):
+    def __init__(self,
+                 key: str,
+                 gguf_loader : GGUFLoader,
+                 config: PretrainedConfig,
+                 orig_module: nn.Module,
+                 prefill_device: str = "cuda",
+                 generate_device: str = "cuda",
+                 chunck_size: int = 1000,
+                 **kwargs):
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
+        self.orig_module.__init__(orig_module.config,
+            orig_module.layer_idx)
+        self.chunck_size = chunck_size # TODO, generate chunck_size automatically.
+
+
+    # Copied from transformers.models.mistral.modeling_mistral.apply_rotary_pos_emb
+    def apply_rotary_pos_emb(self, q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+        """Applies Rotary Position Embedding to the query and key tensors.
+
+        Args:
+            q (`torch.Tensor`): The query tensor.
+            k (`torch.Tensor`): The key tensor.
+            cos (`torch.Tensor`): The cosine part of the rotary embedding.
+            sin (`torch.Tensor`): The sine part of the rotary embedding.
+            position_ids (`torch.Tensor`):
+                Deprecated and unused.
+            unsqueeze_dim (`int`, *optional*, defaults to 1):
+                The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+                sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+                that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+                k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+                cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+                the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+        Returns:
+            `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+        """
+        cos = cos.unsqueeze(unsqueeze_dim)
+        sin = sin.unsqueeze(unsqueeze_dim)
+        q_embed = (q * cos) + (rotate_half(q) * sin)
+        k_embed = (k * cos) + (rotate_half(k) * sin)
+        return q_embed, k_embed
+
+
+    def forward(self,
+                hidden_states: torch.Tensor,
+                kv_cache: KGQACache,
+                position_ids: torch.Tensor,
+                wrapper: flashInferAttn,
+                bsz_tensors: torch.Tensor,
+                page_idx: torch.Tensor,
+                page_offset: torch.Tensor,
+                ):
+        q_len, _ = hidden_states.size()
+
+        bsz_tensors_q = bsz_tensors * self.num_heads
+        bsz_tensors_kv = bsz_tensors * self.num_key_value_heads
+
+        query_states = self.q_norm(self.q_proj(hidden_states, bsz_tensors), bsz_tensors_q)
+        key_states = self.k_norm(self.k_proj(hidden_states, bsz_tensors), bsz_tensors_kv)
+        value_states = self.v_proj(hidden_states, bsz_tensors)
+
+
+        query_states = query_states.view(q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(q_len, self.num_key_value_heads, self.head_dim)
+        value_states = value_states.view(q_len, self.num_key_value_heads, self.head_dim)
+        
+        cos, sin = self.rotary_emb(value_states.unsqueeze(0), position_ids.unsqueeze(0))
+        query_states, key_states = self.apply_rotary_pos_emb(query_states.unsqueeze(0), key_states.unsqueeze(0), cos, sin, unsqueeze_dim=2)
+
+        query_states = query_states.view(q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(
+            q_len, self.num_key_value_heads, self.head_dim
+        )
+        value_states = value_states.view(
+            q_len, self.num_key_value_heads, self.head_dim
+        )
+
+        k_cache = kv_cache.get_k_cache(self.layer_idx)
+        v_cache = kv_cache.get_v_cache(self.layer_idx)
+
+
+        attn_output = wrapper.forward(query_states, k_cache, v_cache, key_states, value_states)
+  
+
+        attn_output = self.o_proj(attn_output.view(q_len, self.num_heads * self.head_dim), bsz_tensors)
+
+        return attn_output

ktransformers/operators/experts.py

@@ -26,7 +26,7 @@ sys.path.append(os.path.join(os.path.dirname(__file__), "..", "ktransformers_ext
 import cpuinfer_ext
 from cpuinfer_ext.moe import MOEConfig, MOE
 import ctypes
-from ktransformers.util.custom_gguf import GGUFLoader
+from ktransformers.util.custom_gguf import GGMLQuantizationType, GGUFLoader
 from ktransformers.util.utils import InferenceState
 from ktransformers.server.config.config import Config
 from transformers.activations import ACT2FN
@@ -141,6 +141,7 @@ class KExpertsCPU(KExpertsBase):
         assert device.lower() == "cpu", "KExpertsCPU can only be loaded on CPU"
         self.n_routed_experts = n_routed_experts
         self.out_device = out_device
+        self.backend = kwargs.get("backend", "llamafile")
 
     def load(self, w: dict | nn.Parameter | tuple | None = None, device:str|None = None, warmup:bool = False):
         if device:
@@ -163,27 +164,64 @@ class KExpertsCPU(KExpertsBase):
         )
         # print(self.gate_qtype, self.up_qtype, self.down_qtype)
         n_routed_experts = self.n_routed_experts
+        self.cpu_infer = KExpertsCPU.CPU_INFER
         # n_routed_experts = len(self.orig_module)
-        moe_config = MOEConfig(
-            n_routed_experts,
-            self.config.num_experts_per_tok,
-            self.config.hidden_size,
-            self.config.moe_intermediate_size,
-            64,
-            10,
-            1024,
-            gate_ptr,
-            up_ptr,
-            down_ptr,
-            self.gate_type,
-            self.up_type,
-            self.down_type,
-            30, # TODO: get from model.dtype
-        )
+        if self.backend == "llamafile":
+            moe_config = MOEConfig(
+                n_routed_experts,
+                self.config.num_experts_per_tok,
+                self.config.hidden_size,
+                self.config.moe_intermediate_size,
+                64,
+                10,
+                1024,
+                gate_ptr,
+                up_ptr,
+                down_ptr,
+                self.gate_type,
+                self.up_type,
+                self.down_type,
+                30, # TODO: get from model.dtype
+            )
+            self.moe = MOE(moe_config)
+        elif self.backend == "AMXBF16":
+            from cpuinfer_ext.moe import AMX_MOEConfig, AMXBF16_MOE
+            assert self.gate_type == GGMLQuantizationType.BF16
+            assert self.up_type == GGMLQuantizationType.BF16
+            assert self.down_type == GGMLQuantizationType.BF16
+            moe_config = AMX_MOEConfig(
+                n_routed_experts,
+                self.config.num_experts_per_tok,
+                self.config.hidden_size,
+                self.config.moe_intermediate_size,
+                25600,
+                gate_ptr,
+                up_ptr,
+                down_ptr,
+            )
+            self.moe = AMXBF16_MOE(moe_config)
+            self.cpu_infer.submit(self.moe.load_weights())
+            self.cpu_infer.sync()
+        elif self.backend == "AMXInt8":
+            from cpuinfer_ext.moe import AMX_MOEConfig, AMXInt8_MOE
+            assert self.gate_type == GGMLQuantizationType.BF16
+            assert self.up_type == GGMLQuantizationType.BF16
+            assert self.down_type == GGMLQuantizationType.BF16
+            moe_config = AMX_MOEConfig(
+                n_routed_experts,
+                self.config.num_experts_per_tok,
+                self.config.hidden_size,
+                self.config.moe_intermediate_size,
+                25600,
+                gate_ptr,
+                up_ptr,
+                down_ptr,
+            )
+            self.moe = AMXInt8_MOE(moe_config)
+            self.cpu_infer.submit(self.moe.load_weights())
+            self.cpu_infer.sync()
         # print(n_routed_experts, hidden_size, moe_intermediate_size)
         num_experts_per_tok = self.config.num_experts_per_tok
-        self.moe = MOE(moe_config)
-        self.cpu_infer = KExpertsCPU.CPU_INFER
         if warmup:
             self.cpu_infer.submit(self.moe.warm_up())
             self.cpu_infer.sync()
@@ -652,6 +690,7 @@ class KTransformersExperts(BaseInjectedModule, KExpertsBase):
 from ktransformers.models.modeling_deepseek import DeepseekV2MoE
 from ktransformers.models.modeling_deepseek_v3 import DeepseekV3MoE
 from ktransformers.models.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+from ktransformers.models.modeling_qwen3_moe import Qwen3MoeSparseMoeBlock
 from ktransformers.models.modeling_mixtral import MixtralSparseMoeBlock
 
 
@@ -1230,3 +1269,229 @@ class KTransformersExpertsV2(BaseInjectedModule, KExpertsBase):
             self.unload()
         else:
             raise ValueError("mode must be either InferenceState.GENERATE, InferenceState.PREFILL or InferenceState.UNLOAD")
+
+class KQwen2MoeSparseMoeBlockV2(BaseInjectedModule, Qwen2MoeSparseMoeBlock):
+    def forward(self, hidden_states, bsz_tensor, cuda_graph_idx=0):
+
+        orig_shape = hidden_states.shape
+        sequence_length = orig_shape[1]
+
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+
+        router_logits = self.gate(hidden_states, bsz_tensor)        
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        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)
+
+        # only for generate phase
+        if hasattr(self.experts.generate_experts, "submit_for_one_decode") and torch.cuda.is_current_stream_capturing(): # TODO: this branch cause jit bug
+            self.experts.generate_experts.submit_for_one_decode(hidden_states, selected_experts, routing_weights, bsz_tensor, cuda_graph_idx)
+            y_ = self.shared_expert(hidden_states, bsz_tensor).squeeze(0)
+            y_ = F.sigmoid(self.shared_expert_gate(hidden_states)) * y_    
+
+            y = self.experts.generate_experts.sync_for_one_decode(cuda_graph_idx).unsqueeze(0)
+            
+            y += y_
+            y.resize_(*orig_shape)
+            return y
+
+        y_ = self.shared_expert(hidden_states, bsz_tensor).squeeze(0)
+        y_ = (
+            F.sigmoid(self.shared_expert_gate(hidden_states)) * y_
+        )
+
+
+        if isinstance(self.experts, KExpertsBase):
+            y = self.moe_on_cpuinfer(hidden_states, selected_experts, routing_weights, bsz_tensor, cuda_graph_idx).view(*orig_shape).to(device=hidden_states.device)
+        elif hidden_states.size(0) > 10:
+            # TODO may bugs here
+            y = (
+                self.moe_infer(hidden_states, selected_experts, routing_weights)
+                .view(*orig_shape)
+                .to(device=hidden_states.device)
+            )
+        else:
+            # TODO may bugs here
+            y = (
+                self.moe_infer_simple(hidden_states, selected_experts, routing_weights)
+                .view(*orig_shape)
+                .to(device=hidden_states.device)
+            ) 
+        y += y_
+        return y
+
+    @torch.no_grad()
+    def moe_on_cpuinfer(self, x: torch.Tensor, topk_ids: torch.Tensor, topk_weight: torch.Tensor, bsz_tensor, cuda_graph_idx=0) -> torch.Tensor:
+        outs = torch.empty_like(x)
+        outs = self.experts(x, topk_ids, topk_weight, bsz_tensor, cuda_graph_idx)
+        return outs
+
+    @torch.no_grad()
+    # TODO may bugs here
+    def moe_infer_simple(
+        self, x: torch.Tensor, topk_ids: torch.Tensor, topk_weight: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        x: [num_tokens, hidden_size]
+        topk_ids, topk_weight: [num_tokens, num_selected_experts]
+        """
+        outs = torch.zeros_like(x)
+        for token_idx in range(topk_ids.size(0)):
+            for expert_idx in range(topk_ids.size(1)):
+                expert = self.experts[topk_ids[token_idx, expert_idx]]
+                outs[token_idx] += (
+                    expert.forward(x[token_idx]) * topk_weight[token_idx, expert_idx]
+                )
+        return outs
+
+    @torch.no_grad()
+    # TODO may bugs here
+    def moe_infer(self, x, topk_ids, topk_weight):
+        cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
+        cnts.scatter_(1, topk_ids, 1)
+        tokens_per_expert = cnts.sum(dim=0)
+        idxs = topk_ids.view(-1).argsort()
+        sorted_tokens = x[idxs // topk_ids.shape[1]]
+        tokens_per_expert = tokens_per_expert.cpu().numpy()
+
+        outputs = []
+        start_idx = 0
+        for i, num_tokens in enumerate(tokens_per_expert):
+            end_idx = start_idx + num_tokens
+            if num_tokens == 0:
+                continue
+            expert = self.experts[i + self.ep_rank * self.experts_per_rank]
+            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
+            expert_out = expert.forward(tokens_for_this_expert)
+            outputs.append(expert_out)
+            start_idx = end_idx
+
+        outs = torch.cat(outputs, dim=0) if len(outputs) else sorted_tokens.new_empty(0)
+
+        new_x = torch.empty_like(outs)
+        new_x[idxs] = outs
+        final_out = (
+            new_x.view(*topk_ids.shape, -1)
+            .type(topk_weight.dtype)
+            .mul_(topk_weight.unsqueeze(dim=-1))
+            .sum(dim=1)
+            .type(new_x.dtype)
+        )
+        return final_out
+
+class KQwen3MoeSparseMoeBlockV2(BaseInjectedModule, Qwen3MoeSparseMoeBlock):
+    def forward(self, hidden_states, bsz_tensor, cuda_graph_idx=0):
+
+        orig_shape = hidden_states.shape
+        sequence_length = orig_shape[1]
+
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+
+        router_logits = self.gate(hidden_states, bsz_tensor)        
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        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)
+
+        # only for generate phase
+        if hasattr(self.experts.generate_experts, "submit_for_one_decode") and torch.cuda.is_current_stream_capturing(): # TODO: this branch cause jit bug
+            self.experts.generate_experts.submit_for_one_decode(hidden_states, selected_experts, routing_weights, bsz_tensor, cuda_graph_idx)
+            # y_ = self.shared_expert(hidden_states, bsz_tensor).squeeze(0)
+            # y_ = F.sigmoid(self.shared_expert_gate(hidden_states)) * y_    
+
+            y = self.experts.generate_experts.sync_for_one_decode(cuda_graph_idx).unsqueeze(0)
+            
+            # y += y_
+            y.resize_(*orig_shape)
+            return y
+
+        # y_ = self.shared_expert(hidden_states, bsz_tensor).squeeze(0)
+        # y_ = (
+        #     F.sigmoid(self.shared_expert_gate(hidden_states)) * y_
+        # )
+
+
+        if isinstance(self.experts, KExpertsBase):
+            y = self.moe_on_cpuinfer(hidden_states, selected_experts, routing_weights, bsz_tensor, cuda_graph_idx).view(*orig_shape).to(device=hidden_states.device)
+        elif hidden_states.size(0) > 10:
+            # TODO may bugs here
+            y = (
+                self.moe_infer(hidden_states, selected_experts, routing_weights)
+                .view(*orig_shape)
+                .to(device=hidden_states.device)
+            )
+        else:
+            # TODO may bugs here
+            y = (
+                self.moe_infer_simple(hidden_states, selected_experts, routing_weights)
+                .view(*orig_shape)
+                .to(device=hidden_states.device)
+            ) 
+        # y += y_
+        return y
+
+    @torch.no_grad()
+    def moe_on_cpuinfer(self, x: torch.Tensor, topk_ids: torch.Tensor, topk_weight: torch.Tensor, bsz_tensor, cuda_graph_idx=0) -> torch.Tensor:
+        outs = torch.empty_like(x)
+        outs = self.experts(x, topk_ids, topk_weight, bsz_tensor, cuda_graph_idx)
+        return outs
+
+    @torch.no_grad()
+    # TODO may bugs here
+    def moe_infer_simple(
+        self, x: torch.Tensor, topk_ids: torch.Tensor, topk_weight: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        x: [num_tokens, hidden_size]
+        topk_ids, topk_weight: [num_tokens, num_selected_experts]
+        """
+        outs = torch.zeros_like(x)
+        for token_idx in range(topk_ids.size(0)):
+            for expert_idx in range(topk_ids.size(1)):
+                expert = self.experts[topk_ids[token_idx, expert_idx]]
+                outs[token_idx] += (
+                    expert.forward(x[token_idx]) * topk_weight[token_idx, expert_idx]
+                )
+        return outs
+
+    @torch.no_grad()
+    # TODO may bugs here
+    def moe_infer(self, x, topk_ids, topk_weight):
+        cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
+        cnts.scatter_(1, topk_ids, 1)
+        tokens_per_expert = cnts.sum(dim=0)
+        idxs = topk_ids.view(-1).argsort()
+        sorted_tokens = x[idxs // topk_ids.shape[1]]
+        tokens_per_expert = tokens_per_expert.cpu().numpy()
+
+        outputs = []
+        start_idx = 0
+        for i, num_tokens in enumerate(tokens_per_expert):
+            end_idx = start_idx + num_tokens
+            if num_tokens == 0:
+                continue
+            expert = self.experts[i + self.ep_rank * self.experts_per_rank]
+            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
+            expert_out = expert.forward(tokens_for_this_expert)
+            outputs.append(expert_out)
+            start_idx = end_idx
+
+        outs = torch.cat(outputs, dim=0) if len(outputs) else sorted_tokens.new_empty(0)
+
+        new_x = torch.empty_like(outs)
+        new_x[idxs] = outs
+        final_out = (
+            new_x.view(*topk_ids.shape, -1)
+            .type(topk_weight.dtype)
+            .mul_(topk_weight.unsqueeze(dim=-1))
+            .sum(dim=1)
+            .type(new_x.dtype)
+        )
+        return final_out
\ No newline at end of file

ktransformers/operators/flashinfer_batch_prefill_wrapper.py

+import torch
+import flashinfer
+import gc
+try:
+    from flash_attn import flash_attn_with_kvcache
+    print("found flash_attn")
+    
+except ImportError:
+    print("flash_attn not found, flashinfer unit test needed it. If you are using balance serve, ignore this.")
+
+from typing import Union, Optional
+
+def setup_seed(seed):
+	torch.manual_seed(seed)
+	torch.cuda.manual_seed_all(seed)
+
+setup_seed(998244353)
+
+torch.set_grad_enabled(False)
+torch.set_default_dtype(torch.bfloat16)
+global_dtype=torch.bfloat16
+global_device=torch.device("cuda",0)
+torch.cuda.set_device(0)
+torch.backends.cudnn.enabled =True
+torch.backends.cudnn.benchmark = True
+
+class flashInferAttn():
+	
+	float_workspace_buffer = None
+	def __init__(self,
+			max_batch_token,
+			max_batch_size,
+			max_pages,
+			device = "cuda:0",
+			kv_layout: str = "NHD",
+			use_cuda_graph: bool = False,
+			) -> None:
+		self.device = device
+		self.max_batch_token = max_batch_token
+		self.kv_layout = kv_layout
+		self.use_cuda_graph = use_cuda_graph
+		if flashInferAttn.float_workspace_buffer is None:
+			flashInferAttn.float_workspace_buffer = torch.empty(1024 * 1024 * 1024, dtype=torch.uint8, device=device)
+		self.qo_indptr_buf = torch.empty((max_batch_size+1,), dtype=torch.int32, device=device)
+		self.paged_kv_indptr_buf = torch.empty((max_batch_size+1,), dtype=torch.int32, device=device)
+		self.paged_kv_indices_buf = torch.empty((max_pages,), dtype=torch.int32, device=device)
+		self.paged_kv_last_page_len_buf = torch.empty((max_batch_size,), dtype=torch.int32, device=device)
+		self.batch_size_tensor_buf = torch.empty((1,), dtype=torch.int32, device=device)
+		self.num_tokens_tensor_buf = torch.empty((1,), dtype=torch.uint32, device=device)
+	
+		# TODO: custom mask
+		self.custom_mask_buf = None
+		self.qk_indptr_buf = None
+		self.warpper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
+			flashInferAttn.float_workspace_buffer,
+			self.kv_layout,
+			use_cuda_graph=self.use_cuda_graph,
+			qo_indptr_buf=self.qo_indptr_buf,
+			paged_kv_indptr_buf=self.paged_kv_indptr_buf,
+			paged_kv_indices_buf=self.paged_kv_indices_buf,
+			paged_kv_last_page_len_buf=self.paged_kv_last_page_len_buf,
+			backend = "fa2",
+		)
+
+	def plan(self,
+		qo_indptr: torch.Tensor,
+		paged_kv_indptr: torch.Tensor,
+		paged_kv_indices: torch.Tensor,
+		paged_kv_last_page_len: torch.Tensor,
+		batch_size_tensor: torch.Tensor,
+		num_tokens_tensor: torch.Tensor,
+		num_qo_heads: int,
+		num_kv_heads: int,
+		head_dim: int,
+		page_size: int,
+		causal: bool = True, 
+		pos_encoding_mode: str = "NONE",
+		q_data_type: Union[str, torch.dtype] = torch.bfloat16,
+		kv_data_type: Optional[Union[str, torch.dtype]] = None):
+		
+		self.batch_size_tensor_buf.copy_(batch_size_tensor, non_blocking=True)
+		self.num_tokens_tensor_buf.copy_(num_tokens_tensor, non_blocking=True)
+		self.page_size = page_size
+		self.warpper.plan(
+			qo_indptr,
+			paged_kv_indptr,
+			paged_kv_indices,
+			paged_kv_last_page_len,
+			num_qo_heads,
+			num_kv_heads,
+			head_dim,
+			page_size,
+			causal = causal,
+			pos_encoding_mode = pos_encoding_mode,
+			q_data_type = q_data_type,
+			kv_data_type = kv_data_type
+			)
+
+	def calc_batch_indices(self, ragged_size = None):
+		if self.use_cuda_graph:
+			self.batch_indices, self.positions = flashinfer.get_batch_indices_positions(
+				self.qo_indptr_buf, flashinfer.get_seq_lens(self.paged_kv_indptr_buf, self.paged_kv_last_page_len_buf, self.page_size), self.batch_size_tensor_buf, self.max_batch_token)
+		else:
+			self.batch_indices, self.positions = flashinfer.get_batch_indices_positions(
+				self.warpper._qo_indptr_buf, flashinfer.get_seq_lens(self.warpper._paged_kv_indptr_buf, self.warpper._paged_kv_last_page_len_buf, self.page_size), self.batch_size_tensor_buf, ragged_size)
+
+	def forward(self, q, k_cache, v_cache, k, v):
+		if self.use_cuda_graph:
+			flashinfer.page.append_paged_kv_cache(k, v, self.batch_indices, self.positions, (k_cache, v_cache), self.paged_kv_indices_buf, self.paged_kv_indptr_buf, self.paged_kv_last_page_len_buf, self.num_tokens_tensor_buf)
+			return self.warpper.run(q, (k_cache, v_cache))
+		else:
+			flashinfer.page.append_paged_kv_cache(k, v, self.batch_indices, self.positions, (k_cache, v_cache), self.warpper._paged_kv_indices_buf, self.warpper._paged_kv_indptr_buf, self.warpper._paged_kv_last_page_len_buf, self.num_tokens_tensor_buf)
+			return self.warpper.run(q, (k_cache, v_cache))
+
+
+def testCudaGraph():
+	
+	# use max batch to create buffer
+	batch_decode = 8
+	prefill_chunk = 48
+	past_kv_0 = 4090
+	past_kv_1 = 4096
+	raged_size = prefill_chunk + batch_decode
+	num_key_value_heads = 8
+	head_dim = 128
+	num_attention_heads = 64
+	page_size = 256
+	num_pages_per_seq = (past_kv_1 + page_size - 1) // page_size
+	total_num_pages = (num_pages_per_seq + 1) * (batch_decode + 1) + prefill_chunk // page_size
+	attn = flashInferAttn(raged_size, batch_decode+1, total_num_pages, use_cuda_graph=True)
+
+	batch_size_tensor = torch.tensor([batch_decode + 1], device=global_device, dtype=torch.int32)
+	
+	k_caches = []	
+	v_caches = []
+	ks = []
+	vs = []
+	qs = []
+	for layer_idx in range(3):
+		k_caches.append(torch.randn(total_num_pages, page_size, num_key_value_heads, head_dim, device=global_device, dtype=torch.bfloat16))
+		v_caches.append(torch.randn(total_num_pages, page_size, num_key_value_heads, head_dim, device=global_device, dtype=torch.bfloat16))
+		ks.append(torch.randn(raged_size, num_key_value_heads, head_dim, device=global_device, dtype=torch.bfloat16))
+		vs.append(torch.randn(raged_size, num_key_value_heads, head_dim, device=global_device, dtype=torch.bfloat16))
+		qs.append(torch.randn(raged_size, num_attention_heads, head_dim, device=global_device, dtype=torch.bfloat16))
+	
+	# warmup and capture small batch
+	past_kv_0 = 250
+	past_kv_1 = 256
+	num_pages_per_seq = (past_kv_1 + page_size - 1) // page_size
+	total_num_pages = (num_pages_per_seq + 1) * (batch_decode + 1) + prefill_chunk // page_size
+	q_indptr = torch.empty((batch_decode + 2,), dtype=torch.int32, device=global_device)
+	q_indptr[0] = 0
+	q_indptr[1:] = torch.arange(prefill_chunk, prefill_chunk + batch_decode + 1, device=global_device, dtype=torch.int32)
+	kv_indptr = torch.arange(0, batch_decode + 2, device=global_device, dtype=torch.int32) * num_pages_per_seq
+	kv_indices = torch.arange(0, total_num_pages, device=global_device, dtype=torch.int32)
+	kv_last_page_len = torch.empty((batch_decode + 1,), dtype=torch.int32, device=global_device)
+	kv_last_page_len[:1+batch_decode//2] = int((past_kv_0 - 1) % page_size + 1)
+	kv_last_page_len[1+batch_decode//2:] = int((past_kv_1 - 1) % page_size + 1)
+
+	print(q_indptr)
+	print(kv_indptr)
+	print(kv_indices)
+	print(kv_last_page_len)
+	attn.plan(q_indptr,
+			kv_indptr,
+			kv_indices,
+			kv_last_page_len,
+			batch_size_tensor,
+			num_attention_heads,
+			num_key_value_heads,
+			head_dim,
+			page_size,
+			causal = True,
+			pos_encoding_mode="NONE",
+			q_data_type=torch.bfloat16)
+
+	attn.calc_batch_indices(raged_size)
+	for layer_idx in range(3):
+		attn.forward(qs[layer_idx], k_caches[layer_idx], v_caches[layer_idx], ks[layer_idx], vs[layer_idx])
+		torch.cuda.synchronize()
+
+	outs = []
+	g = torch.cuda.CUDAGraph()
+	with torch.cuda.graph(g):
+		for layer_idx in range(3):
+			outs.append(attn.forward(qs[layer_idx], k_caches[layer_idx], v_caches[layer_idx], ks[layer_idx], vs[layer_idx]))
+	g.replay()
+	
+	kv_last_page_len[:1+batch_decode//2] = int(past_kv_0)
+	kv_last_page_len[1+batch_decode//2:] = int(past_kv_1)
+	for layer_idx in range(3):
+		for i in range(batch_decode + 1):
+			
+			qi = qs[layer_idx][q_indptr[i] : q_indptr[i + 1]]
+			o_ref_i = flash_attn_with_kvcache(
+				qi.unsqueeze(0),
+				k_caches[layer_idx],
+				v_caches[layer_idx],
+				causal=True,
+				block_table=kv_indices[kv_indptr[i]:kv_indptr[i+1]].unsqueeze(0),
+				cache_seqlens=torch.tensor([past_kv_0 if i < 1+batch_decode//2 else past_kv_1], device=global_device, dtype=torch.int32)
+			)
+			o_i = outs[layer_idx][q_indptr[i] : q_indptr[i + 1]]
+			print(layer_idx, i)
+			torch.testing.assert_close(o_i.unsqueeze(0), o_ref_i, rtol=5e-3, atol=5e-3)
+
+	# run another batch size use capture cuda graph
+	past_kv_0 = 4090
+	past_kv_1 = 4096
+	prefill_chunk = 24
+	batch_decode = 4
+	num_pages_per_seq = (past_kv_1 + page_size - 1) // page_size
+	total_num_pages = (num_pages_per_seq + 1) * (batch_decode + 1) + prefill_chunk // page_size
+	batch_size_tensor = torch.tensor([batch_decode + 1], device=global_device, dtype=torch.int32)
+	num_tokens_tensor = torch.tensor([batch_decode + prefill_chunk], device=global_device, dtype=torch.int32)
+
+	q_indptr = torch.empty((batch_decode + 2,), dtype=torch.int32, device=global_device)
+	q_indptr[0] = 0
+	q_indptr[1:] = torch.arange(prefill_chunk, prefill_chunk + batch_decode + 1, device=global_device, dtype=torch.int32)
+	kv_indptr = torch.arange(0, batch_decode + 2, device=global_device, dtype=torch.int32) * num_pages_per_seq
+	kv_indices = torch.arange(0, total_num_pages, device=global_device, dtype=torch.int32)
+	kv_last_page_len = torch.empty((batch_decode + 1,), dtype=torch.int32, device=global_device)
+	kv_last_page_len[:1+batch_decode//2] = int((past_kv_0 - 1) % page_size + 1)
+	kv_last_page_len[1+batch_decode//2:] = int((past_kv_1 - 1) % page_size + 1)
+	attn.plan(q_indptr,
+			kv_indptr,
+			kv_indices,
+			kv_last_page_len,
+			batch_size_tensor,
+			num_attention_heads,
+			num_key_value_heads,
+			head_dim,
+			page_size,
+			causal = True,
+			pos_encoding_mode="NONE",
+			q_data_type=torch.bfloat16)
+	attn.calc_batch_indices(raged_size)
+	g.replay()
+	
+	kv_last_page_len[:1+batch_decode//2] = int(past_kv_0)
+	kv_last_page_len[1+batch_decode//2:] = int(past_kv_1)
+	for layer_idx in range(3):
+		for i in range(batch_decode + 1):
+			
+			qi = qs[layer_idx][q_indptr[i] : q_indptr[i + 1]]
+			o_ref_i = flash_attn_with_kvcache(
+				qi.unsqueeze(0),
+				k_caches[layer_idx],
+				v_caches[layer_idx],
+				causal=True,
+				block_table=kv_indices[kv_indptr[i]:kv_indptr[i+1]].unsqueeze(0),
+				cache_seqlens=torch.tensor([past_kv_0 if i < 1+batch_decode//2 else past_kv_1], device=global_device, dtype=torch.int32)
+			)
+			o_i = outs[layer_idx][q_indptr[i] : q_indptr[i + 1]]
+			print(layer_idx, i)
+			torch.testing.assert_close(o_i.unsqueeze(0), o_ref_i, rtol=5e-3, atol=5e-3)
+			
+
+
+def testAttentionFlashInfer(	
+	):
+	batch_decode = 32
+	prefill_chunk = 64
+	past_kv_0 = 510
+	past_kv_1 = 512
+	raged_size = prefill_chunk + batch_decode
+	num_key_value_heads = 8
+	head_dim = 128
+	num_attention_heads = 64
+	cases = 1
+	page_size = 32
+	num_pages_per_seq = (past_kv_1 + page_size - 1) // page_size
+	total_num_pages = (num_pages_per_seq + 1) * (batch_decode + 1) + prefill_chunk // page_size
+	workspace_buffer = torch.empty(128 * 1024 * 1024, dtype=torch.uint8, device="cuda:0")
+	qs = []
+	kvs = []
+	q_indptrs = []
+	kv_indptrs = []
+	kv_indicess = []
+	kv_last_page_lens = []
+	wrappers = []
+	for case_id in range(cases):
+		kvs.append(torch.randn(total_num_pages, 2, page_size, num_key_value_heads, head_dim, device=global_device, dtype=torch.bfloat16))
+		qs.append(torch.randn(raged_size, num_attention_heads, head_dim, device=global_device, dtype=torch.bfloat16))
+		q_indptr = torch.empty((batch_decode + 2,), dtype=torch.int32, device=global_device)
+		q_indptr[0] = 0
+		q_indptr[1:] = torch.arange(prefill_chunk, prefill_chunk + batch_decode + 1, device=global_device, dtype=torch.int32)
+		q_indptrs.append(q_indptr)
+		kv_indptrs.append(torch.arange(0, batch_decode + 2, device=global_device, dtype=torch.int32) * num_pages_per_seq)
+		kv_indicess.append(torch.arange(0, total_num_pages, device=global_device, dtype=torch.int32))
+		kv_last_page_len = torch.empty((batch_decode + 1,), dtype=torch.int32, device=global_device)
+		kv_last_page_len[:1+batch_decode//2] = int((past_kv_0 - 1) % page_size + 1)
+		kv_last_page_len[1+batch_decode//2:] = int((past_kv_1 - 1) % page_size + 1)
+		kv_last_page_lens.append(kv_last_page_len)
+		wrappers.append(flashinfer.BatchPrefillWithPagedKVCacheWrapper(
+			workspace_buffer,
+			"NHD",
+			use_cuda_graph=True,
+			qo_indptr_buf=q_indptrs[case_id],
+			paged_kv_indptr_buf=kv_indptrs[case_id],
+			paged_kv_indices_buf=kv_indicess[case_id],
+			paged_kv_last_page_len_buf=kv_last_page_lens[case_id],
+		))
+		wrappers[case_id].plan(
+			q_indptrs[case_id],
+			kv_indptrs[case_id],
+			kv_indicess[case_id],
+			kv_last_page_lens[case_id],
+			num_attention_heads,
+			num_key_value_heads,
+			head_dim,
+			page_size,
+			causal = True,
+			pos_encoding_mode="ROPE_LLAMA",
+			q_data_type=torch.bfloat16
+		)
+					
+	def custom_forward(case_id):
+		out = wrappers[case_id].run(qs[case_id], kvs[case_id])
+	
+	custom_forward(0)
+
+# testCudaGraph()
+# pass
\ No newline at end of file

ktransformers/operators/gate.py

@@ -122,3 +122,72 @@ class KMoEGate(BaseInjectedModule, KMoEGateBase):
             self.e_score_correction_bias = None
 
 
+class KMoEGateQwen2Moe(BaseInjectedModule, KMoEGateBase):
+    def __init__(
+        self,
+        key: str,
+        gguf_loader: GGUFLoader,
+        config: PretrainedConfig,
+        orig_module: nn.Module = None,
+        generate_device: str = "cuda",
+        generate_op: str| None = "KLinearMarlin",
+        prefill_device: str = "cuda",
+        prefill_op: str| None = "KLinearMarlin",
+        use_quant: bool = False,
+        **kwargs,
+    ):
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs)
+        KMoEGateBase.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
+        self.generate_device = generate_device
+        self.prefill_device = prefill_device
+        self.generate_op = generate_op
+        self.prefill_op = prefill_op
+        self.is_windows = os.name == 'nt'
+        self.use_quant = use_quant
+        if not self.is_windows and use_quant:
+            self.gate_linear = nn.Linear(self.gating_dim, self.n_routed_experts, device=generate_device)
+            self.gate_linear = KTransformersLinear(key + ".ffn_gate_inp", 
+                                               gguf_loader, config, self.gate_linear, #orig_module
+                                               generate_device, generate_op, prefill_device, prefill_op)
+        else:
+            self.gate_linear = None
+
+    def forward(self, hidden_states) -> torch.Tensor:
+        if self.is_windows:
+            return self.orig_module.forward(hidden_states)
+        
+        bsz, seq_len, h = hidden_states.shape
+        ### compute gating score
+        hidden_states = hidden_states.view(-1, h)
+        if self.use_quant:
+            logits = self.gate_linear.forward(logits)
+        else:
+            logits = F.linear(
+                hidden_states.type(torch.float32), self.weight.type(torch.float32), None
+            )
+            
+        return grouped_topk(hidden_states, logits,
+                            self.top_k, self.norm_topk_prob,
+                            self.n_group, self.topk_group)
+
+    def load(self, w: dict | nn.Parameter | tuple | None = None, device: str|None = None):
+        if device is None: device = self.device
+        if w is None: w = self.load_weights(device=device)
+        
+        if isinstance(w, dict):
+            self.weight_type = w["weight_type"]
+            self.e_score_correction_bias_type = w["e_score_correction_bias_type"]
+            self.orig_module.weight = nn.Parameter(w["weight"])
+            self.orig_module.e_score_correction_bias = nn.Parameter(w["e_score_correction_bias"])
+        else:
+            raise ValueError("Invalid weight type")
+        self.orig_module.weight = nn.Parameter(self.orig_module.weight.to(device))
+        self.orig_module.e_score_correction_bias = nn.Parameter(self.orig_module.e_score_correction_bias.to(device))
+        if not self.is_windows and self.use_quant:
+            self.gate_linear.load(self.orig_module.weight)
+
+    def unload(self):
+        if self.weight is not None:
+            self.weight = None
+        if self.e_score_correction_bias is not None:
+            self.e_score_correction_bias = None
\ No newline at end of file

ktransformers/operators/layernorm.py

@@ -26,6 +26,8 @@ from transformers import PretrainedConfig
 import torch
 import torch.nn as nn
 from ktransformers.models.modeling_deepseek_v3 import DeepseekV3RMSNorm
+from ktransformers.models.modeling_qwen2_moe import Qwen2MoeRMSNorm
+from ktransformers.models.modeling_qwen3_moe import Qwen3MoeRMSNorm
 from ktransformers.operators.base_operator import BaseInjectedModule
 from ktransformers.util.custom_gguf import GGUFLoader
 from flashinfer.norm import (
@@ -75,4 +77,89 @@ class RMSNorm(DeepseekV3RMSNorm, BaseInjectedModule):
         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)
+        return self.weight * hidden_states.to(input_dtype)
\ No newline at end of file
+
+
+class KQwen2MoeRMSNorm(Qwen2MoeRMSNorm, BaseInjectedModule):
+    def __init__(self,
+                 key: str,
+                 gguf_loader : GGUFLoader,
+                 config: PretrainedConfig,
+                 orig_module: nn.Module,
+                 prefill_device: str = "cuda",
+                 generate_device: str = "cuda",
+                 **kwargs):
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
+        self.orig_module.__init__(config.hidden_size,
+            orig_module.variance_epsilon)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        batch_size_tensor: torch.Tensor = None,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        #return self.forward_native(x, residual)
+        if batch_size_tensor is None:
+            return self.forward_native(x)
+        if residual is not None:
+            fused_add_rmsnorm(x, residual, self.weight.data, batch_size_tensor, self.variance_epsilon)
+            #residual = x + residual
+            #out = rmsnorm(residual, self.weight.data, batch_size_tensor, self.variance_epsilon)
+            return x, residual
+        # print(x.shape, self.weight.data.shape, self.variance_epsilon, x.dtype, self.weight.data.dtype, x.device, self.weight.device, x.is_contiguous(), self.weight.data.is_contiguous())
+        out = rmsnorm(x, self.weight.data, batch_size_tensor,self.variance_epsilon)
+        return out
+
+    def forward_native(
+        self, hidden_states    
+    ):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class KQwen3MoeRMSNorm(Qwen3MoeRMSNorm, BaseInjectedModule):
+    def __init__(self,
+                 key: str,
+                 gguf_loader : GGUFLoader,
+                 config: PretrainedConfig,
+                 orig_module: nn.Module,
+                 prefill_device: str = "cuda",
+                 generate_device: str = "cuda",
+                 **kwargs):
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
+        self.orig_module.__init__(orig_module.hidden_size,
+            orig_module.variance_epsilon)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        batch_size_tensor: torch.Tensor = None,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        #return self.forward_native(x, residual)
+        bsz, hidden_size = x.shape
+        x = x.view(-1, self.orig_module.hidden_size)
+        if batch_size_tensor is None:
+            return self.forward_native(x)
+        if residual is not None:
+            fused_add_rmsnorm(x, residual, self.weight.data, batch_size_tensor, self.variance_epsilon)
+            #residual = x + residual
+            #out = rmsnorm(residual, self.weight.data, batch_size_tensor, self.variance_epsilon)
+            return x, residual
+        # print(x.shape, self.weight.data.shape, self.variance_epsilon, x.dtype, self.weight.data.dtype, x.device, self.weight.device, x.is_contiguous(), self.weight.data.is_contiguous())
+        out = rmsnorm(x, self.weight.data, batch_size_tensor,self.variance_epsilon)
+        out = out.view(bsz, hidden_size)
+        return out
+
+    def forward_native(
+        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)

ktransformers/operators/mlp.py

@@ -4,8 +4,7 @@ from ktransformers.util.custom_gguf import GGUFLoader
 from transformers import PretrainedConfig
 import torch.nn as nn
 from ktransformers.models.modeling_deepseek_v3 import DeepseekV3MLP
-
-
+from ktransformers.models.modeling_qwen2_moe import Qwen2MoeMLP
 class kDeepseekV3MLP(DeepseekV3MLP, BaseInjectedModule):
     def __init__(self,
                  key: str,
@@ -18,6 +17,21 @@ class kDeepseekV3MLP(DeepseekV3MLP, BaseInjectedModule):
         BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
         self.orig_module.__init__(orig_module.config,
             orig_module.hidden_size, orig_module.intermediate_size)
+    def forward(self, x, bsz_tensor):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x, bsz_tensor)) * self.up_proj(x, bsz_tensor), bsz_tensor)
+        return down_proj
+class KQwen2MoeMLP(Qwen2MoeMLP, BaseInjectedModule):
+    def __init__(self,
+                 key: str,
+                 gguf_loader : GGUFLoader,
+                 config: PretrainedConfig,
+                 orig_module: nn.Module,
+                 prefill_device: str = "cuda",
+                 generate_device: str = "cuda",
+                 **kwargs):
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
+        self.orig_module.__init__(orig_module.config,
+            orig_module.intermediate_size)
     def forward(self, x, bsz_tensor):
         down_proj = self.down_proj(self.act_fn(self.gate_proj(x, bsz_tensor)) * self.up_proj(x, bsz_tensor), bsz_tensor)
         return down_proj
\ No newline at end of file

ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-amx.yaml

+- match:
+    class: ktransformers.models.modeling_deepseek_v3.DeepseekV3RotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.YarnRotaryEmbeddingV3
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^lm_head$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+- match:
+    name: "^model\\.layers\\.(?!.*self_attn\\.kv_b_proj).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\..*\\.mlp$"
+    class: ktransformers.models.modeling_deepseek_v3.DeepseekV3MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV3MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- match:
+    class: ktransformers.models.modeling_deepseek_v3.MoEGate
+  replace:
+    class: ktransformers.operators.gate.KMoEGate
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\..*\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op: "KExpertsCPU"
+      out_device: "cuda"
+      backend: "AMXInt8" # or "AMXBF16" or "llamafile" (default)
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\..*\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      absorb_for_prefill: False # change this to True to enable long context(prefill may slower).
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KDeepseekV2Model"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+- match:
+    name: "^model.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cpu"
+      prefill_device: "cpu"
\ No newline at end of file

ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-serve.yaml

@@ -56,7 +56,7 @@
 - match:
     name: "^model\\.layers\\..*\\.self_attn$"
   replace:
-    class: ktransformers.operators.attention.flashinfer_attn # optimized MLA implementation
+    class: ktransformers.operators.balance_serve_attention.flashinfer_attn # optimized MLA implementation
     kwargs:
       generate_device: "cuda"
       prefill_device: "cuda"

ktransformers/optimize/optimize_rules/Moonlight-16B-A3B-serve.yaml

@@ -50,7 +50,7 @@
 - match:
     name: "^model\\.layers\\..*\\.self_attn$"
   replace:
-    class: ktransformers.operators.attention.flashinfer_attn # optimized MLA implementation
+    class: ktransformers.operators.balance_serve_attention.flashinfer_attn # optimized MLA implementation
     kwargs:
       generate_device: "cuda"
       prefill_device: "cuda"

ktransformers/optimize/optimize_rules/Qwen2-serve-amx.yaml

+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.RotaryEmbedding
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^lm_head$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+# - match:
+#     name: "^model\\.layers\\..*$"  # regular expression 
+#     class: torch.nn.Linear  # only match modules matching name and class simultaneously
+#   replace:
+#     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+#     kwargs:
+#       generate_device: "cuda"
+#       prefill_device: "cuda"
+#       generate_op: "VLinearMarlin"
+#       prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\.(?!.*mlp\\.shared_expert_gate).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "VLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\..*\\.mlp$"
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeSparseMoeBlock
+  replace:
+    class: ktransformers.operators.experts.KQwen2MoeSparseMoeBlockV2     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^model\\.layers\\..*\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExpertsV2     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op: "KExpertsCPU"
+      out_device: "cuda"
+      backend: "AMXInt8" # or "AMXBF16" or "llamafile" (default)
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\..*\\.self_attn$"
+  replace:
+    class: ktransformers.operators.balance_serve_attention.KQwen2MoeAttention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KQwen2MoeModel"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+- match:
+    name: "^model.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cpu"
+      prefill_device: "cpu"
+
+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRMSNorm
+  replace:
+    class: ktransformers.operators.layernorm.KQwen2MoeRMSNorm
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeMLP
+  replace:
+    class:  ktransformers.operators.mlp.KQwen2MoeMLP
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
\ No newline at end of file

ktransformers/optimize/optimize_rules/Qwen2-serve.yaml

+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.RotaryEmbedding
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^lm_head$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+# - match:
+#     name: "^model\\.layers\\..*$"  # regular expression 
+#     class: torch.nn.Linear  # only match modules matching name and class simultaneously
+#   replace:
+#     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+#     kwargs:
+#       generate_device: "cuda"
+#       prefill_device: "cuda"
+#       generate_op: "VLinearMarlin"
+#       prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\.(?!.*mlp\\.shared_expert_gate).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "VLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\..*\\.mlp$"
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeSparseMoeBlock
+  replace:
+    class: ktransformers.operators.experts.KQwen2MoeSparseMoeBlockV2     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^model\\.layers\\..*\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExpertsV2     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op: "KExpertsCPU"
+      out_device: "cuda"
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\..*\\.self_attn$"
+  replace:
+    class: ktransformers.operators.balance_serve_attention.KQwen2MoeAttention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KQwen2MoeModel"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+- match:
+    name: "^model.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cpu"
+      prefill_device: "cpu"
+
+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRMSNorm
+  replace:
+    class: ktransformers.operators.layernorm.KQwen2MoeRMSNorm
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeMLP
+  replace:
+    class:  ktransformers.operators.mlp.KQwen2MoeMLP
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
\ No newline at end of file

ktransformers/optimize/optimize_rules/Qwen3Moe-serve-amx.yaml

+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.RotaryEmbedding
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^lm_head$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "VLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+# - match:
+#     name: "^model\\.layers\\..*$"  # regular expression 
+#     class: torch.nn.Linear  # only match modules matching name and class simultaneously
+#   replace:
+#     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+#     kwargs:
+#       generate_device: "cuda"
+#       prefill_device: "cuda"
+#       generate_op: "VLinearMarlin"
+#       prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\.(?!.*mlp\\.shared_expert_gate).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\..*\\.mlp$"
+    class: ktransformers.models.modeling_qwen3_moe.Qwen3MoeSparseMoeBlock
+  replace:
+    class: ktransformers.operators.experts.KQwen3MoeSparseMoeBlockV2     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^model\\.layers\\..*\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExpertsV2     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op: "KExpertsCPU"
+      out_device: "cuda"
+      backend: "AMXBF16" # or "AMXBF16" or "llamafile" (default)
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\..*\\.self_attn$"
+  replace:
+    class: ktransformers.operators.balance_serve_attention.KQwen3MoeAttention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KQwen2MoeModel"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+- match:
+    name: "^model.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cpu"
+      prefill_device: "cpu"
+
+- match:
+    class: ktransformers.models.modeling_qwen3_moe.Qwen3MoeRMSNorm
+  replace:
+    class: ktransformers.operators.layernorm.KQwen3MoeRMSNorm
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    class: ktransformers.models.modeling_qwen3_moe.Qwen3MoeMLP
+  replace:
+    class:  ktransformers.operators.mlp.KQwen2MoeMLP
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
\ No newline at end of file