support deepseekv3; runable but have precition problem

ktransformers/local_chat.py

@@ -15,6 +15,7 @@ from ktransformers.server.args import ArgumentParser
 
 
 from ktransformers.models.modeling_deepseek import DeepseekV2ForCausalLM
+from ktransformers.models.modeling_deepseekv3 import DeepseekV3ForCausalLM
 from ktransformers.models.modeling_qwen2_moe import Qwen2MoeForCausalLM
 from ktransformers.models.modeling_llama import LlamaForCausalLM
 from ktransformers.models.modeling_mixtral import MixtralForCausalLM
@@ -22,6 +23,7 @@ from ktransformers.server.config.config import Config
 
 custom_models = {
     "DeepseekV2ForCausalLM": DeepseekV2ForCausalLM,
+    "DeepseekV3ForCausalLM": DeepseekV3ForCausalLM,
     "Qwen2MoeForCausalLM": Qwen2MoeForCausalLM,
     "LlamaForCausalLM": LlamaForCausalLM,
     "MixtralForCausalLM": MixtralForCausalLM,
@@ -30,6 +32,8 @@ custom_models = {
 ktransformer_rules_dir = os.path.dirname(os.path.abspath(__file__)) + "/optimize/optimize_rules/"
 default_optimize_rules = {
     "DeepseekV2ForCausalLM": ktransformer_rules_dir + "DeepSeek-V2-Chat.yaml",
+    # "DeepseekV3ForCausalLM": ktransformer_rules_dir + "DeepSeek-V2-Chat.yaml",
+    "DeepseekV3ForCausalLM": ktransformer_rules_dir + "DeepSeek-V3-Chat-multi-gpu.yaml",
     "Qwen2MoeForCausalLM": ktransformer_rules_dir + "Qwen2-57B-A14B-Instruct.yaml",
     "LlamaForCausalLM": ktransformer_rules_dir + "Internlm2_5-7b-Chat-1m.yaml",
     "MixtralForCausalLM": ktransformer_rules_dir + "Mixtral.yaml",
@@ -74,8 +78,8 @@ def local_chat():
                 else:
                     content += line + "\n"
         if content == "":
-            if config.prompt_file == None or config.prompt_file == "":
-                content = "Please write a piece of quicksort code in C++."
+            if True: # config.prompt_file == None or config.prompt_file == "":
+                content = "hi"
             else:
                 content = open(config.prompt_file, "r").read()
         elif os.path.isfile(content):

ktransformers/models/configuration_deepseekv3.py

+# coding=utf-8
+# Copyright 2025 bzantium and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on the DeepSeekV3 implementations from the DeepSeek AI team. (https://huggingface.co/deepseek-ai/DeepSeek-V3)
+
+# 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.
+""" DeepSeekV3 model configuration """
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+class DeepseekV3Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DeepSeek-V3.
+    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 129280):
+            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`DeepseekV3Model`]
+        hidden_size (`int`, *optional*, defaults to 7168):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 18432):
+            Dimension of the MLP representations.
+        moe_intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimension of the MoE representations.
+        num_hidden_layers (`int`, *optional*, defaults to 61):
+            Number of hidden layers in the Transformer decoder.
+        num_nextn_predict_layers (`int`, *optional*, defaults to 1):
+            Number of nextn predict layers in the DeepSeekV3 Model.
+        num_attention_heads (`int`, *optional*, defaults to 128):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*, defaults to 128):
+            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
+            `num_attention_heads`.
+        n_shared_experts (`int`, *optional*, defaults to 1):
+            Number of shared experts, None means dense model.
+        n_routed_experts (`int`, *optional*, defaults to 256):
+            Number of routed experts, None means dense model.
+        ep_size (`<fill_type>`, *optional*, defaults to 1): <fill_docstring>
+        routed_scaling_factor (`float`, *optional*, defaults to 2.5):
+            Scaling factor or routed experts.
+        kv_lora_rank (`<fill_type>`, *optional*, defaults to 512): <fill_docstring>
+        q_lora_rank (`<fill_type>`, *optional*, defaults to 1536): <fill_docstring>
+        qk_rope_head_dim (`<fill_type>`, *optional*, defaults to 64): <fill_docstring>
+        v_head_dim (`<fill_type>`, *optional*, defaults to 128): <fill_docstring>
+        qk_nope_head_dim (`<fill_type>`, *optional*, defaults to 128): <fill_docstring>
+        topk_method (`str`, *optional*, defaults to `"noaux_tc"`):
+            Topk method used in routed gate.
+        n_group (`int`, *optional*, defaults to 8):
+            Number of groups for routed experts.
+        topk_group (`int`, *optional*, defaults to 4):
+            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
+        num_experts_per_tok (`int`, *optional*, defaults to 8):
+            Number of selected experts, None means dense model.
+        moe_layer_freq (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
+        first_k_dense_replace (`int`, *optional*, defaults to 3):
+            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
+                                                            \--k dense layers--/
+        norm_topk_prob (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the weights of the routed experts.
+        scoring_func (`str`, *optional*, defaults to `"sigmoid"`):
+            Method of computing expert weights.
+        aux_loss_alpha (`float`, *optional*, defaults to 0.001):
+            Auxiliary loss weight coefficient.
+            Whether to compute the auxiliary loss for each individual sample.
+        seq_aux (`<fill_type>`, *optional*, defaults to `True`): <fill_docstring>
+        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 4096):
+            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`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 0):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 1):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
+            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
+            issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        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. Currently supports two scaling
+            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
+            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
+            `max_position_embeddings` to the expected new maximum.
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    ```python
+    >>> from transformers import DeepseekV3Model, DeepseekV3Config
+    >>> # Initializing a Deepseek-V3 style configuration
+    >>> configuration = DeepseekV3Config()
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "deepseek_v3"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=129280,
+        hidden_size=7168,
+        intermediate_size=18432,
+        moe_intermediate_size = 2048,
+        num_hidden_layers=61,
+        num_nextn_predict_layers=1,
+        num_attention_heads=128,
+        num_key_value_heads=128,
+        n_shared_experts = 1,
+        n_routed_experts = 256,
+        ep_size = 1,
+        routed_scaling_factor = 2.5,
+        kv_lora_rank = 512,
+        q_lora_rank = 1536,
+        qk_rope_head_dim = 64,
+        v_head_dim = 128,
+        qk_nope_head_dim = 128,
+        topk_method = 'noaux_tc',
+        n_group = 8,
+        topk_group = 4,
+        num_experts_per_tok = 8,
+        moe_layer_freq = 1,
+        first_k_dense_replace = 3,
+        norm_topk_prob = True,
+        scoring_func = 'sigmoid',
+        aux_loss_alpha = 0.001,
+        seq_aux = True,
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=0,
+        eos_token_id=1,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.ep_size = ep_size
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.topk_method = topk_method
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_layer_freq = moe_layer_freq
+        self.first_k_dense_replace = first_k_dense_replace
+        self.norm_topk_prob = norm_topk_prob
+        self.scoring_func = scoring_func
+        self.aux_loss_alpha = aux_loss_alpha
+        self.seq_aux = seq_aux
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.mlp_bias = mlp_bias
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["DeepseekV3Config"]
\ No newline at end of file

ktransformers/models/custom_cache.py

@@ -34,9 +34,12 @@ class StaticCache(transformers.StaticCache):
         self.max_batch_size = max_batch_size
         self.max_cache_len = config.max_position_embeddings if max_cache_len is None else max_cache_len
         # Some model define a custom `head_dim` != config.hidden_size // config.num_attention_heads
-        self.head_dim = (
-            config.head_dim if hasattr(config, "head_dim") else config.hidden_size // config.num_attention_heads
-        )
+        if config.architectures[0] == "DeepseekV3ForCausalLM":
+            self.head_dim = config.qk_rope_head_dim
+        else:
+            self.head_dim = (
+                config.head_dim if hasattr(config, "head_dim") else config.hidden_size // config.num_attention_heads
+            )
 
         self.dtype = dtype if dtype is not None else torch.float32
         self.num_key_value_heads = (
@@ -46,7 +49,7 @@ class StaticCache(transformers.StaticCache):
         self.key_cache: List[torch.Tensor] = []
         self.value_cache: List[torch.Tensor] = []
         cache_shape = (max_batch_size, self.num_key_value_heads, self.max_cache_len, self.head_dim)
-        if config.architectures[0] == "DeepseekV2ForCausalLM":
+        if config.architectures[0] == "DeepseekV2ForCausalLM" or config.architectures[0] == "DeepseekV3ForCausalLM":
             # TODO: for deepseek, cache_shape is different whether using Absorbed MLA, check it automatically
             # key_shape = (max_batch_size, self.num_key_value_heads, self.max_cache_len, config.qk_rope_head_dim + config.qk_nope_head_dim)
             # value_shape = (max_batch_size, self.num_key_value_heads, self.max_cache_len, config.v_head_dim)

ktransformers/models/modeling_deepseekv3.py

+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/deepseekv3/modular_deepseekv3.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_deepseekv3.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+import math
+from typing import Callable, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, 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, SequenceClassifierOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from transformers.modeling_utils import PreTrainedModel # ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+# from transformers.processing_utils import Unpack
+from transformers.utils import (
+    # LossKwargs,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+from .configuration_deepseekv3 import DeepseekV3Config
+
+
+logger = logging.get_logger(__name__)
+_CONFIG_FOR_DOC = "DeepseekV3Config"
+
+
+class DeepseekV3RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        DeepseekV3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class DeepseekV3RotaryEmbedding(nn.Module):
+    def __init__(self, config: DeepseekV3Config, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    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)
+
+
+class DeepseekV3MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.moe_intermediate_size
+        # TODO rm hard coding
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)# config.mlp_bias)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)# config.mlp_bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)# config.mlp_bias)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class MoEGate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.scoring_func = config.scoring_func
+        self.seq_aux = config.seq_aux
+        self.topk_method = config.topk_method
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+
+        # topk selection algorithm
+        self.norm_topk_prob = config.norm_topk_prob
+        self.gating_dim = config.hidden_size
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim)))
+        if self.topk_method == "noaux_tc":
+            self.e_score_correction_bias = nn.Parameter(torch.empty((self.n_routed_experts)))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        import torch.nn.init as init
+
+        init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+
+    def forward(self, hidden_states):
+        bsz, seq_len, h = hidden_states.shape
+        ### compute gating score
+        hidden_states = hidden_states.view(-1, h)
+        logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32), None)
+        if self.scoring_func == "sigmoid":
+            scores = logits.sigmoid()
+        else:
+            raise NotImplementedError(f"insupportable scoring function for MoE gating: {self.scoring_func}")
+
+        ### select top-k experts
+        if self.topk_method == "noaux_tc":
+            # assert not self.training
+            scores_for_choice = scores.view(bsz * seq_len, -1) + self.e_score_correction_bias.unsqueeze(0)
+            group_scores = (
+                scores_for_choice.view(bsz * seq_len, self.n_group, -1).topk(2, dim=-1)[0].sum(dim=-1)
+            )  # [n, n_group]
+            group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]  # [n, top_k_group]
+            group_mask = torch.zeros_like(group_scores)  # [n, n_group]
+            group_mask.scatter_(1, group_idx, 1)  # [n, n_group]
+            score_mask = (
+                group_mask.unsqueeze(-1)
+                .expand(bsz * seq_len, self.n_group, self.n_routed_experts // self.n_group)
+                .reshape(bsz * seq_len, -1)
+            )  # [n, e]
+            tmp_scores = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)  # [n, e]
+            _, topk_idx = torch.topk(tmp_scores, k=self.top_k, dim=-1, sorted=False)
+            topk_weight = scores.gather(1, topk_idx)
+        else:
+            raise NotImplementedError(f"insupportable TopK function for MoE gating: {self.topk_method}")
+
+        ### norm gate to sum 1
+        if self.top_k > 1 and self.norm_topk_prob:
+            denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weight = topk_weight / denominator
+        topk_weight = topk_weight * self.routed_scaling_factor  # must multiply the scaling factor
+
+        return topk_idx, topk_weight
+
+
+class DeepseekV3MoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.num_experts_per_tok = config.num_experts_per_tok
+
+        if hasattr(config, "ep_size") and config.ep_size > 1:
+            assert config.ep_size == dist.get_world_size()
+            self.ep_size = config.ep_size
+            self.experts_per_rank = config.n_routed_experts // config.ep_size
+            self.ep_rank = dist.get_rank()
+            self.experts = nn.ModuleList(
+                [
+                    (
+                        DeepseekV3MLP(config, intermediate_size=config.moe_intermediate_size)
+                        if i >= self.ep_rank * self.experts_per_rank and i < (self.ep_rank + 1) * self.experts_per_rank
+                        else None
+                    )
+                    for i in range(config.n_routed_experts)
+                ]
+            )
+        else:
+            self.ep_size = 1
+            self.experts_per_rank = config.n_routed_experts
+            self.ep_rank = 0
+            self.experts = nn.ModuleList(
+                [
+                    DeepseekV3MLP(config)
+                    for i in range(config.n_routed_experts)
+                ]
+            )
+        self.gate = MoEGate(config)
+        if config.n_shared_experts is not None:
+            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
+            self.shared_experts = DeepseekV3MLP(config=config)
+
+    def forward(self, hidden_states):
+        identity = hidden_states
+        orig_shape = hidden_states.shape
+        topk_idx, topk_weight = self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        if not self.training:
+            y = self.moe_infer(hidden_states, topk_idx, topk_weight).view(*orig_shape)
+        if self.config.n_shared_experts is not None:
+            y = y + self.shared_experts(identity)
+        return y
+
+    @torch.no_grad()
+    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]]
+        sorted_tokens_shape = sorted_tokens.shape
+        if self.ep_size > 1:
+            tokens_per_ep_rank = tokens_per_expert.view(self.ep_size, -1).sum(dim=1)
+            tokens_per_expert_group = tokens_per_expert.new_empty(tokens_per_expert.shape[0])
+            dist.all_to_all_single(tokens_per_expert_group, tokens_per_expert)
+            output_splits = tokens_per_expert_group.view(self.ep_size, -1).sum(1).cpu().numpy().tolist()
+            gathered_tokens = sorted_tokens.new_empty(
+                tokens_per_expert_group.sum(dim=0).cpu().item(), sorted_tokens.shape[1]
+            )
+            input_split_sizes = tokens_per_ep_rank.cpu().numpy().tolist()
+            dist.all_to_all(
+                list(gathered_tokens.split(output_splits)),
+                list(sorted_tokens.split(input_split_sizes)),
+            )
+            tokens_per_expert_post_gather = tokens_per_expert_group.view(self.ep_size, self.experts_per_rank).sum(
+                dim=0
+            )
+            gatherd_idxs = np.zeros(shape=(gathered_tokens.shape[0],), dtype=np.int32)
+            s = 0
+            for i, k in enumerate(tokens_per_expert_group.cpu().numpy()):
+                gatherd_idxs[s : s + k] = i % self.experts_per_rank
+                s += k
+            gatherd_idxs = gatherd_idxs.argsort()
+            sorted_tokens = gathered_tokens[gatherd_idxs]
+            tokens_per_expert = tokens_per_expert_post_gather
+        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(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)
+        if self.ep_size > 1:
+            new_x = torch.empty_like(outs)
+            new_x[gatherd_idxs] = outs
+            gathered_tokens = new_x.new_empty(*sorted_tokens_shape)
+            dist.all_to_all(
+                list(gathered_tokens.split(input_split_sizes)),
+                list(new_x.split(output_splits)),
+            )
+            outs = gathered_tokens
+
+        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
+
+
+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
+
+
+# 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)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    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
+
+
+class DeepseekV3Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: DeepseekV3Config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.kv_lora_rank = config.kv_lora_rank
+        self.v_head_dim = config.v_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.q_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
+
+        self.is_causal = True
+
+        if self.q_lora_rank is None:
+            self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.q_head_dim, bias=False)
+        else:
+            self.q_a_proj = nn.Linear(self.hidden_size, config.q_lora_rank, bias=config.attention_bias)
+            self.q_a_layernorm = DeepseekV3RMSNorm(config.q_lora_rank)
+            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.q_head_dim, bias=False)
+
+        self.kv_a_proj_with_mqa = nn.Linear(
+            self.hidden_size,
+            config.kv_lora_rank + config.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kv_a_layernorm = DeepseekV3RMSNorm(config.kv_lora_rank)
+        self.kv_b_proj = nn.Linear(
+            config.kv_lora_rank,
+            self.num_heads * (self.q_head_dim - self.qk_rope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            self.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        self.rotary_emb = DeepseekV3RotaryEmbedding(
+            config=self.config,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+        position_ids: Optional[torch.LongTensor] = None,
+        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]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        if self.q_lora_rank is None:
+            q = self.q_proj(hidden_states)
+        else:
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q = q.view(bsz, q_len, self.num_heads, self.q_head_dim).transpose(1, 2)
+        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)
+        compressed_kv, k_pe = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        k_pe = k_pe.view(bsz, q_len, 1, self.qk_rope_head_dim).transpose(1, 2)
+        kv = (
+            self.kv_b_proj(self.kv_a_layernorm(compressed_kv))
+            .view(bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
+            .transpose(1, 2)
+        )
+
+        k_nope, value_states = torch.split(kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+        kv_seq_len = value_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+
+        q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, position_ids)
+
+        query_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
+        query_states[:, :, :, : self.qk_nope_head_dim] = q_nope
+        query_states[:, :, :, self.qk_nope_head_dim :] = q_pe
+
+        key_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
+        key_states[:, :, :, : self.qk_nope_head_dim] = k_nope
+        key_states[:, :, :, self.qk_nope_head_dim :] = k_pe
+
+        if self.q_head_dim != self.v_head_dim:
+            value_states = F.pad(value_states, [0, self.q_head_dim - self.v_head_dim])
+
+        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:
+                pass
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.v_head_dim)
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class DeepseekV3DecoderLayer(nn.Module):
+    def __init__(self, config: DeepseekV3Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = DeepseekV3Attention(config=config, layer_idx=layer_idx)
+
+        self.mlp = (
+            DeepseekV3MoE(config)
+            if (
+                config.n_routed_experts is not None
+                and layer_idx >= config.first_k_dense_replace
+                and layer_idx % config.moe_layer_freq == 0
+            )
+            else DeepseekV3MLP(config)
+        )
+        self.input_layernorm = DeepseekV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = DeepseekV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        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]]]:
+        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,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+DEEPSEEKV3_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 ([`DeepseekV3Config`]):
+            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 DeepseekV3 Model outputting raw hidden-states without any specific head on top.",
+    DEEPSEEKV3_START_DOCSTRING,
+)
+class DeepseekV3PreTrainedModel(PreTrainedModel):
+    config_class = DeepseekV3Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["DeepseekV3DecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+    _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_()
+
+
+DEEPSEEKV3_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` or `tuple(tuple(torch.FloatTensor))`, *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`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance, see our
+            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            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 DeepseekV3 Model outputting raw hidden-states without any specific head on top.",
+    DEEPSEEKV3_START_DOCSTRING,
+)
+class DeepseekV3Model(DeepseekV3PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`DeepseekV3DecoderLayer`]
+
+    Args:
+        config: DeepseekV3Config
+    """
+
+    def __init__(self, config: DeepseekV3Config):
+        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(
+            [DeepseekV3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = DeepseekV3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = DeepseekV3RotaryEmbedding(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(DEEPSEEKV3_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[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs# : Unpack[FlashAttentionKwargs],
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        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
+
+        for decoder_layer in self.layers[: self.config.num_hidden_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,
+                    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,
+                    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],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        output = BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+        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,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and (attention_mask == 0.0).any():
+                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)
+
+        # 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 and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_length()
+        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],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            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
+            min_dtype = torch.finfo(dtype).min
+            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,
+        **kwargs,
+    ):
+        """
+        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 plcae 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.
+        """
+        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
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            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
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+
+        return causal_mask
+
+
+# class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
+class DeepseekV3ForCausalLM(DeepseekV3PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = DeepseekV3Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    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(DEEPSEEKV3_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[Union[Cache, 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,
+        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"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            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, DeepseekV3ForCausalLM
+
+        >>> model = DeepseekV3ForCausalLM.from_pretrained("meta-deepseekv3/DeepseekV3-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-deepseekv3/DeepseekV3-2-7b-hf")
+
+        >>> 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_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,
+            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=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The DeepseekV3 Model transformer with a sequence classification head on top (linear layer).
+
+    [`DeepseekV3ForSequenceClassification`] 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).
+    """,
+    DEEPSEEKV3_START_DOCSTRING,
+)
+class DeepseekV3ForSequenceClassification(DeepseekV3PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = DeepseekV3Model(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(DEEPSEEKV3_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[Union[Cache, 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,
+        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:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        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,
+        )
\ No newline at end of file

ktransformers/operators/attention.py

@@ -13,6 +13,7 @@ from ktransformers.models.configuration_deepseek import DeepseekV2Config
 from ktransformers.models.configuration_llama import LlamaConfig
 from ktransformers.models.modeling_llama import LlamaRotaryEmbedding
 from ktransformers.models.modeling_deepseek import DeepseekV2Attention, apply_rotary_pos_emb
+from ktransformers.models.modeling_deepseekv3 import DeepseekV3Attention, apply_rotary_pos_emb
 from typing import Optional, Tuple
 from ktransformers.operators.base_operator import BaseInjectedModule
 from ktransformers.util.custom_gguf import GGUFLoader
@@ -20,6 +21,206 @@ import logging
 from transformers.configuration_utils import PretrainedConfig
 from transformers.cache_utils import Cache
 logger = logging.getLogger("attention")
+
+class KDeepseekV3Attention(BaseInjectedModule, DeepseekV3Attention):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+    attn_mask: Optional[torch.Tensor] = None
+
+    def __init__(self,
+                 key: str,
+                 gguf_loader : GGUFLoader,
+                 config: PretrainedConfig,
+                 orig_module: nn.Module,
+                 device: str = "cuda",
+                 chunck_size: int = 1000,
+                 **kwargs):
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
+        self.orig_module.__init__(orig_module.config,
+            orig_module.layer_idx)
+        self.chunck_size = chunck_size # TODO, generate chunck_size automatically.
+        self.softmax_scale = self.q_head_dim ** (-0.5)
+
+    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_chunck(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+        if self.q_lora_rank is None:
+            q = self.q_proj(hidden_states)
+        else:
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q = q.view(bsz, q_len, self.num_heads, self.q_head_dim).transpose(1, 2)
+        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)
+        compressed_kv, k_pe = torch.split(
+            compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+        )
+        compressed_kv = self.kv_a_layernorm(compressed_kv)
+        k_pe = k_pe.view(bsz, q_len, 1, self.qk_rope_head_dim).transpose(1, 2)
+
+        kv_seq_len = k_pe.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+        cos, sin = self.rotary_emb(q_pe, position_ids)
+        q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            compressed_kv = compressed_kv.unsqueeze(1)
+            k_pe, compressed_kv = past_key_value.update(k_pe, compressed_kv, self.layer_idx, cache_kwargs)
+            compressed_kv = compressed_kv.squeeze(1)
+            #if cache_position is not None:  
+            #    compressed_kv = compressed_kv[:,: cache_position[-1] + 1,:]
+            #    k_pe = k_pe[:,:,: cache_position[-1] + 1,:]
+        q_absorb, out_absorb = self.get_absorbed()
+
+        q_nope = torch.matmul(q_nope, q_absorb)
+        attn_weights = (torch.matmul(q_pe, k_pe.mT) + torch.matmul(q_nope, compressed_kv.unsqueeze(-3).mT)) * self.softmax_scale
+        """
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+        assert attention_mask is not None
+        """
+        if attention_mask is not None:
+            """
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            """
+            #causal_mask = attention_mask[:, :, :, : kv_seq_len]
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(
+            attn_weights, dim=-1, dtype=torch.float32
+        ).to(q_pe.dtype)
+        attn_weights = nn.functional.dropout(
+            attn_weights, p=self.attention_dropout, training=self.training
+        )
+        attn_output = torch.einsum('bhql,blc->bhqc', attn_weights, compressed_kv)
+
+        attn_output = torch.matmul(attn_output, out_absorb.mT) 
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.v_head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.v_head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.v_head_dim)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights 
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        bsz, q_len, _ = hidden_states.size()
+        
+        if q_len <= self.chunck_size:
+            return self.forward_chunck(
+                            hidden_states,
+                            attention_mask,
+                            position_ids,
+                            past_key_value,
+                            output_attentions,
+                            use_cache,
+                            cache_position,
+                            **kwargs
+                        )
+
+        assert output_attentions == False, "output_attentions is not supported when using chunked attention"
+        attn_output = None
+        attn_weight = None
+        cur_idx = 0
+        while cur_idx < q_len:
+            if attention_mask is not None:
+                chunk_mask = attention_mask[:, :, cur_idx:min(cur_idx + self.chunck_size, q_len), ...]
+            else:
+                # generate chunk_mask automatically.
+                self.attn_mask = \
+                    torch.zeros(1, 1, self.chunck_size, past_key_value.max_cache_len, device=hidden_states.device) \
+                        if self.attn_mask is None \
+                            else self.attn_mask
+                self.attn_mask[:, :, :, cur_idx:min(cur_idx+self.chunck_size, past_key_value.max_cache_len)] = \
+                    -1e+38 * torch.triu(torch.ones(self.chunck_size, self.chunck_size, device=hidden_states.device), diagonal=1)\
+                        [:,:min(self.chunck_size, min(past_key_value.max_cache_len-cur_idx, self.chunck_size))]
+                self.attn_mask[:, :, :, cur_idx+self.chunck_size:] = -1e+38
+                self.attn_mask[:, :, :, :cur_idx] = 0
+                chunk_mask = torch.narrow(self.attn_mask, 2, 0, min(self.chunck_size, q_len-cur_idx))
+
+            cur_output, cur_attn_weight = self.forward_chunck(
+                            hidden_states[:, cur_idx:min(cur_idx + self.chunck_size, q_len), ...],
+                            chunk_mask,
+                            position_ids[:, cur_idx:min(cur_idx + self.chunck_size, q_len)],
+                            past_key_value,
+                            output_attentions,
+                            use_cache,
+                            cache_position[cur_idx:min(cur_idx + self.chunck_size, q_len)],
+                            **kwargs
+                        )
+            cur_idx += self.chunck_size
+            if attn_output is None:
+                attn_output = cur_output
+                attn_weight = cur_attn_weight
+            else:
+                attn_output = torch.cat((attn_output, cur_output), dim=-2)
+                attn_weight = torch.cat((attn_weight, cur_attn_weight), dim=-2)
+                
+        return attn_output, attn_weight
+
 class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
     attn_mask: Optional[torch.Tensor] = None

ktransformers/operators/experts.py

@@ -519,6 +519,7 @@ class KTransformersExperts(BaseInjectedModule, KExpertsBase):
 
 
 from ktransformers.models.modeling_deepseek import DeepseekV2MoE
+from ktransformers.models.modeling_deepseekv3 import DeepseekV3MoE
 from ktransformers.models.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
 from ktransformers.models.modeling_mixtral import MixtralSparseMoeBlock
 
@@ -727,6 +728,106 @@ class KDeepseekV2MoE(BaseInjectedModule, DeepseekV2MoE):
         )
         return final_out
 
+class KDeepseekV3MoE(BaseInjectedModule, DeepseekV3MoE):
+    
+    def forward(self, hidden_states):
+        identity = hidden_states
+        orig_shape = hidden_states.shape
+        sequence_length = orig_shape[1]
+        topk_idx, topk_weight= self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        
+        if sequence_length == 1 and hasattr(self.experts.generate_experts, "submit_for_one_decode") and torch.cuda.is_current_stream_capturing():
+            self.experts.generate_experts.submit_for_one_decode(hidden_states[0], topk_idx[0], topk_weight[0])
+            if self.config.n_shared_experts is not None:
+                y_ = self.shared_experts(identity).squeeze(0)
+            y = self.experts.generate_experts.sync_for_one_decode().unsqueeze(0)
+            y += y_
+            y.resize_(*orig_shape)
+            return y
+
+        if self.config.n_shared_experts is not None:
+            y_ = self.shared_experts(identity).squeeze(0)
+            
+        if isinstance(self.experts, KExpertsBase):
+            y = self.moe_on_cpuinfer(hidden_states, topk_idx, topk_weight).view(*orig_shape).to(device=hidden_states.device)
+        elif hidden_states.size(0) > 10:
+            # TODO may bugs here
+            y = (
+                self.moe_infer(hidden_states, topk_idx, topk_weight)
+                .view(*orig_shape)
+                .to(device=hidden_states.device)
+            )
+        else:
+            # TODO may bugs here
+            y = (
+                self.moe_infer_simple(hidden_states, topk_idx, topk_weight)
+                .view(*orig_shape)
+                .to(device=hidden_states.device)
+            )
+        if self.config.n_shared_experts is not None:
+            y += y_
+        return y
+
+    @torch.no_grad()
+    def moe_on_cpuinfer(self, x: torch.Tensor, topk_ids: torch.Tensor, topk_weight: torch.Tensor) -> torch.Tensor:
+        outs = torch.empty_like(x)
+        outs = self.experts(x, topk_ids, topk_weight)
+        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 KMistralSparseMoEBlock(BaseInjectedModule, MixtralSparseMoeBlock):
     
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:

ktransformers/operators/gate.py

+
+from typing import Any, Union
+import numpy as np
+import numpy.typing as npt
+from torch import Tensor, nn
+import torch.nn.functional as F
+import torch
+import sys, os
+from ktransformers.operators.base_operator import BaseInjectedModule
+
+sys.path.append(os.path.join(os.path.dirname(__file__), "..", "ktransformers_ext", "build"))
+sys.path.append(os.path.join(os.path.dirname(__file__), "..", "ktransformers_ext", "build", "Release"))
+sys.path.append(os.path.join(os.path.dirname(__file__), "..", "ktransformers_ext", "build", "Debug"))
+import cpuinfer_ext
+from cpuinfer_ext.moe import MOEConfig, MOE
+import ctypes
+from ktransformers.operators.base_operator import BaseInjectedModule
+from ktransformers.util.custom_gguf import GGUFLoader
+from ktransformers.models.modeling_deepseekv3 import MoEGate
+from ktransformers.util.utils import InferenceState
+from ktransformers.server.config.config import Config
+from transformers.activations import ACT2FN
+from transformers.configuration_utils import PretrainedConfig
+from abc import ABC, abstractmethod
+import time
+
+
+# class Base(BaseInjectedModule, ABC):
+class KMoEGateBase(ABC):
+    def __init__(self, 
+                 key: str, 
+                 gguf_loader: GGUFLoader, 
+                 config: PretrainedConfig, 
+                 orig_module: nn.Module, 
+                 device: str = "cuda", 
+                 **kwargs):
+        # super().__init__(key, gguf_loader, config, orig_module, device, **kwargs)
+        super().__init__()
+        self.key = key
+        self.gguf_loader = gguf_loader
+        self.config = config
+        self.device = device
+        self.orig_module = orig_module
+    
+    @abstractmethod
+    def forward(self, input_tensor, expert_ids, weights):
+        pass
+
+    @abstractmethod
+    def load(self, w: dict | nn.Parameter | tuple | None = None, device: str = "cpu", warmup: bool = False):
+        pass
+    
+    @abstractmethod
+    def unload():
+        pass
+
+    def load_weights(self, override_key: str | None = None, device: str = "cpu"):
+        res = {}
+        if override_key is not None:
+            keys = override_key
+        else:
+            keys = [self.key]
+
+        gate = None
+        up = None
+        down = None
+        gate_type = None
+        up_type = None
+        down_type = None
+
+        for key in keys:
+            key = ".".join(key.split(".")[:-1])
+            if key + ".ffn_gate_inp.weight" in self.gguf_loader.tensor_info:
+                targets = [".ffn_gate_inp.weight", ".exp_probs_b.bias"]
+                tensors = self.load_multi(key, targets, device=device)
+                weight = tensors[".ffn_gate_inp.weight"]
+                e_score_correction_bias = tensors[".exp_probs_b.bias"]
+                weight_type = self.gguf_loader.tensor_info[key + ".ffn_gate_inp.weight"]["ggml_type"]
+                e_score_correction_bias_type = self.gguf_loader.tensor_info[key + ".exp_probs_b.bias"]["ggml_type"]
+            else:
+                raise ValueError(f"Experts {key} not found in gguf_loader")
+            res = {"weight": weight, "e_score_correction_bias": e_score_correction_bias,  "weight_type": weight_type, "e_score_correction_bias_type": e_score_correction_bias_type}
+        return res
+    
+    def load_multi(self, key: str, keys: list[str], device: str = "cpu"):
+        tensors = {}
+        for k in keys:
+            tensors[k] = self.gguf_loader.load_gguf_tensor(key + k, device=device)
+        return tensors
+
+
+class KMoEGate(BaseInjectedModule, KMoEGateBase):
+    def __init__(
+        self,
+        key: str,
+        gguf_loader: GGUFLoader,
+        config: PretrainedConfig,
+        orig_module: nn.Module = None,
+        generate_device: str = "cuda",
+        prefill_device: str = "cuda",
+        **kwargs,
+    ):
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
+        KMoEGateBase.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
+
+    def forward(self, hidden_states) -> torch.Tensor:
+        return self.orig_module.forward(hidden_states)
+
+    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 = self.orig_module.weight.to(device)
+        if self.topk_method == "noaux_tc":
+            self.orig_module.e_score_correction_bias = self.orig_module.e_score_correction_bias.to(device)
+
+    def unload(self):
+        if self.weight is not None:
+            self.weight = None
+        if self.topk_method == "noaux_tc":
+            self.e_score_correction_bias = None

ktransformers/operators/linear.py

@@ -54,15 +54,15 @@ class KLinearBase(ABC):
 
         self.has_bias = False
         self.dtype = torch.get_default_dtype()
-        if orig_module is not None:
-            self.in_features = orig_module.in_features
-            self.out_features = orig_module.out_features
-        else:
-            shape = self.gguf_loader.tensor_info[key + ".weight"]["shape"]
-            if len(shape) == 1:
-                print("Warning: orig_module is not set, but has in_features or out_features equals to 1, can't get in_features and out_features from GGUF")
-            self.in_features  = self.gguf_loader.tensor_info[key + ".weight"]["shape"][0]
-            self.out_features = self.gguf_loader.tensor_info[key + ".weight"]["shape"][1]
+        # if orig_module is not None:
+        #     self.in_features = orig_module.in_features
+        #     self.out_features = orig_module.out_features
+        # else:
+        shape = self.gguf_loader.tensor_info[key + ".weight"]["shape"]
+        if len(shape) == 1:
+            print("Warning: orig_module is not set, but has in_features or out_features equals to 1, can't get in_features and out_features from GGUF")
+        self.in_features  = self.gguf_loader.tensor_info[key + ".weight"]["shape"][0]
+        self.out_features = self.gguf_loader.tensor_info[key + ".weight"]["shape"][1]
 
     @abstractmethod
     def forward(self, x: torch.Tensor) -> torch.Tensor:

ktransformers/operators/models.py

@@ -641,6 +641,7 @@ class KDeepseekV2Model(BaseInjectedModule):
 
         if inputs_embeds is None:
             org_device = input_ids.device
+            # TODO move to embed_tokens's device, not hard code to cpu
             input_ids = input_ids.to("cpu")
             inputs_embeds = self.embed_tokens(input_ids)
             input_ids = input_ids.to(org_device)
@@ -737,8 +738,9 @@ class KDeepseekV2Model(BaseInjectedModule):
 
             hidden_states = layer_outputs[0]
 
-            if use_cache:
-                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+            # @@@@@@@ TODO open this notes, tmp close to fit deepseekv3
+            # if use_cache:
+            #     next_decoder_cache = layer_outputs[2 if output_attentions else 1]
 
             if output_attentions:
                 all_self_attns += (layer_outputs[1],)

ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-multi-gpu.yaml

+- match:
+    name: "^model.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+        generate_device: "cpu"
+        prefill_device: "cpu"
+
+- match:
+    name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\."
+    class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.YarnRotaryEmbedding
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([3456][0-9])\\."
+    class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.YarnRotaryEmbedding
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
+
+- match:
+    name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.(?!self_attn\\.kv_b_proj).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+- match:
+    name: "^model\\.layers\\.([3456][0-9])\\.(?!self_attn\\.kv_b_proj).*$"  # 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:1"
+      prefill_device: "cuda:1"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+  
+- match:
+    name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.mlp$"
+    class: ktransformers.models.modeling_deepseekv3.DeepseekV3MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV3MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([3456][0-9])\\.mlp$"
+    class: ktransformers.models.modeling_deepseekv3.DeepseekV3MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV3MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
+
+- match:
+    name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.mlp\\.gate$"
+    class: ktransformers.models.modeling_deepseekv3.MoEGate
+  replace:
+    class: ktransformers.operators.gate.KMoEGate
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([3456][0-9])\\.mlp\\.gate$"
+    class: ktransformers.models.modeling_deepseekv3.MoEGate
+  replace:
+    class: ktransformers.operators.gate.KMoEGate     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
+
+- match:
+    name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda:0"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op:  "KExpertsCPU"
+      out_device: "cuda:0"
+  recursive: False # don't recursively inject submodules of this module
+
+- match:
+    name: "^model\\.layers\\.([3456][0-9])\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda:1"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op:  "KExpertsCPU"
+      out_device: "cuda:1"
+  recursive: False # don't recursively inject submodules of this module
+
+- match:
+    name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV3Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([3456][0-9])\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV3Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KDeepseekV2Model"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+      transfer_map: 
+        30: "cuda:1"
+
+- match:
+    name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\."
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+
+- match:
+    name: "(^model\\.layers\\.([3456][0-9])\\.)|(model.norm)|(lm_head)"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"

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

+- match:
+    class: ktransformers.models.modeling_deepseek.DeepseekV3YarnRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.YarnRotaryEmbedding
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- 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.DeepseekV2MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- match:
+    name: "^model\\.layers\\..*\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op: "KExpertsCPU"
+      out_device: "cuda"
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\..*\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- 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/server/backend/interfaces/ktransformers.py

@@ -46,17 +46,26 @@ class KTransformersInterface(TransformersInterface):
             )
         optimize_and_load_gguf(self.model, optimize_rule_path, gguf_path, config)
 
-        device_map = self.model.gguf_loader.tensor_device_map
-        logger.info(f"{args.model_name} loaded from {args.model_dir} to {device_map}")
+        self.device_map = self.model.gguf_loader.tensor_device_map
+        # logger.info(f"{args.model_name} loaded from {args.model_dir} to {self.device_map}")
         self.cache = StaticCache(
             config=self.model.config,
             max_batch_size=args.batch_size,
             max_cache_len=args.cache_lens,
-            device=device_map,
+            device=self.device_map,
             dtype=self.model.dtype,
         )
-        logger.info(f"StaticCache (length={args.cache_lens}) created at {device_map}, batch size:{args.batch_size}")
-        self.model.generation_config = GenerationConfig.from_pretrained(args.model_dir)
+        # logger.info(f"StaticCache (length={args.cache_lens}), batch size:{args.batch_size}")
+        try:
+            self.model.generation_config = GenerationConfig.from_pretrained(args.model_dir)
+        except:
+            gen_config = GenerationConfig(
+                max_length=128,
+                temperature=0.7,
+                top_p=0.9,
+                do_sample=True
+            )
+            self.model.generation_config = gen_config
         if self.model.generation_config.pad_token_id is None:
             self.model.generation_config.pad_token_id = self.model.generation_config.eos_token_id
         self.streamer = TextStreamer(self.tokenizer)
@@ -102,3 +111,63 @@ class KTransformersInterface(TransformersInterface):
         logits = logits[0, -1, :]
 
         return self.logits_to_token(logits)
+
+
+
+    @torch.no_grad
+    def prefill(self, input_ids: torch.Tensor, is_new: bool):
+        input_ids_length = input_ids.shape[-1]
+        self.profiler.set_counter("prefill", input_ids_length)
+        logger.debug(f"input_ids: {input_ids.shape}")
+
+        device = self.device_map.get("blk.0.self_attn", {}).get("generate_device", "cuda:0")
+
+        if is_new:
+            self.cache.reset()
+            self.ever_generated_ids.clear()
+            former_seq_length = 0
+            self.seq_length = input_ids_length
+            self.generated_ids = torch.zeros(
+                self.args.batch_size,
+                self.seq_length + self.args.max_new_tokens + 1,
+                dtype=torch.int,
+                device=self.args.device,
+            )
+        else:
+            logger.debug(f"generate_ids: {self.generated_ids.shape}")
+            former_seq_length = self.seq_length
+            self.seq_length += input_ids_length
+            expected_length = self.seq_length + self.args.max_new_tokens + 1
+            delta_length = expected_length - self.generated_ids.shape[-1]
+            if delta_length > 0:
+                new_generate_ids = torch.zeros(
+                    self.args.batch_size, delta_length, dtype=torch.int, device=self.args.device
+                )
+                self.generated_ids = torch.cat([self.generated_ids, new_generate_ids], dim=-1)
+        logger.debug(f"cache position: {former_seq_length} to {self.seq_length}")
+        cache_position = torch.arange(former_seq_length, self.seq_length, device=device)
+        self.generated_ids[:, cache_position] = input_ids.to(self.args.device).to(torch.int)
+
+        mask = torch.ones((1, self.seq_length)).to(device)
+        if not (type(self) is TransformersInterface):
+            input_ids = input_ids.to("cpu")
+        inputs_embeds = self.model.model.embed_tokens(input_ids).to(device)
+        if self.use_static_cache:
+            logits = self.model(
+                inputs_embeds=inputs_embeds,
+                cache_position=cache_position,
+                past_key_values=self.cache,
+                return_dict=False,
+                use_cache=True,
+                attention_mask=mask,
+            )[0]
+        else:
+            logits = self.model(inputs_embeds=inputs_embeds, return_dict=False)[0]
+
+        next_token = self.logits_to_token(logits[0, -1, :])
+        yield self.append_new_tokens(next_token)
+
+    @property
+    def active_cache_position(self):
+        device = self.device_map.get("blk.0.self_attn", {}).get("generate_device", "cuda:0")
+        return torch.tensor([self.seq_length - 1], device=device)
\ No newline at end of file

ktransformers/server/backend/interfaces/transformers.py

@@ -134,7 +134,7 @@ class TransformersInterface(BackendInterfaceBase):
 
         self.tokenizer = AutoTokenizer.from_pretrained(args.model_dir)
         self.model = AutoModelForCausalLM.from_pretrained(args.model_dir, device_map=args.device, use_safetensors=True)
-        logger.info(f"{args.model_name} loaded from {args.model_dir} to {args.device}")
+        # logger.info(f"{args.model_name} loaded from {args.model_dir} to {args.device}")
 
         self.cache = StaticCache(
             config=self.model.config,
@@ -143,7 +143,7 @@ class TransformersInterface(BackendInterfaceBase):
             device=args.device,
             dtype=self.model.dtype,
         )
-        logger.info(f"StaticCache (length={args.cache_lens}) created at {args.device}, batch size:{args.batch_size}")
+        # logger.info(f"StaticCache (length={args.cache_lens}) created at {args.device}, batch size:{args.batch_size}")
 
         self.streamer = TextStreamer(self.tokenizer)