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

@@ -0,0 +1,231 @@
+# 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"]

ktransformers/models/custom_cache.py

@@ -34,6 +34,9 @@ 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
+        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
             )
@@ -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/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

@@ -0,0 +1,128 @@
+
+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

@@ -54,10 +54,10 @@ 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:
+        # 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")

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

@@ -0,0 +1,143 @@
+- 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

@@ -0,0 +1,56 @@
+- 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"

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}")
+        # 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)

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)