Merge pull request #36 from kvcache-ai/develop-0.1.2

Release v0.1.2

ktransformers/ktransformers_ext/operators/llamafile/moe.h

@@ -3,7 +3,7 @@
  * @Author       : chenht2022
  * @Date         : 2024-07-22 02:03:22
  * @Version      : 1.0.0
- * @LastEditors  : chenht2022 
+ * @LastEditors  : chenht2022
  * @LastEditTime : 2024-07-25 10:35:10
  * @Copyright (c) 2024 by KVCache.AI, All Rights Reserved.
  **/
@@ -22,6 +22,7 @@
 #include "llama.cpp/ggml-quants.h"
 #include "llama.cpp/ggml.h"
 #include "llamafile/sgemm.h"
+#include "shared_mem_buffer.h"
 
 struct MOEConfig {
     int expert_num;
@@ -48,13 +49,13 @@ struct MOEConfig {
 class MOE {
    public:
     MOE(MOEConfig);
+    ~MOE();
     void warm_up(Backend* backend);
     void forward_one(int k, const uint64_t* expert_ids, const float* weights, const void* input, void* output, Backend* backend);
     void forward_many(int qlen, int k, const uint64_t* expert_ids, const float* weights, const void* input, void* output, Backend* backend);
     void forward(int qlen, int k, const uint64_t* expert_ids, const float* weights, const void* input, void* output, Backend* backend);
 
    private:
-    static uint8_t* buffer_;
     MOEConfig config_;
     void* gate_proj_;  // [expert_num * intermediate_size * hidden_size ( /32 if quantized)]
     void* up_proj_;    // [expert_num * intermediate_size * hidden_size ( /32 if quantized)]

ktransformers/ktransformers_ext/operators/llamafile/shared_mem_buffer.cpp

+/**
+ * @Description  :
+ * @Author       : chenht2022
+ * @Date         : 2024-08-05 04:49:08
+ * @Version      : 1.0.0
+ * @LastEditors  : chenht2022 
+ * @LastEditTime : 2024-08-05 09:21:29
+ * @Copyright (c) 2024 by KVCache.AI, All Rights Reserved.
+ **/
+#include "shared_mem_buffer.h"
+#include <cstdio>
+
+SharedMemBuffer::SharedMemBuffer() {
+    buffer_ = nullptr;
+    size_ = 0;
+}
+
+SharedMemBuffer::~SharedMemBuffer() {
+    if (buffer_) {
+        free(buffer_);
+    }
+}
+
+void SharedMemBuffer::alloc(void* object, std::vector<std::pair<void**, uint64_t>> requests) {
+    uint64_t size = 0;
+    for (auto& request : requests) {
+        size += request.second;
+    }
+    if (size > size_) {
+        if (buffer_) {
+            free(buffer_);
+        }
+        buffer_ = malloc(size);
+        size_ = size;
+        for (auto& obj_requests : hist_requests_) {
+            for (auto& requests : obj_requests.second) {
+                arrange(requests);
+            }
+        }
+    }
+    arrange(requests);
+    hist_requests_[object].push_back(requests);
+}
+
+void SharedMemBuffer::dealloc(void* object) {
+    hist_requests_.erase(object);
+}
+
+void SharedMemBuffer::arrange(std::vector<std::pair<void**, uint64_t>> requests) {
+    uint64_t offset = 0;
+    for (auto& request : requests) {
+        *(request.first) = (uint8_t*)buffer_ + offset;
+        offset += request.second;
+    }
+}

ktransformers/ktransformers_ext/operators/llamafile/shared_mem_buffer.h

+/**
+ * @Description  :
+ * @Author       : chenht2022
+ * @Date         : 2024-08-05 04:49:08
+ * @Version      : 1.0.0
+ * @LastEditors  : chenht2022 
+ * @LastEditTime : 2024-08-05 06:36:41
+ * @Copyright (c) 2024 by KVCache.AI, All Rights Reserved.
+ **/
+
+#ifndef CPUINFER_SHAREDMEMBUFFER_H
+#define CPUINFER_SHAREDMEMBUFFER_H
+
+#include <cstdint>
+#include <cstdlib>
+#include <map>
+#include <vector>
+
+class SharedMemBuffer {
+   public:
+    SharedMemBuffer();
+    ~SharedMemBuffer();
+
+    void alloc(void* object, std::vector<std::pair<void**, uint64_t>> requests);
+    void dealloc(void* object);
+
+   private:
+    void* buffer_;
+    uint64_t size_;
+    std::map<void*, std::vector<std::vector<std::pair<void**, uint64_t>>>> hist_requests_;
+
+    void arrange(std::vector<std::pair<void**, uint64_t>> requests);
+};
+
+static SharedMemBuffer shared_mem_buffer;
+
+#endif
\ No newline at end of file

ktransformers/local_chat.py

@@ -31,18 +31,21 @@ import fire
 from ktransformers.optimize.optimize import optimize_and_load_gguf
 from ktransformers.models.modeling_deepseek import DeepseekV2ForCausalLM
 from ktransformers.models.modeling_qwen2_moe import Qwen2MoeForCausalLM
+from ktransformers.models.modeling_mixtral import MixtralForCausalLM
 from ktransformers.util.utils import prefill_and_generate
 from ktransformers.server.config.config import Config
 
 custom_models = {
     "DeepseekV2ForCausalLM": DeepseekV2ForCausalLM,
     "Qwen2MoeForCausalLM": Qwen2MoeForCausalLM,
+    "MixtralForCausalLM": MixtralForCausalLM,
 }
 
 ktransformer_rules_dir = os.path.dirname(os.path.abspath(__file__)) + "/optimize/optimize_rules/"
 default_optimize_rules ={
     "DeepseekV2ForCausalLM": ktransformer_rules_dir + "DeepSeek-V2-Chat.yaml",
     "Qwen2MoeForCausalLM": ktransformer_rules_dir + "Qwen2-57B-A14B-Instruct.yaml",
+    "MixtralForCausalLM": ktransformer_rules_dir + "Mixtral.yaml",
 }
 
 def local_chat(
@@ -50,7 +53,8 @@ def local_chat(
     optimize_rule_path: str = None,
     gguf_path: str = None,
     max_new_tokens: int = 1000,
-    cpu_infer: int = Config().cpu_infer
+    cpu_infer: int = Config().cpu_infer,
+    use_cuda_graph: bool = True,
 ):
     torch.set_grad_enabled(False)
     
@@ -64,6 +68,8 @@ def local_chat(
             print("using custom modeling_xxx.py.")
             if "Qwen2Moe" in config.architectures[0]: # Qwen2Moe must use flash_attention_2 to avoid overflow.
                 config._attn_implementation = "flash_attention_2"
+            if "Mixtral" in config.architectures[0]: 
+                config._attn_implementation = "flash_attention_2"
             model = custom_models[config.architectures[0]](config)
         else:
             model = AutoModelForCausalLM.from_config(
@@ -100,7 +106,6 @@ def local_chat(
 
     while True:
         content = input("Chat: ")
-        # if content is num
         if content == "":
             content = "Please write a piece of quicksort code in C++." 
 
@@ -109,7 +114,7 @@ def local_chat(
             messages, add_generation_prompt=True, return_tensors="pt"
         )
         torch.set_default_dtype(torch.bfloat16) # TODO: Remove this, replace dtype using config
-        generated = prefill_and_generate(model, tokenizer, input_tensor.cuda(), max_new_tokens)
+        generated = prefill_and_generate(model, tokenizer, input_tensor.cuda(), max_new_tokens, use_cuda_graph)
 
 if __name__ == "__main__":
     fire.Fire(local_chat)
\ No newline at end of file

ktransformers/models/custom_cache.py

@@ -22,13 +22,14 @@ class StaticCache(transformers.StaticCache):
             The maximum batch size with which the model will be used.
         max_cache_len (`int`):
             The maximum sequence length with which the model will be used.
-        device (`torch.device`):
+        device (`torch.device` or `dict`):
             The device on which the cache should be initialized. Should be the same as the layer.
+            If a `dict`, it should contain the `device` key with the device name as the value.
         dtype (*optional*, defaults to `torch.float32`):
             The default `dtype` to use when initializing the layer.
     """
 
-    def __init__(self, config: PretrainedConfig, max_batch_size: int, max_cache_len: int, device, dtype=None) -> None:
+    def __init__(self, config: PretrainedConfig, max_batch_size: int, max_cache_len: int, device: torch.device| dict, dtype=None) -> None:
         Cache.__init__(self)
         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
@@ -57,11 +58,15 @@ class StaticCache(transformers.StaticCache):
 
         self.past_tokens = []
         self.num_hidden_layers = config.num_hidden_layers
-        for _ in range(self.num_hidden_layers):
+        for idx in range(self.num_hidden_layers):
             # Note: `mark_static_address` is used to tag the cache as an fixed data pointer, preventing cuda graph
             # breaks when updating the cache.
-            new_layer_key_cache = torch.zeros(key_shape, dtype=self.dtype, device=device)
-            new_layer_value_cache = torch.zeros(value_shape, dtype=self.dtype, device=device)
+            if isinstance(device, dict):
+                target_device = device[f"blk.{idx}.self_attn"]["generate_device"]
+            else:
+                target_device = device
+            new_layer_key_cache = torch.zeros(key_shape, dtype=self.dtype, device=target_device)
+            new_layer_value_cache = torch.zeros(value_shape, dtype=self.dtype, device=target_device)
             torch._dynamo.mark_static_address(new_layer_key_cache)
             torch._dynamo.mark_static_address(new_layer_value_cache)
             self.key_cache.append(new_layer_key_cache)

ktransformers/models/modeling_deepseek.py

@@ -1048,7 +1048,7 @@ class DeepseekV2FlashAttention2(DeepseekV2Attention):
         """
         Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
         first unpad the input, then computes the attention scores and pad the final attention scores.
-        Args:
+        # Args:
             query_states (`torch.Tensor`):
                 Input query states to be passed to Flash Attention API
             key_states (`torch.Tensor`):
@@ -1245,12 +1245,14 @@ class DeepseekV2DecoderLayer(nn.Module):
             cache_position=cache_position,
             **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,)

ktransformers/models/modeling_mixtral.py

+# coding=utf-8
+'''
+Description  : 
+Author       : kkk1nak0
+Date         : 2024-07-29 02:58:57
+Version      : 1.0.0
+LastEditors  : kkk1nak0
+LastEditTime : 2024-08-02 06:08:34
+'''
+
+# Adapted from 
+# https://github.com/huggingface/transformers/blob/main/src/transformers/models/mixtral/modeling_mixtral.py
+# Copyright 2023 Mistral AI and the HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2024 by KVCache.AI, All Rights Reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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.
+"""PyTorch Mixtral model."""
+
+import inspect 
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, StaticCache
+from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
+    _prepare_4d_causal_attention_mask,
+)
+from transformers.modeling_outputs import (
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_13
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.import_utils import is_torch_fx_available
+from transformers.models.mixtral.configuration_mixtral import MixtralConfig
+
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_varlen_func, flash_attn_func, flash_attn_with_kvcache
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
+
+# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
+# It means that the function will not be traced through and simply appear as a node in the graph.
+if is_torch_fx_available():
+    if not is_torch_greater_or_equal_than_1_13:
+        import torch.fx
+
+    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "MixtralConfig"
+
+
+def load_balancing_loss_func(
+    gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2, attention_mask: Optional[torch.Tensor] = None
+) -> float:
+    r"""
+    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
+
+    See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss
+    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
+    experts is too unbalanced.
+
+    Args:
+        gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
+            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
+            shape [batch_size X sequence_length, num_experts].
+        attention_mask (`torch.Tensor`, None):
+            The attention_mask used in forward function
+            shape [batch_size X sequence_length] if not None.
+        num_experts (`int`, *optional*):
+            Number of experts
+
+    Returns:
+        The auxiliary loss.
+    """
+    if gate_logits is None or not isinstance(gate_logits, tuple):
+        return 0
+
+    if isinstance(gate_logits, tuple):
+        compute_device = gate_logits[0].device
+        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
+
+    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
+
+    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
+
+    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
+
+    if attention_mask is None:
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.mean(routing_weights, dim=0)
+    else:
+        batch_size, sequence_length = attention_mask.shape
+        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
+        expert_attention_mask = (
+            attention_mask[None, :, :, None, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
+            .reshape(-1, top_k, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the percentage of tokens routed to each experts
+        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
+            expert_attention_mask, dim=0
+        )
+
+        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
+        router_per_expert_attention_mask = (
+            attention_mask[None, :, :, None]
+            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
+            .reshape(-1, num_experts)
+            .to(compute_device)
+        )
+
+        # Compute the average probability of routing to these experts
+        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
+            router_per_expert_attention_mask, dim=0
+        )
+
+    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
+    return overall_loss * num_experts
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Mixtral
+class MixtralRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        MixtralRMSNorm 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}"
+
+
+# copied from transformers.models.mistral.modeling_mistral.MistralRotaryEmbedding with Mistral->Mixtral
+# TODO @longjie no longer copied from Mistral after static cache
+class MixtralRotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+        
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+        # Build here to make `torch.jit.trace` work.
+        self.max_seq_len_cached = max_position_embeddings
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        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()
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# 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.mistral.modeling_mistral.apply_rotary_pos_emb
+# TODO @longjie no longer copied from Mistral after static cache
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids, 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`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+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)
+
+
+# copied from transformers.models.mistral.modeling_mistral.MistralAttention with Mistral->Mixtral
+# TODO @longjie no longer copied from Mistral after static cache
+class MixtralAttention(nn.Module):
+    """
+    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
+    and "Generating Long Sequences with Sparse Transformers".
+    """
+
+    def __init__(self, config: MixtralConfig, 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 a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.is_causal = True
+        self.attention_dropout = config.attention_dropout
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        self.rotary_emb = MixtralRotaryEmbedding(
+            self.head_dim,
+            max_position_embeddings=self.max_position_embeddings,
+            base=self.rope_theta,
+        )
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    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,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_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, position_ids)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        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()}"
+            )
+
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.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.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Mixtral
+# TODO @longjie no longer copied from Mistral after static cache
+class MixtralFlashAttention2(MixtralAttention):
+    """
+    Mixtral flash attention module. This module inherits from `MixtralAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    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,
+    ):
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_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, position_ids)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        use_sliding_windows = (
+            _flash_supports_window_size
+            and getattr(self.config, "sliding_window", None) is not None
+            and kv_seq_len > self.config.sliding_window
+            and self.config.use_sliding_window
+        )
+
+        if not _flash_supports_window_size:
+            logger.warning_once(
+                "The current flash attention version does not support sliding window attention, for a more memory efficient implementation"
+                " make sure to upgrade flash-attn library."
+            )
+
+        if past_key_value is not None:
+            # Activate slicing cache only if the config has a value `sliding_windows` attribute
+            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+            if (
+                getattr(self.config, "sliding_window", None) is not None
+                and kv_seq_len > self.config.sliding_window
+                and cache_has_contents
+            ):
+                slicing_tokens = 1 - self.config.sliding_window
+
+                past_key = past_key_value[self.layer_idx][0]
+                past_value = past_key_value[self.layer_idx][1]
+
+                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+                past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+                if past_key.shape[-2] != self.config.sliding_window - 1:
+                    raise ValueError(
+                        f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
+                        f" {past_key.shape}"
+                    )
+
+                if attention_mask is not None:
+                    attention_mask = attention_mask[:, slicing_tokens:]
+                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
+
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+            # we slice the states for static kv cache to be supported in FA2. Not sure it's a must as compile fails
+            # for bsz == 1, avoid using slice to capture cuda graph
+            if cache_position is not None and q_len > 1:
+                key_states = key_states[:, :, : cache_position[-1] + 1, :]
+                value_states = value_states[:, :, : cache_position[-1] + 1, :]
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in float16 just to be sure everything works as expected.
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            position_ids=position_ids,
+            dropout=dropout_rate,
+            sliding_window=getattr(self.config, "sliding_window", None),
+            is_causal=self.is_causal,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+    
+
+    def _flash_attention_forward(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        q_len,
+        position_ids,
+        dropout,
+        sliding_window,
+        is_causal,
+        softmax_scale=None,
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            
+        """
+        
+        # Decide whether to use SWA or not by layer index.
+        # if use_sliding_windows and self.layer_idx >= self.config.max_window_layers:
+        #     use_sliding_windows = False
+        use_sliding_windows = False
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, q_len
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            if not use_sliding_windows:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=is_causal,
+                )
+            else:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=is_causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, q_len)
+        else:
+            if not use_sliding_windows:
+                if q_len == 1:
+                    position_ids = position_ids.to(dtype=torch.int32).squeeze(1)
+                    attn_output = flash_attn_with_kvcache(
+                        query_states,
+                        key_states,
+                        value_states,
+                        cache_seqlens=position_ids,
+                        softmax_scale=softmax_scale,
+                        causal=is_causal,
+                    )   
+                else:
+                    attn_output = flash_attn_func(
+                        query_states,
+                        key_states,
+                        value_states,
+                        dropout,
+                        softmax_scale=softmax_scale,
+                        causal=is_causal,
+                    )
+            else:
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    dropout,
+                    softmax_scale=softmax_scale,
+                    causal=is_causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+        return attn_output
+
+    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
+
+        # On the first iteration we need to properly re-create the padding mask
+        # by slicing it on the proper place
+        if kv_seq_len != attention_mask.shape[-1]:
+            attention_mask_num_tokens = attention_mask.shape[-1]
+            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
+
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+
+        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+
+# copied from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Mixtral
+# TODO @longjie no longer copied from Mistral after static cache
+class MixtralSdpaAttention(MixtralAttention):
+    """
+    Mixtral attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `MixtralAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from MixtralAttention.forward
+    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,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "MixtralModel is using MixtralSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                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,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, position_ids)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+MIXTRAL_ATTENTION_CLASSES = {
+    "eager": MixtralAttention,
+    "flash_attention_2": MixtralFlashAttention2,
+    "sdpa": MixtralSdpaAttention,
+}
+
+
+class MixtralBlockSparseTop2MLP(nn.Module):
+    def __init__(self, config: MixtralConfig):
+        super().__init__()
+        self.ffn_dim = config.intermediate_size
+        self.hidden_dim = config.hidden_size
+
+        self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)  # gate
+        self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)  # down
+        self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)  # up
+
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states):
+        current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
+        current_hidden_states = self.w2(current_hidden_states)
+        return current_hidden_states
+
+
+class MixtralSparseMoeBlock(nn.Module):
+    """
+    This implementation is
+    strictly equivalent to standard MoE with full capacity (no
+    dropped tokens). It's faster since it formulates MoE operations
+    in terms of block-sparse operations to accomodate imbalanced
+    assignments of tokens to experts, whereas standard MoE either
+    (1) drop tokens at the cost of reduced performance or (2) set
+    capacity factor to number of experts and thus waste computation
+    and memory on padding.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_dim = config.hidden_size
+        self.ffn_dim = config.intermediate_size
+        self.num_experts = config.num_local_experts
+        self.top_k = config.num_experts_per_tok
+
+        # gating
+        self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
+
+        self.experts = nn.ModuleList([MixtralBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
+
+        # Jitter parameters
+        self.jitter_noise = config.router_jitter_noise
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        if self.training and self.jitter_noise > 0:
+            hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise)
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits
+
+
+class MixtralDecoderLayer(nn.Module):
+    def __init__(self, config: MixtralConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = MIXTRAL_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
+
+        self.block_sparse_moe = MixtralSparseMoeBlock(config)
+        self.input_layernorm = MixtralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = MixtralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, sequence_length)` where padding elements are indicated by 0.
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_router_logits (`bool`, *optional*):
+                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
+                should not be returned during inference.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = 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,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states, router_logits = self.block_sparse_moe(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+MIXTRAL_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 ([`MixtralConfig`]):
+            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 Mixtral Model outputting raw hidden-states without any specific head on top.",
+    MIXTRAL_START_DOCSTRING,
+)
+# Copied from transformers.models.qwen2.modeling_qwen2.Qwen2PreTrainedModel with Qwen2->Mixtral
+class MixtralPreTrainedModel(PreTrainedModel):
+    config_class = MixtralConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["MixtralDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = 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_()
+
+
+MIXTRAL_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 `decoder_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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            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)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_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.
+        output_router_logits (`bool`, *optional*):
+            Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
+            should not be returned during inference.
+        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 Mixtral Model outputting raw hidden-states without any specific head on top.",
+    MIXTRAL_START_DOCSTRING,
+)
+# copied from transformers.models.mistral.modeling_mistral.MistralModel with MISTRAL->MIXTRAL,Mistral->Mixtral
+# TODO @longjie no longer copied from Mistral after static cache
+class MixtralModel(MixtralPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MixtralDecoderLayer`]
+
+    Args:
+        config: MixtralConfig
+    """
+
+    def __init__(self, config: MixtralConfig):
+        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(
+            [MixtralDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self._attn_implementation = config._attn_implementation
+        self.norm = MixtralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        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
+
+    # Ignore copy
+    @add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, MoeModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
+            )
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        use_legacy_cache = False
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
+            use_legacy_cache = True
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            logger.warning_once(
+                "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. "
+                "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_logits = () if output_router_logits else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    output_router_logits,
+                    use_cache,
+                    cache_position,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            if output_router_logits:
+                all_router_logits += (layer_outputs[-1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = None
+        if use_cache:
+            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
+                if v is not None
+            )
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            router_logits=all_router_logits,
+        )
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
+        # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
+        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # 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
+        min_dtype = torch.finfo(dtype).min
+        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
+            )
+
+        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
+            if attention_mask.max() != 0:
+                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
+            causal_mask = attention_mask
+        else:
+            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(input_tensor.shape[0], 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
+                )
+        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
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+
+class MixtralForCausalLM(MixtralPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = MixtralModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.router_aux_loss_coef = config.router_aux_loss_coef
+        self.num_experts = config.num_local_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    # Ignore copy
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+        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]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, MixtralForCausalLM
+
+        >>> model = MixtralForCausalLM.from_pretrained("mistralai/Mixtral-8x7B-v0.1")
+        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mixtral-8x7B-v0.1")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            output_router_logits=output_router_logits,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        aux_loss = None
+        if output_router_logits:
+            aux_loss = load_balancing_loss_func(
+                outputs.router_logits if return_dict else outputs[-1],
+                self.num_experts,
+                self.num_experts_per_tok,
+                attention_mask,
+            )
+            if labels is not None:
+                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            if output_router_logits:
+                output = (aux_loss,) + output
+            return (loss,) + output if loss is not None else output
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        output_router_logits=False,
+        position_ids=None,
+        use_cache=True,
+        **kwargs,
+    ):
+        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
+        # Exception 1: when passing input_embeds, input_ids may be missing entries
+        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
+        if past_key_values is not None:
+            if inputs_embeds is not None:  # Exception 1
+                input_ids = input_ids[:, -cache_position.shape[0] :]
+            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
+                input_ids = input_ids[:, cache_position]
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and cache_position[0] == 0:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+                "output_router_logits": output_router_logits,
+            }
+        )
+        return model_inputs
+
+
+@add_start_docstrings(
+    """
+    The Mixtral Model transformer with a sequence classification head on top (linear layer).
+
+    [`MixtralForSequenceClassification`] 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).
+    """,
+    MIXTRAL_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Mixtral, LLAMA->MIXTRAL
+class MixtralForSequenceClassification(MixtralPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = MixtralModel(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(MIXTRAL_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:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The Mixtral Model transformer with a token classification head on top (a linear layer on top of the hidden-states
+    output) e.g. for Named-Entity-Recognition (NER) tasks.
+    """,
+    MIXTRAL_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForTokenClassification with Llama->Mixtral, LLAMA->MIXTRAL
+class MixtralForTokenClassification(MixtralPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = MixtralModel(config)
+        if getattr(config, "classifier_dropout", None) is not None:
+            classifier_dropout = config.classifier_dropout
+        elif getattr(config, "hidden_dropout", None) is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.score = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(MIXTRAL_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[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, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        sequence_output = self.dropout(sequence_output)
+        logits = self.score(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
\ No newline at end of file

ktransformers/operators/RoPE.py

@@ -10,6 +10,7 @@ from ktransformers.operators.base_operator import BaseInjectedModule
 from ktransformers.util.custom_gguf import GGUFLoader
 from ktransformers.util.utils import InferenceState
 from transformers.configuration_utils import PretrainedConfig
+
 # Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Qwen2Moe
 class RotaryEmbedding(BaseInjectedModule, DeepseekV2RotaryEmbedding):
     def __init__(self,
@@ -17,12 +18,16 @@ class RotaryEmbedding(BaseInjectedModule, DeepseekV2RotaryEmbedding):
                  gguf_loader : GGUFLoader,
                  config: PretrainedConfig,
                  orig_module: nn.Module,
-                 device: str = "cuda",
+                #  device: str = "cuda",
+                 generate_device: str = "cuda",
+                 prefill_device: str = "cuda",
                  **kwargs):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
         self.orig_module.__init__(orig_module.dim,
             orig_module.max_position_embeddings,
             orig_module.base)
+        self.generate_device = generate_device
+        self.prefill_device = prefill_device
     
     def load(self):
         self.orig_module.__init__(self.orig_module.dim,
@@ -36,9 +41,11 @@ class YarnRotaryEmbedding(BaseInjectedModule, DeepseekV2YarnRotaryEmbedding):
                  gguf_loader : GGUFLoader,
                  config: PretrainedConfig,
                  orig_module: nn.Module,
-                 device: str = "cuda",
+                #  device: str = "cuda",
+                 generate_device: str = "cuda",
+                 prefill_device: str = "cuda",
                  **kwargs):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
         self.orig_module.__init__(orig_module.dim,
             orig_module.max_position_embeddings,
             orig_module.base,
@@ -49,13 +56,15 @@ class YarnRotaryEmbedding(BaseInjectedModule, DeepseekV2YarnRotaryEmbedding):
             orig_module.beta_slow,
             orig_module.mscale,
             orig_module.mscale_all_dim)
+        self.generate_device = generate_device
+        self.prefill_device = prefill_device
         
     
     def load(self):
         self.orig_module.__init__(self.orig_module.dim,
             self.orig_module.max_position_embeddings,
             self.orig_module.base,
-            self.device,
+            self.generate_device,
             self.orig_module.scaling_factor,
             self.orig_module.original_max_position_embeddings,
             self.orig_module.beta_fast,

ktransformers/operators/attention.py

@@ -15,7 +15,7 @@ from ktransformers.util.custom_gguf import GGUFLoader
 from transformers.configuration_utils import PretrainedConfig
 from transformers.cache_utils import Cache
 
-class DeepseekV2AttentionInjected(BaseInjectedModule, DeepseekV2Attention):
+class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
     def __init__(self,

ktransformers/operators/cpuinfer.py

+import sys, os
+from typing import Any
+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 ktransformers.server.config.config import Config
+class CPUInfer:
+    cpu_infer = None
+    def __init__(self, cpu_infer:int = Config().cpu_infer):
+        if CPUInfer.cpu_infer is None:
+            CPUInfer.cpu_infer = cpuinfer_ext.CPUInfer(cpu_infer)
+        
+    def __getattribute__(self, __name: str) -> Any:
+        return CPUInfer.cpu_infer.__getattribute__(__name)
+    
+    def __setattr__(self, __name: str, __value: Any) -> None:
+        return CPUInfer.cpu_infer.__setattr__(__name, __value)
\ No newline at end of file

ktransformers/operators/experts.py

@@ -6,7 +6,7 @@ Author       : Azure-Tang, Boxin Zhang, chenht2022
 Date         : 2024-07-25 11:25:24
 Version      : 0.1.0
 LastEditors  : Azure 
-LastEditTime : 2024-07-26 09:27:41
+LastEditTime : 2024-08-15 02:36:29
 Copyright (c) 2024 by KVCache.AI, All Rights Reserved. 
 '''
 
@@ -31,12 +31,13 @@ from ktransformers.server.config.config import Config
 from transformers.activations import ACT2FN
 from transformers.configuration_utils import PretrainedConfig
 from abc import ABC, abstractmethod
-from ktransformers.operators.linear import QuantizedLinearMarlin, QuantizedLinearTorch, KTransformerLinear
+from ktransformers.operators.linear import KLinearMarlin, KLinearTorch, KTransformersLinear
 import time
+from ktransformers.operators.cpuinfer import CPUInfer
 
 
 # class Base(BaseInjectedModule, ABC):
-class MLPExpertsBase(ABC):
+class KExpertsBase(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)
         self.key = key
@@ -80,6 +81,25 @@ class MLPExpertsBase(ABC):
                 gate_type = self.gguf_loader.tensor_info[key + ".ffn_gate_exps.weight"]["ggml_type"]
                 up_type = self.gguf_loader.tensor_info[key + ".ffn_up_exps.weight"]["ggml_type"]
                 down_type = self.gguf_loader.tensor_info[key + ".ffn_down_exps.weight"]["ggml_type"]
+            elif key + ".ffn_down.0.weight" in self.gguf_loader.tensor_info:
+                # for supporting  Mixtral-8x7B-Instuct  
+                gate = []
+                up = []
+                down = []
+                for i in range(8):
+                    gatei, upi, downi = f".ffn_gate.{i}.weight", f".ffn_up.{i}.weight", f".ffn_down.{i}.weight"
+                    targets = [gatei, upi, downi]
+                    tensors = self.load_multi(key, targets, device=device)
+                    gate_it, up_it, down_it = tensors[gatei], tensors[upi], tensors[downi]
+                    gate.append(gate_it)
+                    up.append(up_it)
+                    down.append(down_it)
+                gate = torch.stack(gate)
+                up = torch.stack(up)
+                down = torch.stack(down)
+                gate_type = self.gguf_loader.tensor_info[key + ".ffn_gate.0.weight"]["ggml_type"]
+                up_type = self.gguf_loader.tensor_info[key + ".ffn_up.0.weight"]["ggml_type"]
+                down_type = self.gguf_loader.tensor_info[key + ".ffn_down.0.weight"]["ggml_type"]
             else:
                 raise ValueError(f"Experts {key} not found in gguf_loader")
             res = {key:{"gate": gate, "up": up, "down": down, "gate_type": gate_type, "up_type": up_type, "down_type": down_type}}
@@ -91,13 +111,14 @@ class MLPExpertsBase(ABC):
             tensors[k] = self.gguf_loader.load_gguf_tensor(key + k, device=device)
         return tensors
 
-class MLPCPUExperts(MLPExpertsBase):
+class KExpertsCPU(KExpertsBase):
     input_tensor_cpu:Tensor = None
     expert_ids_cpu:Tensor = None
     weights_cpu:Tensor = None
     output_cpu:Tensor = None
-    output_gpu:Tensor = None
-    CPU_INFER = cpuinfer_ext.CPUInfer(Config().cpu_infer)
+    output_gpu_map:dict = {} # Manage output tensor buffer on different gpu
+    #stream_map:dict = {} # Manage cuda stream on different gpu
+    CPU_INFER = CPUInfer(Config().cpu_infer)
     def __init__(
         self,
         key: str,
@@ -106,17 +127,17 @@ class MLPCPUExperts(MLPExpertsBase):
         n_routed_experts: int,
         orig_module: nn.Module = None,
         device: str = "cpu",
-        out_device: str = "cuda", # this device mean which device the output should on 
+        out_device: str = "cuda", # this device mean which device the output should on. TODO: support cpu.
         **kwargs
     ):
         super().__init__(key, gguf_loader, config, orig_module, device, **kwargs)
-        assert device.lower() == "cpu", "MLPCPUExperts can only be loaded on CPU"
+        assert device.lower() == "cpu", "KExpertsCPU can only be loaded on CPU"
         self.n_routed_experts = n_routed_experts
         self.out_device = out_device
 
     def load(self, w: dict | nn.Parameter | tuple | None = None, device:str|None = None, warmup:bool = False):
         if device:
-            assert device.lower() == "cpu", "MLPCPUExperts can only be loaded on CPU, Parameter \"device\" can be cpu or None."
+            assert device.lower() == "cpu", "KExpertsCPU can only be loaded on CPU, Parameter \"device\" can be cpu or None."
         if w is None: w = self.load_weights()[self.key]
         self.gate = w["gate"]
         self.up = w["up"]
@@ -155,50 +176,50 @@ class MLPCPUExperts(MLPExpertsBase):
         # print(n_routed_experts, hidden_size, moe_intermediate_size)
         num_experts_per_tok = self.config.num_experts_per_tok
         self.moe = MOE(moe_config)
-        self.cpu_infer = MLPCPUExperts.CPU_INFER
+        self.cpu_infer = KExpertsCPU.CPU_INFER
         if warmup:
-            self.cpu_infer.submit(self.moe.warm_up)
+            self.cpu_infer.submit(self.moe.warm_up())
             self.cpu_infer.sync()
-        if MLPCPUExperts.output_gpu == None:
-            MLPCPUExperts.input_tensor_cpu = torch.empty((self.config.hidden_size), device="cpu", pin_memory=True)
-            MLPCPUExperts.expert_ids_cpu = torch.empty((num_experts_per_tok), device="cpu", dtype=torch.long, pin_memory=True)
-            MLPCPUExperts.weights_cpu = torch.empty((num_experts_per_tok), device="cpu", dtype=torch.float32, pin_memory=True)
-            MLPCPUExperts.output_cpu = torch.empty((self.config.hidden_size), device="cpu", pin_memory=True)
-            MLPCPUExperts.output_gpu = torch.empty((self.config.hidden_size), device=self.out_device)
-
+        if self.out_device not in KExpertsCPU.output_gpu_map:
+            KExpertsCPU.output_gpu_map[self.out_device] = torch.zeros((self.config.hidden_size), device=self.out_device)
+        if KExpertsCPU.input_tensor_cpu == None:
+            KExpertsCPU.input_tensor_cpu = torch.zeros((self.config.hidden_size), device="cpu", pin_memory=True)
+            KExpertsCPU.expert_ids_cpu = torch.zeros((num_experts_per_tok), device="cpu", dtype=torch.long, pin_memory=True)
+            KExpertsCPU.weights_cpu = torch.zeros((num_experts_per_tok), device="cpu", dtype=torch.float32, pin_memory=True)
+            KExpertsCPU.output_cpu = torch.zeros((self.config.hidden_size), device="cpu", pin_memory=True, dtype=torch.bfloat16)
+            
     def submit_for_one_decode(self, input_tensor, expert_ids, weights):
-        MLPCPUExperts.input_tensor_cpu.copy_(input_tensor, non_blocking=True)
-        MLPCPUExperts.expert_ids_cpu.copy_(expert_ids, non_blocking=True)
-        MLPCPUExperts.weights_cpu.copy_(weights, non_blocking=True)
-        self.cpu_infer.submit_with_cuda_stream(torch.cuda.current_stream().cuda_stream, self.moe.forward, 1, expert_ids.size(0), MLPCPUExperts.expert_ids_cpu.data_ptr(), MLPCPUExperts.weights_cpu.data_ptr(), MLPCPUExperts.input_tensor_cpu.data_ptr(), MLPCPUExperts.output_cpu.data_ptr())
-    
+        KExpertsCPU.input_tensor_cpu.copy_(input_tensor, non_blocking=True)
+        KExpertsCPU.expert_ids_cpu.copy_(expert_ids, non_blocking=True)
+        KExpertsCPU.weights_cpu.copy_(weights, non_blocking=True)
+        self.cpu_infer.submit_with_cuda_stream(torch.cuda.current_stream(self.out_device).cuda_stream, self.moe.forward(1, expert_ids.size(0), KExpertsCPU.expert_ids_cpu.data_ptr(), KExpertsCPU.weights_cpu.data_ptr(), KExpertsCPU.input_tensor_cpu.data_ptr(), KExpertsCPU.output_cpu.data_ptr()))
+        
     def sync_for_one_decode(self):
-        self.cpu_infer.sync_with_cuda_stream(torch.cuda.current_stream().cuda_stream)
-        MLPCPUExperts.output_gpu.copy_(MLPCPUExperts.output_cpu, non_blocking=True)
-        #print("capturing experts finish")
-        return MLPCPUExperts.output_gpu
+        self.cpu_infer.sync_with_cuda_stream(torch.cuda.current_stream(self.out_device).cuda_stream)
+        KExpertsCPU.output_gpu_map[self.out_device].copy_(KExpertsCPU.output_cpu, non_blocking=True)
+        return KExpertsCPU.output_gpu_map[self.out_device]
 
     def forward(self, input_tensor, expert_ids, weights):
         # generate, capture and run cuda graph
+        # print(expert_ids)
         if input_tensor.size(0)==1:
             # TODO: this branch is unreachable, but the shape of input_tensor([1,hidden_size]) and input_tensor_cpu([hidden_size]) is not compatible
             #print("capturing experts")
-            MLPCPUExperts.input_tensor_cpu.copy_(input_tensor, non_blocking=True)
-            MLPCPUExperts.expert_ids_cpu.copy_(expert_ids, non_blocking=True)
-            MLPCPUExperts.weights_cpu.copy_(weights, non_blocking=True)
-            self.cpu_infer.submit_with_cuda_stream(torch.cuda.current_stream().cuda_stream, self.moe.forward, 1, expert_ids.size(1), MLPCPUExperts.expert_ids_cpu.data_ptr(), MLPCPUExperts.weights_cpu.data_ptr(), MLPCPUExperts.input_tensor_cpu.data_ptr(), MLPCPUExperts.output_cpu.data_ptr())
+            KExpertsCPU.input_tensor_cpu.copy_(input_tensor, non_blocking=True)
+            KExpertsCPU.expert_ids_cpu.copy_(expert_ids, non_blocking=True)
+            KExpertsCPU.weights_cpu.copy_(weights, non_blocking=True)
+            self.cpu_infer.submit_with_cuda_stream(torch.cuda.current_stream().cuda_stream, self.moe.forward(1, expert_ids.size(1), KExpertsCPU.expert_ids_cpu.data_ptr(), KExpertsCPU.weights_cpu.data_ptr(), KExpertsCPU.input_tensor_cpu.data_ptr(), KExpertsCPU.output_cpu.data_ptr()))
             self.cpu_infer.sync_with_cuda_stream(torch.cuda.current_stream().cuda_stream)
-            MLPCPUExperts.output_gpu.copy_(MLPCPUExperts.output_cpu, non_blocking=True)
-            #print("capturing experts finish")
-            return MLPCPUExperts.output_gpu
+            KExpertsCPU.output_gpu_map[self.out_device].copy_(KExpertsCPU.output_cpu, non_blocking=True)
+            return KExpertsCPU.output_gpu_map[self.out_device]
         else:
             input_tensor = input_tensor.contiguous().cpu()
             expert_ids = expert_ids.contiguous().cpu()
             weights = weights.contiguous().to(torch.float32).cpu()
             output = torch.empty_like(input_tensor).contiguous()
-            self.cpu_infer.submit(self.moe.forward, expert_ids.size(0), expert_ids.size(1), expert_ids.data_ptr(), weights.data_ptr(), input_tensor.data_ptr(), output.data_ptr())
+            self.cpu_infer.submit(self.moe.forward(expert_ids.size(0), expert_ids.size(1), expert_ids.data_ptr(), weights.data_ptr(), input_tensor.data_ptr(), output.data_ptr()))
             self.cpu_infer.sync()
-            return output.to(device=object.__getattribute__(self, "device"))
+            return output.to(device=object.__getattribute__(self, "out_device"))
     
     def unload(self):
         return
@@ -225,12 +246,30 @@ class MLPCPUExperts(MLPExpertsBase):
                 gate_type = self.gguf_loader.tensor_info[key + ".ffn_gate_exps.weight"]["ggml_type"]
                 up_type = self.gguf_loader.tensor_info[key + ".ffn_up_exps.weight"]["ggml_type"]
                 down_type = self.gguf_loader.tensor_info[key + ".ffn_down_exps.weight"]["ggml_type"]
+            elif key + ".ffn_down.0.weight" in self.gguf_loader.tensor_info:
+                # for supporting  Mixtral-8x7B-Instuct  
+                gate = []
+                up = []
+                down = []
+                for i in range(8):
+                    gate_it = self.gguf_loader.get_mmap_tensor(f"{key}.ffn_gate.{i}.weight")
+                    up_it = self.gguf_loader.get_mmap_tensor(f"{key}.ffn_up.{i}.weight")
+                    down_it = self.gguf_loader.get_mmap_tensor(f"{key}.ffn_down.{i}.weight")
+                    gate.append(gate_it)
+                    up.append(up_it)
+                    down.append(down_it)
+                gate = np.stack(gate)
+                up = np.stack(up)
+                down = np.stack(down)
+                gate_type = self.gguf_loader.tensor_info[key + ".ffn_gate.0.weight"]["ggml_type"]
+                up_type = self.gguf_loader.tensor_info[key + ".ffn_up.0.weight"]["ggml_type"]
+                down_type = self.gguf_loader.tensor_info[key + ".ffn_down.0.weight"]["ggml_type"]
             else:
                 raise ValueError(f"Experts {key} not found in gguf_loader")
             res = {key:{"gate": gate, "up": up, "down": down, "gate_type": gate_type, "up_type": up_type, "down_type": down_type}}
         return res
     
-class MLPExpertsMarlin(MLPExpertsBase):
+class KExpertsMarlin(KExpertsBase):
     expert_num: int
     loaded_experts_idx: list[int]
     def __init__(
@@ -251,11 +290,11 @@ class MLPExpertsMarlin(MLPExpertsBase):
         self.device = device
         # create empty marlin experts according to the number of experts per token
         # up
-        self.up_projs = [QuantizedLinearMarlin(key+ "." + "ffn_up_exps", gguf_loader, config, device=device) for i in range(self.expert_num)]
+        self.up_projs = [KLinearMarlin(key+ "." + "ffn_up_exps", gguf_loader, config, device=device) for i in range(self.expert_num)]
         # gate
-        self.gate_projs = [QuantizedLinearMarlin(key+ "." + "ffn_gate_exps", gguf_loader, config, device=device) for i in range(self.expert_num)]
+        self.gate_projs = [KLinearMarlin(key+ "." + "ffn_gate_exps", gguf_loader, config, device=device) for i in range(self.expert_num)]
         # down
-        self.down_projs = [QuantizedLinearMarlin(key+ "." + "ffn_down_exps", gguf_loader, config, device=device) for i in range(self.expert_num)]
+        self.down_projs = [KLinearMarlin(key+ "." + "ffn_down_exps", gguf_loader, config, device=device) for i in range(self.expert_num)]
 
     def load(self, w: dict | nn.Parameter | tuple | None = None, device: str | None = None, warmup: bool = False):
         if device is None: device = self.device
@@ -302,7 +341,7 @@ class MLPExpertsMarlin(MLPExpertsBase):
                 gate_type = self.gguf_loader.tensor_info[key + ".ffn_gate_exps.weight"]["ggml_type"]
                 up_type = self.gguf_loader.tensor_info[key + ".ffn_up_exps.weight"]["ggml_type"]
                 down_type = self.gguf_loader.tensor_info[key + ".ffn_down_exps.weight"]["ggml_type"]
-                # tensors = self.load_multi(key, [".ffn_gate_exps.weight", ".ffn_up_exps.weight", ".ffn_down_exps.weight"])    
+                # tensors = self.load_multi(key, [".ffn_gate_exps.weight", ".ffn_up_exps.weight", ".ffn_down_exps.weight"])
             res = {key:{"gate": gate, "up": up, "down": down, "gate_type": gate_type, "up_type": up_type, "down_type": down_type}}
         return res
 
@@ -320,7 +359,7 @@ class MLPExpertsMarlin(MLPExpertsBase):
         outs = outs.to(device)
         return outs
 
-class MLPExpertsTorch(MLPExpertsBase):
+class KExpertsTorch(KExpertsBase):
     expert_num: int
     loaded_experts_idx: list[int]
     gate: torch.Tensor
@@ -362,12 +401,12 @@ class MLPExpertsTorch(MLPExpertsBase):
             self.down = None
 
     def forward(self, hidden_states_cpu: torch.Tensor, selected_experts_cpu: torch.Tensor, routing_weights_cpu: torch.Tensor) -> torch.Tensor:
-        # TODO: forward should transfer data to gpu, and make the data transfering capturable using pin memory, 
-        # just like CPUInfer MLPCPUExperts. There may be a base class of experts on cpu
-        hidden_states_cpu = hidden_states_cpu.to("cpu")
-        selected_experts_cpu = selected_experts_cpu.to("cpu")
-        routing_weights_cpu = routing_weights_cpu.to("cpu")
 
+        org_device = hidden_states_cpu.device
+        hidden_states_cpu = hidden_states_cpu.to(self.device)
+        selected_experts_cpu = selected_experts_cpu.to(self.device)
+        routing_weights_cpu = routing_weights_cpu.to(self.device)
+        
         batch_sequence_length, hidden_dim = hidden_states_cpu.size()
 
         final_hidden_states = torch.zeros(
@@ -396,37 +435,39 @@ class MLPExpertsTorch(MLPExpertsBase):
             # the `top_x` tensor here.
             final_hidden_states.index_add_(0, top_x, current_hidden_states)
 
-        return final_hidden_states.to(org_dtype)
+
+        return final_hidden_states.to(org_dtype, device=org_device)
 
 EXPERTS_MAP = {
-    "MLPCPUExperts": MLPCPUExperts,
-    "MLPExpertsTorch": MLPExpertsTorch,
-    "MLPExpertsMarlin": MLPExpertsMarlin,
+    "KExpertsCPU": KExpertsCPU,
+    "KExpertsTorch": KExpertsTorch,
+    "KExpertsMarlin": KExpertsMarlin,
 }
-class KTransformersMLPExpert(BaseInjectedModule, MLPExpertsBase):
+
+class KTransformersExperts(BaseInjectedModule, KExpertsBase):
     def __init__(self,
                  key: str,
                  gguf_loader: GGUFLoader,
                  config: PretrainedConfig,
                  orig_module: nn.Module,
-                 device: str = "cuda",
+                #  device: str = "cuda",
                  prefill_device:str = "cuda",
-                 prefill_mlp_type: str | None = "MLPExpertsTorch",
+                 prefill_op: str | None = "KExpertsTorch",
                  generate_device: str = "cpu",
-                 generate_mlp_type: str | None = "MLPCPUExperts",
+                 generate_op: str | None = "KExpertsCPU",
                  **kwargs):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
-        MLPExpertsBase.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
-        if generate_mlp_type is not None:
-            self.generate_experts = EXPERTS_MAP[generate_mlp_type](key, gguf_loader, config, len(orig_module), device=generate_device, **kwargs)
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
+        KExpertsBase.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
+        if generate_op is not None:
+            self.generate_experts = EXPERTS_MAP[generate_op](key, gguf_loader, config, len(orig_module), device=generate_device, **kwargs)
         else:
             self.generate_experts = None
-        if prefill_mlp_type is not None:
-            self.prefill_experts = EXPERTS_MAP[prefill_mlp_type](key, gguf_loader, config, len(orig_module), device=prefill_device, **kwargs)
+        if prefill_op is not None:
+            self.prefill_experts = EXPERTS_MAP[prefill_op](key, gguf_loader, config, len(orig_module), device=prefill_device, **kwargs)
         else:
             self.prefill_experts = None
-        self.gpu_mlp_type = prefill_mlp_type
-        self.cpu_mlp_type = generate_mlp_type
+        self.gpu_mlp_type = prefill_op
+        self.cpu_mlp_type = generate_op
         self.mode = InferenceState.UNLOAD
 
     def load(self, w: dict = None,  mode: InferenceState = None, warmup: bool = True):
@@ -479,9 +520,10 @@ class KTransformersMLPExpert(BaseInjectedModule, MLPExpertsBase):
 
 from ktransformers.models.modeling_deepseek import DeepseekV2MoE
 from ktransformers.models.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+from ktransformers.models.modeling_mixtral import MixtralSparseMoeBlock
 
 
-class Qwen2MoeSparseMoeBlockInjected(BaseInjectedModule, Qwen2MoeSparseMoeBlock):
+class KQwen2MoeSparseMoeBlock(BaseInjectedModule, Qwen2MoeSparseMoeBlock):
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         """ """
         orig_shape = hidden_states.shape
@@ -506,16 +548,16 @@ class Qwen2MoeSparseMoeBlockInjected(BaseInjectedModule, Qwen2MoeSparseMoeBlock)
             y.resize_(*orig_shape)
             return y, router_logits
         
-        hidden_states_expert = hidden_states.to(self.experts.device)  if isinstance(self.experts, MLPExpertsBase) else hidden_states_expert.cpu()
-        selected_experts_expert = selected_experts.to(self.experts.device) if isinstance(self.experts, MLPExpertsBase) else selected_experts_expert.cpu()
-        routing_weights_expert = routing_weights.to(self.experts.device) if isinstance(self.experts, MLPExpertsBase) else routing_weights_expert.cpu()
+        hidden_states_expert = hidden_states.to(self.experts.device)  if isinstance(self.experts, KExpertsBase) else hidden_states_expert.cpu()
+        selected_experts_expert = selected_experts.to(self.experts.device) if isinstance(self.experts, KExpertsBase) else selected_experts_expert.cpu()
+        routing_weights_expert = routing_weights.to(self.experts.device) if isinstance(self.experts, KExpertsBase) else routing_weights_expert.cpu()
 
         shared_expert_output = self.shared_expert(hidden_states)
         shared_expert_output = (
             F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_expert_output
         )
 
-        if isinstance(self.experts, MLPExpertsBase):
+        if isinstance(self.experts, KExpertsBase):
             y = (
                 self.moe_on_cpuinfer(
                     hidden_states_expert, selected_experts_expert, routing_weights_expert
@@ -586,8 +628,7 @@ class Qwen2MoeSparseMoeBlockInjected(BaseInjectedModule, Qwen2MoeSparseMoeBlock)
 
         return final_hidden_states
 
-
-class DeepseekV2MoEInjected(BaseInjectedModule, DeepseekV2MoE):
+class KDeepseekV2MoE(BaseInjectedModule, DeepseekV2MoE):
     def forward(self, hidden_states):
         identity = hidden_states
         orig_shape = hidden_states.shape
@@ -607,7 +648,7 @@ class DeepseekV2MoEInjected(BaseInjectedModule, DeepseekV2MoE):
         if self.config.n_shared_experts is not None:
             y_ = self.shared_experts(identity).squeeze(0)
             
-        if isinstance(self.experts, MLPExpertsBase):
+        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
@@ -685,3 +726,102 @@ class DeepseekV2MoEInjected(BaseInjectedModule, DeepseekV2MoE):
             .type(new_x.dtype)
         )
         return final_out
+
+class KMisrtalSparseMoEBlock(BaseInjectedModule, MixtralSparseMoeBlock):
+    
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        orig_shape = hidden_states.shape
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        if self.training and self.jitter_noise > 0:
+            hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise)
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+        
+        if sequence_length == 1 and hasattr(self.experts.generate_experts, "submit_for_one_decode"):
+            self.experts.generate_experts.submit_for_one_decode(hidden_states[0], selected_experts[0], routing_weights[0])
+            y = self.experts.generate_experts.sync_for_one_decode().unsqueeze(0)
+            y.resize_(*orig_shape)
+            return y, router_logits
+        
+        hidden_states_expert = hidden_states.to(self.experts.device)  if isinstance(self.experts, KExpertsBase) else hidden_states_expert.cpu()
+        selected_experts_expert = selected_experts.to(self.experts.device) if isinstance(self.experts, KExpertsBase) else selected_experts_expert.cpu()
+        routing_weights_expert = routing_weights.to(self.experts.device) if isinstance(self.experts, KExpertsBase) else routing_weights_expert.cpu()
+
+        if isinstance(self.experts, KExpertsBase):
+            y = (
+                self.moe_on_cpuinfer(
+                    hidden_states_expert, selected_experts_expert, routing_weights_expert
+                )
+                .view(*orig_shape)
+                .to(device=hidden_states.device)
+            )
+        elif hidden_states_expert.size(0) > 10:
+            y = self.moe_infer(
+                hidden_states_expert, selected_experts_expert, routing_weights_expert, orig_shape
+            ).to(device=hidden_states.device)
+        else:
+            y = self.moe_infer_simple(
+                hidden_states_expert, selected_experts_expert, routing_weights_expert
+            ).to(device=hidden_states.device)
+            
+        y.resize_(*orig_shape)
+        return y, router_logits
+    
+    @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, hidden_states_cpu: torch.Tensor, selected_experts_cpu: torch.Tensor, routing_weights_cpu: torch.Tensor) -> torch.Tensor:
+        '''
+        hidden_states_cpu: [num_tokens, hidden_size]
+        topk_ids, topk_weight: [num_tokens, num_selected_experts]
+        '''
+        outs = torch.zeros_like(hidden_states_cpu)
+        for token_idx in range(selected_experts_cpu.size(0)):
+            for expert_idx in range(selected_experts_cpu.size(1)):
+                expert = self.experts[selected_experts_cpu[token_idx, expert_idx]]
+                outs[token_idx] += expert.forward(hidden_states_cpu[token_idx]) * routing_weights_cpu[token_idx, expert_idx]
+        return outs
+    
+    @torch.no_grad()
+    # TODO may bugs here
+    def moe_infer(self, hidden_states_cpu: torch.Tensor, selected_experts_cpu: torch.Tensor, routing_weights_cpu: torch.Tensor, orig_shape: tuple) -> torch.Tensor:
+        
+        batch_size, sequence_length, hidden_dim = orig_shape
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states_cpu.dtype, device=hidden_states_cpu.device
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts_cpu, num_classes=self.num_experts).permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states_cpu[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer.forward(current_state) * routing_weights_cpu[top_x, idx, None]
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states_cpu.dtype))
+
+        return final_hidden_states
\ No newline at end of file

ktransformers/operators/linear.py

@@ -6,13 +6,14 @@ Author       : Azure-Tang, Boxin Zhang
 Date         : 2024-07-25 11:25:24
 Version      : 0.1.0
 LastEditors  : Azure 
-LastEditTime : 2024-07-26 09:27:53
+LastEditTime : 2024-08-14 14:57:04
 Copyright (c) 2024 by KVCache.AI, All Rights Reserved. 
 '''
 
 
+import ctypes
 import torch
-from torch import nn
+from torch import Tensor, nn
 import KTransformersOps 
 from ktransformers.util.custom_gguf import GGUFLoader
 from ktransformers.util.utils import InferenceState
@@ -25,10 +26,16 @@ from ktransformers.ktransformers_ext.operators.custom_marlin.quantize.utils.marl
 from ktransformers.operators.base_operator import BaseInjectedModule
 from transformers.configuration_utils import PretrainedConfig
 from abc import ABC, abstractmethod
+import sys, os
+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 ktransformers.operators.cpuinfer import CPUInfer
+from ktransformers.server.config.config import Config
 
-
-#class QuantizedLinearBase(BaseInjectedModule, ABC):
-class QuantizedLinearBase(ABC):
+#class KLinearBase(BaseInjectedModule, ABC):
+class KLinearBase(ABC):
     def __init__(
         self,
         key: str,
@@ -99,7 +106,7 @@ class QuantizedLinearBase(ABC):
         pass
 
 
-class QuantizedLinearTorch(QuantizedLinearBase):
+class KLinearTorch(KLinearBase):
     def __init__(
         self,
         key: str,
@@ -118,6 +125,7 @@ class QuantizedLinearTorch(QuantizedLinearBase):
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         dtype = x.dtype
         out_device = x.device
+        # TODO: support CUDA Graph when using cpu, but CPUInfer is recommended.
         x = x.to(device=self.device, dtype=self.dtype)
         x = x @ self.w
         if self.has_bias:
@@ -128,7 +136,7 @@ class QuantizedLinearTorch(QuantizedLinearBase):
     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_weight(device=device)
-
+        
         if isinstance(w, nn.Parameter):
             self.w = w.to(dtype=self.dtype).view(self.out_features, self.in_features).T
             self.has_bias = False
@@ -150,7 +158,7 @@ class QuantizedLinearTorch(QuantizedLinearBase):
             self.bias = None
 
 
-class QuantizedLinearMarlin(QuantizedLinearBase):
+class KLinearMarlin(KLinearBase):
     marlin_q_w: torch.Tensor
     marlin_s: torch.Tensor
     g_idx: torch.Tensor
@@ -176,7 +184,7 @@ class QuantizedLinearMarlin(QuantizedLinearBase):
         self.act_order = act_order
         self.is_k_full = is_k_full
 
-    def load(self, w: dict | nn.Parameter | tuple | None = None, device: str|None = "cuda"):
+    def load(self, w: dict | nn.Parameter | tuple | None = None, device: str|None = None):
         if device is None: device = self.device
         assert device.lower() != "cpu", "Marlin quantized linear only supports GPU device"
         if w is None: w = self.load_weight(device=device)
@@ -200,7 +208,7 @@ class QuantizedLinearMarlin(QuantizedLinearBase):
             weight, self.num_bits, self.group_size, self.act_order
         )
         self.workspace = MarlinWorkspace(
-            self.out_features, GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL
+            self.out_features, GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL,self.device
         )
         self.marlin_q_w = marlin_q_w
         self.marlin_s = marlin_s
@@ -243,36 +251,138 @@ class QuantizedLinearMarlin(QuantizedLinearBase):
         self.g_idx = None
         self.sort_indices = None
         self.workspace = None
-    
+
+class KLinearCPUInfer(KLinearBase):
+    CPU_INFER = CPUInfer(Config().cpu_infer)
+    def __init__(
+        self,
+        key: str,
+        gguf_loader: GGUFLoader,
+        config: PretrainedConfig,
+        orig_module: nn.Module = None,
+        device: str = "cpu",
+        out_device: str = "cuda", # this device mean which device the output should on. TODO: support cpu.
+        stride = 16,
+        group_max_len = 1024,
+        **kwargs,
+    ):
+        super().__init__(key, gguf_loader, config, orig_module, device, **kwargs)
+        self.has_bias = False
+        self.dtype = torch.get_default_dtype()
+        self.w = None
+        self.has_bias = False
+        self.stride = stride
+        self.group_max_len = group_max_len
+        self.out_device = out_device
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        origin_shape = x.shape # [batch_size, q_len, hidden_size]
+        if origin_shape[1] == 1:
+            out_device = x.device
+            self.input_tensor_cpu.copy_(x, non_blocking=True)
+            qlen = origin_shape[1]
+            KLinearCPUInfer.CPU_INFER.submit_with_cuda_stream(
+                torch.cuda.current_stream().cuda_stream,
+                self.linear.forward(
+                    qlen, 
+                    self.input_tensor_cpu.data_ptr(), 
+                    self.output_cpu.data_ptr()
+                )
+            )
+            KLinearCPUInfer.CPU_INFER.sync_with_cuda_stream(torch.cuda.current_stream().cuda_stream)
+            self.output_gpu.copy_(self.output_cpu, non_blocking=True)
+            if self.has_bias:
+                self.output_gpu += self.bias
+            return self.output_gpu
+        else:
+            dtype = x.dtype
+            out_device = x.device
+            x = x.to(device=self.device)
+            qlen = origin_shape[1]
+            output_shape = (*origin_shape[:-1], self.out_features)
+            output = torch.empty(output_shape, device=x.device, dtype=x.dtype)
+            KLinearCPUInfer.CPU_INFER.submit(
+                self.linear.forward(
+                    qlen, 
+                    x.data_ptr(), 
+                    output.data_ptr()
+                )
+            )
+            KLinearCPUInfer.CPU_INFER.sync()
+            if self.has_bias:
+                output = output + self.bias
+            output = output.to(dtype=dtype, device=out_device)
+            return output
+
+    def load(self, w: dict | nn.Parameter | tuple | None = None, device: str|None = None, warmup:bool = True):
+        print(f"loading {self.key} to {self.device} using CPUInfer")
+        if device is None: device = self.device
+        self.load_weights(w=w, device=device)
+        if self.bias is not None:
+            self.has_bias = True
+            self.bias = self.bias.to(device)
+            
+        weight_ptr = ctypes.addressof(
+            ctypes.cast(self.weight.ctypes.data, ctypes.POINTER(ctypes.c_uint64)).contents
+        )
+        config = cpuinfer_ext.linear.LinearConfig(self.in_features, self.out_features, self.stride, self.group_max_len, weight_ptr, self.weight_type, 30)
+        self.linear = cpuinfer_ext.linear.Linear(config)
+        
+        if warmup:
+            KLinearCPUInfer.CPU_INFER.submit(self.linear.warm_up())
+            KLinearCPUInfer.CPU_INFER.sync()
+        self.input_tensor_cpu = torch.zeros((1, 1, self.in_features), device="cpu", pin_memory=True)
+        self.output_cpu = torch.zeros((1, 1, self.out_features), device="cpu", pin_memory=True, dtype=torch.bfloat16)
+        self.output_gpu = torch.zeros((1, 1, self.out_features), device=self.out_device)
+
+    def load_weights(self, w: dict | nn.Parameter | tuple | None = None, device: str = "cpu"):
+        if self.key + ".weight" in self.gguf_loader.tensor_info:
+            if self.key + ".bias" in self.gguf_loader.tensor_file_map:
+                self.weight = self.gguf_loader.get_mmap_tensor(self.key + ".weight")
+                self.weight_type = self.gguf_loader.tensor_info[self.key + ".weight"]["ggml_type"]
+                self.bias = self.gguf_loader.load_gguf_tensor(self.key + ".bias", device=device)
+            else:
+                self.weight = self.gguf_loader.get_mmap_tensor(self.key + ".weight")
+                self.weight_type = self.gguf_loader.tensor_info[self.key + ".weight"]["ggml_type"]
+                self.bias = None
+        else:
+            raise ValueError(f"Linear {self.key} not found in gguf_loader")
+
+    def unload(self):
+        if self.w is not None:
+            self.w = None
+        if self.has_bias:
+            self.bias = None        
+
 LINEAR_MAP = {
-    "QuantizedLinearMarlin": QuantizedLinearMarlin,
-    "QuantizedLinearTorch": QuantizedLinearTorch,
-    "QuantizedLinearTorch": QuantizedLinearTorch,
+    "KLinearMarlin": KLinearMarlin,
+    "KLinearTorch": KLinearTorch,
+    "KLinearCPUInfer": KLinearCPUInfer
 }
 
-class KTransformerLinear(BaseInjectedModule, QuantizedLinearBase):
+class KTransformersLinear(BaseInjectedModule, KLinearBase):
     def __init__(
         self,
         key: str,
         gguf_loader: GGUFLoader,
         config: PretrainedConfig,
         orig_module: nn.Module,
-        device: str = "cuda",
+        # device: str = "cuda",
         generate_device: str = "cuda",
-        generate_op: str| None = "QuantizedLinearMarlin",
+        generate_op: str| None = "KLinearMarlin",
         prefill_device: str = "cuda",
-        prefill_op: str| None = "QuantizedLinearTorch",
+        prefill_op: str| None = "KLinearTorch",
         **kwargs,
     ):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
-        QuantizedLinearBase.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
+        KLinearBase.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
         # build all the linear operators
         if prefill_op is not None:
             assert prefill_op in LINEAR_MAP, f"linear_type {prefill_op} not supported"
-            if prefill_op == "QuantizedLinearMarlin" and (orig_module.in_features%GPTQ_MARLIN_MIN_THREAD_N!=0 or orig_module.out_features%GPTQ_MARLIN_MIN_THREAD_N!=0):
-                print(f"This linear module's in_features or out_features is not divisible by GPTQ_MARLIN_MIN_THREAD_N({GPTQ_MARLIN_MIN_THREAD_N}), using QuantizedLinearTorch instead.")
+            if prefill_op == "KLinearMarlin" and (orig_module.in_features%GPTQ_MARLIN_MIN_THREAD_N!=0 or orig_module.out_features%GPTQ_MARLIN_MIN_THREAD_N!=0):
+                print(f"This linear module's in_features or out_features is not divisible by GPTQ_MARLIN_MIN_THREAD_N({GPTQ_MARLIN_MIN_THREAD_N}), using KLinearTorch instead.")
                 print(f"module info: key:{key} orig_module:{orig_module}")
-                self.prefill_linear = QuantizedLinearTorch(key, gguf_loader, config, orig_module, prefill_device, **kwargs)
+                self.prefill_linear = KLinearTorch(key, gguf_loader, config, orig_module, prefill_device, **kwargs)
             else:
                 self.prefill_linear = LINEAR_MAP[prefill_op](key, gguf_loader, config, orig_module, prefill_device, **kwargs)
         else:
@@ -280,16 +390,15 @@ class KTransformerLinear(BaseInjectedModule, QuantizedLinearBase):
 
         if generate_op is not None:
             assert generate_op in LINEAR_MAP, f"linear_type {generate_op} not supported"
-            if generate_op == "QuantizedLinearMarlin" and (orig_module.in_features%GPTQ_MARLIN_MIN_THREAD_N!=0 or orig_module.out_features%GPTQ_MARLIN_MIN_THREAD_N!=0):
-                print(f"This linear module's in_features or out_features is not divisible by GPTQ_MARLIN_MIN_THREAD_N({GPTQ_MARLIN_MIN_THREAD_N}), using QuantizedLinearTorch instead.")
+            if generate_op == "KLinearMarlin" and (orig_module.in_features%GPTQ_MARLIN_MIN_THREAD_N!=0 or orig_module.out_features%GPTQ_MARLIN_MIN_THREAD_N!=0):
+                print(f"This linear module's in_features or out_features is not divisible by GPTQ_MARLIN_MIN_THREAD_N({GPTQ_MARLIN_MIN_THREAD_N}), using KLinearTorch instead.")
                 print(f"module info: key:{key} orig_module:{orig_module}")
-                self.generate_op = "QuantizedLinearTorch"
-                self.generate_linear = QuantizedLinearTorch(key, gguf_loader, config, orig_module, generate_device, **kwargs)
+                self.generate_op = "KLinearTorch"
+                self.generate_linear = KLinearTorch(key, gguf_loader, config, orig_module, generate_device, **kwargs)
             else:
                 self.generate_linear = LINEAR_MAP[generate_op](key, gguf_loader, config, orig_module, generate_device, **kwargs)
         else:
             self.generate_linear = None
-        self.device = device
         self.mode = InferenceState.UNLOAD
 
     def forward(self, x):

ktransformers/operators/layer_wise_prefill.py → ktransformers/operators/models.py

@@ -6,7 +6,7 @@ Author       : Azure-Tang
 Date         : 2024-07-25 11:25:24
 Version      : 1.0.0
 LastEditors  : Azure 
-LastEditTime : 2024-07-26 09:27:48
+LastEditTime : 2024-08-14 14:53:05
 Copyright (c) 2024 by KVCache.AI, All Rights Reserved. 
 '''
 
@@ -45,6 +45,8 @@ from ktransformers.models.modeling_deepseek import BaseModelOutputWithPast, Deep
 from transformers.models.qwen2_moe.configuration_qwen2_moe import Qwen2MoeConfig
 from ktransformers.operators.base_operator import BaseInjectedModule
 from ktransformers.util.utils import InferenceState
+from ktransformers.util.custom_gguf import GGUFLoader
+from transformers.configuration_utils import PretrainedConfig
 
 if is_flash_attn_2_available():
     from flash_attn import flash_attn_func, flash_attn_varlen_func
@@ -73,34 +75,6 @@ QWEN2MOE_START_DOCSTRING = r"""
             [`~PreTrainedModel.from_pretrained`] method to load the model weights.
 """
 
-
-@add_start_docstrings(
-    "The bare Qwen2MoE Model outputting raw hidden-states without any specific head on top.",
-    QWEN2MOE_START_DOCSTRING,
-)
-class Qwen2MoePreTrainedModel(PreTrainedModel):
-    config_class = Qwen2MoeConfig
-    base_model_prefix = "model"
-    supports_gradient_checkpointing = True
-    _no_split_modules = ["Qwen2MoeDecoderLayer"]
-    _skip_keys_device_placement = "past_key_values"
-    _supports_flash_attn_2 = True
-    _supports_sdpa = True
-    _supports_cache_class = True
-    _supports_static_cache = 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_()
-
-
 QWEN2MOE_INPUTS_DOCSTRING = r"""
     Args:
         input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
@@ -177,13 +151,11 @@ QWEN2MOE_INPUTS_DOCSTRING = r"""
             the complete sequence length.
 """
 
-from ktransformers.util.custom_gguf import GGUFLoader
-from transformers.configuration_utils import PretrainedConfig
 @add_start_docstrings(
     "The bare Qwen2MoE Model outputting raw hidden-states without any specific head on top.",
     QWEN2MOE_START_DOCSTRING,
 )
-class Qwen2MoeModelPerLayerPrefill(BaseInjectedModule):
+class KQwen2MoeModel(BaseInjectedModule):
     """
     Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2MoeDecoderLayer`]
 
@@ -198,10 +170,13 @@ class Qwen2MoeModelPerLayerPrefill(BaseInjectedModule):
         orig_module: nn.Module,
         device: str = "cuda",
         per_layer_prefill_intput_threshold: int = 30000, # if None, no per-layer prefill
+        transfer_map: dict = None,
         **kwargs,
     ):
         BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
         self.per_layer_prefill_intput_threshold = per_layer_prefill_intput_threshold
+        self.transfer_map = transfer_map
+        self.stream_device_map = dict()
 
     @add_start_docstrings_to_model_forward(QWEN2MOE_INPUTS_DOCSTRING)
     def forward(
@@ -287,7 +262,20 @@ class Qwen2MoeModelPerLayerPrefill(BaseInjectedModule):
         all_router_logits = () if output_router_logits else None
         next_decoder_cache = None
 
-        for decoder_layer in self.layers:
+        for i, decoder_layer in enumerate(self.layers):
+            if self.transfer_map is not None and i in self.transfer_map: 
+                prev_stream = torch.cuda.current_stream()
+                cur_device = self.transfer_map[i]
+                if cur_device not in self.stream_device_map:
+                    self.stream_device_map[cur_device] = torch.cuda.Stream(cur_device)
+                torch.cuda.set_device(cur_device)
+                self.stream_device_map[cur_device].wait_stream(prev_stream)
+                torch.cuda.set_stream(self.stream_device_map[cur_device])
+                hidden_states = hidden_states.to(self.transfer_map[i], non_blocking = True)
+                causal_mask = causal_mask.to(self.transfer_map[i], non_blocking = True) if causal_mask is not None else None
+                position_ids = position_ids.to(self.transfer_map[i], non_blocking = True) if position_ids is not None else None
+                cache_position = cache_position.to(self.transfer_map[i], non_blocking = True) if cache_position is not None else None
+                
             if output_hidden_states:
                 all_hidden_states += (hidden_states,)
 
@@ -463,7 +451,7 @@ DeepseekV2_INPUTS_DOCSTRING = r"""
 """
 
 
-class DeepseekV2ModelPerLayerPrefill(BaseInjectedModule):
+class KDeepseekV2Model(BaseInjectedModule):
     """
     Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`DeepseekV2DecoderLayer`]
 
@@ -478,10 +466,13 @@ class DeepseekV2ModelPerLayerPrefill(BaseInjectedModule):
         orig_module: nn.Module,
         device: str = "cuda",
         per_layer_prefill_intput_threshold: int = 30000, # if None, no per-layer prefill
+        transfer_map: dict = None,
         **kwargs,
     ):
         BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
         self.per_layer_prefill_intput_threshold = per_layer_prefill_intput_threshold
+        self.transfer_map = transfer_map
+        self.stream_device_map = dict()
 
     @add_start_docstrings_to_model_forward(DeepseekV2_INPUTS_DOCSTRING)
     def forward(
@@ -584,7 +575,20 @@ class DeepseekV2ModelPerLayerPrefill(BaseInjectedModule):
         t_cpu = 0
         t_f = 0
 
-        for decoder_layer in self.layers:
+        for i, decoder_layer in enumerate(self.layers):
+            if self.transfer_map is not None and i in self.transfer_map: 
+                prev_stream = torch.cuda.current_stream()
+                cur_device = self.transfer_map[i]
+                if cur_device not in self.stream_device_map:
+                    self.stream_device_map[cur_device] = torch.cuda.Stream(cur_device)
+                torch.cuda.set_device(cur_device)
+                self.stream_device_map[cur_device].wait_stream(prev_stream)
+                torch.cuda.set_stream(self.stream_device_map[cur_device])
+                hidden_states = hidden_states.to(self.transfer_map[i], non_blocking = True)
+                causal_mask = causal_mask.to(self.transfer_map[i], non_blocking = True) if causal_mask is not None else None
+                position_ids = position_ids.to(self.transfer_map[i], non_blocking = True) if position_ids is not None else None
+                cache_position = cache_position.to(self.transfer_map[i], non_blocking = True) if cache_position is not None else None
+
             if output_hidden_states:
                 all_hidden_states += (hidden_states,)
 

ktransformers/optimize/optimize.py

 '''
 Description  :  
-Author       : Boxin Zhang
+Author       : Boxin Zhang, Azure-Tang
 Version      : 0.1.0
 Copyright (c) 2024 by KVCache.AI, All Rights Reserved. 
 '''
@@ -15,6 +15,7 @@ from transformers.configuration_utils import PretrainedConfig
 from ktransformers.util.custom_gguf import GGUFLoader, translate_name_to_gguf
 from ktransformers.util.utils import set_module, load_weights
 import itertools
+import copy
 
 def inject(module, local_optimization_dict, model_config:AutoConfig ,gguf_loader:GGUFLoader, prefix=''):
     for name, child in module._modules.items():
@@ -22,18 +23,20 @@ def inject(module, local_optimization_dict, model_config:AutoConfig ,gguf_loader
             child_prefix = prefix + name
             if child_prefix in local_optimization_dict:
                 inject_module_meta=local_optimization_dict[child_prefix]
-                if isinstance(inject_module_meta, Mapping):
+                if inject_module_meta["class"] != "default":
                     import_path = inject_module_meta["class"].split(".")
                     import_module_name = ".".join(import_path[:-1])
+                    gguf_loader.tensor_device_map[inject_module_meta["key"]] = inject_module_meta["kwargs"] if "kwargs" in inject_module_meta else dict()
                     import_class_name = import_path[-1]
                     module_cls=getattr(__import__(import_module_name, fromlist=[""]), import_class_name)
                     print(f"Injecting {child_prefix} as", import_module_name, ".", import_class_name)
-                    inject_module=module_cls(key = inject_module_meta["key"], gguf_loader = gguf_loader, config = model_config, orig_module=child, device = inject_module_meta["device"], **inject_module_meta["kwargs"])
+                    inject_module=module_cls(key = inject_module_meta["key"], gguf_loader = gguf_loader, config = model_config, orig_module=child, **inject_module_meta["kwargs"])
                     set_module(module, name, inject_module)
-                elif isinstance(inject_module_meta, str):
-                    assert inject_module_meta=="default", "for str inject_module_meta, only support \"default\"."
+                elif inject_module_meta["class"] == "default":
+                    print(f"Injecting {child_prefix} as default")
+                    gguf_loader.tensor_device_map[inject_module_meta["key"]] = inject_module_meta["kwargs"] if "kwargs" in inject_module_meta else dict()
                 else:
-                    raise Exception("inject_module_meta must be a dict or str")
+                    raise Exception("inject_module_meta[\"class\"] must be \"default\" or a class path")
                 child_prefix += "."
                 child_optimization_dict = {k: v for k, v in local_optimization_dict.items() if k.startswith(child_prefix)}
                 inject(child, child_optimization_dict, model_config, gguf_loader, child_prefix)
@@ -55,8 +58,9 @@ def gen_optimize_config(module: nn.Module, out_data: Mapping, rule_list: List, p
     #print("gen_optimize_config", prefix, module_name, translated_name)
     recursive = True
     for rule in rule_list:
-        #print(rule)
         match_meta = rule["match"]
+        if "class" not in match_meta and "name" not in match_meta:
+            raise Exception("match must have at least one of \"class\" and \"name\"")
         if "class" in match_meta:
             import_path = match_meta["class"].split(".")
             import_module_name = ".".join(import_path[:-1])
@@ -67,16 +71,30 @@ def gen_optimize_config(module: nn.Module, out_data: Mapping, rule_list: List, p
         if "name" in match_meta:
             if re.search(match_meta["name"], module_name) is None:
                 continue
-        replace_meta = rule["replace"]
-        out_data[module_name]={"key": translated_name,
-                               "class": replace_meta["class"],
-                               "device": replace_meta["device"] if "device" in replace_meta else default_device,
-                               "kwargs": replace_meta["kwargs"] if "kwargs" in replace_meta else dict()}
+        if "replace" not in rule:
+            raise Exception("replace must be in rule")
+        if "replace" in rule:
+            replace_meta = rule["replace"]
+            if module_name not in out_data:
+                out_data[module_name]={"key": translated_name,
+                                    "class": replace_meta["class"] if "class" in replace_meta else "default",
+                                    # "device": replace_meta["device"] if "device" in replace_meta else default_device,
+                                    "kwargs": copy.deepcopy(replace_meta["kwargs"]) if "kwargs" in replace_meta else dict()}
+            else:
+                if out_data[module_name]["class"] == "default":
+                    out_data[module_name]["class"] = replace_meta["class"] if "class" in replace_meta else "default"
+                out_data[module_name]["kwargs"].update(copy.deepcopy(replace_meta["kwargs"]) if "kwargs" in replace_meta else dict())
         if "recursive" in rule:
             recursive = bool(rule["recursive"])
+        break
             
     if module_name not in out_data:
-        out_data[module_name]="default"
+        out_data[module_name]= {
+            "class": "default",
+            "key": translated_name,
+            "kwargs": {"generate_device": default_device,
+                       "prefill_device": default_device}
+        }
 
     #print(out_data[module_name])
     #input()
@@ -88,6 +106,14 @@ def gen_optimize_config(module: nn.Module, out_data: Mapping, rule_list: List, p
                 gen_optimize_config(child, out_data, rule_list, child_prefix)
     
 
+def translate_model_config(model_config: PretrainedConfig):
+    # for supporting some special model 
+    if model_config.model_type == "mixtral":
+        model_config.moe_intermediate_size = model_config.intermediate_size
+    
+    return model_config
+
+
 def optimize_and_load_gguf(module: nn.Module, rule_file: str, gguf_path: str, model_config: PretrainedConfig, default_device: str = "cuda:0"):
     with open(rule_file, 'r', encoding='utf-8') as f:
         rule_list = yaml.load(f.read(), Loader=yaml.FullLoader)
@@ -95,8 +121,12 @@ def optimize_and_load_gguf(module: nn.Module, rule_file: str, gguf_path: str, mo
     optimize_config = dict()
     gen_optimize_config(module, optimize_config, rule_list, default_device = default_device)
     
+    model_config = translate_model_config(model_config)
+
     gguf_loader=GGUFLoader(gguf_path)
     with torch.device("meta"):
         inject(module, optimize_config, model_config, gguf_loader)
     load_weights(module, gguf_loader)
+    module.gguf_loader = gguf_loader
     del_meta(module)
+    torch.cuda.empty_cache()

ktransformers/optimize/optimize_rules/DeepSeek-V2-Chat-multi-gpu-4.yaml

+- match:
+    name: "^model.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+        generate_device: "cpu"
+        prefill_device: "cpu"
+
+- match:
+    name: "^model\\.layers\\.([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\\.([1][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\\.([2][0-9])\\."
+    class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.YarnRotaryEmbedding
+    kwargs:
+      generate_device: "cuda:2"
+      prefill_device: "cuda:2"
+- match:
+    name: "^model\\.layers\\.([345][0-9])\\."
+    class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.YarnRotaryEmbedding
+    kwargs:
+      generate_device: "cuda:3"
+      prefill_device: "cuda:3"
+
+- match:
+    name: "^model\\.layers\\.([0-9])\\.(?!self_attn).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\.([1][0-9])\\.(?!self_attn).*$"  # 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\\.([2][0-9])\\.(?!self_attn).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda:2"
+      prefill_device: "cuda:2"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\.([345][0-9])\\.(?!self_attn).*$"  # 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:3"
+      prefill_device: "cuda:3"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+- match:
+    name: "^model\\.layers\\.([0-9])\\.mlp$"
+    class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([1][0-9])\\.mlp$"
+    class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
+- match:
+    name: "^model\\.layers\\.([2][0-9])\\.mlp$"
+    class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:2"
+      prefill_device: "cuda:2"
+- match:
+    name: "^model\\.layers\\.([345][0-9])\\.mlp$"
+    class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:3"
+      prefill_device: "cuda:3"
+
+- match:
+    name: "^model\\.layers\\.([0-9])\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda:0"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op:  "KExpertsCPU"
+      out_device: "cuda:0"
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\.([1][0-9])\\.mlp\\.experts$"
+  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\\.([2][0-9])\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda:2"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op:  "KExpertsCPU"
+      out_device: "cuda:2"
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\.([345][0-9])\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda:3"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op:  "KExpertsCPU"
+      out_device: "cuda:3"
+  recursive: False # don't recursively inject submodules of this module
+
+- match:
+    name: "^model\\.layers\\.([0-9])\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([1][0-9])\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
+- match:
+    name: "^model\\.layers\\.([2][0-9])\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda:2"
+      prefill_device: "cuda:2"
+- match:
+    name: "^model\\.layers\\.([345][0-9])\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda:3"
+      prefill_device: "cuda:3"
+
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KDeepseekV2Model"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+      transfer_map: 
+        10: "cuda:1"
+        20: "cuda:2"
+        30: "cuda:3"
+
+- match:
+    name: "^model\\.layers\\.([0-9])\\."
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "(^model\\.layers\\.([1][0-9])\\.)"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
+- match:
+    name: "(^model\\.layers\\.([2][0-9])\\.)"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:2"
+      prefill_device: "cuda:2"
+- match:
+    name: "(^model\\.layers\\.([345][0-9])\\.)|(^model.norm)|(^lm_head)"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:3"
+      prefill_device: "cuda:3"
\ No newline at end of file

ktransformers/optimize/optimize_rules/DeepSeek-V2-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\\.([345][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).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+- match:
+    name: "^model\\.layers\\.([345][0-9])\\.(?!self_attn).*$"  # 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$"
+    class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([345][0-9])\\.mlp$"
+    class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # 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$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda:0"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op:  "KExpertsCPU"
+      out_device: "cuda:0"
+  recursive: False # don't recursively inject submodules of this module
+
+- match:
+    name: "^model\\.layers\\.([345][0-9])\\.mlp\\.experts$"
+  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$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([345][0-9])\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # 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\\.([345][0-9])\\.)|(model.norm)|(lm_head)"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
\ No newline at end of file

ktransformers/optimize/optimize_rules/DeepSeek-V2-Chat.yaml

@@ -2,40 +2,49 @@
     class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
   replace:
     class: ktransformers.operators.RoPE.YarnRotaryEmbedding
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
 - match:
     name: "^model\\.layers\\.(?!.*self_attn).*$"  # regular expression 
     class: torch.nn.Linear  # only match modules matching name and class simultaneously
   replace:
-    class: ktransformers.operators.linear.KTransformerLinear  # optimized Kernel on quantized data types
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
     kwargs:
       generate_device: "cuda"
       prefill_device: "cuda"
-      generate_op: "QuantizedLinearMarlin"
-      prefill_op: "QuantizedLinearTorch"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
 - match:
     name: "^model\\.layers\\..*\\.mlp$"
     class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
   replace:
-    class: ktransformers.operators.experts.DeepseekV2MoEInjected     # mlp module with custom forward function
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
 - match:
     name: "^model\\.layers\\..*\\.mlp\\.experts$"
   replace:
-    class: ktransformers.operators.experts.KTransformersMLPExpert     # custom MoE Kernel with expert paralleism
-    device: "cpu"   # which devices to load this module when initializing
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
     kwargs:
       prefill_device: "cuda"
-      prefill_mlp_type: "MLPExpertsTorch"
+      prefill_op: "KExpertsTorch"
       generate_device: "cpu"
-      generate_mlp_type: "MLPCPUExperts"
+      generate_op: "KExpertsCPU"
       out_device: "cuda"
   recursive: False # don't recursively inject submodules of this module
 - match:
     name: "^model\\.layers\\..*\\.self_attn$"
   replace:
-    class: ktransformers.operators.attention.DeepseekV2AttentionInjected # optimized MLA implementation
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
 - match:
-    name: "^model$"
+    name: "^model.embed_tokens"
   replace:
-    class: "ktransformers.operators.layer_wise_prefill.DeepseekV2ModelPerLayerPrefill"
+    class: "default"
     kwargs:
-      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+      generate_device: "cpu"
+      prefill_device: "cpu"
\ No newline at end of file

ktransformers/optimize/optimize_rules/DeepSeek-V2-Lite-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])\\."
+    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\\.([12][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])\\.(?!self_attn).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+- match:
+    name: "^model\\.layers\\.([12][0-9])\\.(?!self_attn).*$"  # 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])\\.mlp$"
+    class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([12][0-9])\\.mlp$"
+    class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
+  replace:
+    class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
+
+- match:
+    name: "^model\\.layers\\.(0|[1-9])\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda:0"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op:  "KExpertsCPU"
+      out_device: "cuda:0"
+  recursive: False # don't recursively inject submodules of this module
+
+- match:
+    name: "^model\\.layers\\.([12][0-9])\\.mlp\\.experts$"
+  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])\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([12][0-9])\\.self_attn$"
+  replace:
+    class: ktransformers.operators.attention.KDeepseekV2Attention # 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: 
+        10: "cuda:1"
+
+- match:
+    name: "^model\\.layers\\.(0|[1-9])\\."
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+
+- match:
+    name: "(^model\\.layers\\.([12][0-9])\\.)|(model.norm)|(lm_head)"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
\ No newline at end of file

ktransformers/optimize/optimize_rules/Mixtral.yaml

+- match:
+    class: ktransformers.models.modeling_mixtral.MixtralRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.RotaryEmbedding
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- match:
+    name: "^model\\.layers\\..*$"
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\..*\\.block_sparse_moe$"
+    class: ktransformers.models.modeling_mixtral.MixtralSparseMoeBlock
+  replace: 
+    class: ktransformers.operators.experts.KMisrtalSparseMoEBlock
+- match:
+    name: "^model\\.layers\\..*\\.block_sparse_moe\\.experts$"
+  replace: 
+    class: ktransformers.operators.experts.KTransformersExperts
+    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.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+        generate_device: "cpu"
+        prefill_device: "cpu"
+
+- match:
+    name: "^model\\.layers\\..*\\."
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
\ No newline at end of file

ktransformers/optimize/optimize_rules/Qwen2-57B-A14B-Instruct-multi-gpu.yaml

+- match:
+    name: "^model\\.layers\\.([012])\\."
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.RotaryEmbedding
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+- match:
+    name: "^model\\.layers\\.([012])$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\.([012])\\.mlp$"
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeSparseMoeBlock
+  replace:
+    class: ktransformers.operators.experts.KQwen2MoeSparseMoeBlock     # mlp module with custom forward function
+- match:
+    name: "^model\\.layers\\.([012])\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    # device: "cpu"   # which devices to load this module when initializing
+    kwargs:
+      prefill_device: "cuda:0"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op:  "KExpertsCPU"
+      out_device: "cuda:0"
+  recursive: False # don't recursively inject submodules of this module
+
+- match:
+    name: "^model\\.layers\\.([12][0-9]|[3-9])\\."
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.RotaryEmbedding
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
+- match:
+    name: "^model\\.layers\\.([12][0-9]|[3-9])$"  # 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\\.([12][0-9]|[3-9])\\.mlp$"
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeSparseMoeBlock
+  replace:
+    class: ktransformers.operators.experts.KQwen2MoeSparseMoeBlock     # mlp module with custom forward function
+- match:
+    name: "^model\\.layers\\.([12][0-9]|[3-9])\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
+    # device: "cpu"   # which devices to load this module when initializing
+    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.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+        generate_device: "cpu"
+        prefill_device: "cpu"
+
+- match:
+    name: "(^model.norm)|(^lm_head)"
+  replace:
+    class: "default"
+    kwargs:
+        generate_device: "cuda:1"
+        prefill_device: "cuda:1"
+
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KQwen2MoeModel"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+      transfer_map: 
+        3: "cuda:1"
+
+- match:
+    name: "^model\\.layers\\.([012])\\."
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:0"
+      prefill_device: "cuda:0"
+
+- match:
+    name: "^model\\.layers\\.([12][0-9]|[3-9])\\."
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cuda:1"
+      prefill_device: "cuda:1"
\ No newline at end of file