support FlashAttention2

README.md

@@ -12,11 +12,13 @@
 
 ## Changelog
 
+[23/09/10] Now we support using **[FlashAttention](https://github.com/Dao-AILab/flash-attention)** for the LLaMA models. Try `--flash_attn` argument to enable FlashAttention-2 if you are using RTX4090, A100 or H100 GPUs (experimental feature).
+
 [23/08/18] Now we support **resuming training**, upgrade `transformers` to `4.31.0` to enjoy this feature.
 
 [23/08/12] Now we support **RoPE scaling** to extend the context length of the LLaMA models. Try `--rope_scaling linear` argument in training and `--rope_scaling dynamic` argument at inference to extrapolate the position embeddings.
 
-[23/08/11] Now we support **[DPO training](https://arxiv.org/abs/2305.18290)** for instruction-tuned models. See [this example](#dpo-training) to train your models (experimental feature).
+[23/08/11] Now we support **[DPO training](https://arxiv.org/abs/2305.18290)** for instruction-tuned models. See [this example](#dpo-training) to train your models.
 
 [23/08/03] Now we support training the **Qwen-7B** model in this repo. Try `--model_name_or_path Qwen/Qwen-7B-Chat` and `--lora_target c_attn` arguments to train the Qwen-7B model. Remember to use `--template chatml` argument when you are using the Qwen-7B-Chat model.
 
@@ -62,8 +64,11 @@
 | [XVERSE](https://github.com/xverse-ai/XVERSE-13B)        | 13B                         | q_proj,v_proj     | xverse    |
 | [ChatGLM2](https://github.com/THUDM/ChatGLM2-6B)         | 6B                          | query_key_value   | chatglm2  |
 
-- **Default module** is used for the `--lora_target` argument, you can use `--lora_target all` to specify all the available modules.
-- For the "base" models, the `--template` argument can be chosen from `default`, `alpaca`, `vicuna` etc. But make sure to use the corresponding template for the "chat" models.
+> **Note**
+>
+> **Default module** is used for the `--lora_target` argument, you can use `--lora_target all` to specify all the available modules.
+>
+> For the "base" models, the `--template` argument can be chosen from `default`, `alpaca`, `vicuna` etc. But make sure to use the corresponding template for the "chat" models.
 
 ## Supported Training Approaches
 
@@ -75,7 +80,9 @@
 | PPO Training           |                    |                    | :white_check_mark: | :white_check_mark: |
 | DPO Training           | :white_check_mark: |                    | :white_check_mark: | :white_check_mark: |
 
-- Use `--quantization_bit 4/8` argument to enable QLoRA.
+> **Note**
+>
+> Use `--quantization_bit 4/8` argument to enable QLoRA.
 
 ## Provided Datasets
 
@@ -138,7 +145,9 @@ And **powerful GPUs**!
 
 Please refer to `data/example_dataset` for checking the details about the format of dataset files. You can either use a single `.json` file or a [dataset loading script](https://huggingface.co/docs/datasets/dataset_script) with multiple files to create a custom dataset.
 
-Note: please update `data/dataset_info.json` to use your custom dataset. About the format of this file, please refer to `data/README.md`.
+> **Note**
+>
+> Please update `data/dataset_info.json` to use your custom dataset. About the format of this file, please refer to `data/README.md`.
 
 ### Dependence Installation (optional)
 
@@ -164,10 +173,16 @@ CUDA_VISIBLE_DEVICES=0 python src/train_web.py
 
 We strongly recommend using the all-in-one Web UI for newcomers since it can also generate training scripts **automatically**.
 
-Currently the web UI only supports training on **a single GPU**.
+> **Warning**
+>
+> Currently the web UI only supports training on **a single GPU**.
 
 ### Train on a single GPU
 
+> **Warning**
+>
+> If you want to train models on multiple GPUs, please refer to [#distributed-training](Distributed Training).
+
 #### Pre-Training
 
 ```bash
@@ -300,19 +315,13 @@ accelerate config # configure the environment
 accelerate launch src/train_bash.py # arguments (same as above)
 ```
 
-<details><summary>Example config.yaml for training with DeepSpeed ZeRO-2</summary>
+<details><summary>Example config for LoRA training</summary>
 
 ```yaml
 compute_environment: LOCAL_MACHINE
-deepspeed_config:
-  gradient_accumulation_steps: 4
-  gradient_clipping: 0.5
-  offload_optimizer_device: none
-  offload_param_device: none
-  zero3_init_flag: false
-  zero_stage: 2
-distributed_type: DEEPSPEED
+distributed_type: MULTI_GPU
 downcast_bf16: 'no'
+gpu_ids: all
 machine_rank: 0
 main_training_function: main
 mixed_precision: fp16
@@ -336,7 +345,7 @@ deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
     ... # arguments (same as above)
 ```
 
-<details><summary>Example ds_config.json for training with DeepSpeed ZeRO-2</summary>
+<details><summary>Example config for full-parameter training with DeepSpeed ZeRO-2</summary>
 
 ```json
 {
@@ -387,7 +396,9 @@ python src/api_demo.py \
     --checkpoint_dir path_to_checkpoint
 ```
 
-Visit `http://localhost:8000/docs` for API documentation.
+> **Note**
+>
+> Visit `http://localhost:8000/docs` for API documentation.
 
 ### CLI Demo
 
@@ -426,7 +437,9 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --predict_with_generate
 ```
 
-We recommend using `--per_device_eval_batch_size=1` and `--max_target_length 128` at 4/8-bit evaluation.
+> **Note**
+>
+> We recommend using `--per_device_eval_batch_size=1` and `--max_target_length 128` at 4/8-bit evaluation.
 
 ### Predict
 
@@ -445,12 +458,6 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --predict_with_generate
 ```
 
-## TODO
-
-- [ ] Supporting flash attention ([torch](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html) / [xformers](https://github.com/facebookresearch/xformers) / [flashattn](https://github.com/Dao-AILab/flash-attention)).
-- [ ] Implementing multi-query attention for faster inference.
-- [ ] Supporting full-parameter RLHF training.
-
 ## License
 
 This repository is licensed under the [Apache-2.0 License](LICENSE).

README_zh.md

@@ -12,11 +12,13 @@
 
 ## 更新日志
 
+[23/09/10] 现在我们支持了 LLaMA 模型的 **[FlashAttention](https://github.com/Dao-AILab/flash-attention)**。如果您使用的是 RTX4090、A100 或 H100 GPU，请使用 `--flash_attn` 参数以启用 FlashAttention-2（实验性功能）。
+
 [23/08/18] 现在我们支持了**训练状态恢复**，请将 `transformers` 升级至 `4.31.0` 以启用此功能。
 
-[23/08/12] 现在我们支持了 **RoPE 插值**来扩展 LLaMA 模型的上下文长度。请尝试使用 `--rope_scaling linear` 参数训练模型或使用 `--rope_scaling dynamic` 参数评估模型。
+[23/08/12] 现在我们支持了 **RoPE 插值**来扩展 LLaMA 模型的上下文长度。请使用 `--rope_scaling linear` 参数训练模型或使用 `--rope_scaling dynamic` 参数评估模型。
 
-[23/08/11] 现在我们支持了指令模型的 **[DPO 训练](https://arxiv.org/abs/2305.18290)**。详情请参阅[此示例](#dpo-训练)（实验性功能）。
+[23/08/11] 现在我们支持了指令模型的 **[DPO 训练](https://arxiv.org/abs/2305.18290)**。详情请参阅[此示例](#dpo-训练)。
 
 [23/08/03] 现在我们支持了 **Qwen-7B** 模型的训练。请尝试使用 `--model_name_or_path Qwen/Qwen-7B-Chat` 和 `--lora_target c_attn` 参数。使用 Qwen-7B-Chat 模型时请添加 `--template chatml` 参数。
 
@@ -62,8 +64,11 @@
 | [XVERSE](https://github.com/xverse-ai/XVERSE-13B)        | 13B                         | q_proj,v_proj     | xverse    |
 | [ChatGLM2](https://github.com/THUDM/ChatGLM2-6B)         | 6B                          | query_key_value   | chatglm2  |
 
-- **默认模块**应作为 `--lora_target` 参数的默认值，可使用 `--lora_target all` 参数指定全部模块。
-- 对于所有“基座”（Base）模型，`--template` 参数可以是 `default`, `alpaca`, `vicuna` 等任意值。但“对话”（Chat）模型请务必使用对应的模板。
+> **Note**
+>
+> **默认模块**应作为 `--lora_target` 参数的默认值，可使用 `--lora_target all` 参数指定全部模块。
+>
+> 对于所有“基座”（Base）模型，`--template` 参数可以是 `default`, `alpaca`, `vicuna` 等任意值。但“对话”（Chat）模型请务必使用对应的模板。
 
 ## 训练方法
 
@@ -75,7 +80,9 @@
 | PPO 训练               |                    |                    | :white_check_mark: | :white_check_mark: |
 | DPO 训练               | :white_check_mark: |                    | :white_check_mark: | :white_check_mark: |
 
-- 使用 `--quantization_bit 4/8` 参数来启用 QLoRA 训练。
+> **Note**
+>
+> 请使用 `--quantization_bit 4/8` 参数来启用 QLoRA 训练。
 
 ## 数据集
 
@@ -138,7 +145,9 @@ huggingface-cli login
 
 关于数据集文件的格式，请参考 `data/example_dataset` 文件夹的内容。构建自定义数据集时，既可以使用单个 `.json` 文件，也可以使用一个[数据加载脚本](https://huggingface.co/docs/datasets/dataset_script)和多个文件。
 
-注意：使用自定义数据集时，请更新 `data/dataset_info.json` 文件，该文件的格式请参考 `data/README.md`。
+> **Note**
+>
+> 使用自定义数据集时，请更新 `data/dataset_info.json` 文件，该文件的格式请参考 `data/README.md`。
 
 ### 环境搭建（可跳过）
 
@@ -164,10 +173,16 @@ CUDA_VISIBLE_DEVICES=0 python src/train_web.py
 
 我们极力推荐新手使用浏览器一体化界面，因为它还可以**自动**生成运行所需的命令行脚本。
 
-目前网页 UI 仅支持**单卡训练**。
+> **Warning**
+>
+> 目前网页 UI 仅支持**单卡训练**。
 
 ### 单 GPU 训练
 
+> **Warning**
+>
+> 如果您使用多张 GPU 训练模型，请移步[多 GPU 分布式训练](#多-gpu-分布式训练)部分。
+
 #### 预训练
 
 ```bash
@@ -299,19 +314,13 @@ accelerate config # 首先配置分布式环境
 accelerate launch src/train_bash.py # 参数同上
 ```
 
-<details><summary>使用 DeepSpeed ZeRO-2 进行全参数微调的 Accelerate 配置示例</summary>
+<details><summary>LoRA 训练的 Accelerate 配置示例</summary>
 
 ```yaml
 compute_environment: LOCAL_MACHINE
-deepspeed_config:
-  gradient_accumulation_steps: 4
-  gradient_clipping: 0.5
-  offload_optimizer_device: none
-  offload_param_device: none
-  zero3_init_flag: false
-  zero_stage: 2
-distributed_type: DEEPSPEED
+distributed_type: MULTI_GPU
 downcast_bf16: 'no'
+gpu_ids: all
 machine_rank: 0
 main_training_function: main
 mixed_precision: fp16
@@ -335,7 +344,7 @@ deepspeed --num_gpus 8 --master_port=9901 src/train_bash.py \
     ... # 参数同上
 ```
 
-<details><summary>使用 DeepSpeed ZeRO-2 进行全参数微调的 DeepSpeed 配置示例</summary>
+<details><summary>使用 DeepSpeed ZeRO-2 进行全参数训练的 DeepSpeed 配置示例</summary>
 
 ```json
 {
@@ -386,7 +395,9 @@ python src/api_demo.py \
     --checkpoint_dir path_to_checkpoint
 ```
 
-关于 API 文档请见 `http://localhost:8000/docs`。
+> **Note**
+>
+> 关于 API 文档请见 `http://localhost:8000/docs`。
 
 ### 命令行测试
 
@@ -425,7 +436,9 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --predict_with_generate
 ```
 
-我们建议在量化模型的评估中使用 `--per_device_eval_batch_size=1` 和 `--max_target_length 128`。
+> **Note**
+>
+> 我们建议在量化模型的评估中使用 `--per_device_eval_batch_size=1` 和 `--max_target_length 128`。
 
 ### 模型预测
 
@@ -444,12 +457,6 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --predict_with_generate
 ```
 
-## TODO
-
-- [ ] 实现 flash attention ([torch](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html) / [xformers](https://github.com/facebookresearch/xformers) / [flashattn](https://github.com/Dao-AILab/flash-attention))。
-- [ ] 在推理阶段使用 Multi-query attention 进行加速。
-- [ ] 支持 RLHF 的全参数微调。
-
 ## 协议
 
 本仓库的代码依照 [Apache-2.0](LICENSE) 协议开源。

src/llmtuner/extras/models/flash_llama.py

@@ -0,0 +1,771 @@
+# coding=utf-8
+# Modified from:
+# [1] https://huggingface.co/Birchlabs/flash_llama/blob/main/modeling_flash_llama.py
+# [2] https://huggingface.co/togethercomputer/LLaMA-2-7B-32K/blob/main/modeling_flash_llama.py
+# [3] https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
+# With fix from Alex Birch: https://huggingface.co/togethercomputer/LLaMA-2-7B-32K/discussions/17
+
+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 CrossEntropyLoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from transformers.models.llama.configuration_llama import LlamaConfig
+
+
+try:
+    from flash_attn.flash_attn_interface import (
+        flash_attn_kvpacked_func,
+        flash_attn_varlen_kvpacked_func,
+    )
+    from flash_attn.bert_padding import unpad_input, pad_input
+    flash_attn_v2_installed = True
+    print('>>>> Flash Attention installed')
+except ImportError:
+    flash_attn_v2_installed = False
+    raise ImportError('Please install Flash Attention: `pip install flash-attn --no-build-isolation`')
+
+try:
+    from flash_attn.layers.rotary import apply_rotary_emb_func
+    flash_rope_installed = True
+    print('>>>> Flash RoPE installed')
+except ImportError:
+    flash_rope_installed = False
+    raise ImportError('Please install RoPE kernels: `pip install git+https://github.com/HazyResearch/flash-attention.git#subdirectory=csrc/rotary`')
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlamaConfig"
+
+
+def rmsnorm_func(hidden_states, weight, variance_epsilon):
+    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 + variance_epsilon)
+    return (weight * hidden_states).to(input_dtype)
+    
+
+class LlamaRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        LlamaRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.register_buffer(
+            "variance_epsilon",
+            torch.tensor(eps),
+            persistent=False,
+        )
+        
+    def forward(self, hidden_states):
+        return rmsnorm_func(hidden_states, self.weight, self.variance_epsilon)
+
+
+class FlashRotaryEmbedding(torch.nn.Module):
+    """
+    The rotary position embeddings from RoFormer_ (Su et. al).
+    A crucial insight from the method is that the query and keys are
+    transformed by rotation matrices which depend on the relative positions.
+
+    Other implementations are available in the Rotary Transformer repo_ and in
+    GPT-NeoX_, GPT-NeoX was an inspiration
+
+    .. _RoFormer: https://arxiv.org/abs/2104.09864
+    .. _repo: https://github.com/ZhuiyiTechnology/roformer
+    .. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
+
+    If scale_base is not None, this implements XPos (Sun et al., https://arxiv.org/abs/2212.10554).
+    A recommended value for scale_base is 512: https://github.com/HazyResearch/flash-attention/issues/96
+    Reference: https://github.com/sunyt32/torchscale/blob/main/torchscale/component/xpos_relative_position.py
+    """
+
+    def __init__(self, dim: int, base=10000.0, interleaved=False, scale_base=None,
+                 scaling_factor=1.0, pos_idx_in_fp32=True, device=None):
+        """
+            interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
+                of 1st half and 2nd half (GPT-NeoX style).
+            pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
+                otherwise they might be in lower precision.
+                This option was added because previously (before 2023-07-02), when we construct
+                the position indices, we use the dtype of self.inv_freq. In most cases this would
+                be fp32, but if the model is trained in pure bf16 (not mixed precision), then
+                self.inv_freq would be bf16, and the position indices are also in bf16.
+                Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
+                embeddings for some positions will coincide.
+                To maintain compatibility with models previously trained in pure bf16,
+                we add this option.
+            scaling_factor: RotaryEmbedding extended with linear scaling.
+        """
+        super().__init__()
+        self.dim = dim
+        self.base = float(base)
+        self.pos_idx_in_fp32 = pos_idx_in_fp32
+        # Generate and save the inverse frequency buffer (non trainable)
+        inv_freq = self._compute_inv_freq(device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.interleaved = interleaved
+        self.scale_base = scale_base
+        self.scaling_factor = scaling_factor
+        scale = ((torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim)
+                 / (1.4 * dim) if scale_base is not None else None)
+        self.register_buffer("scale", scale)
+
+        self._seq_len_cached = 0
+        self._cos_cached = None
+        self._sin_cached = None
+        self._cos_k_cached = None
+        self._sin_k_cached = None
+
+    def _compute_inv_freq(self, device=None):
+        return 1 / (self.base ** (torch.arange(0, self.dim, 2, device=device,
+                                                 dtype=torch.float32) / self.dim))
+
+
+    def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
+        # Reset the tables if the sequence length has changed,
+        # if we're on a new device (possibly due to tracing for instance),
+        # or if we're switching from inference mode to training
+        if (seqlen > self._seq_len_cached or self._cos_cached.device != device
+            or self._cos_cached.dtype != dtype
+            or (self.training and self._cos_cached.is_inference())):
+            self._seq_len_cached = seqlen
+            # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
+            # And the output of arange can be quite large, so bf16 would lose a lot of precision.
+            # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
+            if self.pos_idx_in_fp32:
+                t = torch.arange(seqlen, device=device, dtype=torch.float32)
+                t /= self.scaling_factor
+                # We want fp32 here as well since inv_freq will be multiplied with t, and the output
+                # will be large. Having it in bf16 will lose a lot of precision and cause the
+                # cos & sin output to change significantly.
+                # We want to recompute self.inv_freq if it was not loaded in fp32
+                if self.inv_freq.dtype != torch.float32:
+                    inv_freq = self.inv_freq.to(torch.float32)
+                else:
+                    inv_freq = self.inv_freq
+            else:
+                t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
+                t /= self.scaling_factor
+                inv_freq = self.inv_freq
+            # Don't do einsum, it converts fp32 to fp16 under AMP
+            # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            freqs = torch.outer(t, inv_freq)
+            if self.scale is None:
+                self._cos_cached = torch.cos(freqs).to(dtype)
+                self._sin_cached = torch.sin(freqs).to(dtype)
+            else:
+                power = ((torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device)
+                          - seqlen // 2) / self.scale_base)
+                scale = self.scale.to(device=power.device) ** power.unsqueeze(-1)
+                # We want the multiplication by scale to happen in fp32
+                self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
+                self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
+                self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
+                self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
+
+    def forward(self, q: torch.Tensor, k: torch.Tensor, seqlen_offset: int = 0) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        q: (batch, seqlen, nheads, headdim)
+        k: (batch, seqlen, nheads, headdim)
+        seqlen_offset: can be used in generation where the qkv being passed in is only the last
+        token in the batch.
+        """
+        self._update_cos_sin_cache(q.shape[1] + seqlen_offset, device=q.device, dtype=q.dtype)
+        if self.scale is None:
+            return apply_rotary_emb_func(
+                q, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:],
+                self.interleaved, True # inplace=True
+            ), apply_rotary_emb_func(
+                k, self._cos_cached[seqlen_offset:], self._sin_cached[seqlen_offset:],
+                self.interleaved, True # inplace=True
+            )
+        else:
+            assert False
+
+class LlamaMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        if self.config.pretraining_tp > 1:
+            slice = self.intermediate_size // self.config.pretraining_tp
+            gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
+            up_proj_slices = self.up_proj.weight.split(slice, dim=0)
+            down_proj_slices = self.down_proj.weight.split(slice, dim=1)
+
+            gate_proj = torch.cat(
+                [F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
+            )
+            up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
+
+            intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
+            down_proj = [
+                F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
+            ]
+            down_proj = sum(down_proj)
+        else:
+            down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+        return down_proj
+
+
+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, slen, _, num_key_value_heads, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, :, :, None, :].expand(batch, slen, 2, num_key_value_heads, n_rep, head_dim)
+    return hidden_states.reshape(batch, slen, 2, num_key_value_heads * n_rep, head_dim)
+
+
+class LlamaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.config = config
+        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
+
+        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.register_buffer(
+            "norm_factor",
+            torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype()),
+            persistent=False,
+        )
+
+        if self.config.rope_scaling is None:
+            scaling_factor = 1
+        else:
+            scaling_type = self.config.rope_scaling["type"]
+            scaling_factor = self.config.rope_scaling["factor"]
+            assert scaling_type == 'linear'
+        
+        self.rotary_emb = FlashRotaryEmbedding(
+            self.head_dim, base=10000, interleaved=False, scaling_factor=scaling_factor,
+        )
+
+    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[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        is_padded_inputs: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, h_size = hidden_states.size()
+
+        has_layer_past = past_key_value is not None
+
+        if has_layer_past:
+            past_kv = past_key_value[0]
+            past_len = past_key_value[1]
+        else:
+            past_len = 0
+
+        if self.config.pretraining_tp > 1:
+            key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
+            query_slices = self.q_proj.weight.split(
+                (self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
+            )
+            key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+            value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+            q = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
+            q = torch.cat(q, dim=-1)
+
+            k = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
+            k = torch.cat(k, dim=-1)
+
+            v = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
+            v = torch.cat(v, dim=-1)
+
+        else:
+            q = self.q_proj(hidden_states)
+            k = self.k_proj(hidden_states)
+            v = self.v_proj(hidden_states) 
+
+        q = q.view(bsz, q_len, self.num_heads, self.head_dim)
+        k = k.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
+        v = v.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
+        
+        q, k = self.rotary_emb(q, k, past_len)
+        
+        kv = torch.stack([k, v], 2)
+        kv = repeat_kv(kv, self.num_key_value_groups)
+
+        # Cache QKV values
+        if has_layer_past:
+            new_len = past_len+q.size(1)
+            if new_len > past_kv.size(1):
+                past_kv = torch.cat([past_kv, torch.empty(bsz, 256, 2, kv.size(3), kv.size(4), dtype=kv.dtype, device=kv.device)], 1)
+            past_kv[:, past_len:new_len] = kv
+            kv = past_kv[:, :new_len]
+        else:
+            past_kv = kv
+
+        past_key_value = (past_kv, past_len+q.size(1)) if use_cache else None
+
+        if is_padded_inputs:
+
+            # varlen, ignore padding tokens, efficient for large batch with many paddings
+            logger.warning_once("padded")
+
+            assert attention_mask is not None
+            
+            unpadded_kv, indices_k, cu_seqlens_k, max_seqlen_k = unpad_input(kv, attention_mask)
+            unpadded_q, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, attention_mask[:, -q.size(1):])
+            attn_outputs = flash_attn_varlen_kvpacked_func(
+                unpadded_q, unpadded_kv, cu_seqlens_q, cu_seqlens_k, 
+                max_seqlen_q, max_seqlen_k,
+                dropout_p=0.0, softmax_scale=1.0/self.norm_factor, 
+                causal=(not has_layer_past), return_attn_probs=output_attentions
+            )
+
+            attn_output = attn_outputs[0] if output_attentions else attn_outputs
+            attn_output = pad_input(
+                attn_output, indices_q, bsz, q_len
+            ).reshape(bsz, q_len, h_size)
+            attn_weights = attn_outputs[2] if output_attentions else None
+            
+        else:
+
+            # no padding tokens, more efficient
+            
+            attn_outputs = flash_attn_kvpacked_func(
+                q, kv, dropout_p=0.0, softmax_scale=1.0/self.norm_factor, causal=(not has_layer_past), return_attn_probs=output_attentions)
+
+            attn_output = attn_outputs[0] if output_attentions else attn_outputs
+            attn_output = attn_output.reshape(bsz, q_len, h_size)
+            attn_weights = attn_outputs[2] if output_attentions else None
+
+        if self.config.pretraining_tp > 1:
+            attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
+            o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
+            attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
+        else:
+            attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class LlamaDecoderLayer(nn.Module):
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = LlamaAttention(config=config)
+        self.mlp = LlamaMLP(config)
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    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,
+        is_padded_inputs: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ) -> 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, 1, tgt_len, src_len)` where padding elements are indicated by very large negative 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.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        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,
+            is_padded_inputs=is_padded_inputs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+LLAMA_START_DOCSTRING, LLAMA_INPUTS_DOCSTRING = "", ""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaPreTrainedModel(PreTrainedModel):
+    config_class = LlamaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamaDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+
+    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_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, LlamaModel):
+            module.gradient_checkpointing = value
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaModel(LlamaPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
+
+    Args:
+        config: LlamaConfig
+    """
+
+    def __init__(self, config: LlamaConfig):
+        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([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = LlamaRMSNorm(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
+
+    @add_start_docstrings_to_model_forward(LLAMA_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,
+        return_dict: Optional[bool] = None,
+        is_padded_inputs: Optional[bool] = False,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        position_ids = None
+        
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        hidden_states = inputs_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    None,
+                    is_padded_inputs
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    is_padded_inputs=is_padded_inputs,
+                )
+
+            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],)
+
+        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 = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class LlamaForCausalLM(LlamaPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LlamaModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(LLAMA_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,
+        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,
+        is_padded_inputs: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        is_padded_inputs = ((attention_mask is not None) and (not attention_mask.all().item()))
+        
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs: "CausalLMOutputWithPast" = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_padded_inputs=is_padded_inputs,
+        )
+
+        hidden_states = outputs[0]
+        if self.config.pretraining_tp > 1:
+            lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.config.pretraining_tp, dim=0)
+            logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.config.pretraining_tp)]
+            logits = torch.cat(logits, dim=-1)
+        else:
+            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)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        position_ids = kwargs.get("position_ids", None)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+                "is_padded_inputs": ((attention_mask is not None) and (not attention_mask.all().item()))
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past

src/llmtuner/extras/template.py

@@ -206,9 +206,6 @@ def get_template_and_fix_tokenizer(
     name: str,
     tokenizer: "PreTrainedTokenizer"
 ) -> Template:
-    template = templates.get(name, None)
-    assert template is not None, "Template {} does not exist.".format(name)
-
     if tokenizer.eos_token_id is None:
         tokenizer.eos_token = "<|endoftext|>"
         logger.info("Add eos token: {}".format(tokenizer.eos_token))
@@ -217,6 +214,11 @@ def get_template_and_fix_tokenizer(
         tokenizer.pad_token = tokenizer.eos_token
         logger.info("Add pad token: {}".format(tokenizer.pad_token))
 
+    if name is None:
+        return None
+
+    template = templates.get(name, None)
+    assert template is not None, "Template {} does not exist.".format(name)
     tokenizer.add_special_tokens(
         dict(additional_special_tokens=template.stop_words),
         replace_additional_special_tokens=False

src/llmtuner/hparams/model_args.py

@@ -43,6 +43,10 @@ class ModelArguments:
         default=None,
         metadata={"help": "Adopt scaled rotary positional embeddings."}
     )
+    flash_attn: Optional[bool] = field(
+        default=False,
+        metadata={"help": "Enable flash attention for faster training."}
+    )
     checkpoint_dir: Optional[str] = field(
         default=None,
         metadata={"help": "Path to the directory(s) containing the delta model checkpoints as well as the configurations."}

src/llmtuner/tuner/core/loader.py

@@ -4,6 +4,7 @@ import torch
 from types import MethodType
 from typing import TYPE_CHECKING, Literal, Optional, Tuple
 
+import transformers
 from transformers import (
     AutoConfig,
     AutoModelForCausalLM,
@@ -84,7 +85,8 @@ def load_model_and_tokenizer(
 
     config = AutoConfig.from_pretrained(model_to_load, **config_kwargs)
 
-    if is_trainable and hasattr(config, "fp16") and hasattr(config, "bf16"): # fix Qwen config
+    # Fix config (for Qwen)
+    if is_trainable and hasattr(config, "fp16") and hasattr(config, "bf16"):
         if model_args.compute_dtype == torch.bfloat16:
             setattr(config, "bf16", True)
         else:
@@ -105,6 +107,7 @@ def load_model_and_tokenizer(
 
             if is_trainable:
                 if model_args.rope_scaling == "dynamic":
+                    assert not model_args.flash_attn, "Flash attention does not support dynamic rope scaling."
                     logger.warning(
                         "Dynamic NTK may not work well with fine-tuning. "
                         "See: https://github.com/huggingface/transformers/pull/24653"
@@ -127,6 +130,15 @@ def load_model_and_tokenizer(
         else:
             logger.warning("Current model does not support RoPE scaling.")
 
+    # Set flash attention
+    if model_args.flash_attn and getattr(config, "model_type", None) == "llama":
+        from llmtuner.extras.models.flash_llama import LlamaForCausalLM
+        transformers.models.llama.modeling_llama.LlamaForCausalLM = LlamaForCausalLM
+        if not hasattr(config, "num_key_value_heads"):
+            setattr(config, "num_key_value_heads", getattr(config, "num_attention_heads"))
+        if getattr(config, "pretraining_tp", 1) != 1:
+            setattr(config, "pretraining_tp", 1)
+
     # Quantization configurations (using bitsandbytes library).
     is_mergeable = True
     if model_args.quantization_bit is not None:

src/llmtuner/tuner/sft/trainer.py

@@ -33,6 +33,9 @@ class Seq2SeqPeftTrainer(PeftTrainer):
 
         Subclass and override to inject custom behavior.
         """
+        if self.args.predict_with_generate:
+            assert self.tokenizer.padding_side == "left", "This method only accepts left-padded tensor."
+            assert self.tokenizer.pad_token_id is not None, "Pad token is required."
             prompt_len, label_len = inputs["input_ids"].size(-1), inputs["labels"].size(-1)
             if prompt_len > label_len:
                 inputs["labels"] = self._pad_tensors_to_target_len(inputs["labels"], inputs["input_ids"])
@@ -50,10 +53,8 @@ class Seq2SeqPeftTrainer(PeftTrainer):
         loss, generated_tokens, labels = super().prediction_step(
             model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys
         )
-        if generated_tokens is not None:
-            generated_tokens[:, :max(prompt_len, label_len)] = (
-                self.tokenizer.pad_token_id * torch.ones_like(generated_tokens[:, :max(prompt_len, label_len)])
-            )
+        if generated_tokens is not None and self.args.predict_with_generate:
+            generated_tokens[:, :max(prompt_len, label_len)] = self.tokenizer.pad_token_id
             generated_tokens = generated_tokens.contiguous()
 
         return loss, generated_tokens, labels
@@ -66,16 +67,8 @@ class Seq2SeqPeftTrainer(PeftTrainer):
     ) -> torch.Tensor:
         r"""
         Pads the tensor to the same length as the target tensor.
-
-        Should only be called when predict_with_generate=True.
         """
-        if pad_token_id is None:
-            if self.tokenizer is not None and hasattr(self.tokenizer, "pad_token_id"):
-                assert self.tokenizer.padding_side == "left", "This method only accepts left-padded tensor."
-                pad_token_id = self.tokenizer.pad_token_id
-            else:
-                raise ValueError("PAD token is required.")
-
+        pad_token_id = pad_token_id if pad_token_id is not None else self.tokenizer.pad_token_id
         padded_tensor = pad_token_id * torch.ones_like(tgt_tensor)
         padded_tensor[:, -src_tensor.shape[-1]:] = src_tensor # adopt left-padding
         return padded_tensor.contiguous() # in contiguous memory

tests/template_encode.py

@@ -1,36 +0,0 @@
-# Test Template Encode
-# Usage: python .\tests\template_encode.py --model_name_and_path D:\llm\chinese-alpaca-2-7b
-#                                          --template llama2_zh --query 'how are you?'
-#                                          --history '[[\"Hello!\",\"Hi，I am llama2.\"]]'
-
-import sys
-import fire
-from typing import List, Optional, Tuple
-from transformers import AutoTokenizer
-sys.path.append("./src")
-from llmtuner.extras.template import get_template_and_fix_tokenizer
-
-
-def encode(
-        model_name_and_path: str,
-        template: str,
-        query: str,
-        resp: Optional[str] = "",
-        history: Optional[List[Tuple[str, str]]] = None,
-        system: Optional[str] = None):
-    tokenizer = AutoTokenizer.from_pretrained(
-        model_name_and_path,
-        trust_remote_code=True
-    )
-
-    template = get_template_and_fix_tokenizer(template, tokenizer)
-
-    encoded_pairs = template.encode_multiturn(tokenizer, query, resp, history, system)
-    for prompt_ids, answer_ids in encoded_pairs:
-        print("="*50)
-        print("prompt_ids: {}, answer_ids: {}".format(prompt_ids, answer_ids))
-        print("prompt decode: {}".format(tokenizer.decode(prompt_ids)))
-
-
-if __name__ == '__main__':
-    fire.Fire(encode)