reenable sdpa and fast tok by default

README.md

@@ -72,8 +72,6 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 [24/04/21] We supported **[Mixture-of-Depths](https://arxiv.org/abs/2404.02258)** according to [AstraMindAI's implementation](https://github.com/astramind-ai/Mixture-of-depths). See `examples/extras/mod` for usage.
 
-[24/04/19] We supported **Meta Llama 3** model series.
-
 [24/04/16] We supported **[BAdam](https://arxiv.org/abs/2404.02827)**. See `examples/extras/badam` for usage.
 
 [24/04/16] We supported **[unsloth](https://github.com/unslothai/unsloth)**'s long-sequence training (Llama-2-7B-56k within 24GB). It achieves **117%** speed and **50%** memory compared with FlashAttention-2, more benchmarks can be found in [this page](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison).
@@ -112,7 +110,7 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 [23/09/23] We integrated MMLU, C-Eval and CMMLU benchmarks in this repo. See [this example](#evaluation) to evaluate your models.
 
-[23/09/10] We supported **[FlashAttention-2](https://github.com/Dao-AILab/flash-attention)**. Try `--flash_attn` argument to enable FlashAttention-2 if you are using RTX4090, A100 or H100 GPUs.
+[23/09/10] We supported **[FlashAttention-2](https://github.com/Dao-AILab/flash-attention)**. Try `--flash_attn fa2` argument to enable FlashAttention-2 if you are using RTX4090, A100 or H100 GPUs.
 
 [23/08/12] We supported **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.
 

README_zh.md

@@ -72,8 +72,6 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 [24/04/21] 我们基于 [AstraMindAI 的仓库](https://github.com/astramind-ai/Mixture-of-depths)支持了 **[混合深度训练](https://arxiv.org/abs/2404.02258)**。详细用法请参照 `examples/extras/mod`。
 
-[24/04/19] 我们支持了 **Meta Llama 3** 系列模型。
-
 [24/04/16] 我们支持了 **[BAdam](https://arxiv.org/abs/2404.02827)**。详细用法请参照 `examples/extras/badam`。
 
 [24/04/16] 我们支持了 **[unsloth](https://github.com/unslothai/unsloth)** 的长序列训练（24GB 可训练 Llama-2-7B-56k）。该方法相比 FlashAttention-2 提供了 **117%** 的训练速度和 **50%** 的显存节约。更多数据请见[此页面](https://github.com/hiyouga/LLaMA-Factory/wiki/Performance-comparison)。
@@ -112,7 +110,7 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/ec36a9dd-37f4-4f72-81bd
 
 [23/09/23] 我们在项目中集成了 MMLU、C-Eval 和 CMMLU 评估集。使用方法请参阅[此示例](#模型评估)。
 
-[23/09/10] 我们支持了 **[FlashAttention-2](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。
+[23/09/10] 我们支持了 **[FlashAttention-2](https://github.com/Dao-AILab/flash-attention)**。如果您使用的是 RTX4090、A100 或 H100 GPU，请使用 `--flash_attn fa2` 参数以启用 FlashAttention-2。
 
 [23/08/12] 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。
 

requirements.txt

@@ -15,3 +15,4 @@ fastapi
 sse-starlette
 matplotlib
 fire
+packaging

src/llmtuner/extras/packages.py

@@ -1,16 +1,23 @@
 import importlib.metadata
 import importlib.util
+from typing import TYPE_CHECKING
+
+from packaging import version
+
+
+if TYPE_CHECKING:
+    from packaging.version import Version
 
 
 def _is_package_available(name: str) -> bool:
     return importlib.util.find_spec(name) is not None
 
 
-def _get_package_version(name: str) -> str:
+def _get_package_version(name: str) -> "Version":
     try:
-        return importlib.metadata.version(name)
+        return version.parse(importlib.metadata.version(name))
     except Exception:
-        return "0.0.0"
+        return version.parse("0.0.0")
 
 
 def is_fastapi_availble():
@@ -18,7 +25,7 @@ def is_fastapi_availble():
 
 
 def is_flash_attn2_available():
-    return _is_package_available("flash_attn") and _get_package_version("flash_attn").startswith("2")
+    return _is_package_available("flash_attn") and _get_package_version("flash_attn") > version.parse("2.0.0")
 
 
 def is_galore_available():
@@ -49,6 +56,10 @@ def is_rouge_available():
     return _is_package_available("rouge_chinese")
 
 
+def is_sdpa_available():
+    return _get_package_version("torch") > version.parse("2.1.1")
+
+
 def is_starlette_available():
     return _is_package_available("sse_starlette")
 

src/llmtuner/hparams/model_args.py

@@ -22,7 +22,7 @@ class ModelArguments:
         metadata={"help": "Where to store the pre-trained models downloaded from huggingface.co or modelscope.cn."},
     )
     use_fast_tokenizer: bool = field(
-        default=False,
+        default=True,
         metadata={"help": "Whether or not to use one of the fast tokenizer (backed by the tokenizers library)."},
     )
     resize_vocab: bool = field(
@@ -61,9 +61,9 @@ class ModelArguments:
         default=None,
         metadata={"help": "Which scaling strategy should be adopted for the RoPE embeddings."},
     )
-    flash_attn: bool = field(
+    flash_attn: Literal["off", "sdpa", "fa2", "auto"] = field(
-        default=False,
+        default="auto",
-        metadata={"help": "Enable FlashAttention for faster training."},
+        metadata={"help": "Enable FlashAttention for faster training and inference."},
     )
     shift_attn: bool = field(
         default=False,

src/llmtuner/model/patcher.py

@@ -15,7 +15,7 @@ from transformers.utils.versions import require_version
 from ..extras.constants import FILEEXT2TYPE, LAYERNORM_NAMES
 from ..extras.logging import get_logger
 from ..extras.misc import get_current_device, infer_optim_dtype
-from ..extras.packages import is_flash_attn2_available
+from ..extras.packages import is_flash_attn2_available, is_sdpa_available
 from ..extras.patches.llama_patch import apply_llama_patch
 from .utils import QuantizationMethod, add_z3_leaf_module, gradient_checkpointing_enable
 
@@ -62,18 +62,45 @@ def _get_quantization_dataset(tokenizer: "PreTrainedTokenizer", model_args: "Mod
 
 
 def _configure_attn_implementation(config: "PretrainedConfig", model_args: "ModelArguments") -> None:
-    if model_args.flash_attn:
+    if model_args.flash_attn == "auto":
-        if not is_flash_attn2_available():
-            logger.warning("FlashAttention2 is not installed.")
         return
 
-        logger.info("Using FlashAttention-2 for faster training and inference.")
+    elif model_args.flash_attn == "off":
+        requested_attn_implementation = "eager"
+
+    elif model_args.flash_attn == "sdpa":
+        if not is_sdpa_available():
+            logger.warning("Torch>=2.1.1 is required for SDPA attention.")
+            return
+
+        requested_attn_implementation = "sdpa"
+    elif model_args.flash_attn == "fa2":
+        if not is_flash_attn2_available():
+            logger.warning("FlashAttention-2 is not installed.")
+            return
+
+        requested_attn_implementation = "flash_attention_2"
+    else:
+        raise NotImplementedError("Unknown attention type: {}".format(model_args.flash_attn))
+
     if getattr(config, "model_type", None) == "internlm2":  # special case for custom models
-            setattr(config, "attn_implementation", "flash_attention_2")
+        setattr(config, "attn_implementation", requested_attn_implementation)
     else:
-            setattr(config, "_attn_implementation", "flash_attention_2")
+        setattr(config, "_attn_implementation", requested_attn_implementation)
+
+
+def _print_attn_implementation(config: "PretrainedConfig") -> None:
+    if getattr(config, "model_type", None) == "internlm2":  # special case for custom models
+        attn_implementation = getattr(config, "attn_implementation", None)
     else:
-        setattr(config, "_attn_implementation", "eager")
+        attn_implementation = getattr(config, "_attn_implementation", None)
+
+    if attn_implementation == "flash_attention_2":
+        logger.info("Using FlashAttention-2 for faster training and inference.")
+    elif attn_implementation == "sdpa":
+        logger.info("Using torch SDPA for faster training and inference.")
+    else:
+        logger.info("Using vanilla Attention implementation.")
 
 
 def _configure_rope(config: "PretrainedConfig", model_args: "ModelArguments", is_trainable: bool) -> None:
@@ -365,6 +392,8 @@ def patch_model(
 
         add_z3_leaf_module(model, Qwen2MoeSparseMoeBlock)
 
+    _print_attn_implementation(model.config)
+
     try:
         model.add_model_tags(["llama-factory"])
     except Exception:

src/llmtuner/webui/components/top.py

@@ -33,7 +33,7 @@ def create_top() -> Dict[str, "Component"]:
             quantization_bit = gr.Dropdown(choices=["none", "8", "4"], value="none")
             template = gr.Dropdown(choices=list(templates.keys()), value="default")
             rope_scaling = gr.Radio(choices=["none", "linear", "dynamic"], value="none")
-            booster = gr.Radio(choices=["none", "flashattn", "unsloth"], value="none")
+            booster = gr.Radio(choices=["none", "flashattn2", "unsloth"], value="none")
 
     model_name.change(list_adapters, [model_name, finetuning_type], [adapter_path], queue=False).then(
         get_model_path, [model_name], [model_path], queue=False

src/llmtuner/webui/runner.py

@@ -67,7 +67,7 @@ class Runner:
         if not model_path:
             return ALERTS["err_no_path"][lang]
 
-        if len(dataset) == 0:
+        if not dataset:
             return ALERTS["err_no_dataset"][lang]
 
         if not from_preview and self.demo_mode:
@@ -122,7 +122,7 @@ class Runner:
             quantization_bit=int(get("top.quantization_bit")) if get("top.quantization_bit") in ["8", "4"] else None,
             template=get("top.template"),
             rope_scaling=get("top.rope_scaling") if get("top.rope_scaling") in ["linear", "dynamic"] else None,
-            flash_attn=(get("top.booster") == "flashattn"),
+            flash_attn="fa2" if get("top.booster") == "flashattn2" else "auto",
             use_unsloth=(get("top.booster") == "unsloth"),
             dataset_dir=get("train.dataset_dir"),
             dataset=",".join(get("train.dataset")),