[trainer] fix pt loss (#7748)

* fix pt loss * robust * fix * test
2025-08-01 11:12:50 +08:00 · 2025-04-17 03:15:35 +08:00 · 2025-04-17 03:15:35 +08:00 · 610f164c69
commit 610f164c69
parent 0a0cfeb782
12 changed files with 34 additions and 34 deletions
--- a/README.md
+++ b/README.md
@ -107,20 +107,22 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/

 ## Changelog

+[25/04/16] We supported fine-tuning the **[InternVL3](https://huggingface.co/OpenGVLab/InternVL3-8B)** model. See [PR #7258](https://github.com/hiyouga/LLaMA-Factory/pull/7258) to get started.
+
 [25/04/14] We supported fine-tuning the **[GLM-Z1](https://huggingface.co/THUDM/GLM-Z1-9B-0414)** and **[Kimi-VL](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct)** models.

 [25/04/06] We supported fine-tuning the **[Llama 4](https://ai.meta.com/blog/llama-4-multimodal-intelligence/)** model. See [PR #7611](https://github.com/hiyouga/LLaMA-Factory/pull/7611) to get started.

 [25/03/31] We supported fine-tuning the **[Qwen2.5 Omni](https://qwenlm.github.io/blog/qwen2.5-omni/)** model. See [PR #7537](https://github.com/hiyouga/LLaMA-Factory/pull/7537) to get started.

+<details><summary>Full Changelog</summary>
+
 [25/03/15] We supported **[SGLang](https://github.com/sgl-project/sglang)** as inference backend. Try `infer_backend: sglang` to accelerate inference.

 [25/03/12] We supported fine-tuning the **[Gemma 3](https://huggingface.co/blog/gemma3)** model.

 [25/02/24] Announcing **[EasyR1](https://github.com/hiyouga/EasyR1)**, an efficient, scalable and multi-modality RL training framework for efficient GRPO training.

-<details><summary>Full Changelog</summary>
-
 [25/02/11] We supported saving the **[Ollama](https://github.com/ollama/ollama)** modelfile when exporting the model checkpoints. See [examples](examples/README.md) for usage.

 [25/02/05] We supported fine-tuning the **[Qwen2-Audio](Qwen/Qwen2-Audio-7B-Instruct)** and **[MiniCPM-o-2.6](https://huggingface.co/openbmb/MiniCPM-o-2_6)** on audio understanding tasks.
@ -247,7 +249,7 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 | [Hunyuan](https://huggingface.co/tencent/)                        | 7B                               | hunyuan             |
 | [Index](https://huggingface.co/IndexTeam)                         | 1.9B                             | index               |
 | [InternLM 2-3](https://huggingface.co/internlm)                   | 7B/8B/20B                        | intern2             |
-| [InternVL2.5-3](https://huggingface.co/OpenGVLab/InternVL)\*\*    | 1B/2B/4B/8B/9B/14B/26B/38B/78B   | intern_vl           |
+| [InternVL 2.5-3](https://huggingface.co/OpenGVLab)\*\*            | 1B/2B/4B/8B/9B/14B/26B/38B/78B   | intern_vl           |
 | [Kimi-VL](https://huggingface.co/moonshotai)                      | 16B                              | kimi_vl             |
 | [Llama](https://github.com/facebookresearch/llama)                | 7B/13B/33B/65B                   | -                   |
 | [Llama 2](https://huggingface.co/meta-llama)                      | 7B/13B/70B                       | llama2              |
--- a/README_zh.md
+++ b/README_zh.md
@ -110,20 +110,22 @@ https://github.com/user-attachments/assets/43b700c6-a178-41db-b1f8-8190a5d3fcfc

 ## 更新日志

+[25/04/16] 我们支持了 **[InternVL3](https://huggingface.co/OpenGVLab/InternVL3-8B)** 模型的微调。查看 [PR #7258](https://github.com/hiyouga/LLaMA-Factory/pull/7258) 以使用。
+
 [25/04/14] 我们支持了 **[GLM-Z1](https://huggingface.co/THUDM/GLM-Z1-9B-0414)** 和 **[Kimi-VL](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct)** 模型的微调。

 [25/04/06] 我们支持了 **[Llama 4](https://ai.meta.com/blog/llama-4-multimodal-intelligence/)** 模型的微调。查看 [PR #7611](https://github.com/hiyouga/LLaMA-Factory/pull/7611) 以使用。

 [25/03/31] 我们支持了 **[Qwen2.5 Omni](https://qwenlm.github.io/blog/qwen2.5-omni/)** 模型的微调。查看 [PR #7537](https://github.com/hiyouga/LLaMA-Factory/pull/7537) 以使用。

+<details><summary>展开日志</summary>
+
 [25/03/15] 我们支持了 **[SGLang](https://github.com/sgl-project/sglang)** 推理后端，请使用 `infer_backend: sglang` 启用。

 [25/03/12] 我们支持了 **[Gemma 3](https://huggingface.co/blog/gemma3)** 模型的微调。

 [25/02/24] 我们宣布开源 **[EasyR1](https://github.com/hiyouga/EasyR1)**，一个高效可扩展的多模态强化学习框架，支持高效的 GRPO 训练。

-<details><summary>展开日志</summary>
-
 [25/02/11] 我们支持了在导出模型时保存 **[Ollama](https://github.com/ollama/ollama)** 配置文件。详细用法请参照 [examples](examples/README_zh.md)。

 [25/02/05] 我们支持了在语音理解任务上微调 **[Qwen2-Audio](Qwen/Qwen2-Audio-7B-Instruct)** 和 **[MiniCPM-o-2.6](https://huggingface.co/openbmb/MiniCPM-o-2_6)** 模型。
@ -250,7 +252,7 @@ https://github.com/user-attachments/assets/43b700c6-a178-41db-b1f8-8190a5d3fcfc
 | [Hunyuan](https://huggingface.co/tencent/)                        | 7B                               | hunyuan             |
 | [Index](https://huggingface.co/IndexTeam)                         | 1.9B                             | index               |
 | [InternLM 2-3](https://huggingface.co/internlm)                   | 7B/8B/20B                        | intern2             |
-| [InternVL2.5-3](https://huggingface.co/OpenGVLab/InternVL)\*\*    | 1B/2B/4B/8B/9B/14B/26B/38B/78B   | intern_vl           |
+| [InternVL 2.5-3](https://huggingface.co/OpenGVLab)\*\*            | 1B/2B/4B/8B/9B/14B/26B/38B/78B   | intern_vl           |
 | [Kimi-VL](https://huggingface.co/moonshotai)                      | 16B                              | kimi_vl             |
 | [Llama](https://github.com/facebookresearch/llama)                | 7B/13B/33B/65B                   | -                   |
 | [Llama 2](https://huggingface.co/meta-llama)                      | 7B/13B/70B                       | llama2              |
--- a/assets/wechat.jpg
+++ b/assets/wechat.jpg
--- a/assets/wechat_npu.jpg
+++ b/assets/wechat_npu.jpg
--- a/src/llamafactory/init.py
+++ b/src/llamafactory/init.py
@ -17,16 +17,8 @@ r"""Efficient fine-tuning of large language models.
 Level:
  api, webui > chat, eval, train > data, model > hparams > extras

-Dependency graph:
-  transformers>=4.41.2,<=4.43.0,!=4.46.*,!=4.47.*,!=4.48.0
-  datasets>=2.16.0,<=3.5.0
-  accelerate>=0.34.0,<=1.6.0
-  peft>=0.14.0,<=0.15.1
-  trl>=0.8.6,<=0.9.6
-
 Disable version checking: DISABLE_VERSION_CHECK=1
 Enable VRAM recording: RECORD_VRAM=1
-Force check imports: FORCE_CHECK_IMPORTS=1
 Force using torchrun: FORCE_TORCHRUN=1
 Set logging verbosity: LLAMAFACTORY_VERBOSITY=WARN
 Use modelscope: USE_MODELSCOPE_HUB=1
--- a/src/llamafactory/data/mm_plugin.py
+++ b/src/llamafactory/data/mm_plugin.py
@ -21,7 +21,7 @@ import re
 from copy import deepcopy
 from dataclasses import dataclass
 from io import BytesIO
-from typing import TYPE_CHECKING, BinaryIO, Literal, Optional, TypedDict, Union
+from typing import TYPE_CHECKING, Any, BinaryIO, Literal, Optional, TypedDict, Union

 import numpy as np
 import torch
@ -86,7 +86,7 @@ if TYPE_CHECKING:
            pass


-def _concatenate_list(input_list):
+def _concatenate_list(input_list: list[Any]) -> Union[list[Any], "NDArray", "torch.Tensor"]:
    r"""Concatenate a list of lists, numpy arrays or torch tensors.

    Returns:
--- a/src/llamafactory/extras/misc.py
+++ b/src/llamafactory/extras/misc.py
@ -89,7 +89,7 @@ def check_version(requirement: str, mandatory: bool = False) -> None:

 def check_dependencies() -> None:
    r"""Check the version of the required packages."""
-    check_version("transformers>=4.43.0,<=4.51.3,!=4.46.0,!=4.46.1,!=4.46.2,!=4.46.3,!=4.47.0,!=4.47.1,!=4.48.0")
+    check_version("transformers>=4.45.0,<=4.51.3,!=4.46.0,!=4.46.1,!=4.46.2,!=4.46.3,!=4.47.0,!=4.47.1,!=4.48.0")
    check_version("datasets>=2.16.0,<=3.5.0")
    check_version("accelerate>=0.34.0,<=1.6.0")
    check_version("peft>=0.14.0,<=0.15.1")
--- a/src/llamafactory/model/model_utils/longlora.py
+++ b/src/llamafactory/model/model_utils/longlora.py
@ -23,7 +23,6 @@ from typing import TYPE_CHECKING, Optional
 import torch
 import torch.nn as nn
 import transformers
-from transformers.models.llama.modeling_llama import Cache, apply_rotary_pos_emb, repeat_kv

 from ...extras import logging
 from ...extras.constants import SUPPORTED_CLASS_FOR_S2ATTN
@ -32,7 +31,15 @@ from ...extras.packages import is_transformers_version_greater_than


 if not is_transformers_version_greater_than("4.48.0"):
-    from transformers.models.llama.modeling_llama import LlamaAttention, LlamaFlashAttention2, LlamaSdpaAttention
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+    from transformers.models.llama.modeling_llama import (
+        Cache,
+        LlamaAttention,
+        LlamaFlashAttention2,
+        LlamaSdpaAttention,
+        apply_rotary_pos_emb,
+        repeat_kv,
+    )


 if TYPE_CHECKING:
@ -206,9 +213,6 @@ def llama_flash_attention_2_forward(
        if attention_mask is not None:
            attention_mask = attention_mask[:, :groupsz].repeat(num_groups, 1)

-    if is_transformers_version_greater_than("4.43.0"):
-        from transformers.modeling_flash_attention_utils import _flash_attention_forward
-
        attn_output: torch.Tensor = _flash_attention_forward(
            query_states,
            key_states,
@ -220,10 +224,6 @@ def llama_flash_attention_2_forward(
            use_top_left_mask=self._flash_attn_uses_top_left_mask,
            is_causal=self.is_causal,
        )
-    else:
-        attn_output: torch.Tensor = self._flash_attention_forward(
-            query_states, key_states, value_states, attention_mask, query_states.size(1), dropout=dropout_rate
-        )

    if getattr(self.config, "group_size_ratio", None) and self.training:  # shift back
        attn_output.reshape(bsz, q_len, self.num_heads, self.head_dim)
@ -350,7 +350,7 @@ def llama_sdpa_attention_forward(


 def _apply_llama_patch() -> None:
-    check_version("transformers>=4.43.0,<4.48.0", mandatory=True)
+    check_version("transformers>=4.45.0,<4.48.0", mandatory=True)
    LlamaAttention.forward = llama_attention_forward
    LlamaFlashAttention2.forward = llama_flash_attention_2_forward
    LlamaSdpaAttention.forward = llama_sdpa_attention_forward
--- a/src/llamafactory/model/model_utils/packing.py
+++ b/src/llamafactory/model/model_utils/packing.py
@ -43,11 +43,6 @@ import torch
 import torch.nn.functional as F

 from ...extras import logging
-from ...extras.packages import is_transformers_version_greater_than
-
-
-if is_transformers_version_greater_than("4.43.0"):
-    import transformers.modeling_flash_attention_utils


 if TYPE_CHECKING:
@ -116,5 +111,7 @@ def configure_packing(model_args: "ModelArguments", is_trainable: bool) -> None:
    if not is_trainable or not model_args.block_diag_attn:
        return

+    import transformers.modeling_flash_attention_utils
+
    transformers.modeling_flash_attention_utils._get_unpad_data = get_unpad_data
    logger.info_rank0("Using block diagonal attention for sequence packing without cross-attention.")
--- a/src/llamafactory/train/pt/trainer.py
+++ b/src/llamafactory/train/pt/trainer.py
@ -40,6 +40,11 @@ class CustomTrainer(Trainer):
            kwargs["processing_class"] = kwargs.pop("tokenizer")

        super().__init__(**kwargs)
+        if processor is not None:
+            # avoid wrong loss under gradient accumulation
+            # https://github.com/huggingface/transformers/pull/36044#issuecomment-2746657112
+            self.model_accepts_loss_kwargs = False
+
        self.finetuning_args = finetuning_args

        if processor is not None:
--- a/src/llamafactory/train/sft/trainer.py
+++ b/src/llamafactory/train/sft/trainer.py
@ -60,6 +60,8 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):

        super().__init__(**kwargs)
        if processor is not None:
+            # avoid wrong loss under gradient accumulation
+            # https://github.com/huggingface/transformers/pull/36044#issuecomment-2746657112
            self.model_accepts_loss_kwargs = False

        self.finetuning_args = finetuning_args
--- a/tests/version.txt
+++ b/tests/version.txt
@ -1,2 +1,2 @@
-# change if test fails
-0.9.3.103
+# change if test fails or cache is outdated
+0.9.3.103