Merge branch 'hiyouga:main' into pixtral-patch

Former-commit-id: 93a441a6b746e9a933dad8c45553fb5b68bf2b34

README.md

@@ -30,6 +30,9 @@ Choose your path:
 - **Local machine**: Please refer to [usage](#getting-started)
 - **Documentation (WIP)**: https://llamafactory.readthedocs.io/zh-cn/latest/
 
+> [!NOTE]
+> Except for the above links, all other websites are unauthorized third-party websites. Please carefully use them.
+
 ## Table of Contents
 
 - [Features](#features)

README_zh.md

@@ -31,6 +31,9 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272
 - **入门教程**：https://zhuanlan.zhihu.com/p/695287607
 - **框架文档**：https://llamafactory.readthedocs.io/zh-cn/latest/
 
+> [!NOTE]
+> 除上述链接以外的其他网站均为未经许可的第三方网站，请小心甄别。
+
 ## 目录
 
 - [项目特色](#项目特色)

requirements.txt

@@ -1,4 +1,4 @@
-transformers>=4.41.2,<=4.45.0
+transformers>=4.41.2,<=4.45.2
 datasets>=2.16.0,<=2.21.0
 accelerate>=0.30.1,<=0.34.2
 peft>=0.11.1,<=0.12.0

setup.py

@@ -54,7 +54,7 @@ extra_require = {
     "gptq": ["optimum>=1.17.0", "auto-gptq>=0.5.0"],
     "awq": ["autoawq"],
     "aqlm": ["aqlm[gpu]>=1.1.0"],
-    "vllm": ["vllm>=0.4.3,<=0.6.0"],
+    "vllm": ["vllm>=0.4.3,<=0.6.2"],
     "galore": ["galore-torch"],
     "badam": ["badam>=1.2.1"],
     "adam-mini": ["adam-mini"],

src/llamafactory/__init__.py

@@ -20,7 +20,7 @@ Level:
 
 Dependency graph:
   main:
-    transformers>=4.41.2,<=4.45.0
+    transformers>=4.41.2,<=4.45.2
     datasets>=2.16.0,<=2.21.0
     accelerate>=0.30.1,<=0.34.2
     peft>=0.11.1,<=0.12.0
@@ -28,9 +28,9 @@ Dependency graph:
   attention:
     transformers>=4.42.4 (gemma+fa2)
   longlora:
-    transformers>=4.41.2,<=4.45.0
+    transformers>=4.41.2,<=4.45.2
   packing:
-    transformers>=4.41.2,<=4.45.0
+    transformers>=4.41.2,<=4.45.2
 
 Disable version checking: DISABLE_VERSION_CHECK=1
 Enable VRAM recording: RECORD_VRAM=1

src/llamafactory/data/template.py

@@ -357,9 +357,7 @@ def get_template_and_fix_tokenizer(tokenizer: "PreTrainedTokenizer", data_args:
     Gets chat template and fixes the tokenizer.
     """
     if data_args.template in ["llava", "paligemma", "qwen2_vl"]:
-        require_version(
+        require_version("transformers>=4.45.0", "To fix: pip install transformers>=4.45.0")
-            "transformers>=4.45.0.dev0", "To fix: pip install git+https://github.com/huggingface/transformers.git"
-        )
         require_version("accelerate>=0.34.0", "To fix: pip install accelerate>=0.34.0")
 
     if data_args.template is None:

src/llamafactory/extras/misc.py

@@ -79,7 +79,7 @@ def check_dependencies() -> None:
     if os.environ.get("DISABLE_VERSION_CHECK", "0").lower() in ["true", "1"]:
         logger.warning("Version checking has been disabled, may lead to unexpected behaviors.")
     else:
-        require_version("transformers>=4.41.2,<=4.45.0", "To fix: pip install transformers>=4.41.2,<=4.45.0")
+        require_version("transformers>=4.41.2,<=4.45.2", "To fix: pip install transformers>=4.41.2,<=4.45.2")
         require_version("datasets>=2.16.0,<=2.21.0", "To fix: pip install datasets>=2.16.0,<=2.21.0")
         require_version("accelerate>=0.30.1,<=0.34.2", "To fix: pip install accelerate>=0.30.1,<=0.34.2")
         require_version("peft>=0.11.1,<=0.12.0", "To fix: pip install peft>=0.11.1,<=0.12.0")

src/llamafactory/hparams/parser.py

@@ -57,7 +57,7 @@ def _parse_args(parser: "HfArgumentParser", args: Optional[Dict[str, Any]] = Non
     if args is not None:
         return parser.parse_dict(args)
 
-    if len(sys.argv) == 2 and sys.argv[1].endswith(".yaml"):
+    if len(sys.argv) == 2 and (sys.argv[1].endswith(".yaml") or sys.argv[1].endswith(".yml")):
         return parser.parse_yaml_file(os.path.abspath(sys.argv[1]))
 
     if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
@@ -123,7 +123,7 @@ def _check_extra_dependencies(
         require_version("mixture-of-depth>=1.1.6", "To fix: pip install mixture-of-depth>=1.1.6")
 
     if model_args.infer_backend == "vllm":
-        require_version("vllm>=0.4.3,<=0.6.0", "To fix: pip install vllm>=0.4.3,<=0.6.0")
+        require_version("vllm>=0.4.3,<=0.6.2", "To fix: pip install vllm>=0.4.3,<=0.6.2")
 
     if finetuning_args.use_galore:
         require_version("galore_torch", "To fix: pip install galore_torch")

src/llamafactory/model/loader.py

@@ -82,6 +82,8 @@ def load_tokenizer(model_args: "ModelArguments") -> "TokenizerModule":
             padding_side="right",
             **init_kwargs,
         )
+    except Exception as e:
+        raise OSError("Failed to load tokenizer.") from e
 
     if model_args.new_special_tokens is not None:
         num_added_tokens = tokenizer.add_special_tokens(
@@ -97,12 +99,13 @@ def load_tokenizer(model_args: "ModelArguments") -> "TokenizerModule":
     try:
         processor = AutoProcessor.from_pretrained(model_args.model_name_or_path, **init_kwargs)
         patch_processor(processor, config, tokenizer, model_args)
-    except Exception:
+    except Exception as e:
+        logger.warning("Processor was not found: {}.".format(e))
         processor = None
 
     # Avoid load tokenizer, see:
     # https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/models/auto/processing_auto.py#L324
-    if "Processor" not in processor.__class__.__name__:
+    if processor is not None and "Processor" not in processor.__class__.__name__:
         processor = None
 
     return {"tokenizer": tokenizer, "processor": processor}

src/llamafactory/model/model_utils/longlora.py

@@ -353,7 +353,7 @@ def llama_sdpa_attention_forward(
 
 
 def _apply_llama_patch() -> None:
-    require_version("transformers>=4.41.2,<=4.45.0", "To fix: pip install transformers>=4.41.2,<=4.45.0")
+    require_version("transformers>=4.41.2,<=4.45.2", "To fix: pip install transformers>=4.41.2,<=4.45.2")
     LlamaAttention.forward = llama_attention_forward
     LlamaFlashAttention2.forward = llama_flash_attention_2_forward
     LlamaSdpaAttention.forward = llama_sdpa_attention_forward

src/llamafactory/model/model_utils/packing.py

@@ -114,7 +114,7 @@ def get_unpad_data(attention_mask: "torch.Tensor") -> Tuple["torch.Tensor", "tor
 
 
 def _patch_for_block_diag_attn(model_type: str) -> None:
-    require_version("transformers>=4.41.2,<=4.45.0", "To fix: pip install transformers>=4.41.2,<=4.45.0")
+    require_version("transformers>=4.41.2,<=4.45.2", "To fix: pip install transformers>=4.41.2,<=4.45.2")
     if is_transformers_version_greater_than_4_43():
         import transformers.modeling_flash_attention_utils
 

src/llamafactory/model/patcher.py

@@ -110,6 +110,9 @@ def patch_config(
     if getattr(config, "model_type", None) == "qwen2" and is_trainable and model_args.flash_attn == "fa2":
         setattr(config, "use_cache", False)  # qwen2 does not support use_cache when using flash attn
 
+    if "LlavaLlamaForCausalLM" in getattr(config, "architectures", []):
+        raise ValueError("Please download llava models with hf-compatible format: https://huggingface.co/llava-hf")
+
     # deepspeed zero3 is not compatible with low_cpu_mem_usage
     init_kwargs["low_cpu_mem_usage"] = model_args.low_cpu_mem_usage and (not is_deepspeed_zero3_enabled())
 

src/llamafactory/train/tuner.py

@@ -139,5 +139,5 @@ def export_model(args: Optional[Dict[str, Any]] = None) -> None:
                     model_args.export_hub_model_id, token=model_args.hf_hub_token
                 )
 
-    except Exception:
+    except Exception as e:
-        logger.warning("Cannot save tokenizer, please copy the files manually.")
+        logger.warning("Cannot save tokenizer, please copy the files manually: {}.".format(e))