[breaking] bump transformers to 4.45.0 & improve ci (#7746)

* update ci

* fix

* fix

* fix

* fix

* fix

src/llamafactory/model/model_utils/attention.py

@@ -18,7 +18,6 @@ from transformers.utils import is_flash_attn_2_available, is_torch_sdpa_availabl
 
 from ...extras import logging
 from ...extras.constants import AttentionFunction
-from ...extras.misc import check_version
 
 
 if TYPE_CHECKING:
@@ -36,8 +35,6 @@ def configure_attn_implementation(
     if getattr(config, "model_type", None) == "gemma2" and is_trainable:
         if model_args.flash_attn == AttentionFunction.AUTO or model_args.flash_attn == AttentionFunction.FA2:
             if is_flash_attn_2_available():
-                check_version("transformers>=4.42.4")
-                check_version("flash_attn>=2.6.3")
                 if model_args.flash_attn != AttentionFunction.FA2:
                     logger.warning_rank0("Gemma 2 should use flash attention 2, change `flash_attn` to fa2.")
                     model_args.flash_attn = AttentionFunction.FA2

src/llamafactory/model/model_utils/longlora.py

@@ -350,7 +350,7 @@ def llama_sdpa_attention_forward(
 
 
 def _apply_llama_patch() -> None:
-    check_version("transformers>=4.41.2,<4.48.0")
+    check_version("transformers>=4.43.0,<4.48.0", mandatory=True)
     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

@@ -43,7 +43,6 @@ import torch
 import torch.nn.functional as F
 
 from ...extras import logging
-from ...extras.misc import check_version
 from ...extras.packages import is_transformers_version_greater_than
 
 
@@ -117,6 +116,5 @@ def configure_packing(model_args: "ModelArguments", is_trainable: bool) -> None:
     if not is_trainable or not model_args.block_diag_attn:
         return
 
-    check_version("transformers>=4.43.0")
     transformers.modeling_flash_attention_utils._get_unpad_data = get_unpad_data
     logger.info_rank0("Using block diagonal attention for sequence packing without cross-attention.")