src/llamafactory/extras/packages.py

@@ -110,10 +110,6 @@ def is_starlette_available():
 def is_transformers_version_greater_than(content: str):
     return _get_package_version("transformers") >= version.parse(content)
 
-@lru_cache
-def is_torch_version_greater_than(content: str):
-    return _get_package_version("torch") >= version.parse(content)
-
 
 def is_uvicorn_available():
     return _is_package_available("uvicorn")

src/llamafactory/model/model_utils/attention.py

@@ -16,7 +16,6 @@ from typing import TYPE_CHECKING
 
 from ...extras import logging
 from ...extras.constants import AttentionFunction
-from ...extras.packages import is_torch_version_greater_than
 
 
 if TYPE_CHECKING:
@@ -52,14 +51,15 @@ def configure_attn_implementation(config: "PretrainedConfig", model_args: "Model
         requested_attn_implementation = "eager"
 
     elif model_args.flash_attn == AttentionFunction.SDPA:
-        if not is_torch_version_greater_than("2.1.1"):
+        from transformers.utils import is_torch_sdpa_available
+
+        if not is_torch_sdpa_available():
             logger.warning_rank0("torch>=2.1.1 is required for SDPA attention.")
             return
 
         requested_attn_implementation = "sdpa"
     elif model_args.flash_attn == AttentionFunction.FA2:
-        from transformers import is_torch_npu_available
-        if not (is_flash_attn_2_available() or is_torch_npu_available()):
+        if not is_flash_attn_2_available():
             logger.warning_rank0("FlashAttention-2 is not installed.")
             return
 

src/llamafactory/model/model_utils/visual.py

@@ -355,7 +355,7 @@ _register_composite_model(
 _register_composite_model(
     model_type="qwen3_vl",
     projector_key="visual.merger",
-    vision_model_keys=["visual.patch_embed", "visual.blocks", "visual.deepstack_merger_list"],
+    vision_model_keys=["visual.patch_embed", "visual.blocks"],
     language_model_keys=["language_model", "lm_head"],
     lora_conflict_keys=["patch_embed"],
 )
@@ -364,7 +364,7 @@ _register_composite_model(
 _register_composite_model(
     model_type="qwen3_vl_moe",
     projector_key="visual.merger",
-    vision_model_keys=["visual.patch_embed", "visual.blocks", "visual.deepstack_merger_list"],
+    vision_model_keys=["visual.patch_embed", "visual.blocks"],
     language_model_keys=["language_model", "lm_head"],
     lora_conflict_keys=["patch_embed"],
 )
@@ -373,7 +373,7 @@ _register_composite_model(
 _register_composite_model(
     model_type="qwen3_omni_moe_thinker",
     projector_key="visual.merger",
-    vision_model_keys=["visual.patch_embed", "visual.blocks", "visual.deepstack_merger_list", "audio_tower"],
+    vision_model_keys=["visual.patch_embed", "visual.blocks", "audio_tower"],
     language_model_keys=["model", "lm_head"],
     lora_conflict_keys=["patch_embed"],
 )

src/llamafactory/train/dpo/trainer.py

@@ -203,7 +203,7 @@ class CustomDPOTrainer(DPOTrainer):
                 bco_losses = self.bco_loss(
                     policy_chosen_logps, policy_rejected_logps, reference_chosen_logps, reference_rejected_logps
                 )
-                losses = (losses + bco_losses * self.bco_gemma) / (1.0 + self.bco_gemma)  # re-weight W_p and W_q
+                losses += bco_losses * self.bco_gemma
 
         return losses, chosen_rewards, rejected_rewards
 
@@ -284,6 +284,9 @@ class CustomDPOTrainer(DPOTrainer):
         sft_loss = -policy_chosen_logps_avg
         if self.ftx_gamma > 1e-6:
             losses += self.ftx_gamma * sft_loss
+            if self.bco_gemma > 1e-6:
+                # re-weigthing for MPO
+                losses /= self.ftx_gamma + self.bco_gemma + 1.0
 
         prefix = "eval_" if train_eval == "eval" else ""
         metrics[f"{prefix}rewards/chosen"] = chosen_rewards.mean().item()