diff --git a/pyproject.toml b/pyproject.toml
index 8c60cb5866..160182c2a2 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -37,7 +37,7 @@ triton = [
 ]
 
 huggingface = [
-    "unsloth_zoo>=2025.8.9",
+    "unsloth_zoo>=2025.9.1",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1",
@@ -453,7 +453,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3",
 ]
 colab-new = [
-    "unsloth_zoo>=2025.8.9",
+    "unsloth_zoo>=2025.9.1",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.47.0,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1",
diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index 1b2a9310ff..25a54165b7 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -214,7 +214,7 @@ elif DEVICE_TYPE == "xpu":
 # Check for unsloth_zoo
 try:
     unsloth_zoo_version = importlib_version("unsloth_zoo")
-    if Version(unsloth_zoo_version) < Version("2025.8.8"):
+    if Version(unsloth_zoo_version) < Version("2025.9.1"):
         print(
             "Unsloth: Please update Unsloth and Unsloth-Zoo to the latest version!\n"\
             "Do this via `pip install --upgrade --force-reinstall --no-cache-dir --no-deps unsloth unsloth_zoo`"
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index cf2ca75f75..f978060c9c 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -1236,7 +1236,7 @@ def CausalLM_fast_forward(fast_forward_inference):
             # < 1024 Normal Unsloth uses less VRAM!
             if bsz*q_len <= 1024: RETURN_LOGITS = True
 
-            if not RETURN_LOGITS and HAS_CUT_CROSS_ENTROPY and labels is not None:
+            if not RETURN_LOGITS and labels is not None:
 
                 n_items = kwargs.get("num_items_in_batch", None) or kwargs.get("n_items", None)
 
@@ -1259,7 +1259,7 @@ def CausalLM_fast_forward(fast_forward_inference):
                     mask                 = None,
                     n_items              = n_items,
                     scaling              = getattr(self, "accelerator_scaler", None),
-                    target_gb            = 1,
+                    target_gb            = None,
                     torch_compile        = True,
                     logit_softcapping    = logit_softcapping,
                 )
diff --git a/unsloth/models/mistral.py b/unsloth/models/mistral.py
index 6274f2e5df..faab2d30b1 100644
--- a/unsloth/models/mistral.py
+++ b/unsloth/models/mistral.py
@@ -300,17 +300,30 @@ def MistralForCausalLM_fast_forward(
         # < 1024 Normal Unsloth uses less VRAM!
         if bsz * q_len <= 1024: RETURN_LOGITS = True
 
-        if not RETURN_LOGITS and HAS_CUT_CROSS_ENTROPY and os.environ.get("UNSLOTH_ENABLE_CCE", "1") != "0" and labels is not None:
+        if not RETURN_LOGITS and labels is not None:
             n_items = kwargs.get("num_items_in_batch", None) or kwargs.get("n_items", None)
             logit_softcapping = getattr(self.config, "final_logit_softcapping", 0)
-            loss = fused_linear_cross_entropy(
-                hidden_states = hidden_states,
-                lm_weight = lm_head,
-                labels = labels,
-                num_items_in_batch = n_items,
-                logit_softcapping = logit_softcapping,
-            )
 
+            # loss = fused_linear_cross_entropy(
+            #     hidden_states = hidden_states,
+            #     lm_weight = lm_head,
+            #     labels = labels,
+            #     num_items_in_batch = n_items,
+            #     logit_softcapping = logit_softcapping,
+            # )
+            loss = unsloth_fused_ce_loss(
+                trainer              = None,
+                hidden_states        = hidden_states,
+                lm_head_weight       = lm_head,
+                lm_head_bias         = None,
+                labels               = labels,
+                mask                 = None,
+                n_items              = n_items,
+                scaling              = getattr(self, "accelerator_scaler", None),
+                target_gb            = None,
+                torch_compile        = True,
+                logit_softcapping    = logit_softcapping,
+            )
             if not return_dict:
                 output = (logits,) + outputs[1:]
                 return (loss,) + output if loss is not None else output
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index fe9951ddf0..f342a4d86b 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -513,7 +513,7 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
         "fp16"                          : False,
         "include_tokens_per_second"     : False,
         "include_num_input_tokens_seen" : False,
-        "auto_find_batch_size"          : True, # Auto /2 batch size
+        "auto_find_batch_size"          : False, # Auto /2 batch size - too many people complained so removing
         "dataloader_pin_memory"         : True,
         # Might fail so disable for now
         # "dataloader_persistent_workers" : True, # Keeps dataloader in RAM