From 7ecb049ac9d0d13525d17950dbe251e6ab5655eb Mon Sep 17 00:00:00 2001
From: Daniel Han-Chen <danielhanchen@users.noreply.github.com>
Date: Sun, 8 Feb 2026 16:20:03 +0000
Subject: [PATCH 1/3] Make bitsandbytes optional on ROCm and add bf16 helper

---
 unsloth/__init__.py               |  6 +++++
 unsloth/import_fixes.py           |  5 +++-
 unsloth/kernels/utils.py          | 43 +++++++++++++++++++------------
 unsloth/models/_utils.py          | 11 ++++++--
 unsloth/models/granite.py         | 17 +++++++++---
 unsloth/models/llama.py           |  8 ++++++
 unsloth/models/rl_replacements.py |  7 +++++
 unsloth/save.py                   | 15 ++++++++---
 8 files changed, 87 insertions(+), 25 deletions(-)

diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index 4357ad63aa..19b433a680 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -205,6 +205,12 @@ elif DEVICE_TYPE == "xpu":
     # set SUPPORTS_BFLOAT16 as torch.xpu.is_bf16_supported()
     SUPPORTS_BFLOAT16 = torch.xpu.is_bf16_supported()
 
+# Backwards compatibility: some notebooks import `unsloth.is_bf16_supported`.
+# Ensure it exists on all backends (HIP / XPU) and has a stable signature.
+if "is_bf16_supported" not in globals():
+    def is_bf16_supported(including_emulation = False):
+        return SUPPORTS_BFLOAT16
+
 # For Gradio HF Spaces?
 # if "SPACE_AUTHOR_NAME" not in os.environ and "SPACE_REPO_NAME" not in os.environ:
 import triton
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index 97e74dfb57..af90de1bdb 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -309,7 +309,10 @@ def fix_vllm_aimv2_issue():
     spec = importlib.util.find_spec("vllm")
     if spec is None:
         return
-    vllm_version = importlib_version("vllm")
+    try:
+        vllm_version = importlib_version("vllm")
+    except Exception:
+        return
     if Version(vllm_version) < Version("0.10.1"):
         vllm_location = spec.origin
         if vllm_location is None:
diff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.py
index 5dcc7c232c..63b1966b32 100644
--- a/unsloth/kernels/utils.py
+++ b/unsloth/kernels/utils.py
@@ -108,11 +108,14 @@ def calculate_settings(
 
 
 HAS_CUDA_STREAM = False
-import bitsandbytes as bnb
-
-# https://github.com/bitsandbytes-foundation/bitsandbytes/pull/1330/files
-HAS_CUDA_STREAM = Version(bnb.__version__) > Version("0.43.3")
-get_ptr = bnb.functional.get_ptr
+try:
+    import bitsandbytes as bnb
+    # https://github.com/bitsandbytes-foundation/bitsandbytes/pull/1330/files
+    HAS_CUDA_STREAM = Version(bnb.__version__) > Version("0.43.3")
+    get_ptr = bnb.functional.get_ptr
+except Exception:
+    bnb = None
+    get_ptr = None
 
 if DEVICE_TYPE == "xpu":
     HAS_XPU_STREAM = True
@@ -180,21 +183,29 @@ else:
     CUDA_STREAMS = tuple(CUDA_STREAMS)
     del _CUDA_STREAMS
 
-# Bitsandbytes operations
+# Bitsandbytes operations (optional)
 ctypes_c_int = ctypes.c_int
 ctypes_c_int32 = ctypes.c_int32
-cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
-cdequantize_blockwise_fp16_nf4 = bnb.functional.lib.cdequantize_blockwise_fp16_nf4
-cdequantize_blockwise_bf16_nf4 = bnb.functional.lib.cdequantize_blockwise_bf16_nf4
 
-if DEVICE_TYPE == "xpu":
-    # https://github.com/bitsandbytes-foundation/bitsandbytes/blob/c3b8de268fdb55a88f92feada23fc811a1e6877a/bitsandbytes/backends/xpu/ops.py#L115
-    # for xpu, inference gemv using above link
-    cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemv_4bit_inference_fp16
-    cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemv_4bit_inference_bf16
+if bnb is not None:
+    cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
+    cdequantize_blockwise_fp16_nf4 = bnb.functional.lib.cdequantize_blockwise_fp16_nf4
+    cdequantize_blockwise_bf16_nf4 = bnb.functional.lib.cdequantize_blockwise_bf16_nf4
+
+    if DEVICE_TYPE == "xpu":
+        # https://github.com/bitsandbytes-foundation/bitsandbytes/blob/c3b8de268fdb55a88f92feada23fc811a1e6877a/bitsandbytes/backends/xpu/ops.py#L115
+        cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemv_4bit_inference_fp16
+        cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemv_4bit_inference_bf16
+    else:
+        cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemm_4bit_inference_naive_fp16
+        cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemm_4bit_inference_naive_bf16
 else:
-    cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemm_4bit_inference_naive_fp16
-    cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemm_4bit_inference_naive_bf16
+    cdequantize_blockwise_fp32 = None
+    cdequantize_blockwise_fp16_nf4 = None
+    cdequantize_blockwise_bf16_nf4 = None
+    cgemm_4bit_inference_naive_fp16 = None
+    cgemm_4bit_inference_naive_bf16 = None
+
 
 
 torch_device_stream = (
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 48e5683076..3d48438800 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -723,7 +723,10 @@ if is_openai_available():
 
 # =============================================
 # Get Flash Attention v2 if Ampere (RTX 30xx, A100)
-import bitsandbytes as bnb
+try:
+    import bitsandbytes as bnb
+except Exception:
+    bnb = None
 
 from transformers import AutoTokenizer
 from transformers.utils.import_utils import _is_package_available
@@ -1908,7 +1911,11 @@ def patch_tokenizer(model, tokenizer):
 
 
 def patch_fast_lora():
-    import peft.tuners.lora.bnb
+    try:
+        import peft.tuners.lora.bnb
+    except Exception as e:
+        print("Unsloth: bitsandbytes/peft bnb not available - skipping 4bit LoRA patch.", repr(e))
+        return
 
     peft.tuners.lora.bnb.Linear4bit.forward = fast_lora_forward
 
diff --git a/unsloth/models/granite.py b/unsloth/models/granite.py
index aae746aed1..e28c21233d 100644
--- a/unsloth/models/granite.py
+++ b/unsloth/models/granite.py
@@ -30,8 +30,19 @@ from .llama import (
     LlamaLinearScalingRotaryEmbedding,
 )
 from .mistral import *
-from bitsandbytes.nn import Linear4bit as Bnb_Linear4bit
-from peft.tuners.lora import Linear4bit as Peft_Linear4bit
+try:
+    from bitsandbytes.nn import Linear4bit as Bnb_Linear4bit
+except Exception:
+    Bnb_Linear4bit = None
+
+try:
+    from peft.tuners.lora import Linear4bit as Peft_Linear4bit
+except Exception:
+    Peft_Linear4bit = None
+
+_BNB_LINEAR_TYPES = tuple(
+    t for t in (Bnb_Linear4bit, Peft_Linear4bit) if t is not None
+)
 
 try:
     from transformers.models.granite.modeling_granite import (
@@ -575,7 +586,7 @@ class FastGraniteModel(FastLlamaModel):
         correct_dtype = lm_head.weight.dtype
 
         for name, module in model.named_modules():
-            if isinstance(module, (Bnb_Linear4bit, Peft_Linear4bit)):
+            if _BNB_LINEAR_TYPES and isinstance(module, _BNB_LINEAR_TYPES):
                 weight = module.weight
                 quant_state = weight.quant_state
 
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index f18a07ac3c..63390a4927 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -3001,6 +3001,14 @@ class FastLlamaModel:
         if not SUPPORTS_RSLORA:
             del arguments["use_rslora"]
 
+        # PEFT API compatibility: only pass kwargs supported by the installed peft version.
+        try:
+            import inspect as _inspect
+            if "ensure_weight_tying" not in _inspect.signature(LoraConfig.__init__).parameters:
+                arguments.pop("ensure_weight_tying", None)
+        except Exception:
+            arguments.pop("ensure_weight_tying", None)
+
         _saved_temp_tokenizer = model._saved_temp_tokenizer
 
         lora_config = LoraConfig(**arguments)
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 410eee66e6..237a584bdd 100755
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -1143,6 +1143,13 @@ def grpo_trainer_compute_loss(function_name, function):
             if x.shape[1] == 1:  # when importance_sampling_level == "sequence"
                 return x.mean()
             else:
+                # Align mask/coef lengths when left-padding adds extra tokens.
+                if x.shape[1] != completion_mask.shape[1]:
+                    min_len = min(x.shape[1], completion_mask.shape[1])
+                    x = x[:, -min_len:]
+                    cm = completion_mask[:, -min_len:]
+                    denom = cm.sum().clamp(min = 1.0)
+                    return (x * cm).sum() / denom
                 return (x * completion_mask).sum() / completion_token_count
 
         if advantages.dim() == 1:
diff --git a/unsloth/save.py b/unsloth/save.py
index 071e032c53..0187ded794 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -23,8 +23,15 @@ from unsloth_zoo.llama_cpp import (
     check_llama_cpp,
     _download_convert_hf_to_gguf,
 )
-from bitsandbytes.nn import Linear4bit as Bnb_Linear4bit
-from peft.tuners.lora import Linear4bit as Peft_Linear4bit
+try:
+    from bitsandbytes.nn import Linear4bit as Bnb_Linear4bit
+except Exception:
+    Bnb_Linear4bit = None
+
+try:
+    from peft.tuners.lora import Linear4bit as Peft_Linear4bit
+except Exception:
+    Peft_Linear4bit = None
 from peft.tuners.lora import Linear as Peft_Linear
 from typing import Optional, Callable, Union, List
 import sys
@@ -58,6 +65,8 @@ except:
 from pathlib import Path
 from peft import PeftModelForCausalLM, PeftModel
 
+_MERGE_LORA_LINEAR_TYPES = tuple(t for t in (Bnb_Linear4bit, Peft_Linear4bit, Peft_Linear) if t is not None)
+
 __all__ = [
     "print_quantization_methods",
     "unsloth_save_model",
@@ -188,7 +197,7 @@ def _free_cached_model(model):
 
 def _merge_lora(layer, name):
     bias = getattr(layer, "bias", None)
-    if isinstance(layer, (Bnb_Linear4bit, Peft_Linear4bit, Peft_Linear)):
+    if _MERGE_LORA_LINEAR_TYPES and isinstance(layer, _MERGE_LORA_LINEAR_TYPES):
         # Is LoRA so we need to merge!
         W, quant_state, A, B, s, bias = get_lora_parameters_bias(layer)
         if quant_state is not None:

From e8d897466ee842e873197999e5ba0e4e4b4956af Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Sun, 8 Feb 2026 16:21:23 +0000
Subject: [PATCH 2/3] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 unsloth/__init__.py       |  2 ++
 unsloth/kernels/utils.py  | 10 +++++++---
 unsloth/models/_utils.py  |  5 ++++-
 unsloth/models/granite.py |  5 ++---
 unsloth/models/llama.py   |  6 +++++-
 unsloth/save.py           |  5 ++++-
 6 files changed, 24 insertions(+), 9 deletions(-)

diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index 19b433a680..19639a11c7 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -208,9 +208,11 @@ elif DEVICE_TYPE == "xpu":
 # Backwards compatibility: some notebooks import `unsloth.is_bf16_supported`.
 # Ensure it exists on all backends (HIP / XPU) and has a stable signature.
 if "is_bf16_supported" not in globals():
+
     def is_bf16_supported(including_emulation = False):
         return SUPPORTS_BFLOAT16
 
+
 # For Gradio HF Spaces?
 # if "SPACE_AUTHOR_NAME" not in os.environ and "SPACE_REPO_NAME" not in os.environ:
 import triton
diff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.py
index 63b1966b32..29d47c35e3 100644
--- a/unsloth/kernels/utils.py
+++ b/unsloth/kernels/utils.py
@@ -110,6 +110,7 @@ def calculate_settings(
 HAS_CUDA_STREAM = False
 try:
     import bitsandbytes as bnb
+
     # https://github.com/bitsandbytes-foundation/bitsandbytes/pull/1330/files
     HAS_CUDA_STREAM = Version(bnb.__version__) > Version("0.43.3")
     get_ptr = bnb.functional.get_ptr
@@ -197,8 +198,12 @@ if bnb is not None:
         cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemv_4bit_inference_fp16
         cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemv_4bit_inference_bf16
     else:
-        cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemm_4bit_inference_naive_fp16
-        cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemm_4bit_inference_naive_bf16
+        cgemm_4bit_inference_naive_fp16 = (
+            bnb.functional.lib.cgemm_4bit_inference_naive_fp16
+        )
+        cgemm_4bit_inference_naive_bf16 = (
+            bnb.functional.lib.cgemm_4bit_inference_naive_bf16
+        )
 else:
     cdequantize_blockwise_fp32 = None
     cdequantize_blockwise_fp16_nf4 = None
@@ -207,7 +212,6 @@ else:
     cgemm_4bit_inference_naive_bf16 = None
 
 
-
 torch_device_stream = (
     torch.xpu.current_stream if DEVICE_TYPE == "xpu" else torch.cuda.current_stream
 )
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 3d48438800..4c75b9f64f 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -1914,7 +1914,10 @@ def patch_fast_lora():
     try:
         import peft.tuners.lora.bnb
     except Exception as e:
-        print("Unsloth: bitsandbytes/peft bnb not available - skipping 4bit LoRA patch.", repr(e))
+        print(
+            "Unsloth: bitsandbytes/peft bnb not available - skipping 4bit LoRA patch.",
+            repr(e),
+        )
         return
 
     peft.tuners.lora.bnb.Linear4bit.forward = fast_lora_forward
diff --git a/unsloth/models/granite.py b/unsloth/models/granite.py
index e28c21233d..c40be8dfdb 100644
--- a/unsloth/models/granite.py
+++ b/unsloth/models/granite.py
@@ -30,6 +30,7 @@ from .llama import (
     LlamaLinearScalingRotaryEmbedding,
 )
 from .mistral import *
+
 try:
     from bitsandbytes.nn import Linear4bit as Bnb_Linear4bit
 except Exception:
@@ -40,9 +41,7 @@ try:
 except Exception:
     Peft_Linear4bit = None
 
-_BNB_LINEAR_TYPES = tuple(
-    t for t in (Bnb_Linear4bit, Peft_Linear4bit) if t is not None
-)
+_BNB_LINEAR_TYPES = tuple(t for t in (Bnb_Linear4bit, Peft_Linear4bit) if t is not None)
 
 try:
     from transformers.models.granite.modeling_granite import (
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 63390a4927..9d3ab0ecaf 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -3004,7 +3004,11 @@ class FastLlamaModel:
         # PEFT API compatibility: only pass kwargs supported by the installed peft version.
         try:
             import inspect as _inspect
-            if "ensure_weight_tying" not in _inspect.signature(LoraConfig.__init__).parameters:
+
+            if (
+                "ensure_weight_tying"
+                not in _inspect.signature(LoraConfig.__init__).parameters
+            ):
                 arguments.pop("ensure_weight_tying", None)
         except Exception:
             arguments.pop("ensure_weight_tying", None)
diff --git a/unsloth/save.py b/unsloth/save.py
index 0187ded794..5f26e0a210 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -23,6 +23,7 @@ from unsloth_zoo.llama_cpp import (
     check_llama_cpp,
     _download_convert_hf_to_gguf,
 )
+
 try:
     from bitsandbytes.nn import Linear4bit as Bnb_Linear4bit
 except Exception:
@@ -65,7 +66,9 @@ except:
 from pathlib import Path
 from peft import PeftModelForCausalLM, PeftModel
 
-_MERGE_LORA_LINEAR_TYPES = tuple(t for t in (Bnb_Linear4bit, Peft_Linear4bit, Peft_Linear) if t is not None)
+_MERGE_LORA_LINEAR_TYPES = tuple(
+    t for t in (Bnb_Linear4bit, Peft_Linear4bit, Peft_Linear) if t is not None
+)
 
 __all__ = [
     "print_quantization_methods",

From c3d86a89839c9be64852d79f08eab03d640fe7e6 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Sun, 31 May 2026 10:12:59 +0000
Subject: [PATCH 3/3] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 unsloth/_gpu_init.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index d7a7b2ca6e..b35e18cf75 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -241,6 +241,7 @@ if "is_bf16_supported" not in globals():
     def is_bf16_supported(including_emulation = False):
         return SUPPORTS_BFLOAT16
 
+
 # For Gradio HF Spaces?
 # if "SPACE_AUTHOR_NAME" not in os.environ and "SPACE_REPO_NAME" not in os.environ:
 import triton