[pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

unsloth/models/_utils.py

@@ -82,6 +82,7 @@ __all__ = [
     "_get_inference_mode_context_manager",
     "hf_login",
     "is_moe_model",
+    "is_gemma4_shared_kv_model",
     "get_moe_target_parameters",
     "make_fast_generate_wrapper",
 ]
@@ -3280,6 +3281,30 @@ def is_moe_model(model) -> bool:
     return num_experts is not None and num_experts > 0
 
 
+def is_gemma4_shared_kv_model(model) -> bool:
+    """
+    Detect Gemma 4 E-series models (E2B / E4B) that share KV across layers.
+
+    These have num_kv_shared_layers > 0, and in transformers 5.5.0 their
+    KV cache generation path is broken (huggingface/transformers#45242):
+    generating with a KV cache (use_cache=True / cache_implementation set)
+    yields different, garbled logits to the cache-free path. Gemma 4 31B /
+    26B-A4B have num_kv_shared_layers == 0 and are unaffected.
+    """
+    config = getattr(model, "config", model)
+    # Use _config_get so a dict config (some serialization / init paths) is read
+    # correctly; plain getattr would silently miss the field on a dict.
+    text_config = _config_get(config, "text_config", None) or config
+    model_type = _config_get(config, "model_type", "") or ""
+    text_model_type = _config_get(text_config, "model_type", "") or ""
+    if not (
+        (isinstance(model_type, str) and model_type.startswith("gemma4"))
+        or (isinstance(text_model_type, str) and text_model_type.startswith("gemma4"))
+    ):
+        return False
+    return (_config_get(text_config, "num_kv_shared_layers", 0) or 0) > 0
+
+
 def _resolve_moe_parameter_name(model, default_name: str, alternate_name: str) -> str:
     """
     Resolve the actual parameter path for MoE expert weights.

unsloth/models/llama.py

@@ -2156,6 +2156,14 @@ def unsloth_fast_generate(self, *args, **kwargs):
 
     # For newer HF
     kwargs["cache_implementation"] = "dynamic"
+    # Gemma 4 E-series (E2B / E4B) share KV across layers and their KV cache
+    # generation path is broken in transformers 5.5.0
+    # (huggingface/transformers#45242): a cached forward diverges from the
+    # cache-free forward, so default greedy decode emits garbage. Default to the
+    # correct cache-free path unless the caller explicitly opts into a cache.
+    if "use_cache" not in kwargs and is_gemma4_shared_kv_model(self):
+        kwargs["use_cache"] = False
+        kwargs["cache_implementation"] = None
     # transformers 4.50 renamed num_logits_to_keep -> logits_to_keep; pop both,
     # re-emit under the spelling forward() accepts.
     _provided_num = kwargs.pop("num_logits_to_keep", None)

unsloth/models/vision.py

@@ -405,6 +405,15 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
     if do_bfloat16_mixed_precision:
         cache_implementation = None
 
+    # Gemma 4 E-series (E2B / E4B) share KV across layers and their KV cache
+    # generation path is broken in transformers 5.5.0
+    # (huggingface/transformers#45242): a cached forward diverges from the
+    # cache-free forward, so default greedy decode emits garbage. Default to the
+    # correct cache-free path unless the caller explicitly opts into a cache.
+    if "use_cache" not in kwargs and is_gemma4_shared_kv_model(self):
+        kwargs["use_cache"] = False
+        cache_implementation = None
+
     if "generation_config" in kwargs:
         kwargs["generation_config"].cache_implementation = cache_implementation
         if cache_implementation is not None: