FastGemmaModel

unsloth/kernels/fast_lora.py

@@ -90,8 +90,6 @@ class LoRA_MLP(torch.autograd.Function):
 
         e = matmul_lora(X, gateW, gateW_quant, gateA, gateB, gateS)
         g = matmul_lora(X,   upW,   upW_quant,   upA,   upB,   upS)
-        # f = torch.nn.functional.silu(e)
-        # h = f * g
         h = _forward_function(e, g)
         i = matmul_lora(h, downW, downW_quant, downA, downB, downS)
 

unsloth/models/llama.py

@@ -891,9 +891,10 @@ class FastLlamaModel:
         device_map     = "sequential",
         rope_scaling   = None,
         fix_tokenizer  = True,
-        model_patcher  = FastLlamaModel,
+        model_patcher  = None,
         **kwargs,
     ):
+        if model_patcher is None: model_patcher = FastLlamaModel
         SUPPORTS_BFLOAT16 = torch.cuda.is_bf16_supported()
         gpu_stats = torch.cuda.get_device_properties(0)
         max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)

unsloth/models/mistral.py

@@ -293,9 +293,10 @@ class FastMistralModel(FastLlamaModel):
         device_map     = "sequential",
         rope_scaling   = None, # Mistral does not support RoPE scaling
         fix_tokenizer  = True,
-        model_patcher  = FastMistralModel,
+        model_patcher  = None,
         **kwargs,
     ):
+        if model_patcher is None: model_patcher = FastMistralModel
         # Mistral does NOT support RoPE Scaling!
         if rope_scaling is not None:
             logger.warning_once("Unsloth: Mistral models do not support RoPE scaling.")