Bugs

unsloth/kernels/fast_lora.py

@@ -87,6 +87,7 @@ class LoRA_MLP(torch.autograd.Function):
                 downW, downW_quant, downA, downB, downS,
                 _forward_function, _backward_function,):
         dtype = X.dtype
+        print("Forward " + dtype)
 
         e = matmul_lora(X, gateW, gateW_quant, gateA, gateB, gateS)
         g = matmul_lora(X,   upW,   upW_quant,   upA,   upB,   upS)
@@ -122,6 +123,7 @@ class LoRA_MLP(torch.autograd.Function):
         e  = e .view(-1, e .shape[-1])
         g  = g .view(-1, g .shape[-1])
         dtype = X.dtype
+        print("Backward " + dtype)
 
         DW = matmul_lora(dY, downW.t(), downW_quant, downB, downA, downS)
         DW, e, g = _backward_function(DW, e, g)

unsloth/save.py

@@ -684,6 +684,9 @@ def save_to_gguf(
 ):
     from transformers.models.llama.modeling_llama import logger
 
+    if quantization_method.startswith("iq2"):
+        raise RuntimeError("Unsloth: Currently iq2 type quantizations aren't supported yet - sorry!")
+
     # Careful convert.py is only for Llama / Mistral based archs
     use_fast_convert = False
     if   model_type == "llama":   use_fast_convert = True
@@ -743,8 +746,11 @@ def save_to_gguf(
         if   first_conversion == "f32" : pass
         elif first_conversion == "f16" : pass
         elif first_conversion == "q8_0":
-            logger.warning_once("Unsloth: We must use f16 for quantization first.")
-            first_conversion = "f16"
+            logger.warning_once(
+                "Unsloth: Using q8_0 for the `first_conversion` will lose a bit of accuracy, "\
+                "but saves disk space!"
+            )
+            # first_conversion = "f16"
         pass
     pass