dtype

unsloth/kernels/fast_lora.py

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

unsloth/models/llama.py

@@ -505,13 +505,10 @@ def LlamaModel_fast_forward(
             position_ids = position_ids.repeat((batch_size, 1))
     pass
 
-    # embed positions
+    # Embed positions
     if inputs_embeds is None:
         inputs_embeds = self.embed_tokens(input_ids)
 
-    # Downcast to the correct dtype ie float32 to float16
-    inputs_embeds = inputs_embeds.to(self.config.torch_dtype)
-
     # Normalized from Gemma
     IS_GEMMA = self.config.model_type == "gemma"
     train_embed_tokens = self.embed_tokens.weight.requires_grad
@@ -759,6 +756,7 @@ def CausalLM_fast_forward(fast_forward_inference):
         else:
             logits = self.lm_head(hidden_states)
         pass
+        print(logits.dtype)
 
         loss = None
         if labels is not None: