add dtype <-> fp16 cast

bitsandbytes/autograd/_functions.py

@@ -213,6 +213,10 @@ class MatMul8bitLt(torch.autograd.Function):
             else:
                 return torch.empty(A.shape[:-1]+B.shape[:1], dtype=torch.float16, device=A.device)
 
+        # Cast A to fp16
+        A_dtype = A.dtype
+        A = A.to(torch.float16)
+
         # 1. Quantize A
         # 2. Quantize B
         # 3. Matmul
@@ -322,14 +326,21 @@ class MatMul8bitLt(torch.autograd.Function):
             ctx.tensor_states = (None, None)
             ctx.save_for_backward(None, None)
 
+        # Cast fp16 output back to A.dtype
+        output = output.to(A_dtype)
+
         clone_func = torch.clone if len(output_shape) == 3 else lambda x : x
-        #clone_func = torch.clone
         return clone_func(output.view(output_shape))
 
     def backward(ctx, grad_output):
         if ctx.is_empty:
             bias_grad = (None if ctx.bias is None else torch.zeros_like(ctx.bias))
             return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None
+
+        # Cast grad_output to fp16
+        grad_output_dtype = grad_output.dtype
+        grad_output.to(torch.float16)
+
         req_gradA, req_gradB, req_gradBias = ctx.req_grads
         assert not req_gradB, "TODO: support weight updates as well"
         state = ctx.state
@@ -350,6 +361,9 @@ class MatMul8bitLt(torch.autograd.Function):
         if req_gradBias:
             grad_bias = grad_output.sum(0)
 
+        # Cast grad_A back to grad_output_dtype
+        grad_output.to(grad_output_dtype)
+
         return grad_A, grad_B, None, grad_bias, None