Merge pull request #33 from dbaranchuk/memory-efficient-backward

Memory efficient backward

bitsandbytes/autograd/_functions.py

 import operator
+import warnings
+
 import torch
 import bitsandbytes.functional as F
 
@@ -184,6 +186,7 @@ class MatmulLtState:
     idx = None
     is_training = True
     has_fp16_weights = True
+    memory_efficient_backward = False
     use_pool = False
     formatB = F.get_special_format_str()
 
@@ -209,31 +212,29 @@ class MatMul8bitLt(torch.autograd.Function):
             ctx.B = B
             ctx.bias = bias
             if A.shape[-1] == B.shape[0]:
-                return torch.empty(A.shape[:-1]+B.shape[1:], dtype=torch.float16, device=A.device)
+                return torch.empty(A.shape[:-1]+B.shape[1:], dtype=A.dtype, device=A.device)
             else:
-                return torch.empty(A.shape[:-1]+B.shape[:1], dtype=torch.float16, device=A.device)
+                return torch.empty(A.shape[:-1]+B.shape[:1], dtype=A.dtype, device=A.device)
 
         # 1. Quantize A
         # 2. Quantize B
         # 3. Matmul
         # 4. Mixed-precision decomposition matmul
         # 5. Save state
-        requires_gradA = A.requires_grad
-        requires_gradB = B.requires_grad
-        requires_gradBias = bias is not None and bias.requires_grad
         formatB = state.formatB
         input_shape = A.shape
         if state.outlier_pool is None:
             state.outlier_pool = GlobalOutlierPooler.get_instance()
-        assert (
-            A.dtype == torch.float16
-        ), f"The input data type needs to be fp16 but {A.dtype} was found!"
+
+        # Cast A to fp16
+        if A.dtype != torch.float16:
+            warnings.warn(f"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization")
 
         # 1. Quantize A
         if len(A.shape) == 3:
             A = A.view(-1, A.shape[-1]).contiguous()
         CA, CAt, SCA, SCAt, coo_tensorA = F.double_quant(
-            A, threshold=state.threshold
+            A.to(torch.float16), threshold=state.threshold
         )
 
         if state.threshold > 0.0 and coo_tensorA is not None:
@@ -269,7 +270,7 @@ class MatMul8bitLt(torch.autograd.Function):
                     state.SCB,
                     state.SCBt,
                     coo_tensorB,
-                ) = F.double_quant(B)
+                ) = F.double_quant(B.to(torch.float16))
                 state.CxB, state.SB = F.transform(CB, to_order=formatB)
         else:
             has_grad = False
@@ -290,7 +291,7 @@ class MatMul8bitLt(torch.autograd.Function):
                 (outliers * state.SCB.view(-1, 1) / 127.0)
                 .t()
                 .contiguous()
-                .half()
+                .to(A.dtype)
             )
             CA[:, state.idx.long()] = 0
             CAt[:, state.idx.long()] = 0
@@ -307,7 +308,13 @@ class MatMul8bitLt(torch.autograd.Function):
         C32A, SA = F.transform(CA, "col32")
         out32, Sout32 = F.igemmlt(C32A, state.CxB, SA, state.SB)
         # we apply the fused bias here
+
+        if bias is None or bias.dtype == torch.float16:
             output = F.mm_dequant(out32, Sout32, SCA, state.SCB, bias=bias)
+            output = output.to(A.dtype)
+        else:  # apply bias separately
+            output = F.mm_dequant(out32, Sout32, SCA, state.SCB, bias=None)
+            output = output.to(A.dtype).add_(bias)
 
         # 4. Mixed-precision decomposition matmul
         if coo_tensorA is not None and subA is not None:
@@ -318,9 +325,9 @@ class MatMul8bitLt(torch.autograd.Function):
 
         ctx.formatB = formatB
         ctx.grad_shape = input_shape
-        ctx.req_grads = [requires_gradA, requires_gradB, requires_gradBias]
+        ctx.dtype_A, ctx.dtype_B, ctx.dtype_bias = A.dtype, B.dtype, None if bias is None else bias.dtype
 
-        if requires_gradA or requires_gradB:
+        if any(ctx.needs_input_grad[:2]):
             ctx.tensors = (CAt, subA)
             ctx.tensor_states = (SCAt, state.idx)
         else:
@@ -328,8 +335,8 @@ class MatMul8bitLt(torch.autograd.Function):
             ctx.tensor_states = (None, None)
             ctx.save_for_backward(None, None)
 
+
         clone_func = torch.clone if len(output_shape) == 3 else lambda x : x
-        #clone_func = torch.clone
         return clone_func(output.view(output_shape))
 
     @staticmethod
@@ -337,23 +344,24 @@ class MatMul8bitLt(torch.autograd.Function):
         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
-        req_gradA, req_gradB, req_gradBias = ctx.req_grads
+        req_gradA, req_gradB, _, req_gradBias, _ = ctx.needs_input_grad
         CAt, subA = ctx.tensors
         SCAt, idx = ctx.tensor_states
         formatB = ctx.formatB
         state = ctx.state
-        assert (
-            state.has_fp16_weights
-        ), "Backprop only supported for fp16 weights."
+        grad_A = grad_B = grad_bias = None
 
+        if req_gradBias:
+            # compute grad_bias first before changing grad_output dtype
+            grad_bias = grad_output.sum(0, dtype=ctx.dtype_bias)
+
+        # Cast grad_output to fp16
         if len(grad_output.shape) == 3:
-            grad_output = grad_output.view(
+            grad_output = grad_output.reshape(
                 -1, grad_output.shape[-1]
             ).contiguous()
 
-        grad_A = grad_B = grad_bias = None
-
-        Cgrad, Cgradt, SCgrad, SCgradt, coo_tensor = F.double_quant(grad_output)
+        Cgrad, Cgradt, SCgrad, SCgradt, coo_tensor = F.double_quant(grad_output.to(torch.float16))
         if req_gradB:
             CxAt, SAt = F.transform(CAt, formatB, transpose=True)
             C32grad, Sgrad = F.transform(Cgradt, "col32", transpose=True)
@@ -363,16 +371,20 @@ class MatMul8bitLt(torch.autograd.Function):
                 grad_B[:, idx] += torch.matmul(grad_output.t(), subA)
 
         if req_gradA:
+            if state.CBt is not None:
                 C32grad, Sgrad = F.transform(Cgrad, "col32")
                 if state.CxBt is None:
                     state.CxBt, state.SBt = F.transform(
                         state.CBt, to_order=formatB, transpose=True
                     )
                 gradA32, SgradA32 = F.igemmlt(C32grad, state.CxBt, Sgrad, state.SBt)
-            grad_A = F.mm_dequant(gradA32, SgradA32, SCgrad, state.SCBt).view(ctx.grad_shape)
+                grad_A = F.mm_dequant(gradA32, SgradA32, SCgrad, state.SCBt).view(ctx.grad_shape).to(ctx.dtype_A)
 
-        if req_gradBias:
-            grad_bias = grad_output.sum(0)
+            elif state.CB is not None:
+                CB = state.CB.to(ctx.dtype_A, copy=True).mul_(state.SCB.unsqueeze(1).mul(1. / 127.0))
+                grad_A = torch.matmul(grad_output, CB).view(ctx.grad_shape).to(ctx.dtype_A)
+            else:
+                raise Exception('State must contain either CBt or CB matrix for backward')
 
         return grad_A, grad_B, None, grad_bias, None
 

bitsandbytes/nn/modules.py

@@ -221,6 +221,7 @@ class Linear8bitLt(nn.Linear):
         output_features,
         bias=True,
         has_fp16_weights=True,
+        memory_efficient_backward=False,
         threshold=0.0,
         index=None,
     ):
@@ -232,10 +233,13 @@ class Linear8bitLt(nn.Linear):
 
         self.state.threshold = threshold
         self.state.has_fp16_weights = has_fp16_weights
+        self.state.memory_efficient_backward = memory_efficient_backward
         if threshold > 0.0 and not has_fp16_weights:
             self.state.use_pool = True
 
-        self.weight = Int8Params(self.weight.data, has_fp16_weights=has_fp16_weights)
+        self.weight = Int8Params(
+            self.weight.data, has_fp16_weights=has_fp16_weights, requires_grad=has_fp16_weights
+        )
 
     def init_8bit_state(self):
         self.state.CB = self.weight.CB
@@ -255,11 +259,16 @@ class Linear8bitLt(nn.Linear):
 
         out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state)
 
-        if not self.state.has_fp16_weights and self.state.CB is not None:
+        if not self.state.has_fp16_weights:
+            if not self.state.memory_efficient_backward and self.state.CB is not None:
                 # we converted 8-bit row major to turing/ampere format in the first inference pass
                 # we no longer need the row-major weight
                 del self.state.CB
                 self.weight.data = self.state.CxB
+            elif self.state.memory_efficient_backward and self.state.CxB is not None:
+                # For memory efficient backward, we convert 8-bit row major to turing/ampere format at each inference pass.
+                # Thus, we delete CxB from the state. 
+                del self.state.CxB
 
         return out
 

tests/test_autograd.py

@@ -253,7 +253,7 @@ for c in req_grad:
 
 transpose = [(False, True), (False, False)]
 str_transpose = ["NT", "NN"]
-dtype = [torch.float16]
+dtype = [torch.float16, torch.bfloat16, torch.float32]
 has_fp16_weights = [True, False]
 has_bias = [True, False]
 values = list(
@@ -354,7 +354,7 @@ def test_matmullt(
                     state.SCB,
                     SCBt,
                     coo_tensorB,
-                ) = bnb.functional.double_quant(B2)
+                ) = bnb.functional.double_quant(B2.to(torch.float16))
                 B2 = state.CB
 
             if not transpose[0] and transpose[1]:
@@ -367,11 +367,14 @@ def test_matmullt(
             if has_bias:
                 out_torch += bias
 
+            assert out_bnb.dtype == A.dtype, f"bnb matmullt received {A.dtype} but returned {out_bnb.dtype}"
+
             n = out_bnb.numel()
             err = torch.abs(out_bnb - out_torch).mean().item()
             # print(f'abs error {err:.4f}')
+
             idx = torch.isclose(out_bnb, out_torch, atol=0.01, rtol=0.1)
-            assert (idx == 0).sum().item() <= n * 0.0175
+            assert (idx == 0).sum().item() <= n * (0.0175 if dtype == torch.float16 else 0.021)
             idx = torch.isclose(out_bnb, out_torch, atol=0.035, rtol=0.2)
             assert (idx == 0).sum().item() <= n * 0.001
 

tests/test_modules.py

@@ -14,13 +14,15 @@ class MockArgs(object):
 
 
 class MLP8bit(torch.nn.Module):
-    def __init__(self, dim1, dim2, has_fp16_weights=True, threshold=0.0):
+    def __init__(self, dim1, dim2, has_fp16_weights=True, memory_efficient_backward=False, threshold=0.0):
         super(MLP8bit, self).__init__()
         self.fc1 = bnb.nn.Linear8bitLt(
-            dim1, dim2, has_fp16_weights=has_fp16_weights, threshold=threshold
+            dim1, dim2, has_fp16_weights=has_fp16_weights, memory_efficient_backward=memory_efficient_backward,
+            threshold=threshold
         )
         self.fc2 = bnb.nn.Linear8bitLt(
-            dim2, dim1, has_fp16_weights=has_fp16_weights, threshold=threshold
+            dim2, dim1, has_fp16_weights=has_fp16_weights, memory_efficient_backward=memory_efficient_backward,
+            threshold=threshold
         )
 
     def forward(self, x):
@@ -451,9 +453,12 @@ names = ["threshold_{0}".format(vals) for vals in values]
 
 
 @pytest.mark.parametrize("threshold", values, ids=names)
-def test_linear8bitlt_no_fp16_weights(threshold):
+@pytest.mark.parametrize("memory_efficient_backward", [True, False])
+def test_linear8bitlt_no_fp16_weights(threshold, memory_efficient_backward):
     l1 = (
-        bnb.nn.Linear8bitLt(32, 64, threshold=threshold, has_fp16_weights=False)
+        bnb.nn.Linear8bitLt(
+            32, 64, threshold=threshold, has_fp16_weights=False, memory_efficient_backward=memory_efficient_backward
+        )
         .cuda()
         .half()
     )
@@ -513,7 +518,9 @@ def test_linear8bitlt_no_fp16_weights(threshold):
     assert mlp.fc2.weight.dtype == torch.int8
 
     mlp = (
-        MLP8bit(32, 64, threshold=threshold, has_fp16_weights=False)
+        MLP8bit(
+            32, 64, threshold=threshold, has_fp16_weights=False, memory_efficient_backward=memory_efficient_backward
+        )
         .half()
         .to("cuda")
     )
@@ -531,11 +538,11 @@ def test_linear8bitlt_no_fp16_weights(threshold):
     assert mlp.fc1.weight.device.type == "cuda"
     assert mlp.fc2.weight.device.type == "cuda"
 
-    mlp = (
-        MLP8bit(32, 64, threshold=threshold, has_fp16_weights=False)
-        .to(torch.float16)
-        .to("cuda")
+    mlp = MLP8bit(
+            32, 64, threshold=threshold, has_fp16_weights=False, memory_efficient_backward=memory_efficient_backward
         )
+    w1, w2 = mlp.fc1.weight.clone().cuda(), mlp.fc2.weight.clone().cuda()  # grab weights before quantization,
+    mlp = mlp.cuda().half()  # and this line triggers quantization
 
     for i in range(100):
         b1 = torch.randn(16, 8, 32, device="cuda").half()
@@ -545,11 +552,28 @@ def test_linear8bitlt_no_fp16_weights(threshold):
             assert mlp.fc1.state.idx is not None
         if threshold > 0:
             assert mlp.fc2.state.idx is not None
+
     assert mlp.fc1.weight.dtype == torch.int8
     assert mlp.fc2.weight.dtype == torch.int8
     assert mlp.fc1.weight.device.type == "cuda"
     assert mlp.fc2.weight.device.type == "cuda"
 
+    if memory_efficient_backward:
+        b1 = torch.randn(16, 8, 32, device="cuda", requires_grad=True, dtype=torch.half)
+        o1 = mlp(b1)
+        assert o1.dtype == torch.float16
+        assert o1.requires_grad
+        grad_proj = torch.randn_like(o1)
+
+        mlp.zero_grad()
+        (o1 * grad_proj).sum().backward()
+        grad_ref = grad_proj.flatten(2) @ w2.half() @ w1.half()
+        scale = grad_ref.abs().mean()
+
+        torch.testing.assert_allclose(b1.grad, grad_ref, rtol=0, atol=0.05 * scale)
+        idx = torch.isclose(b1.grad, grad_ref, atol=0.01 * scale, rtol=0.1)
+        assert (idx == 0).sum().item() <= b1.numel() * 0.005
+
 
 def test_linear8bitlt_fp32_bias():
     # casts model to fp16 -> int8 automatically