Add support for splitting in_features in linear layers (#8715)

init

Author: metascroy

examples/apple/coreml/llama/export.py

@@ -1,6 +1,8 @@
-# (c) Meta Platforms, Inc. and affiliates. Confidential and proprietary.
-
-# pyre-strict
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
 
 import argparse
 
@@ -24,55 +26,7 @@
 
 sys.path.insert(0, ".")
 from llama_transformer import InputManager, load_model
-
-
-class SplitLinearModule(torch.nn.Module):
-    def __init__(self, in_features, out_features, target_split_size, max_splits):
-        super(SplitLinearModule, self).__init__()
-        num_splits = max(out_features // target_split_size, 1)
-        if num_splits > max_splits:
-            num_splits = max_splits
-
-        self.split_size = out_features // num_splits
-        self.split_remainder = out_features % num_splits
-        self.splits = torch.nn.ModuleList(
-            [torch.nn.Linear(in_features, self.split_size) for _ in range(num_splits)]
-        )
-        print(
-            f"Splitting out_features={out_features} into {num_splits} of size {self.split_size}"
-        )
-        if self.split_remainder > 0:
-            print(
-                f"Warning: remainder {self.split_remainder} after splitting out_features={out_features} into {num_splits} of size {self.split_size}"
-            )
-            self.splits.append(torch.nn.Linear(in_features, self.split_remainder))
-
-    def split_sizes(self):
-        return [split.out_features for split in self.splits]
-
-    def forward(self, x):
-        return torch.cat([split(x) for split in self.splits], dim=-1)
-
-
-def replace_linear_with_split_linear(model, target_split_size, max_splits):
-    for name, module in model.named_children():
-        if isinstance(module, torch.nn.Linear):
-            new_module = SplitLinearModule(
-                module.in_features, module.out_features, target_split_size, max_splits
-            )
-            split_sizes = new_module.split_sizes()
-            if module.bias is not None:
-                split_bias = module.bias.split(split_sizes)
-            split_weights = module.weight.split(split_sizes, dim=0)
-            for i, split in enumerate(new_module.splits):
-                split.weight = torch.nn.Parameter(split_weights[i])
-                if module.bias is not None:
-                    split.bias = torch.nn.Parameter(split_bias[i])
-                else:
-                    split.bias = None
-            setattr(model, name, new_module)
-        else:
-            replace_linear_with_split_linear(module, target_split_size, max_splits)
+from utils import replace_linear_with_split_linear
 
 
 def main() -> None:
@@ -175,7 +129,13 @@ def main() -> None:
 
     if export_args.target_split_size is not None:
         replace_linear_with_split_linear(
-            model, export_args.target_split_size, export_args.max_splits
+            model,
+            out_target_split_size=export_args.target_split_size,
+            out_max_splits=export_args.max_splits,
+            # I have not found splitting on in_features to be beneficial,
+            # and it often leads to OOM so I set in_max_splits to 1
+            in_target_split_size=1,
+            in_max_splits=1,
         )
 
     model.eval()
@@ -241,6 +201,7 @@ def main() -> None:
         ep,
         preserve_ops=[
             torch.ops.aten.scaled_dot_product_attention.default,
+            # preserve norm op for numerical stability
             torch.ops.aten.linalg_vector_norm.default,
         ],
         compile_config=EdgeCompileConfig(

examples/apple/coreml/llama/readme.md

@@ -38,8 +38,9 @@ The runner can also be used to run an eager model model to compare with CoreML n
 
 We are actively experimenting with different settings.  But here are ones that we've found work well for Llama1B on iPhone 15 Pro:
 
-* Set use_cache_list
-* Split linear layers with target_split_size=1024, max_splits=8
-* Use seq_length=32 or seq_length=64, both of which offer reasonable tradeoffs for prefill and decode performance.  seq_length=32 is better at decode and seq_length=64 is better at prefill.
-
-In our tests, we set max_seq_length=1024, but if your application allows for it, performance can improve with max_seq_length=512 or by keeping max_seq_length=1024 and setting cache_size=512-seq_length.
+* Set use_cache_list.
+* Use seq_length = 32, which offers a good balance between prefill/decode performance.
+* Split out_features in linear layers with target_split_size=1024, max_splits=8.
+* For ANE, set dtype = fp16, coreml-quantize = c4w.  The requires doing QAT on Llama1B for good accuracy.
+* Set embedding-quantize to "4,32".
+* Set max_seq_length to 128, 256, 512, 1024, and 2048, depending on needed context.  Note that performance drops with max_seq_length.  More specifically, performance drops with cache_size, and the best experience may require a good cache eviction policy.  The python runner in run.py uses a last-in-last-out policy when cache_size is specified.

examples/apple/coreml/llama/run.py

@@ -1,3 +1,9 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
 import argparse
 import sys
 

examples/apple/coreml/llama/test.py

@@ -0,0 +1,48 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import sys
+
+sys.path.insert(0, ".")
+import copy
+
+import torch
+from utils import replace_linear_with_split_linear
+
+
+def get_split_model(
+    model,
+    out_target_split_size=1,
+    out_max_splits=1,
+    in_target_split_size=1,
+    in_max_splits=1,
+):
+    model_copy = copy.deepcopy(model)
+    replace_linear_with_split_linear(
+        model_copy,
+        out_target_split_size,
+        out_max_splits,
+        in_target_split_size,
+        in_max_splits,
+    )
+    return model_copy
+
+
+def test_split_model():
+    inputs = torch.randn(10, 5, 1, 512)
+
+    model = torch.nn.Sequential(*[torch.nn.Linear(512, 1024, bias=False)])
+    model1 = get_split_model(model, 64, 2, 64, 1000)
+    model2 = get_split_model(model, 64, 2, 64, 1)
+    model3 = get_split_model(model, 64, 1, 64, 1000)
+
+    assert torch.allclose(model(inputs), model1(inputs), atol=1e-5)
+    assert torch.allclose(model(inputs), model2(inputs), atol=1e-5)
+    assert torch.allclose(model(inputs), model3(inputs), atol=1e-5)
+
+
+if __name__ == "__main__":
+    test_split_model()

examples/apple/coreml/llama/utils.py

@@ -0,0 +1,116 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+
+class SplitLinearModule(torch.nn.Module):
+    def __init__(
+        self,
+        in_features,
+        out_features,
+        out_target_split_size=1,
+        out_max_splits=1,
+        in_target_split_size=1,
+        in_max_splits=1,
+    ):
+        super(SplitLinearModule, self).__init__()
+        self.out_split_sizes = self._get_split_sizes(
+            out_features, out_target_split_size, out_max_splits
+        )
+        self.in_split_sizes = self._get_split_sizes(
+            in_features, in_target_split_size, in_max_splits
+        )
+        print(
+            f"Splitting out_features={out_features} into {len(self.out_split_sizes)} of size {self.out_split_sizes[0]}."
+        )
+        print(
+            f"Splitting in_features={in_features} into {len(self.in_split_sizes)} of size {self.in_split_sizes[0]}."
+        )
+
+        # self.ops contains a list of linear ops for different pieces of the output matrix
+        # The index of an op at (in_idx, out_idx) is given by self.op_index(in_idx, out_idx)
+        self.ops = torch.nn.ModuleList()
+        for idx_out, s_out in enumerate(self.out_split_sizes):
+            for idx_in, s_in in enumerate(self.in_split_sizes):
+                assert len(self.ops) == self.op_index(idx_in, idx_out)
+                self.ops.append(torch.nn.Linear(s_in, s_out, bias=False))
+
+    def op_index(self, in_index, out_index):
+        idx = out_index * len(self.in_split_sizes) + in_index
+        return idx
+
+    def _get_split_sizes(self, n_features, target_split_size, max_splits):
+        num_splits = max(n_features // target_split_size, 1)
+        if num_splits > max_splits:
+            num_splits = max_splits
+
+        split_size = n_features // num_splits
+        split_remainder = n_features % num_splits
+        if split_remainder > 0:
+            raise ValueError(
+                f"Cannot split {n_features} with target_split_size={target_split_size} and max_splits={max_splits} because it leaves a remainder of {split_remainder}."
+            )
+
+        ret = [split_size for _ in range(num_splits)]
+        return ret
+
+    def set_params(self, weight):
+        split_weights = []
+        for w_out in weight.split(self.out_split_sizes, dim=0):
+            for w in w_out.split(self.in_split_sizes, dim=1):
+                split_weights.append(w)
+
+        for i, split in enumerate(self.ops):
+            split.weight = torch.nn.Parameter(split_weights[i])
+
+    def forward(self, x):
+        if len(self.in_split_sizes) == 1:
+            out_chunks = [op(x) for op in self.ops]
+        else:
+            x_splits = x.split(self.in_split_sizes, dim=-1)
+            out_chunks = [
+                torch.sum(
+                    torch.stack(
+                        [
+                            self.ops[self.op_index(in_idx, out_idx)].forward(
+                                x_splits[in_idx]
+                            )
+                            for in_idx in range(len(self.in_split_sizes))
+                        ],
+                    ),
+                    dim=0,
+                )
+                for out_idx in range(len(self.out_split_sizes))
+            ]
+
+        return torch.concat(out_chunks, dim=-1)
+
+
+def replace_linear_with_split_linear(
+    model, out_target_split_size, out_max_splits, in_target_split_size, in_max_splits=1
+):
+    for name, module in model.named_children():
+        if isinstance(module, torch.nn.Linear):
+            assert module.bias is None, "SplitLinearModule does not support bias"
+            new_module = SplitLinearModule(
+                module.in_features,
+                module.out_features,
+                out_target_split_size,
+                out_max_splits,
+                in_target_split_size,
+                in_max_splits,
+            )
+            new_module.set_params(module.weight)
+            setattr(model, name, new_module)
+        else:
+            replace_linear_with_split_linear(
+                module,
+                out_target_split_size,
+                out_max_splits,
+                in_target_split_size,
+                in_max_splits,
+            )