Add pruning-aware training in torchao.prototype.pat

Author: lisjin

test/prototype/test_pat.py

@@ -0,0 +1,279 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD 3-Clause license found in the
+# LICENSE file in the root directory of this source tree.
+
+import random
+import unittest
+
+import torch
+from torch import nn
+from torch.testing._internal import common_utils
+
+from torchao.prototype.pat.group import (
+    AttentionHeadGrouperDim0,
+    AttentionHeadGrouperDim1,
+    PackedSVDGrouper,
+    QKGrouper,
+    QKSVDGrouper,
+    SVDGrouper,
+)
+from torchao.prototype.pat.layers.masked_layernorm import MaskedLayerNorm
+from torchao.prototype.pat.optim import ProxGroupLasso, ProxNuclearNorm, PruneOptimizer
+from torchao.prototype.pat.utils import get_param_groups
+
+
+class TestMaskedLayerNorm(common_utils.TestCase):
+    @common_utils.parametrize("batch", [1, 4])
+    @common_utils.parametrize("seq_len", [2, 8])
+    @common_utils.parametrize("embed_dim", [16, 64])
+    def test_masked_layernorm(self, batch=1, seq_len=2, embed_dim=16):
+        dim2_nz = embed_dim // 2
+        embed = torch.randn(batch, seq_len, embed_dim)
+        embed[..., dim2_nz:] = 0
+
+        masked_layer_norm = MaskedLayerNorm(embed_dim)
+        layer_norm = nn.LayerNorm(dim2_nz)
+        with torch.no_grad():
+            layer_norm.weight.copy_(masked_layer_norm.weight[:dim2_nz])
+            layer_norm.bias.copy_(masked_layer_norm.bias[:dim2_nz])
+
+        out = masked_layer_norm(embed)
+        expected_out = layer_norm(embed[..., :dim2_nz])
+        torch.testing.assert_close(out[..., :dim2_nz], expected_out)
+
+
+class MHADummyModel(nn.Module):
+    def __init__(self, embed_dim, num_heads, n_cls):
+        super().__init__()
+        self.mha = nn.MultiheadAttention(embed_dim, num_heads, bias=False)
+        self.classifier = nn.Linear(embed_dim, n_cls)
+
+    def forward(self, x):
+        attn_output, _ = self.mha(x, x, x)
+        out = self.classifier(attn_output)
+        return out
+
+
+class TestQKGrouper(common_utils.TestCase):
+    def __init__(self, methodName):
+        super(TestQKGrouper, self).__init__(methodName)
+        self.reg_lambda = 1.0
+        self.prox_map = ProxGroupLasso(self.reg_lambda)
+
+    @staticmethod
+    def _get_qk(p, embed_dim, qk_reg_index):
+        qk = p[:embed_dim] if qk_reg_index == 0 else p[embed_dim : (embed_dim * 2)]
+        return qk
+
+    def get_gamma(self, p):
+        """Heuristic that uses the mean of the group to set gamma."""
+        p_col = p[:, 0]
+        gamma = (1 - p_col.mean()) * torch.linalg.vector_norm(p_col)
+        gamma.div_(self.prox_map.tau(p_col))
+        return gamma
+
+    def _test_post_prune(self, p, qk_orig, embed_dim, qk_reg_index, gamma):
+        qk = self._get_qk(p, embed_dim, qk_reg_index)
+        nz_mask = qk.sum(dim=0).ne(0)
+        self.assertTrue(nz_mask.eq(0).any(), "No columns of Q/K were pruned")
+
+        # original columns that are <= gamma are pruned
+        expect_nz_mask = qk_orig.gt(gamma).all(dim=0)
+        torch.testing.assert_close(nz_mask > 1, expect_nz_mask, atol=0, rtol=0)
+
+    def _test_mha_inner(self, p, embed_dim, qk_reg_index):
+        qk_orig = self._get_qk(p, embed_dim, qk_reg_index).clone()
+        qk_no_prune = self._get_qk(p, embed_dim, int(not qk_reg_index)).clone()
+        v_orig = p[(embed_dim * 2) :].clone()
+        qk_pack_dim = 0
+        with QKGrouper(p, qk_pack_dim, qk_reg_index) as grouper:
+            self.assertTrue(grouper.p.equal(qk_orig))
+
+            gamma = self.get_gamma(grouper.p)
+            _ = torch.vmap(
+                self.prox_map.apply_, in_dims=(grouper.in_dims, None), out_dims=0
+            )(grouper.p, gamma)
+
+        self._test_post_prune(p, qk_orig, embed_dim, qk_reg_index, gamma)
+
+        # unregularized query or key was not modified
+        no_prune = self._get_qk(p, embed_dim, int(not qk_reg_index))
+        torch.testing.assert_close(no_prune, qk_no_prune, atol=0, rtol=0)
+
+        # value was not modified
+        v = p[(embed_dim * 2) :]
+        torch.testing.assert_close(v, v_orig, atol=0, rtol=0)
+
+    @common_utils.parametrize("embed_dim", [16, 64])
+    @common_utils.parametrize("num_heads", [2, 4])
+    @common_utils.parametrize("qk_reg_index", [0, 1])
+    def test_pytorch_mha(self, embed_dim=16, num_heads=4, qk_reg_index=0):
+        assert embed_dim % num_heads == 0, (
+            f"{embed_dim=} must be divisible by {num_heads=}"
+        )
+
+        # single in_proj_weight of shape (embed_dim * 3, embed_dim)
+        model = nn.MultiheadAttention(embed_dim, num_heads, bias=False)
+        p = model.in_proj_weight.detach()
+        self._test_mha_inner(p, embed_dim, qk_reg_index)
+
+    @common_utils.parametrize("qk_reg_index", [0, 1])
+    def test_e2e_optimizer(self, embed_dim=64, qk_reg_index=0):
+        n_cls = 3
+        model = MHADummyModel(embed_dim, num_heads=4, n_cls=n_cls)
+        prune_config = {
+            "mha.in_proj_weight": {
+                "group_type": "QKGrouper",
+                "prox_type": "ProxGroupLasso",
+                "qk_pack_dim": 0,
+                "qk_reg_index": qk_reg_index,
+            }
+        }
+        param_groups = get_param_groups(model, prune_config, verbose=False)
+        self.assertEqual(len(param_groups), 3)
+
+        p = model.mha.in_proj_weight.detach()
+        qk_orig = self._get_qk(p, embed_dim, qk_reg_index).clone()
+
+        # set lr to gamma since we run a single step
+        gamma = self.get_gamma(qk_orig)
+        optimizer = PruneOptimizer(
+            torch.optim.SGD(param_groups, lr=gamma), reg_lambda=self.reg_lambda
+        )
+
+        data = torch.randn(1, 8, embed_dim)
+        label = torch.arange(0, n_cls) * data.mean(axis=-1, keepdim=True)
+        output = model(data)
+        loss = nn.functional.mse_loss(output, label)
+
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+
+        self._test_post_prune(p, qk_orig, embed_dim, qk_reg_index, gamma)
+
+
+class TestAttentionHeadGrouper(common_utils.TestCase):
+    def __init__(self, methodName):
+        super(TestAttentionHeadGrouper, self).__init__(methodName)
+        self.reg_lambda = 1.0
+        self.prox_map = ProxGroupLasso(self.reg_lambda)
+
+    @staticmethod
+    def _get_view_shape_reduce_dim(dim, num_heads, head_pack_dim):
+        if head_pack_dim == 0:
+            view_shape = (num_heads, -1, dim)
+            reduce_dim = (1, 2)
+        else:
+            view_shape = (dim, num_heads, -1)
+            reduce_dim = (0, 2)
+        return view_shape, reduce_dim
+
+    def _test_post_prune(self, p, p_orig, head_pack_dim, view_shape, reduce_dim, gamma):
+        nz_mask = p.view(*view_shape).sum(dim=reduce_dim).ne(0)
+        self.assertTrue(nz_mask.eq(0).any(), "No groups of p were pruned")
+
+        # original groups that are <= gamma are pruned
+        expect_nz_mask = p_orig.view(*view_shape).gt(gamma).all(dim=reduce_dim)
+        torch.testing.assert_close(nz_mask > 1, expect_nz_mask, atol=0, rtol=0)
+
+    def get_gamma(self, p, head_pack_dim, view_shape):
+        """Heuristic that uses the mean of the group to set gamma."""
+        p = p.view(*view_shape)
+        p_group = p[0] if head_pack_dim == 0 else p[:, 0]
+        gamma = (1 - p_group.mean()) * torch.linalg.vector_norm(p_group)
+        gamma.div_(self.prox_map.tau(p_group))
+        return gamma
+
+    @common_utils.parametrize("dim", [64, 128])
+    @common_utils.parametrize("head_pack_dim", [0, 1])
+    def test_head_grouper(self, dim=16, head_pack_dim=0, head_dim_ratio=8):
+        assert dim % head_dim_ratio == 0, (
+            f"{dim=} must be divisible by {head_dim_ratio=}"
+        )
+        num_heads = dim // 8
+        packed_dim = dim * num_heads
+        shape = (dim, packed_dim) if head_pack_dim == 0 else (packed_dim, dim)
+        model = nn.Linear(*shape, bias=False)
+        p = model.weight.detach()
+        p_orig = p.clone()
+        view_shape, reduce_dim = self._get_view_shape_reduce_dim(
+            dim, num_heads, head_pack_dim
+        )
+        grouper_cls = (
+            AttentionHeadGrouperDim0 if head_pack_dim == 0 else AttentionHeadGrouperDim1
+        )
+        with grouper_cls(p, num_heads) as grouper:
+            gamma = self.get_gamma(grouper.p, head_pack_dim, view_shape)
+            _ = torch.vmap(
+                self.prox_map.apply_, in_dims=(grouper.in_dims, None), out_dims=0
+            )(grouper.p, gamma)
+            self.assertEqual(grouper.p.size(head_pack_dim), num_heads)
+        self._test_post_prune(p, p_orig, head_pack_dim, view_shape, reduce_dim, gamma)
+
+
+class TestSVDGrouper(common_utils.TestCase):
+    def __init__(self, methodName):
+        super(TestSVDGrouper, self).__init__(methodName)
+        self.reg_lambda = 1.0
+        self.prox_map = ProxNuclearNorm(self.reg_lambda)
+
+    @common_utils.parametrize("embed_dim", (16, 64))
+    def test_grouper(self, embed_dim=16):
+        model = torch.nn.Linear(embed_dim, embed_dim)
+        p = model.weight
+        with SVDGrouper(p) as grouper:
+            gamma = grouper.p.mean()
+            p_orig = grouper.p.clone()
+            torch.vmap(
+                self.prox_map.apply_, in_dims=(grouper.in_dims, None), out_dims=0
+            )(grouper.p, gamma)
+            expect_nz_mask = p_orig.gt(gamma)
+            torch.testing.assert_close(grouper.p.ne(0), expect_nz_mask, atol=0, rtol=0)
+
+    @common_utils.parametrize("embed_dim", (16, 64))
+    @common_utils.parametrize("pack_dim", (0, 1))
+    def test_qk_grouper(self, embed_dim=16, pack_dim=0):
+        shape = [embed_dim, embed_dim]
+        shape[int(not pack_dim)] *= 3
+        model = torch.nn.Linear(*shape)
+        p = model.weight
+        with QKSVDGrouper(p, pack_dim=pack_dim) as grouper:
+            gamma = grouper.p.mean()
+            p_orig = grouper.p.clone()
+            torch.vmap(
+                self.prox_map.apply_, in_dims=(grouper.in_dims, None), out_dims=0
+            )(grouper.p, gamma)
+            expect_nz_mask = p_orig.gt(gamma)
+            torch.testing.assert_close(grouper.p.ne(0), expect_nz_mask, atol=0, rtol=0)
+
+    @common_utils.parametrize("embed_dim", (16, 64))
+    @common_utils.parametrize("pack_dim", (0, 1))
+    def test_packed_grouper(self, embed_dim=16, npack=3, pack_dim=0):
+        shape = [embed_dim, embed_dim]
+        shape[int(not pack_dim)] *= npack
+        model = torch.nn.Linear(*shape)
+        p = model.weight
+        with PackedSVDGrouper(p, npack, pack_dim=pack_dim) as grouper:
+            gamma = grouper.p.mean(0).mean()
+            p_orig = grouper.p.clone()
+            torch.vmap(
+                self.prox_map.apply_, in_dims=(grouper.in_dims, None), out_dims=0
+            )(grouper.p.flatten(), gamma)
+            torch.testing.assert_close(
+                grouper.p.ne(0), p_orig.gt(gamma), atol=0, rtol=0
+            )
+            self.assertEqual(p.data_ptr(), grouper._p.data_ptr())
+
+
+common_utils.instantiate_parametrized_tests(TestMaskedLayerNorm)
+common_utils.instantiate_parametrized_tests(TestQKGrouper)
+common_utils.instantiate_parametrized_tests(TestAttentionHeadGrouper)
+common_utils.instantiate_parametrized_tests(TestSVDGrouper)
+
+if __name__ == "__main__":
+    random.seed(0)
+    torch.manual_seed(0)
+    unittest.main()

torchao/prototype/pat/README.md

@@ -0,0 +1,50 @@
+# PAT: Pruning-Aware Training
+
+PAT is a library based on group Lasso regularization. It directly induces structured sparsity during training, removing the need for custom pruning metrics and/or multiple rounds of training.
+
+PAT's simple optimizer-only interface supports easy integration into existing training pipelines. The code is organized into two main components:
+* grouper: defines the granularity of pruning (e.g., filter, channel, layer)
+* proximal mapping: projects groups of weights onto sparse values
+
+## Optimizer-only interface
+
+This package provides a `PruneOptimizer` that simply wraps around a base optimizer inheriting from `torch.optim.Optimizer`. The following code snippet illustrates how to set up PAT:
+
+```python
+from pat.optim import PruneOptimizer
+
+model = torchvision.models.resnet18().cuda()
+
+# split params into prunable and non-prunable groups
+weights = [p for name, p in model.named_parameters() if name.endswith("weight")]
+others = [p for name, p in model.named_parameters() if not name.endswith("weight")]
+
+# apply row-wise group Lasso regularization to the weights
+param_groups = [
+    {
+        "params": weights",
+        "group_type": "pat.group.Dim0Grouper",
+        "prox_type": "pat.prox.ProxGroupLasso",
+        "reg_lambda": 2e-4,
+    },
+    {"params": others},
+]
+
+# create base optimizer (SGD, Adam or AdamW)
+base_optimizer = torch.optim.SGD(
+    param_groups, lr=0.1, momentum=0.9, weight_decay=1e-4
+)
+
+# create PruneOptimizer
+optimizer = PruneOptimizer(base_optimizer)
+```
+
+After creating `PruneOptimizer`, one can use it as a regular PyTorch optimizer.
+
+## Pruning configuration
+
+Pruning configs are dictionaries that define which parameter groups to prune and how to prune them. Each key-value pair in the config maps to a prunable parameter group of `PruneOptimizer`. The keys are used to match model parameters, while the values specify the pruning granularity and proximal map. The key can be one of the following types:
+
+- parameter name (string): e.g., `blocks.0.attn.qkv.weight`
+- regex pattern (string): e.g., `:.*attn\.qkv\.weight`
+- module type and parameter name suffix ((class, string) tuple): e.g., `(torch.nn.Linear, 'weight')`

torchao/prototype/pat/__init__.py

@@ -0,0 +1,27 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD 3-Clause license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .group import (  # noqa: F401
+    AttentionHeadGrouperDim0,
+    AttentionHeadGrouperDim1,
+    ConvFilterGrouper,
+    Dim0Grouper,
+    Dim1Grouper,
+    ElemGrouper,
+    LayerGrouper,
+    PackedSVDGrouper,
+    QKGrouper,
+    QKSVDGrouper,
+    SVDGrouper,
+)
+from .optim import (  # noqa: F401
+    NMSGDOptimizer,
+    ProxGroupLasso,
+    ProxGroupLassoReduce,
+    ProxLasso,
+    ProxNuclearNorm,
+    PruneOptimizer,
+)

torchao/prototype/pat/group/__init__.py

@@ -0,0 +1,19 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD 3-Clause license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .attention import (  # noqa: F401
+    AttentionHeadGrouperDim0,
+    AttentionHeadGrouperDim1,
+    QKGrouper,
+)
+from .conv import ConvFilterGrouper  # noqa: F401
+from .dim import Dim0Grouper, Dim1Grouper  # noqa: F401
+from .grouper import (  # noqa: F401
+    ElemGrouper,
+    Grouper,
+    LayerGrouper,
+)
+from .low_rank import PackedSVDGrouper, QKSVDGrouper, SVDGrouper  # noqa: F401