Add option to unfuse Wqkv (#1367)

* yo

* logging

* bro

* yo

* yo

* yo

* lint

* if

* datest

* webacc

* liny

llmfoundry/models/layers/attention.py

@@ -411,6 +411,7 @@ def __init__(
         clip_qkv: Optional[float] = None,
         qk_ln: bool = False,
         qk_gn: bool = False,
+        fused_qkv: bool = True,
         softmax_scale: Optional[float] = None,
         attn_pdrop: float = 0.0,
         norm_type: str = 'low_precision_layernorm',
@@ -426,6 +427,7 @@ def __init__(
         self.clip_qkv = clip_qkv
         self.qk_ln = qk_ln
         self.qk_gn = qk_gn
+        self.fused_qkv = fused_qkv
 
         self.d_model = d_model
         self.n_heads = n_heads
@@ -462,7 +464,17 @@ def __init__(
             self.softmax_scale = 1 / math.sqrt(self.d_model / self.n_heads)
         self.attn_dropout_p = attn_pdrop
 
-        if self.reuse_kv_layer_idx is None:
+        if self.reuse_kv_layer_idx is not None:
+            self.Wq = build_fc(
+                name=fc_type_name,
+                in_features=self.d_model,
+                out_features=self.d_model,
+                fc_kwargs=fc_type,
+            )
+            # for param init fn; enables shape based init of fused layers
+            fuse_splits = [i * self.head_dim for i in range(1, self.n_heads)]
+            self.Wq._fused = (0, fuse_splits)
+        elif self.fused_qkv:
             self.Wqkv = build_fc(
                 name=fc_type_name,
                 in_features=self.d_model,
@@ -482,9 +494,26 @@ def __init__(
                 out_features=self.d_model,
                 fc_kwargs=fc_type,
             )
+            self.Wk = build_fc(
+                name=fc_type_name,
+                in_features=self.d_model,
+                out_features=self.kv_n_heads * self.head_dim,
+                fc_kwargs=fc_type,
+            )
+            self.Wv = build_fc(
+                name=fc_type_name,
+                in_features=self.d_model,
+                out_features=self.kv_n_heads * self.head_dim,
+                fc_kwargs=fc_type,
+            )
             # for param init fn; enables shape based init of fused layers
-            fuse_splits = [i * self.head_dim for i in range(1, self.n_heads)]
-            self.Wq._fused = (0, fuse_splits)
+            q_fuse_splits = [i * self.head_dim for i in range(1, self.n_heads)]
+            kv_fuse_splits = [
+                i * self.head_dim for i in range(1, self.kv_n_heads)
+            ]
+            self.Wq._fused = (0, q_fuse_splits)
+            self.Wk._fused = (0, kv_fuse_splits)
+            self.Wv._fused = (0, kv_fuse_splits)
 
         if self.qk_ln or self.qk_gn:
             norm_size = self.head_dim if qk_gn else d_model
@@ -601,19 +630,29 @@ def get_qkv(
                 query = self.q_ln(query).to(dtype).view(q_shape)
             return query, key, value
 
-        qkv = self.Wqkv(x)
+        if self.fused_qkv:
+            qkv = self.Wqkv(x)
 
-        if self.clip_qkv:
-            qkv = qkv.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+            if self.clip_qkv:
+                qkv = qkv.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+
+            query, key, value = qkv.split(
+                [
+                    self.d_model,
+                    self.kv_n_heads * self.head_dim,
+                    self.kv_n_heads * self.head_dim,
+                ],
+                dim=2,
+            )
+        else:
+            query = self.Wq(x)
+            key = self.Wk(x)
+            value = self.Wv(x)
 
-        query, key, value = qkv.split(
-            [
-                self.d_model,
-                self.kv_n_heads * self.head_dim,
-                self.kv_n_heads * self.head_dim,
-            ],
-            dim=2,
-        )
+            if self.clip_qkv:
+                query = query.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+                key = key.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+                value = value.clamp(min=-self.clip_qkv, max=self.clip_qkv)
 
         if self.qk_ln or self.qk_gn:
             # Applying layernorm to qk
@@ -753,6 +792,7 @@ def __init__(
         clip_qkv: Optional[float] = None,
         qk_ln: bool = False,
         qk_gn: bool = False,
+        fused_qkv: bool = True,
         softmax_scale: Optional[float] = None,
         attn_pdrop: float = 0.0,
         norm_type: str = 'low_precision_layernorm',
@@ -770,6 +810,7 @@ def __init__(
             clip_qkv=clip_qkv,
             qk_ln=qk_ln,
             qk_gn=qk_gn,
+            fused_qkv=fused_qkv,
             softmax_scale=softmax_scale,
             attn_pdrop=attn_pdrop,
             norm_type=norm_type,
@@ -796,6 +837,7 @@ def __init__(
         clip_qkv: Optional[float] = None,
         qk_ln: bool = False,
         qk_gn: bool = False,
+        fused_qkv: bool = True,
         softmax_scale: Optional[float] = None,
         attn_pdrop: float = 0.0,
         norm_type: str = 'low_precision_layernorm',
@@ -813,6 +855,7 @@ def __init__(
             clip_qkv=clip_qkv,
             qk_ln=qk_ln,
             qk_gn=qk_gn,
+            fused_qkv=fused_qkv,
             softmax_scale=softmax_scale,
             attn_pdrop=attn_pdrop,
             norm_type=norm_type,

llmfoundry/models/mpt/configuration_mpt.py

@@ -70,6 +70,8 @@ def __init__(
                 attn_impl (str): The attention implementation to use. One of 'torch' or 'flash'.
                 qk_ln (bool): Whether to apply layer normalization to the queries and keys in the attention layer.
                 qk_gn (bool): Whether to apply group normalization to the queries and keys in the attention layer.
+                fused_qkv (bool): Whether to fuse the Wq, Wk, and Wv weight matrices in the attention layer. If True, the weights are fused into a single
+                    Wqkv matrix, which can be faster for matmuls. If False, the weights are kept separate. Defaults to True.
                 clip_qkv (Optional[float]): If not None, clip the queries, keys, and values in the attention layer to
                     this value.
                 softmax_scale (Optional[float]): If not None, scale the softmax in the attention layer by this value. If None,

llmfoundry/models/utils/config_defaults.py

@@ -15,6 +15,7 @@
     'attn_impl': 'flash',
     'qk_ln': False,
     'qk_gn': False,
+    'fused_qkv': True,
     'clip_qkv': None,
     'softmax_scale': None,
     'attn_uses_sequence_id': False,

tests/models/layers/test_attention.py

@@ -0,0 +1,160 @@
+# Copyright 2024 MosaicML LLM Foundry authors
+# SPDX-License-Identifier: Apache-2.0
+
+import pytest
+import torch
+
+from llmfoundry.models.layers.layer_builders import build_attention_layer
+
+
+@pytest.mark.parametrize(
+    'attn_name',
+    ['multihead_attention', 'grouped_query_attention', 'multiquery_attention'],
+)
+@pytest.mark.parametrize('dim', [1024])
+def test_unfused_wqkv(attn_name: str, dim: int):
+    d_head = 128
+    n_heads = dim // d_head
+
+    generic_attn_kwargs = {
+        'd_model': dim,
+        'n_heads': n_heads,
+        'fc_type': {
+            'name': 'torch',
+        },
+        'device': 'cpu',
+        'attn_pdrop': 0.0,
+        'attn_impl': 'torch',
+        'qk_ln': False,
+        'qk_gn': False,
+        'clip_qkv': None,
+        'softmax_scale': None,
+        'sliding_window_size': -1,
+    }
+
+    if attn_name == 'grouped_query_attention':
+        kv_n_heads = 2
+        generic_attn_kwargs['kv_n_heads'] = kv_n_heads
+    elif attn_name == 'multiquery_attention':
+        kv_n_heads = 1
+    elif attn_name == 'multihead_attention':
+        kv_n_heads = n_heads
+    else:
+        raise ValueError(f'Unknown attention name: {attn_name}')
+
+    attn_config_fused = generic_attn_kwargs.copy()
+    attn_config_fused['fused_qkv'] = True
+
+    attn_config_unfused = generic_attn_kwargs.copy()
+    attn_config_unfused['fused_qkv'] = False
+
+    attn_fused = build_attention_layer(
+        name=attn_name,
+        attn_kwargs=attn_config_fused,
+    )
+    attn_unfused = build_attention_layer(
+        name=attn_name,
+        attn_kwargs=attn_config_unfused,
+    )
+
+    # Make sure unfused attention has the same params as the fused one.
+    fused_wqkv = attn_fused.Wqkv.weight.detach().clone()
+    kv_heads_len = (fused_wqkv.shape[0] - dim) // 2
+    Wq_shape_before = (attn_unfused.Wq.weight.shape, attn_unfused.Wq.bias.shape)
+    Wk_shape_before = (attn_unfused.Wk.weight.shape, attn_unfused.Wk.bias.shape)
+    Wv_shape_before = (attn_unfused.Wv.weight.shape, attn_unfused.Wv.bias.shape)
+
+    attn_unfused.Wq.weight.data = fused_wqkv[:dim, :]
+    attn_unfused.Wk.weight.data = fused_wqkv[dim:dim + kv_heads_len, :]
+    attn_unfused.Wv.weight.data = fused_wqkv[dim + kv_heads_len:, :]
+    attn_unfused.out_proj.weight.data = attn_fused.out_proj.weight
+    attn_unfused.Wq.bias.data = attn_fused.Wqkv.bias[:dim]
+    attn_unfused.Wk.bias.data = attn_fused.Wqkv.bias[dim:dim + kv_heads_len]
+    attn_unfused.Wv.bias.data = attn_fused.Wqkv.bias[dim + kv_heads_len:]
+    attn_unfused.out_proj.bias.data = attn_fused.out_proj.bias
+
+    # Make sure initialization fuse splits are as expected.
+    all_fuse_splits = (
+        0,
+        [i * d_head for i in range(1, n_heads + 2 * kv_n_heads)],
+    )
+    q_fuse_splits = (0, [i * d_head for i in range(1, n_heads)])
+    kv_fuse_splits = (0, [i * d_head for i in range(1, kv_n_heads)])
+
+    assert attn_fused.Wqkv._fused == all_fuse_splits
+    assert attn_unfused.Wq._fused == q_fuse_splits
+    assert attn_unfused.Wk._fused == kv_fuse_splits
+    assert attn_unfused.Wv._fused == kv_fuse_splits
+
+    assert torch.allclose(
+        attn_fused.Wqkv.weight,
+        torch.cat(
+            [
+                attn_unfused.Wq.weight,
+                attn_unfused.Wk.weight,
+                attn_unfused.Wv.weight,
+            ],
+            dim=0,
+        ),
+    )
+    assert torch.allclose(
+        attn_fused.Wqkv.bias,
+        torch.cat(
+            [
+                attn_unfused.Wq.bias,
+                attn_unfused.Wk.bias,
+                attn_unfused.Wv.bias,
+            ],
+            dim=0,
+        ),
+    )
+    assert torch.allclose(
+        attn_fused.out_proj.weight,
+        attn_unfused.out_proj.weight,
+    )
+    assert torch.allclose(attn_fused.out_proj.bias, attn_unfused.out_proj.bias)
+
+    assert Wq_shape_before == (
+        attn_unfused.Wq.weight.shape,
+        attn_unfused.Wq.bias.shape,
+    )
+    assert Wk_shape_before == (
+        attn_unfused.Wk.weight.shape,
+        attn_unfused.Wk.bias.shape,
+    )
+    assert Wv_shape_before == (
+        attn_unfused.Wv.weight.shape,
+        attn_unfused.Wv.bias.shape,
+    )
+
+    x1 = torch.randn(1, 1, dim)
+    x2 = x1.detach().clone()
+    x1.requires_grad = True
+    x2.requires_grad = True
+
+    out_fused, _, _ = attn_fused(x1)
+    out_unfused, _, _ = attn_unfused(x2)
+
+    assert torch.allclose(out_fused, out_unfused)
+
+    # Dummy loss function is simply the sum.
+    loss_fused = out_fused.sum()
+    loss_fused.backward()
+
+    loss_unfused = out_unfused.sum()
+    loss_unfused.backward()
+
+    assert isinstance(x1.grad, torch.Tensor)
+    assert isinstance(x2.grad, torch.Tensor)
+    assert torch.allclose(x1.grad, x2.grad)
+    combined_grad = torch.concat(
+        [
+            attn_unfused.Wq.weight.grad,
+            attn_unfused.Wk.weight.grad,
+            attn_unfused.Wv.weight.grad,
+        ],
+        dim=0,
+    )
+    assert isinstance(attn_fused.Wqkv.weight.grad, torch.Tensor)
+    assert isinstance(combined_grad, torch.Tensor)
+    assert torch.allclose(attn_fused.Wqkv.weight.grad, combined_grad)