File size: 13,201 Bytes
c7f3ffb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
import torch
import torch.distributed as dist


def zeropower_via_newtonschulz5(G, steps: int):
    """
    Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
    quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
    of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
    zero even beyond the point where the iteration no longer converges all the way to one everywhere
    on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
    where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
    performance at all relative to UV^T, where USV^T = G is the SVD.
    """
    assert G.ndim >= 2 # batched Muon implementation by @scottjmaddox, and put into practice in the record by @YouJiacheng
    a, b, c = (3.4445, -4.7750,  2.0315)
    X = G.bfloat16()
    if G.size(-2) > G.size(-1):
        X = X.mT

    # Ensure spectral norm is at most 1
    X = X / (X.norm(dim=(-2, -1), keepdim=True) + 1e-7)
    # Perform the NS iterations
    for _ in range(steps):
        A = X @ X.mT
        B = b * A + c * A @ A # quintic computation strategy adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
        X = a * X + B @ X
    
    if G.size(-2) > G.size(-1):
        X = X.mT
    return X


def muon_update(grad, momentum, beta=0.95, ns_steps=5, nesterov=True):
    momentum.lerp_(grad, 1 - beta)
    update = grad.lerp_(momentum, beta) if nesterov else momentum
    if update.ndim == 4: # for the case of conv filters
        update = update.view(len(update), -1)
    update = zeropower_via_newtonschulz5(update, steps=ns_steps)
    update *= max(1, grad.size(-2) / grad.size(-1))**0.5
    return update


class Muon(torch.optim.Optimizer):
    """
    Muon - MomentUm Orthogonalized by Newton-schulz

    https://kellerjordan.github.io/posts/muon/

    Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
    processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
    matrix. For efficient orthogonalization we use a Newton-Schulz iteration, which has the
    advantage that it can be stably run in bfloat16 on the GPU.

    Muon should only be used for hidden weight layers. The input embedding, final output layer,
    and any internal gains or biases should be optimized using a standard method such as AdamW.
    Hidden convolutional weights can be trained using Muon by viewing them as 2D and then
    collapsing their last 3 dimensions.

    Arguments:
        lr: The learning rate, in units of spectral norm per update.
        weight_decay: The AdamW-style weight decay.
        momentum: The momentum. A value of 0.95 here is usually fine.
    """
    def __init__(self, params, lr=0.02, weight_decay=0, momentum=0.95):
        defaults = dict(lr=lr, weight_decay=weight_decay, momentum=momentum)
        assert isinstance(params, list) and len(params) >= 1 and isinstance(params[0], torch.nn.Parameter)
        params = sorted(params, key=lambda x: x.size(), reverse=True)
        super().__init__(params, defaults)

    @torch.no_grad()
    def step(self, closure=None):

        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            params = group["params"]
            params_pad = params + [torch.empty_like(params[-1])] * (dist.get_world_size() - len(params) % dist.get_world_size())
            for base_i in range(len(params))[::dist.get_world_size()]:
                if base_i + dist.get_rank() < len(params):
                    p = params[base_i + dist.get_rank()]
                    if p.grad is None:
                        # continue
                        p.grad = torch.zeros_like(p)  # Force synchronization
                    state = self.state[p]
                    if len(state) == 0:
                        state["momentum_buffer"] = torch.zeros_like(p)
                    update = muon_update(p.grad, state["momentum_buffer"], beta=group["momentum"])
                    p.mul_(1 - group["lr"] * group["weight_decay"])
                    p.add_(update.reshape(p.shape), alpha=-group["lr"])
                dist.all_gather(params_pad[base_i:base_i + dist.get_world_size()], params_pad[base_i + dist.get_rank()])

        return loss


class SingleDeviceMuon(torch.optim.Optimizer):
    """
    Muon variant for usage in non-distributed settings.
    """
    def __init__(self, params, lr=0.02, weight_decay=0, momentum=0.95):
        defaults = dict(lr=lr, weight_decay=weight_decay, momentum=momentum)
        super().__init__(params, defaults)

    @torch.no_grad()
    def step(self, closure=None):

        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            for p in group["params"]:
                if p.grad is None:
                    # continue
                    p.grad = torch.zeros_like(p)  # Force synchronization
                state = self.state[p]
                if len(state) == 0:
                    state["momentum_buffer"] = torch.zeros_like(p)
                update = muon_update(p.grad, state["momentum_buffer"], beta=group["momentum"])
                p.mul_(1 - group["lr"] * group["weight_decay"])
                p.add_(update.reshape(p.shape), alpha=-group["lr"])

        return loss


def adam_update(grad, buf1, buf2, step, betas, eps):
    buf1.lerp_(grad, 1 - betas[0])
    buf2.lerp_(grad.square(), 1 - betas[1])
    buf1c = buf1 / (1 - betas[0]**step)
    buf2c = buf2 / (1 - betas[1]**step)
    return buf1c / (buf2c.sqrt() + eps)


class MuonWithAuxAdam(torch.optim.Optimizer):
    """
    Distributed Muon variant that can be used for all parameters in the network, since it runs an
    internal AdamW for the parameters that are not compatible with Muon. The user must manually
    specify which parameters shall be optimized with Muon and which with Adam by passing in a
    list of param_groups with the `use_muon` flag set.

    The point of this class is to allow the user to have a single optimizer in their code, rather
    than having both a Muon and an Adam which each need to be stepped.

    You can see an example usage below:

    https://github.com/KellerJordan/modded-nanogpt/blob/master/records/052525_MuonWithAuxAdamExample/b01550f9-03d8-4a9c-86fe-4ab434f1c5e0.txt#L470
    ```
    hidden_matrix_params = [p for n, p in model.blocks.named_parameters() if p.ndim >= 2 and "embed" not in n]
    embed_params = [p for n, p in model.named_parameters() if "embed" in n]
    scalar_params = [p for p in model.parameters() if p.ndim < 2]
    head_params = [model.lm_head.weight]

    from muon import MuonWithAuxAdam
    adam_groups = [dict(params=head_params, lr=0.22), dict(params=embed_params, lr=0.6), dict(params=scalar_params, lr=0.04)]
    adam_groups = [dict(**g, betas=(0.8, 0.95), eps=1e-10, use_muon=False) for g in adam_groups]
    muon_group = dict(params=hidden_matrix_params, lr=0.05, momentum=0.95, use_muon=True)
    param_groups = [*adam_groups, muon_group]
    optimizer = MuonWithAuxAdam(param_groups)
    ```
    """
    def __init__(self, param_groups):
        for group in param_groups:
            assert "use_muon" in group
            if group["use_muon"]:
                group["params"] = sorted(group["params"], key=lambda x: x.size(), reverse=True)
                # defaults
                group["lr"] = group.get("lr", 0.02)
                group["momentum"] = group.get("momentum", 0.95)
                group["weight_decay"] = group.get("weight_decay", 0)
                assert set(group.keys()) == set(["params", "lr", "momentum", "weight_decay", "use_muon"])
            else:
                # defaults
                group["lr"] = group.get("lr", 3e-4)
                group["betas"] = group.get("betas", (0.9, 0.95))
                group["eps"] = group.get("eps", 1e-10)
                group["weight_decay"] = group.get("weight_decay", 0)
                assert set(group.keys()) == set(["params", "lr", "betas", "eps", "weight_decay", "use_muon"])
        super().__init__(param_groups, dict())

    @torch.no_grad()
    def step(self, closure=None):

        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            if group["use_muon"]:
                params = group["params"]
                params_pad = params + [torch.empty_like(params[-1])] * (dist.get_world_size() - len(params) % dist.get_world_size())
                for base_i in range(len(params))[::dist.get_world_size()]:
                    if base_i + dist.get_rank() < len(params):
                        p = params[base_i + dist.get_rank()]
                        if p.grad is None:
                            # continue
                            p.grad = torch.zeros_like(p)  # Force synchronization
                        state = self.state[p]
                        if len(state) == 0:
                            state["momentum_buffer"] = torch.zeros_like(p)
                        update = muon_update(p.grad, state["momentum_buffer"], beta=group["momentum"])
                        p.mul_(1 - group["lr"] * group["weight_decay"])
                        p.add_(update.reshape(p.shape), alpha=-group["lr"])
                    dist.all_gather(params_pad[base_i:base_i + dist.get_world_size()], params_pad[base_i + dist.get_rank()])
            else:
                for p in group["params"]:
                    if p.grad is None:
                        # continue
                        p.grad = torch.zeros_like(p)  # Force synchronization
                    state = self.state[p]
                    if len(state) == 0:
                        state["exp_avg"] = torch.zeros_like(p)
                        state["exp_avg_sq"] = torch.zeros_like(p)
                        state["step"] = 0
                    state["step"] += 1
                    update = adam_update(p.grad, state["exp_avg"], state["exp_avg_sq"],
                                         state["step"], group["betas"], group["eps"])
                    p.mul_(1 - group["lr"] * group["weight_decay"])
                    p.add_(update, alpha=-group["lr"])

        return loss


class SingleDeviceMuonWithAuxAdam(torch.optim.Optimizer):
    """
    Non-distributed variant of MuonWithAuxAdam.
    """
    def __init__(self, param_groups):
        for group in param_groups:
            assert "use_muon" in group
            if group["use_muon"]:
                # defaults
                group["lr"] = group.get("lr", 0.02)
                group["momentum"] = group.get("momentum", 0.95)
                group["weight_decay"] = group.get("weight_decay", 0)
                assert set(group.keys()) == set(["params", "lr", "momentum", "weight_decay", "use_muon"])
            else:
                # defaults
                group["lr"] = group.get("lr", 3e-4)
                group["betas"] = group.get("betas", (0.9, 0.95))
                group["eps"] = group.get("eps", 1e-10)
                group["weight_decay"] = group.get("weight_decay", 0)
                assert set(group.keys()) == set(["params", "lr", "betas", "eps", "weight_decay", "use_muon"])
        super().__init__(param_groups, dict())

    @torch.no_grad()
    def step(self, closure=None):

        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            if group["use_muon"]:
                for p in group["params"]:
                    if p.grad is None:
                        # continue
                        p.grad = torch.zeros_like(p)  # Force synchronization
                    state = self.state[p]
                    if len(state) == 0:
                        state["momentum_buffer"] = torch.zeros_like(p)
                    update = muon_update(p.grad, state["momentum_buffer"], beta=group["momentum"])
                    p.mul_(1 - group["lr"] * group["weight_decay"])
                    p.add_(update.reshape(p.shape), alpha=-group["lr"])
            else:
                for p in group["params"]:
                    if p.grad is None:
                        # continue
                        p.grad = torch.zeros_like(p)  # Force synchronization
                    state = self.state[p]
                    if len(state) == 0:
                        state["exp_avg"] = torch.zeros_like(p)
                        state["exp_avg_sq"] = torch.zeros_like(p)
                        state["step"] = 0
                    state["step"] += 1
                    update = adam_update(p.grad, state["exp_avg"], state["exp_avg_sq"],
                                         state["step"], group["betas"], group["eps"])
                    p.mul_(1 - group["lr"] * group["weight_decay"])
                    p.add_(update, alpha=-group["lr"])

        return loss