Commit
·
8535e80
1
Parent(s):
cf531ba
chore: initial commit
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- .gitattributes +3 -0
- README.md +33 -0
- build.toml +23 -0
- build/torch26-cxx11-cu118-x86_64-linux/optimizer/__init__.py +5 -0
- build/torch26-cxx11-cu118-x86_64-linux/optimizer/_ops.py +9 -0
- build/torch26-cxx11-cu118-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so +3 -0
- build/torch26-cxx11-cu118-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so +3 -0
- build/torch26-cxx11-cu118-x86_64-linux/optimizer/muon.py +458 -0
- build/torch26-cxx11-cu124-x86_64-linux/optimizer/__init__.py +5 -0
- build/torch26-cxx11-cu124-x86_64-linux/optimizer/_ops.py +9 -0
- build/torch26-cxx11-cu124-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so +3 -0
- build/torch26-cxx11-cu124-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so +3 -0
- build/torch26-cxx11-cu124-x86_64-linux/optimizer/muon.py +458 -0
- build/torch26-cxx11-cu126-x86_64-linux/optimizer/__init__.py +5 -0
- build/torch26-cxx11-cu126-x86_64-linux/optimizer/_ops.py +9 -0
- build/torch26-cxx11-cu126-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so +3 -0
- build/torch26-cxx11-cu126-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so +3 -0
- build/torch26-cxx11-cu126-x86_64-linux/optimizer/muon.py +458 -0
- build/torch26-cxx11-rocm62-x86_64-linux/optimizer/__init__.py +5 -0
- build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_ops.py +9 -0
- build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_optimizer_20250614121529.abi3.so +3 -0
- build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_optimizer_20250614123843.abi3.so +3 -0
- build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so +3 -0
- build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so +3 -0
- build/torch26-cxx11-rocm62-x86_64-linux/optimizer/muon.py +458 -0
- build/torch26-cxx98-cu118-x86_64-linux/optimizer/__init__.py +5 -0
- build/torch26-cxx98-cu118-x86_64-linux/optimizer/_ops.py +9 -0
- build/torch26-cxx98-cu118-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so +3 -0
- build/torch26-cxx98-cu118-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so +3 -0
- build/torch26-cxx98-cu118-x86_64-linux/optimizer/muon.py +458 -0
- build/torch26-cxx98-cu124-x86_64-linux/optimizer/__init__.py +5 -0
- build/torch26-cxx98-cu124-x86_64-linux/optimizer/_ops.py +9 -0
- build/torch26-cxx98-cu124-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so +3 -0
- build/torch26-cxx98-cu124-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so +3 -0
- build/torch26-cxx98-cu124-x86_64-linux/optimizer/muon.py +458 -0
- build/torch26-cxx98-cu126-x86_64-linux/optimizer/__init__.py +5 -0
- build/torch26-cxx98-cu126-x86_64-linux/optimizer/_ops.py +9 -0
- build/torch26-cxx98-cu126-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so +3 -0
- build/torch26-cxx98-cu126-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so +3 -0
- build/torch26-cxx98-cu126-x86_64-linux/optimizer/muon.py +458 -0
- build/torch27-cxx11-cu118-x86_64-linux/optimizer/__init__.py +5 -0
- build/torch27-cxx11-cu118-x86_64-linux/optimizer/_ops.py +9 -0
- build/torch27-cxx11-cu118-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so +3 -0
- build/torch27-cxx11-cu118-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so +3 -0
- build/torch27-cxx11-cu118-x86_64-linux/optimizer/muon.py +458 -0
- build/torch27-cxx11-cu126-x86_64-linux/optimizer/__init__.py +5 -0
- build/torch27-cxx11-cu126-x86_64-linux/optimizer/_ops.py +9 -0
- build/torch27-cxx11-cu126-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so +3 -0
- build/torch27-cxx11-cu126-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so +3 -0
- build/torch27-cxx11-cu126-x86_64-linux/optimizer/muon.py +458 -0
.gitattributes
CHANGED
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
*.so filter=lfs diff=lfs merge=lfs -text
|
37 |
+
*.png filter=lfs diff=lfs merge=lfs -text
|
38 |
+
*.pdf filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
tags:
|
3 |
+
- kernel
|
4 |
+
---
|
5 |
+
|
6 |
+
# Optimizer
|
7 |
+
|
8 |
+
Optimizer is a python package that provides:
|
9 |
+
- PyTorch implementation of recent optimizer algorithms
|
10 |
+
- with support for parallelism techniques for efficient large-scale training.
|
11 |
+
|
12 |
+
### Currently implemented
|
13 |
+
- [Parallel Muon with FSDP2](./docs/muon/parallel_muon.pdf)
|
14 |
+
|
15 |
+
## Usage
|
16 |
+
|
17 |
+
```python
|
18 |
+
import torch
|
19 |
+
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
|
20 |
+
from kernels import get_kernel
|
21 |
+
|
22 |
+
optimizer = get_kernel("motif-technologies/optimizer")
|
23 |
+
|
24 |
+
model = None # your model here
|
25 |
+
fsdp_model = FSDP(model)
|
26 |
+
|
27 |
+
optim = optimizer.Muon(
|
28 |
+
fsdp_model.parameters(),
|
29 |
+
lr=0.01,
|
30 |
+
momentum=0.9,
|
31 |
+
weight_decay=1e-4,
|
32 |
+
)
|
33 |
+
```
|
build.toml
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[general]
|
2 |
+
name = "optimizer"
|
3 |
+
universal = false
|
4 |
+
|
5 |
+
[torch]
|
6 |
+
src = [
|
7 |
+
"torch-ext/torch_binding.cpp",
|
8 |
+
"torch-ext/torch_binding.h",
|
9 |
+
]
|
10 |
+
|
11 |
+
[kernel.activation]
|
12 |
+
backend = "rocm"
|
13 |
+
src = [
|
14 |
+
"optimizer/dummy.cu",
|
15 |
+
]
|
16 |
+
depends = [ "torch" ]
|
17 |
+
|
18 |
+
[kernel.activation_cuda]
|
19 |
+
backend = "cuda"
|
20 |
+
src = [
|
21 |
+
"optimizer/dummy.cu",
|
22 |
+
]
|
23 |
+
depends = [ "torch" ]
|
build/torch26-cxx11-cu118-x86_64-linux/optimizer/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .muon import Muon
|
2 |
+
|
3 |
+
__all__ = [
|
4 |
+
"Muon",
|
5 |
+
]
|
build/torch26-cxx11-cu118-x86_64-linux/optimizer/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _optimizer_b4b3752_dirty
|
3 |
+
ops = torch.ops._optimizer_b4b3752_dirty
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_optimizer_b4b3752_dirty::{op_name}"
|
build/torch26-cxx11-cu118-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:66ca698639fff584999fe65f8f10cc4436c197829e936be2741bf53db685caa0
|
3 |
+
size 1787272
|
build/torch26-cxx11-cu118-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f8325d12959ef4f31b6c6340eca29176f5077abeaa10f3a6651db55ccf3c634f
|
3 |
+
size 1787272
|
build/torch26-cxx11-cu118-x86_64-linux/optimizer/muon.py
ADDED
@@ -0,0 +1,458 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.distributed as dist
|
6 |
+
from torch.distributed._tensor import DTensor
|
7 |
+
|
8 |
+
|
9 |
+
# TODO leave original url and consider LICENSE
|
10 |
+
# This code snippet is a modified version adapted from the following GitHub repository:
|
11 |
+
# https://github.com/KellerJordan/Muon/blob/master/muon.py
|
12 |
+
def _zeropower_via_newtonschulz5(G, steps):
|
13 |
+
"""
|
14 |
+
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
|
15 |
+
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
|
16 |
+
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
|
17 |
+
zero even beyond the point where the iteration no longer converges all the way to one everywhere
|
18 |
+
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
|
19 |
+
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
|
20 |
+
performance at all relative to UV^T, where USV^T = G is the SVD.
|
21 |
+
"""
|
22 |
+
assert len(G.shape) == 2
|
23 |
+
a, b, c = (3.4445, -4.7750, 2.0315)
|
24 |
+
X = G # no manual typecast
|
25 |
+
if G.size(0) > G.size(1):
|
26 |
+
X = X.T
|
27 |
+
# Ensure spectral norm is at most 1
|
28 |
+
X = X / (X.norm() + 1e-7)
|
29 |
+
X = X.bfloat16()
|
30 |
+
# Perform the NS iterations
|
31 |
+
for _ in range(steps):
|
32 |
+
A = X @ X.T
|
33 |
+
# B = (
|
34 |
+
# b * A + c * A @ A
|
35 |
+
# ) # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
|
36 |
+
B = torch.addmm(A, A, A, alpha=c, beta=b)
|
37 |
+
# X = a * X + B @ X
|
38 |
+
X = torch.addmm(X, B, X, alpha=1.0, beta=a)
|
39 |
+
|
40 |
+
if G.size(0) > G.size(1):
|
41 |
+
X = X.T
|
42 |
+
return X.to(G.dtype)
|
43 |
+
|
44 |
+
|
45 |
+
@dataclass
|
46 |
+
class _muon_state:
|
47 |
+
# TODO: use Optional
|
48 |
+
worker_rank: int | None = None
|
49 |
+
gathered_grad: torch.Tensor | None = None
|
50 |
+
computed_u: torch.Tensor | None = None
|
51 |
+
scattered_u: torch.Tensor | None = None
|
52 |
+
gather_event: torch.cuda.Event | None = None
|
53 |
+
compute_event: torch.cuda.Event | None = None
|
54 |
+
|
55 |
+
|
56 |
+
def _gather(p, state, rank, comm_stream):
|
57 |
+
g = p.grad
|
58 |
+
mesh = g.device_mesh
|
59 |
+
|
60 |
+
if rank == state.worker_rank:
|
61 |
+
gather_list = [torch.empty_like(g.to_local()) for _ in range(mesh.mesh.numel())]
|
62 |
+
else:
|
63 |
+
gather_list = None
|
64 |
+
|
65 |
+
with torch.cuda.stream(comm_stream):
|
66 |
+
torch.distributed.gather(
|
67 |
+
g.to_local(),
|
68 |
+
dst=state.worker_rank,
|
69 |
+
gather_list=gather_list,
|
70 |
+
group=mesh.get_group(),
|
71 |
+
)
|
72 |
+
if rank == state.worker_rank:
|
73 |
+
# TODO: Consider ,,,
|
74 |
+
if state.gathered_grad is not None:
|
75 |
+
raise RuntimeError(
|
76 |
+
"Gather event already exists, which should not happen."
|
77 |
+
)
|
78 |
+
state.gathered_grad = torch.cat(gather_list, dim=0)
|
79 |
+
state.gather_event = torch.cuda.Event()
|
80 |
+
state.gather_event.record()
|
81 |
+
else:
|
82 |
+
state.gathered_grad = None
|
83 |
+
state.gather_event = None
|
84 |
+
|
85 |
+
|
86 |
+
def _compute_u(state, steps, rank, compute_stream):
|
87 |
+
with torch.cuda.stream(compute_stream):
|
88 |
+
if rank == state.worker_rank:
|
89 |
+
if state.gather_event is None:
|
90 |
+
raise RuntimeError("Gather event must be set before compute.")
|
91 |
+
compute_stream.wait_event(state.gather_event)
|
92 |
+
u = _zeropower_via_newtonschulz5(state.gathered_grad, steps)
|
93 |
+
state.computed_u = u
|
94 |
+
state.compute_event = torch.cuda.Event()
|
95 |
+
state.compute_event.record()
|
96 |
+
else:
|
97 |
+
state.computed_u = None
|
98 |
+
state.compute_event = None
|
99 |
+
|
100 |
+
|
101 |
+
def _scatter(p, state, rank, comm_stream):
|
102 |
+
u = state.computed_u
|
103 |
+
mesh = p.device_mesh
|
104 |
+
|
105 |
+
with torch.cuda.stream(comm_stream):
|
106 |
+
if rank == state.worker_rank:
|
107 |
+
if state.compute_event is None:
|
108 |
+
raise RuntimeError("Compute event must be set before scatter.")
|
109 |
+
comm_stream.wait_event(state.compute_event)
|
110 |
+
scatter_list = list(torch.split(u, p.size(0) // mesh.mesh.numel(), dim=0))
|
111 |
+
else:
|
112 |
+
scatter_list = None
|
113 |
+
|
114 |
+
u = torch.empty_like(p.to_local())
|
115 |
+
torch.distributed.scatter(
|
116 |
+
u,
|
117 |
+
scatter_list=scatter_list,
|
118 |
+
src=state.worker_rank,
|
119 |
+
group=mesh.get_group(),
|
120 |
+
)
|
121 |
+
u = DTensor.from_local(
|
122 |
+
u,
|
123 |
+
placements=p.placements,
|
124 |
+
device_mesh=mesh,
|
125 |
+
)
|
126 |
+
|
127 |
+
state.scattered_u = u
|
128 |
+
|
129 |
+
|
130 |
+
class Muon(torch.optim.Optimizer):
|
131 |
+
"""
|
132 |
+
Muon - MomentUm Orthogonalized by Newton-schulz
|
133 |
+
|
134 |
+
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
|
135 |
+
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
|
136 |
+
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
|
137 |
+
the advantage that it can be stably run in bfloat16 on the GPU.
|
138 |
+
|
139 |
+
Some warnings:
|
140 |
+
- We believe this optimizer is unlikely to work well for training with small batch size.
|
141 |
+
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
|
142 |
+
|
143 |
+
Arguments:
|
144 |
+
muon_params: The parameters to be optimized by Muon.
|
145 |
+
lr: The learning rate. The updates will have spectral norm of `lr`. (0.02 is a good default)
|
146 |
+
momentum: The momentum used by the internal SGD. (0.95 is a good default)
|
147 |
+
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
|
148 |
+
ns_steps: The number of Newton-Schulz iterations to run. (6 is probably always enough)
|
149 |
+
adamw_params: The parameters to be optimized by AdamW. Any parameters in `muon_params` which are
|
150 |
+
{0, 1}-D or are detected as being the embed or lm_head will be optimized by AdamW as well.
|
151 |
+
adamw_lr: The learning rate for the internal AdamW.
|
152 |
+
adamw_betas: The betas for the internal AdamW.
|
153 |
+
adamw_eps: The epsilon for the internal AdamW.
|
154 |
+
adamw_wd: The weight decay for the internal AdamW.
|
155 |
+
"""
|
156 |
+
|
157 |
+
def __init__(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
is_muon_func,
|
161 |
+
lr=1e-3,
|
162 |
+
momentum=0.95,
|
163 |
+
nesterov=True,
|
164 |
+
ns_steps=5,
|
165 |
+
adamw_wd=0.1,
|
166 |
+
adamw_betas=(0.9, 0.95),
|
167 |
+
adamw_eps=1e-8,
|
168 |
+
debug=False,
|
169 |
+
):
|
170 |
+
defaults = dict(
|
171 |
+
lr=lr,
|
172 |
+
wd=adamw_wd,
|
173 |
+
momentum=momentum,
|
174 |
+
nesterov=nesterov,
|
175 |
+
ns_steps=ns_steps,
|
176 |
+
adamw_betas=adamw_betas,
|
177 |
+
adamw_eps=adamw_eps,
|
178 |
+
)
|
179 |
+
|
180 |
+
super().__init__(model.parameters(), defaults)
|
181 |
+
self.is_muon_func = is_muon_func
|
182 |
+
self.model = model
|
183 |
+
|
184 |
+
if not dist.is_initialized():
|
185 |
+
raise RuntimeError(
|
186 |
+
"Muon optimizer requires distributed training to be initialized."
|
187 |
+
)
|
188 |
+
|
189 |
+
self.rank = dist.get_rank()
|
190 |
+
|
191 |
+
self.comm_stream = torch.cuda.Stream()
|
192 |
+
self.compute_stream = torch.cuda.Stream()
|
193 |
+
self.debug = debug
|
194 |
+
|
195 |
+
def __setstate__(self, state):
|
196 |
+
# Sort parameters into those for which we will use Muon, and those for which we will not
|
197 |
+
super().__setstate__(state)
|
198 |
+
for name, p in self.model.named_parameters():
|
199 |
+
if self.is_muon_func(p, name):
|
200 |
+
# Use Muon for every parameter in muon_params which is >= 2D and doesn't look like an embedding or head layer
|
201 |
+
assert p.ndim == 2, p.ndim
|
202 |
+
self.state[p]["use_muon"] = True
|
203 |
+
self.state[p]["orig_shape"] = p.shape
|
204 |
+
else:
|
205 |
+
# Do not use Muon for parameters in adamw_params
|
206 |
+
self.state[p]["use_muon"] = False
|
207 |
+
|
208 |
+
def _calc_flops(self, G, steps):
|
209 |
+
assert len(G.shape) == 2
|
210 |
+
M, N = G.shape
|
211 |
+
if M > N:
|
212 |
+
M, N = N, M
|
213 |
+
|
214 |
+
return steps * ((M**3) * 2 + (M**2 * N) * 4 + M * N * 2 + M**2 * 3)
|
215 |
+
|
216 |
+
def adjust_lr_for_muon(self, lr, param_shape):
|
217 |
+
A, B = param_shape[:2]
|
218 |
+
# We adjust the learning rate and weight decay based on the size of the parameter matrix
|
219 |
+
# as describted in the paper
|
220 |
+
adjusted_ratio = 0.2 * math.sqrt(max(A, B))
|
221 |
+
adjusted_lr = lr * adjusted_ratio
|
222 |
+
return adjusted_lr
|
223 |
+
|
224 |
+
def init_state_and_assign_params(self, params, group):
|
225 |
+
param_to_state = {}
|
226 |
+
param_to_flops = {}
|
227 |
+
|
228 |
+
total_flops = 0
|
229 |
+
for p in params:
|
230 |
+
g = p.grad
|
231 |
+
if g is None:
|
232 |
+
continue
|
233 |
+
assert g.ndim == 2, "Muon only supports 2D parameters."
|
234 |
+
|
235 |
+
flops = self._calc_flops(g, group["ns_steps"])
|
236 |
+
param_to_flops[id(p)] = flops
|
237 |
+
total_flops += flops
|
238 |
+
|
239 |
+
if self.debug:
|
240 |
+
print(f"Total TFLOPs for Muon: {total_flops / 1e12:.2f} TFLOPs", flush=True)
|
241 |
+
|
242 |
+
ordered_params = sorted(
|
243 |
+
params, key=lambda p: param_to_flops[id(p)], reverse=True
|
244 |
+
)
|
245 |
+
|
246 |
+
round_robin = 0
|
247 |
+
mesh = None
|
248 |
+
for p in ordered_params:
|
249 |
+
if mesh is None:
|
250 |
+
mesh = p.device_mesh
|
251 |
+
if mesh.ndim != 1:
|
252 |
+
raise NotImplementedError(
|
253 |
+
"Muon requires a 1D mesh for distributed training yet."
|
254 |
+
)
|
255 |
+
elif mesh != p.device_mesh:
|
256 |
+
raise ValueError("All parameters must be on the same mesh.")
|
257 |
+
|
258 |
+
param_to_state[id(p)] = _muon_state()
|
259 |
+
param_to_state[id(p)].worker_rank = mesh.mesh[round_robin].item()
|
260 |
+
|
261 |
+
round_robin = (round_robin + 1) % mesh.mesh.numel()
|
262 |
+
|
263 |
+
return param_to_state, ordered_params
|
264 |
+
|
265 |
+
def base(self, params, group, lr, wd, momentum):
|
266 |
+
# generate weight updates in distributed fashion
|
267 |
+
for p in params:
|
268 |
+
g = p.grad
|
269 |
+
if g is None:
|
270 |
+
continue
|
271 |
+
if g.ndim > 2:
|
272 |
+
g = g.view(g.size(0), -1)
|
273 |
+
assert g is not None
|
274 |
+
|
275 |
+
# calc update
|
276 |
+
state = self.state[p]
|
277 |
+
if "momentum_buffer" not in state:
|
278 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
279 |
+
buf = state["momentum_buffer"]
|
280 |
+
buf.mul_(momentum).add_(g)
|
281 |
+
if group["nesterov"]:
|
282 |
+
g = g.add(buf, alpha=momentum)
|
283 |
+
else:
|
284 |
+
g = buf
|
285 |
+
|
286 |
+
u = _zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
|
287 |
+
|
288 |
+
# scale update
|
289 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
290 |
+
|
291 |
+
# apply weight decay
|
292 |
+
p.data.mul_(1 - lr * wd)
|
293 |
+
|
294 |
+
# apply update
|
295 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
296 |
+
|
297 |
+
def _update_g(self, p, g, group, momentum):
|
298 |
+
# calc update
|
299 |
+
state = self.state[p]
|
300 |
+
if "momentum_buffer" not in state:
|
301 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
302 |
+
buf = state["momentum_buffer"]
|
303 |
+
buf.mul_(momentum).add_(g)
|
304 |
+
if group["nesterov"]:
|
305 |
+
g = g.add(buf, alpha=momentum)
|
306 |
+
else:
|
307 |
+
g = buf
|
308 |
+
return g
|
309 |
+
|
310 |
+
def _update_p(self, p, u, lr, wd):
|
311 |
+
# scale update
|
312 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
313 |
+
# apply weight decay
|
314 |
+
p.data.mul_(1 - lr * wd)
|
315 |
+
# apply update
|
316 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
317 |
+
|
318 |
+
def parallel(self, params, group, lr, wd, momentum):
|
319 |
+
"""
|
320 |
+
Perform a parallel optimization step using Muon.
|
321 |
+
"""
|
322 |
+
|
323 |
+
for p in params:
|
324 |
+
g = p.grad
|
325 |
+
if g is None:
|
326 |
+
continue
|
327 |
+
if g.ndim > 2:
|
328 |
+
g = g.view(g.size(0), -1)
|
329 |
+
|
330 |
+
# Update g in the local rank
|
331 |
+
g = self._update_g(
|
332 |
+
p,
|
333 |
+
g,
|
334 |
+
group,
|
335 |
+
momentum=momentum,
|
336 |
+
)
|
337 |
+
p.grad = g
|
338 |
+
|
339 |
+
param_to_state, ordered_params = self.init_state_and_assign_params(
|
340 |
+
params, group
|
341 |
+
)
|
342 |
+
|
343 |
+
def enqueue_gathers(start_idx, chunk_size):
|
344 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
345 |
+
state = param_to_state[id(p)]
|
346 |
+
_gather(p, state, self.rank, self.comm_stream)
|
347 |
+
|
348 |
+
def enqueue_computes(start_idx, chunk_size):
|
349 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
350 |
+
state = param_to_state[id(p)]
|
351 |
+
_compute_u(state, group["ns_steps"], self.rank, self.compute_stream)
|
352 |
+
|
353 |
+
def enqueue_scatters(start_idx, chunk_size):
|
354 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
355 |
+
state = param_to_state[id(p)]
|
356 |
+
_scatter(p, state, self.rank, self.comm_stream)
|
357 |
+
|
358 |
+
chunk_size = params[0].device_mesh.mesh.numel()
|
359 |
+
|
360 |
+
# Wait grad update
|
361 |
+
self.comm_stream.wait_stream(torch.cuda.current_stream())
|
362 |
+
|
363 |
+
enqueue_gathers(0, chunk_size)
|
364 |
+
for i in range(0, len(params) + chunk_size - 1, chunk_size):
|
365 |
+
enqueue_computes(i, chunk_size)
|
366 |
+
enqueue_gathers(i + chunk_size, chunk_size)
|
367 |
+
enqueue_scatters(i, chunk_size)
|
368 |
+
|
369 |
+
torch.cuda.current_stream().wait_stream(self.comm_stream)
|
370 |
+
|
371 |
+
for p in params:
|
372 |
+
g = p.grad
|
373 |
+
if g is None:
|
374 |
+
continue
|
375 |
+
|
376 |
+
# Update p with sharded u
|
377 |
+
state = param_to_state[id(p)]
|
378 |
+
self._update_p(
|
379 |
+
p,
|
380 |
+
state.scattered_u,
|
381 |
+
lr=lr,
|
382 |
+
wd=wd,
|
383 |
+
)
|
384 |
+
|
385 |
+
def step(self, closure=None):
|
386 |
+
"""Perform a single optimization step.
|
387 |
+
|
388 |
+
Args:
|
389 |
+
closure (Callable, optional): A closure that reevaluates the model
|
390 |
+
and returns the loss.
|
391 |
+
"""
|
392 |
+
loss = None
|
393 |
+
if closure is not None:
|
394 |
+
with torch.enable_grad():
|
395 |
+
loss = closure()
|
396 |
+
|
397 |
+
for group in self.param_groups:
|
398 |
+
############################
|
399 |
+
# Muon #
|
400 |
+
############################
|
401 |
+
|
402 |
+
params = [p for p in group["params"] if self.state[p]["use_muon"]]
|
403 |
+
lr = group["lr"]
|
404 |
+
wd = group["wd"]
|
405 |
+
momentum = group["momentum"]
|
406 |
+
|
407 |
+
if isinstance(params[0].data, DTensor):
|
408 |
+
self.parallel(
|
409 |
+
params,
|
410 |
+
group,
|
411 |
+
lr=lr,
|
412 |
+
wd=wd,
|
413 |
+
momentum=momentum,
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
self.base(
|
417 |
+
params,
|
418 |
+
group,
|
419 |
+
lr=lr,
|
420 |
+
wd=wd,
|
421 |
+
momentum=momentum,
|
422 |
+
)
|
423 |
+
|
424 |
+
############################
|
425 |
+
# AdamW backup #
|
426 |
+
############################
|
427 |
+
|
428 |
+
params = [p for p in group["params"] if not self.state[p]["use_muon"]]
|
429 |
+
lr = group["lr"]
|
430 |
+
beta1, beta2 = group["adamw_betas"]
|
431 |
+
eps = group["adamw_eps"]
|
432 |
+
weight_decay = group["wd"]
|
433 |
+
|
434 |
+
for p in params:
|
435 |
+
g = p.grad
|
436 |
+
if g is None:
|
437 |
+
continue
|
438 |
+
state = self.state[p]
|
439 |
+
if "step" not in state:
|
440 |
+
state["step"] = 0
|
441 |
+
state["moment1"] = torch.zeros_like(g)
|
442 |
+
state["moment2"] = torch.zeros_like(g)
|
443 |
+
state["step"] += 1
|
444 |
+
step = state["step"]
|
445 |
+
buf1 = state["moment1"]
|
446 |
+
buf2 = state["moment2"]
|
447 |
+
buf1.lerp_(g, 1 - beta1)
|
448 |
+
buf2.lerp_(g.square(), 1 - beta2)
|
449 |
+
|
450 |
+
g = buf1 / (eps + buf2.sqrt())
|
451 |
+
|
452 |
+
bias_correction1 = 1 - beta1**step
|
453 |
+
bias_correction2 = 1 - beta2**step
|
454 |
+
scale = bias_correction1 / bias_correction2**0.5
|
455 |
+
p.data.mul_(1 - lr * weight_decay)
|
456 |
+
p.data.add_(g, alpha=-lr / scale)
|
457 |
+
|
458 |
+
return loss
|
build/torch26-cxx11-cu124-x86_64-linux/optimizer/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .muon import Muon
|
2 |
+
|
3 |
+
__all__ = [
|
4 |
+
"Muon",
|
5 |
+
]
|
build/torch26-cxx11-cu124-x86_64-linux/optimizer/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _optimizer_b4b3752_dirty
|
3 |
+
ops = torch.ops._optimizer_b4b3752_dirty
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_optimizer_b4b3752_dirty::{op_name}"
|
build/torch26-cxx11-cu124-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e89cd7d514bfe92598684ae3cfc2d35ac2d021340846e09c0b6c880c3d55bfa0
|
3 |
+
size 1820136
|
build/torch26-cxx11-cu124-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9cbffc2cf8039069831a57afb8e2f64fa684f1a44bec79bb4b72dbb57d9ac607
|
3 |
+
size 1824224
|
build/torch26-cxx11-cu124-x86_64-linux/optimizer/muon.py
ADDED
@@ -0,0 +1,458 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.distributed as dist
|
6 |
+
from torch.distributed._tensor import DTensor
|
7 |
+
|
8 |
+
|
9 |
+
# TODO leave original url and consider LICENSE
|
10 |
+
# This code snippet is a modified version adapted from the following GitHub repository:
|
11 |
+
# https://github.com/KellerJordan/Muon/blob/master/muon.py
|
12 |
+
def _zeropower_via_newtonschulz5(G, steps):
|
13 |
+
"""
|
14 |
+
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
|
15 |
+
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
|
16 |
+
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
|
17 |
+
zero even beyond the point where the iteration no longer converges all the way to one everywhere
|
18 |
+
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
|
19 |
+
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
|
20 |
+
performance at all relative to UV^T, where USV^T = G is the SVD.
|
21 |
+
"""
|
22 |
+
assert len(G.shape) == 2
|
23 |
+
a, b, c = (3.4445, -4.7750, 2.0315)
|
24 |
+
X = G # no manual typecast
|
25 |
+
if G.size(0) > G.size(1):
|
26 |
+
X = X.T
|
27 |
+
# Ensure spectral norm is at most 1
|
28 |
+
X = X / (X.norm() + 1e-7)
|
29 |
+
X = X.bfloat16()
|
30 |
+
# Perform the NS iterations
|
31 |
+
for _ in range(steps):
|
32 |
+
A = X @ X.T
|
33 |
+
# B = (
|
34 |
+
# b * A + c * A @ A
|
35 |
+
# ) # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
|
36 |
+
B = torch.addmm(A, A, A, alpha=c, beta=b)
|
37 |
+
# X = a * X + B @ X
|
38 |
+
X = torch.addmm(X, B, X, alpha=1.0, beta=a)
|
39 |
+
|
40 |
+
if G.size(0) > G.size(1):
|
41 |
+
X = X.T
|
42 |
+
return X.to(G.dtype)
|
43 |
+
|
44 |
+
|
45 |
+
@dataclass
|
46 |
+
class _muon_state:
|
47 |
+
# TODO: use Optional
|
48 |
+
worker_rank: int | None = None
|
49 |
+
gathered_grad: torch.Tensor | None = None
|
50 |
+
computed_u: torch.Tensor | None = None
|
51 |
+
scattered_u: torch.Tensor | None = None
|
52 |
+
gather_event: torch.cuda.Event | None = None
|
53 |
+
compute_event: torch.cuda.Event | None = None
|
54 |
+
|
55 |
+
|
56 |
+
def _gather(p, state, rank, comm_stream):
|
57 |
+
g = p.grad
|
58 |
+
mesh = g.device_mesh
|
59 |
+
|
60 |
+
if rank == state.worker_rank:
|
61 |
+
gather_list = [torch.empty_like(g.to_local()) for _ in range(mesh.mesh.numel())]
|
62 |
+
else:
|
63 |
+
gather_list = None
|
64 |
+
|
65 |
+
with torch.cuda.stream(comm_stream):
|
66 |
+
torch.distributed.gather(
|
67 |
+
g.to_local(),
|
68 |
+
dst=state.worker_rank,
|
69 |
+
gather_list=gather_list,
|
70 |
+
group=mesh.get_group(),
|
71 |
+
)
|
72 |
+
if rank == state.worker_rank:
|
73 |
+
# TODO: Consider ,,,
|
74 |
+
if state.gathered_grad is not None:
|
75 |
+
raise RuntimeError(
|
76 |
+
"Gather event already exists, which should not happen."
|
77 |
+
)
|
78 |
+
state.gathered_grad = torch.cat(gather_list, dim=0)
|
79 |
+
state.gather_event = torch.cuda.Event()
|
80 |
+
state.gather_event.record()
|
81 |
+
else:
|
82 |
+
state.gathered_grad = None
|
83 |
+
state.gather_event = None
|
84 |
+
|
85 |
+
|
86 |
+
def _compute_u(state, steps, rank, compute_stream):
|
87 |
+
with torch.cuda.stream(compute_stream):
|
88 |
+
if rank == state.worker_rank:
|
89 |
+
if state.gather_event is None:
|
90 |
+
raise RuntimeError("Gather event must be set before compute.")
|
91 |
+
compute_stream.wait_event(state.gather_event)
|
92 |
+
u = _zeropower_via_newtonschulz5(state.gathered_grad, steps)
|
93 |
+
state.computed_u = u
|
94 |
+
state.compute_event = torch.cuda.Event()
|
95 |
+
state.compute_event.record()
|
96 |
+
else:
|
97 |
+
state.computed_u = None
|
98 |
+
state.compute_event = None
|
99 |
+
|
100 |
+
|
101 |
+
def _scatter(p, state, rank, comm_stream):
|
102 |
+
u = state.computed_u
|
103 |
+
mesh = p.device_mesh
|
104 |
+
|
105 |
+
with torch.cuda.stream(comm_stream):
|
106 |
+
if rank == state.worker_rank:
|
107 |
+
if state.compute_event is None:
|
108 |
+
raise RuntimeError("Compute event must be set before scatter.")
|
109 |
+
comm_stream.wait_event(state.compute_event)
|
110 |
+
scatter_list = list(torch.split(u, p.size(0) // mesh.mesh.numel(), dim=0))
|
111 |
+
else:
|
112 |
+
scatter_list = None
|
113 |
+
|
114 |
+
u = torch.empty_like(p.to_local())
|
115 |
+
torch.distributed.scatter(
|
116 |
+
u,
|
117 |
+
scatter_list=scatter_list,
|
118 |
+
src=state.worker_rank,
|
119 |
+
group=mesh.get_group(),
|
120 |
+
)
|
121 |
+
u = DTensor.from_local(
|
122 |
+
u,
|
123 |
+
placements=p.placements,
|
124 |
+
device_mesh=mesh,
|
125 |
+
)
|
126 |
+
|
127 |
+
state.scattered_u = u
|
128 |
+
|
129 |
+
|
130 |
+
class Muon(torch.optim.Optimizer):
|
131 |
+
"""
|
132 |
+
Muon - MomentUm Orthogonalized by Newton-schulz
|
133 |
+
|
134 |
+
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
|
135 |
+
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
|
136 |
+
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
|
137 |
+
the advantage that it can be stably run in bfloat16 on the GPU.
|
138 |
+
|
139 |
+
Some warnings:
|
140 |
+
- We believe this optimizer is unlikely to work well for training with small batch size.
|
141 |
+
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
|
142 |
+
|
143 |
+
Arguments:
|
144 |
+
muon_params: The parameters to be optimized by Muon.
|
145 |
+
lr: The learning rate. The updates will have spectral norm of `lr`. (0.02 is a good default)
|
146 |
+
momentum: The momentum used by the internal SGD. (0.95 is a good default)
|
147 |
+
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
|
148 |
+
ns_steps: The number of Newton-Schulz iterations to run. (6 is probably always enough)
|
149 |
+
adamw_params: The parameters to be optimized by AdamW. Any parameters in `muon_params` which are
|
150 |
+
{0, 1}-D or are detected as being the embed or lm_head will be optimized by AdamW as well.
|
151 |
+
adamw_lr: The learning rate for the internal AdamW.
|
152 |
+
adamw_betas: The betas for the internal AdamW.
|
153 |
+
adamw_eps: The epsilon for the internal AdamW.
|
154 |
+
adamw_wd: The weight decay for the internal AdamW.
|
155 |
+
"""
|
156 |
+
|
157 |
+
def __init__(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
is_muon_func,
|
161 |
+
lr=1e-3,
|
162 |
+
momentum=0.95,
|
163 |
+
nesterov=True,
|
164 |
+
ns_steps=5,
|
165 |
+
adamw_wd=0.1,
|
166 |
+
adamw_betas=(0.9, 0.95),
|
167 |
+
adamw_eps=1e-8,
|
168 |
+
debug=False,
|
169 |
+
):
|
170 |
+
defaults = dict(
|
171 |
+
lr=lr,
|
172 |
+
wd=adamw_wd,
|
173 |
+
momentum=momentum,
|
174 |
+
nesterov=nesterov,
|
175 |
+
ns_steps=ns_steps,
|
176 |
+
adamw_betas=adamw_betas,
|
177 |
+
adamw_eps=adamw_eps,
|
178 |
+
)
|
179 |
+
|
180 |
+
super().__init__(model.parameters(), defaults)
|
181 |
+
self.is_muon_func = is_muon_func
|
182 |
+
self.model = model
|
183 |
+
|
184 |
+
if not dist.is_initialized():
|
185 |
+
raise RuntimeError(
|
186 |
+
"Muon optimizer requires distributed training to be initialized."
|
187 |
+
)
|
188 |
+
|
189 |
+
self.rank = dist.get_rank()
|
190 |
+
|
191 |
+
self.comm_stream = torch.cuda.Stream()
|
192 |
+
self.compute_stream = torch.cuda.Stream()
|
193 |
+
self.debug = debug
|
194 |
+
|
195 |
+
def __setstate__(self, state):
|
196 |
+
# Sort parameters into those for which we will use Muon, and those for which we will not
|
197 |
+
super().__setstate__(state)
|
198 |
+
for name, p in self.model.named_parameters():
|
199 |
+
if self.is_muon_func(p, name):
|
200 |
+
# Use Muon for every parameter in muon_params which is >= 2D and doesn't look like an embedding or head layer
|
201 |
+
assert p.ndim == 2, p.ndim
|
202 |
+
self.state[p]["use_muon"] = True
|
203 |
+
self.state[p]["orig_shape"] = p.shape
|
204 |
+
else:
|
205 |
+
# Do not use Muon for parameters in adamw_params
|
206 |
+
self.state[p]["use_muon"] = False
|
207 |
+
|
208 |
+
def _calc_flops(self, G, steps):
|
209 |
+
assert len(G.shape) == 2
|
210 |
+
M, N = G.shape
|
211 |
+
if M > N:
|
212 |
+
M, N = N, M
|
213 |
+
|
214 |
+
return steps * ((M**3) * 2 + (M**2 * N) * 4 + M * N * 2 + M**2 * 3)
|
215 |
+
|
216 |
+
def adjust_lr_for_muon(self, lr, param_shape):
|
217 |
+
A, B = param_shape[:2]
|
218 |
+
# We adjust the learning rate and weight decay based on the size of the parameter matrix
|
219 |
+
# as describted in the paper
|
220 |
+
adjusted_ratio = 0.2 * math.sqrt(max(A, B))
|
221 |
+
adjusted_lr = lr * adjusted_ratio
|
222 |
+
return adjusted_lr
|
223 |
+
|
224 |
+
def init_state_and_assign_params(self, params, group):
|
225 |
+
param_to_state = {}
|
226 |
+
param_to_flops = {}
|
227 |
+
|
228 |
+
total_flops = 0
|
229 |
+
for p in params:
|
230 |
+
g = p.grad
|
231 |
+
if g is None:
|
232 |
+
continue
|
233 |
+
assert g.ndim == 2, "Muon only supports 2D parameters."
|
234 |
+
|
235 |
+
flops = self._calc_flops(g, group["ns_steps"])
|
236 |
+
param_to_flops[id(p)] = flops
|
237 |
+
total_flops += flops
|
238 |
+
|
239 |
+
if self.debug:
|
240 |
+
print(f"Total TFLOPs for Muon: {total_flops / 1e12:.2f} TFLOPs", flush=True)
|
241 |
+
|
242 |
+
ordered_params = sorted(
|
243 |
+
params, key=lambda p: param_to_flops[id(p)], reverse=True
|
244 |
+
)
|
245 |
+
|
246 |
+
round_robin = 0
|
247 |
+
mesh = None
|
248 |
+
for p in ordered_params:
|
249 |
+
if mesh is None:
|
250 |
+
mesh = p.device_mesh
|
251 |
+
if mesh.ndim != 1:
|
252 |
+
raise NotImplementedError(
|
253 |
+
"Muon requires a 1D mesh for distributed training yet."
|
254 |
+
)
|
255 |
+
elif mesh != p.device_mesh:
|
256 |
+
raise ValueError("All parameters must be on the same mesh.")
|
257 |
+
|
258 |
+
param_to_state[id(p)] = _muon_state()
|
259 |
+
param_to_state[id(p)].worker_rank = mesh.mesh[round_robin].item()
|
260 |
+
|
261 |
+
round_robin = (round_robin + 1) % mesh.mesh.numel()
|
262 |
+
|
263 |
+
return param_to_state, ordered_params
|
264 |
+
|
265 |
+
def base(self, params, group, lr, wd, momentum):
|
266 |
+
# generate weight updates in distributed fashion
|
267 |
+
for p in params:
|
268 |
+
g = p.grad
|
269 |
+
if g is None:
|
270 |
+
continue
|
271 |
+
if g.ndim > 2:
|
272 |
+
g = g.view(g.size(0), -1)
|
273 |
+
assert g is not None
|
274 |
+
|
275 |
+
# calc update
|
276 |
+
state = self.state[p]
|
277 |
+
if "momentum_buffer" not in state:
|
278 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
279 |
+
buf = state["momentum_buffer"]
|
280 |
+
buf.mul_(momentum).add_(g)
|
281 |
+
if group["nesterov"]:
|
282 |
+
g = g.add(buf, alpha=momentum)
|
283 |
+
else:
|
284 |
+
g = buf
|
285 |
+
|
286 |
+
u = _zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
|
287 |
+
|
288 |
+
# scale update
|
289 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
290 |
+
|
291 |
+
# apply weight decay
|
292 |
+
p.data.mul_(1 - lr * wd)
|
293 |
+
|
294 |
+
# apply update
|
295 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
296 |
+
|
297 |
+
def _update_g(self, p, g, group, momentum):
|
298 |
+
# calc update
|
299 |
+
state = self.state[p]
|
300 |
+
if "momentum_buffer" not in state:
|
301 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
302 |
+
buf = state["momentum_buffer"]
|
303 |
+
buf.mul_(momentum).add_(g)
|
304 |
+
if group["nesterov"]:
|
305 |
+
g = g.add(buf, alpha=momentum)
|
306 |
+
else:
|
307 |
+
g = buf
|
308 |
+
return g
|
309 |
+
|
310 |
+
def _update_p(self, p, u, lr, wd):
|
311 |
+
# scale update
|
312 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
313 |
+
# apply weight decay
|
314 |
+
p.data.mul_(1 - lr * wd)
|
315 |
+
# apply update
|
316 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
317 |
+
|
318 |
+
def parallel(self, params, group, lr, wd, momentum):
|
319 |
+
"""
|
320 |
+
Perform a parallel optimization step using Muon.
|
321 |
+
"""
|
322 |
+
|
323 |
+
for p in params:
|
324 |
+
g = p.grad
|
325 |
+
if g is None:
|
326 |
+
continue
|
327 |
+
if g.ndim > 2:
|
328 |
+
g = g.view(g.size(0), -1)
|
329 |
+
|
330 |
+
# Update g in the local rank
|
331 |
+
g = self._update_g(
|
332 |
+
p,
|
333 |
+
g,
|
334 |
+
group,
|
335 |
+
momentum=momentum,
|
336 |
+
)
|
337 |
+
p.grad = g
|
338 |
+
|
339 |
+
param_to_state, ordered_params = self.init_state_and_assign_params(
|
340 |
+
params, group
|
341 |
+
)
|
342 |
+
|
343 |
+
def enqueue_gathers(start_idx, chunk_size):
|
344 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
345 |
+
state = param_to_state[id(p)]
|
346 |
+
_gather(p, state, self.rank, self.comm_stream)
|
347 |
+
|
348 |
+
def enqueue_computes(start_idx, chunk_size):
|
349 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
350 |
+
state = param_to_state[id(p)]
|
351 |
+
_compute_u(state, group["ns_steps"], self.rank, self.compute_stream)
|
352 |
+
|
353 |
+
def enqueue_scatters(start_idx, chunk_size):
|
354 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
355 |
+
state = param_to_state[id(p)]
|
356 |
+
_scatter(p, state, self.rank, self.comm_stream)
|
357 |
+
|
358 |
+
chunk_size = params[0].device_mesh.mesh.numel()
|
359 |
+
|
360 |
+
# Wait grad update
|
361 |
+
self.comm_stream.wait_stream(torch.cuda.current_stream())
|
362 |
+
|
363 |
+
enqueue_gathers(0, chunk_size)
|
364 |
+
for i in range(0, len(params) + chunk_size - 1, chunk_size):
|
365 |
+
enqueue_computes(i, chunk_size)
|
366 |
+
enqueue_gathers(i + chunk_size, chunk_size)
|
367 |
+
enqueue_scatters(i, chunk_size)
|
368 |
+
|
369 |
+
torch.cuda.current_stream().wait_stream(self.comm_stream)
|
370 |
+
|
371 |
+
for p in params:
|
372 |
+
g = p.grad
|
373 |
+
if g is None:
|
374 |
+
continue
|
375 |
+
|
376 |
+
# Update p with sharded u
|
377 |
+
state = param_to_state[id(p)]
|
378 |
+
self._update_p(
|
379 |
+
p,
|
380 |
+
state.scattered_u,
|
381 |
+
lr=lr,
|
382 |
+
wd=wd,
|
383 |
+
)
|
384 |
+
|
385 |
+
def step(self, closure=None):
|
386 |
+
"""Perform a single optimization step.
|
387 |
+
|
388 |
+
Args:
|
389 |
+
closure (Callable, optional): A closure that reevaluates the model
|
390 |
+
and returns the loss.
|
391 |
+
"""
|
392 |
+
loss = None
|
393 |
+
if closure is not None:
|
394 |
+
with torch.enable_grad():
|
395 |
+
loss = closure()
|
396 |
+
|
397 |
+
for group in self.param_groups:
|
398 |
+
############################
|
399 |
+
# Muon #
|
400 |
+
############################
|
401 |
+
|
402 |
+
params = [p for p in group["params"] if self.state[p]["use_muon"]]
|
403 |
+
lr = group["lr"]
|
404 |
+
wd = group["wd"]
|
405 |
+
momentum = group["momentum"]
|
406 |
+
|
407 |
+
if isinstance(params[0].data, DTensor):
|
408 |
+
self.parallel(
|
409 |
+
params,
|
410 |
+
group,
|
411 |
+
lr=lr,
|
412 |
+
wd=wd,
|
413 |
+
momentum=momentum,
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
self.base(
|
417 |
+
params,
|
418 |
+
group,
|
419 |
+
lr=lr,
|
420 |
+
wd=wd,
|
421 |
+
momentum=momentum,
|
422 |
+
)
|
423 |
+
|
424 |
+
############################
|
425 |
+
# AdamW backup #
|
426 |
+
############################
|
427 |
+
|
428 |
+
params = [p for p in group["params"] if not self.state[p]["use_muon"]]
|
429 |
+
lr = group["lr"]
|
430 |
+
beta1, beta2 = group["adamw_betas"]
|
431 |
+
eps = group["adamw_eps"]
|
432 |
+
weight_decay = group["wd"]
|
433 |
+
|
434 |
+
for p in params:
|
435 |
+
g = p.grad
|
436 |
+
if g is None:
|
437 |
+
continue
|
438 |
+
state = self.state[p]
|
439 |
+
if "step" not in state:
|
440 |
+
state["step"] = 0
|
441 |
+
state["moment1"] = torch.zeros_like(g)
|
442 |
+
state["moment2"] = torch.zeros_like(g)
|
443 |
+
state["step"] += 1
|
444 |
+
step = state["step"]
|
445 |
+
buf1 = state["moment1"]
|
446 |
+
buf2 = state["moment2"]
|
447 |
+
buf1.lerp_(g, 1 - beta1)
|
448 |
+
buf2.lerp_(g.square(), 1 - beta2)
|
449 |
+
|
450 |
+
g = buf1 / (eps + buf2.sqrt())
|
451 |
+
|
452 |
+
bias_correction1 = 1 - beta1**step
|
453 |
+
bias_correction2 = 1 - beta2**step
|
454 |
+
scale = bias_correction1 / bias_correction2**0.5
|
455 |
+
p.data.mul_(1 - lr * weight_decay)
|
456 |
+
p.data.add_(g, alpha=-lr / scale)
|
457 |
+
|
458 |
+
return loss
|
build/torch26-cxx11-cu126-x86_64-linux/optimizer/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .muon import Muon
|
2 |
+
|
3 |
+
__all__ = [
|
4 |
+
"Muon",
|
5 |
+
]
|
build/torch26-cxx11-cu126-x86_64-linux/optimizer/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _optimizer_b4b3752_dirty
|
3 |
+
ops = torch.ops._optimizer_b4b3752_dirty
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_optimizer_b4b3752_dirty::{op_name}"
|
build/torch26-cxx11-cu126-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:7f5dce62d3038e879e688fffa9bbc70f3e82db20b2e7ae3ba09040e0319acb71
|
3 |
+
size 1820136
|
build/torch26-cxx11-cu126-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:58162f994df84868dbf62ae70e39d3c14e3390fc827f152eece83dfae7f51503
|
3 |
+
size 1824224
|
build/torch26-cxx11-cu126-x86_64-linux/optimizer/muon.py
ADDED
@@ -0,0 +1,458 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.distributed as dist
|
6 |
+
from torch.distributed._tensor import DTensor
|
7 |
+
|
8 |
+
|
9 |
+
# TODO leave original url and consider LICENSE
|
10 |
+
# This code snippet is a modified version adapted from the following GitHub repository:
|
11 |
+
# https://github.com/KellerJordan/Muon/blob/master/muon.py
|
12 |
+
def _zeropower_via_newtonschulz5(G, steps):
|
13 |
+
"""
|
14 |
+
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
|
15 |
+
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
|
16 |
+
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
|
17 |
+
zero even beyond the point where the iteration no longer converges all the way to one everywhere
|
18 |
+
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
|
19 |
+
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
|
20 |
+
performance at all relative to UV^T, where USV^T = G is the SVD.
|
21 |
+
"""
|
22 |
+
assert len(G.shape) == 2
|
23 |
+
a, b, c = (3.4445, -4.7750, 2.0315)
|
24 |
+
X = G # no manual typecast
|
25 |
+
if G.size(0) > G.size(1):
|
26 |
+
X = X.T
|
27 |
+
# Ensure spectral norm is at most 1
|
28 |
+
X = X / (X.norm() + 1e-7)
|
29 |
+
X = X.bfloat16()
|
30 |
+
# Perform the NS iterations
|
31 |
+
for _ in range(steps):
|
32 |
+
A = X @ X.T
|
33 |
+
# B = (
|
34 |
+
# b * A + c * A @ A
|
35 |
+
# ) # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
|
36 |
+
B = torch.addmm(A, A, A, alpha=c, beta=b)
|
37 |
+
# X = a * X + B @ X
|
38 |
+
X = torch.addmm(X, B, X, alpha=1.0, beta=a)
|
39 |
+
|
40 |
+
if G.size(0) > G.size(1):
|
41 |
+
X = X.T
|
42 |
+
return X.to(G.dtype)
|
43 |
+
|
44 |
+
|
45 |
+
@dataclass
|
46 |
+
class _muon_state:
|
47 |
+
# TODO: use Optional
|
48 |
+
worker_rank: int | None = None
|
49 |
+
gathered_grad: torch.Tensor | None = None
|
50 |
+
computed_u: torch.Tensor | None = None
|
51 |
+
scattered_u: torch.Tensor | None = None
|
52 |
+
gather_event: torch.cuda.Event | None = None
|
53 |
+
compute_event: torch.cuda.Event | None = None
|
54 |
+
|
55 |
+
|
56 |
+
def _gather(p, state, rank, comm_stream):
|
57 |
+
g = p.grad
|
58 |
+
mesh = g.device_mesh
|
59 |
+
|
60 |
+
if rank == state.worker_rank:
|
61 |
+
gather_list = [torch.empty_like(g.to_local()) for _ in range(mesh.mesh.numel())]
|
62 |
+
else:
|
63 |
+
gather_list = None
|
64 |
+
|
65 |
+
with torch.cuda.stream(comm_stream):
|
66 |
+
torch.distributed.gather(
|
67 |
+
g.to_local(),
|
68 |
+
dst=state.worker_rank,
|
69 |
+
gather_list=gather_list,
|
70 |
+
group=mesh.get_group(),
|
71 |
+
)
|
72 |
+
if rank == state.worker_rank:
|
73 |
+
# TODO: Consider ,,,
|
74 |
+
if state.gathered_grad is not None:
|
75 |
+
raise RuntimeError(
|
76 |
+
"Gather event already exists, which should not happen."
|
77 |
+
)
|
78 |
+
state.gathered_grad = torch.cat(gather_list, dim=0)
|
79 |
+
state.gather_event = torch.cuda.Event()
|
80 |
+
state.gather_event.record()
|
81 |
+
else:
|
82 |
+
state.gathered_grad = None
|
83 |
+
state.gather_event = None
|
84 |
+
|
85 |
+
|
86 |
+
def _compute_u(state, steps, rank, compute_stream):
|
87 |
+
with torch.cuda.stream(compute_stream):
|
88 |
+
if rank == state.worker_rank:
|
89 |
+
if state.gather_event is None:
|
90 |
+
raise RuntimeError("Gather event must be set before compute.")
|
91 |
+
compute_stream.wait_event(state.gather_event)
|
92 |
+
u = _zeropower_via_newtonschulz5(state.gathered_grad, steps)
|
93 |
+
state.computed_u = u
|
94 |
+
state.compute_event = torch.cuda.Event()
|
95 |
+
state.compute_event.record()
|
96 |
+
else:
|
97 |
+
state.computed_u = None
|
98 |
+
state.compute_event = None
|
99 |
+
|
100 |
+
|
101 |
+
def _scatter(p, state, rank, comm_stream):
|
102 |
+
u = state.computed_u
|
103 |
+
mesh = p.device_mesh
|
104 |
+
|
105 |
+
with torch.cuda.stream(comm_stream):
|
106 |
+
if rank == state.worker_rank:
|
107 |
+
if state.compute_event is None:
|
108 |
+
raise RuntimeError("Compute event must be set before scatter.")
|
109 |
+
comm_stream.wait_event(state.compute_event)
|
110 |
+
scatter_list = list(torch.split(u, p.size(0) // mesh.mesh.numel(), dim=0))
|
111 |
+
else:
|
112 |
+
scatter_list = None
|
113 |
+
|
114 |
+
u = torch.empty_like(p.to_local())
|
115 |
+
torch.distributed.scatter(
|
116 |
+
u,
|
117 |
+
scatter_list=scatter_list,
|
118 |
+
src=state.worker_rank,
|
119 |
+
group=mesh.get_group(),
|
120 |
+
)
|
121 |
+
u = DTensor.from_local(
|
122 |
+
u,
|
123 |
+
placements=p.placements,
|
124 |
+
device_mesh=mesh,
|
125 |
+
)
|
126 |
+
|
127 |
+
state.scattered_u = u
|
128 |
+
|
129 |
+
|
130 |
+
class Muon(torch.optim.Optimizer):
|
131 |
+
"""
|
132 |
+
Muon - MomentUm Orthogonalized by Newton-schulz
|
133 |
+
|
134 |
+
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
|
135 |
+
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
|
136 |
+
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
|
137 |
+
the advantage that it can be stably run in bfloat16 on the GPU.
|
138 |
+
|
139 |
+
Some warnings:
|
140 |
+
- We believe this optimizer is unlikely to work well for training with small batch size.
|
141 |
+
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
|
142 |
+
|
143 |
+
Arguments:
|
144 |
+
muon_params: The parameters to be optimized by Muon.
|
145 |
+
lr: The learning rate. The updates will have spectral norm of `lr`. (0.02 is a good default)
|
146 |
+
momentum: The momentum used by the internal SGD. (0.95 is a good default)
|
147 |
+
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
|
148 |
+
ns_steps: The number of Newton-Schulz iterations to run. (6 is probably always enough)
|
149 |
+
adamw_params: The parameters to be optimized by AdamW. Any parameters in `muon_params` which are
|
150 |
+
{0, 1}-D or are detected as being the embed or lm_head will be optimized by AdamW as well.
|
151 |
+
adamw_lr: The learning rate for the internal AdamW.
|
152 |
+
adamw_betas: The betas for the internal AdamW.
|
153 |
+
adamw_eps: The epsilon for the internal AdamW.
|
154 |
+
adamw_wd: The weight decay for the internal AdamW.
|
155 |
+
"""
|
156 |
+
|
157 |
+
def __init__(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
is_muon_func,
|
161 |
+
lr=1e-3,
|
162 |
+
momentum=0.95,
|
163 |
+
nesterov=True,
|
164 |
+
ns_steps=5,
|
165 |
+
adamw_wd=0.1,
|
166 |
+
adamw_betas=(0.9, 0.95),
|
167 |
+
adamw_eps=1e-8,
|
168 |
+
debug=False,
|
169 |
+
):
|
170 |
+
defaults = dict(
|
171 |
+
lr=lr,
|
172 |
+
wd=adamw_wd,
|
173 |
+
momentum=momentum,
|
174 |
+
nesterov=nesterov,
|
175 |
+
ns_steps=ns_steps,
|
176 |
+
adamw_betas=adamw_betas,
|
177 |
+
adamw_eps=adamw_eps,
|
178 |
+
)
|
179 |
+
|
180 |
+
super().__init__(model.parameters(), defaults)
|
181 |
+
self.is_muon_func = is_muon_func
|
182 |
+
self.model = model
|
183 |
+
|
184 |
+
if not dist.is_initialized():
|
185 |
+
raise RuntimeError(
|
186 |
+
"Muon optimizer requires distributed training to be initialized."
|
187 |
+
)
|
188 |
+
|
189 |
+
self.rank = dist.get_rank()
|
190 |
+
|
191 |
+
self.comm_stream = torch.cuda.Stream()
|
192 |
+
self.compute_stream = torch.cuda.Stream()
|
193 |
+
self.debug = debug
|
194 |
+
|
195 |
+
def __setstate__(self, state):
|
196 |
+
# Sort parameters into those for which we will use Muon, and those for which we will not
|
197 |
+
super().__setstate__(state)
|
198 |
+
for name, p in self.model.named_parameters():
|
199 |
+
if self.is_muon_func(p, name):
|
200 |
+
# Use Muon for every parameter in muon_params which is >= 2D and doesn't look like an embedding or head layer
|
201 |
+
assert p.ndim == 2, p.ndim
|
202 |
+
self.state[p]["use_muon"] = True
|
203 |
+
self.state[p]["orig_shape"] = p.shape
|
204 |
+
else:
|
205 |
+
# Do not use Muon for parameters in adamw_params
|
206 |
+
self.state[p]["use_muon"] = False
|
207 |
+
|
208 |
+
def _calc_flops(self, G, steps):
|
209 |
+
assert len(G.shape) == 2
|
210 |
+
M, N = G.shape
|
211 |
+
if M > N:
|
212 |
+
M, N = N, M
|
213 |
+
|
214 |
+
return steps * ((M**3) * 2 + (M**2 * N) * 4 + M * N * 2 + M**2 * 3)
|
215 |
+
|
216 |
+
def adjust_lr_for_muon(self, lr, param_shape):
|
217 |
+
A, B = param_shape[:2]
|
218 |
+
# We adjust the learning rate and weight decay based on the size of the parameter matrix
|
219 |
+
# as describted in the paper
|
220 |
+
adjusted_ratio = 0.2 * math.sqrt(max(A, B))
|
221 |
+
adjusted_lr = lr * adjusted_ratio
|
222 |
+
return adjusted_lr
|
223 |
+
|
224 |
+
def init_state_and_assign_params(self, params, group):
|
225 |
+
param_to_state = {}
|
226 |
+
param_to_flops = {}
|
227 |
+
|
228 |
+
total_flops = 0
|
229 |
+
for p in params:
|
230 |
+
g = p.grad
|
231 |
+
if g is None:
|
232 |
+
continue
|
233 |
+
assert g.ndim == 2, "Muon only supports 2D parameters."
|
234 |
+
|
235 |
+
flops = self._calc_flops(g, group["ns_steps"])
|
236 |
+
param_to_flops[id(p)] = flops
|
237 |
+
total_flops += flops
|
238 |
+
|
239 |
+
if self.debug:
|
240 |
+
print(f"Total TFLOPs for Muon: {total_flops / 1e12:.2f} TFLOPs", flush=True)
|
241 |
+
|
242 |
+
ordered_params = sorted(
|
243 |
+
params, key=lambda p: param_to_flops[id(p)], reverse=True
|
244 |
+
)
|
245 |
+
|
246 |
+
round_robin = 0
|
247 |
+
mesh = None
|
248 |
+
for p in ordered_params:
|
249 |
+
if mesh is None:
|
250 |
+
mesh = p.device_mesh
|
251 |
+
if mesh.ndim != 1:
|
252 |
+
raise NotImplementedError(
|
253 |
+
"Muon requires a 1D mesh for distributed training yet."
|
254 |
+
)
|
255 |
+
elif mesh != p.device_mesh:
|
256 |
+
raise ValueError("All parameters must be on the same mesh.")
|
257 |
+
|
258 |
+
param_to_state[id(p)] = _muon_state()
|
259 |
+
param_to_state[id(p)].worker_rank = mesh.mesh[round_robin].item()
|
260 |
+
|
261 |
+
round_robin = (round_robin + 1) % mesh.mesh.numel()
|
262 |
+
|
263 |
+
return param_to_state, ordered_params
|
264 |
+
|
265 |
+
def base(self, params, group, lr, wd, momentum):
|
266 |
+
# generate weight updates in distributed fashion
|
267 |
+
for p in params:
|
268 |
+
g = p.grad
|
269 |
+
if g is None:
|
270 |
+
continue
|
271 |
+
if g.ndim > 2:
|
272 |
+
g = g.view(g.size(0), -1)
|
273 |
+
assert g is not None
|
274 |
+
|
275 |
+
# calc update
|
276 |
+
state = self.state[p]
|
277 |
+
if "momentum_buffer" not in state:
|
278 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
279 |
+
buf = state["momentum_buffer"]
|
280 |
+
buf.mul_(momentum).add_(g)
|
281 |
+
if group["nesterov"]:
|
282 |
+
g = g.add(buf, alpha=momentum)
|
283 |
+
else:
|
284 |
+
g = buf
|
285 |
+
|
286 |
+
u = _zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
|
287 |
+
|
288 |
+
# scale update
|
289 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
290 |
+
|
291 |
+
# apply weight decay
|
292 |
+
p.data.mul_(1 - lr * wd)
|
293 |
+
|
294 |
+
# apply update
|
295 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
296 |
+
|
297 |
+
def _update_g(self, p, g, group, momentum):
|
298 |
+
# calc update
|
299 |
+
state = self.state[p]
|
300 |
+
if "momentum_buffer" not in state:
|
301 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
302 |
+
buf = state["momentum_buffer"]
|
303 |
+
buf.mul_(momentum).add_(g)
|
304 |
+
if group["nesterov"]:
|
305 |
+
g = g.add(buf, alpha=momentum)
|
306 |
+
else:
|
307 |
+
g = buf
|
308 |
+
return g
|
309 |
+
|
310 |
+
def _update_p(self, p, u, lr, wd):
|
311 |
+
# scale update
|
312 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
313 |
+
# apply weight decay
|
314 |
+
p.data.mul_(1 - lr * wd)
|
315 |
+
# apply update
|
316 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
317 |
+
|
318 |
+
def parallel(self, params, group, lr, wd, momentum):
|
319 |
+
"""
|
320 |
+
Perform a parallel optimization step using Muon.
|
321 |
+
"""
|
322 |
+
|
323 |
+
for p in params:
|
324 |
+
g = p.grad
|
325 |
+
if g is None:
|
326 |
+
continue
|
327 |
+
if g.ndim > 2:
|
328 |
+
g = g.view(g.size(0), -1)
|
329 |
+
|
330 |
+
# Update g in the local rank
|
331 |
+
g = self._update_g(
|
332 |
+
p,
|
333 |
+
g,
|
334 |
+
group,
|
335 |
+
momentum=momentum,
|
336 |
+
)
|
337 |
+
p.grad = g
|
338 |
+
|
339 |
+
param_to_state, ordered_params = self.init_state_and_assign_params(
|
340 |
+
params, group
|
341 |
+
)
|
342 |
+
|
343 |
+
def enqueue_gathers(start_idx, chunk_size):
|
344 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
345 |
+
state = param_to_state[id(p)]
|
346 |
+
_gather(p, state, self.rank, self.comm_stream)
|
347 |
+
|
348 |
+
def enqueue_computes(start_idx, chunk_size):
|
349 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
350 |
+
state = param_to_state[id(p)]
|
351 |
+
_compute_u(state, group["ns_steps"], self.rank, self.compute_stream)
|
352 |
+
|
353 |
+
def enqueue_scatters(start_idx, chunk_size):
|
354 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
355 |
+
state = param_to_state[id(p)]
|
356 |
+
_scatter(p, state, self.rank, self.comm_stream)
|
357 |
+
|
358 |
+
chunk_size = params[0].device_mesh.mesh.numel()
|
359 |
+
|
360 |
+
# Wait grad update
|
361 |
+
self.comm_stream.wait_stream(torch.cuda.current_stream())
|
362 |
+
|
363 |
+
enqueue_gathers(0, chunk_size)
|
364 |
+
for i in range(0, len(params) + chunk_size - 1, chunk_size):
|
365 |
+
enqueue_computes(i, chunk_size)
|
366 |
+
enqueue_gathers(i + chunk_size, chunk_size)
|
367 |
+
enqueue_scatters(i, chunk_size)
|
368 |
+
|
369 |
+
torch.cuda.current_stream().wait_stream(self.comm_stream)
|
370 |
+
|
371 |
+
for p in params:
|
372 |
+
g = p.grad
|
373 |
+
if g is None:
|
374 |
+
continue
|
375 |
+
|
376 |
+
# Update p with sharded u
|
377 |
+
state = param_to_state[id(p)]
|
378 |
+
self._update_p(
|
379 |
+
p,
|
380 |
+
state.scattered_u,
|
381 |
+
lr=lr,
|
382 |
+
wd=wd,
|
383 |
+
)
|
384 |
+
|
385 |
+
def step(self, closure=None):
|
386 |
+
"""Perform a single optimization step.
|
387 |
+
|
388 |
+
Args:
|
389 |
+
closure (Callable, optional): A closure that reevaluates the model
|
390 |
+
and returns the loss.
|
391 |
+
"""
|
392 |
+
loss = None
|
393 |
+
if closure is not None:
|
394 |
+
with torch.enable_grad():
|
395 |
+
loss = closure()
|
396 |
+
|
397 |
+
for group in self.param_groups:
|
398 |
+
############################
|
399 |
+
# Muon #
|
400 |
+
############################
|
401 |
+
|
402 |
+
params = [p for p in group["params"] if self.state[p]["use_muon"]]
|
403 |
+
lr = group["lr"]
|
404 |
+
wd = group["wd"]
|
405 |
+
momentum = group["momentum"]
|
406 |
+
|
407 |
+
if isinstance(params[0].data, DTensor):
|
408 |
+
self.parallel(
|
409 |
+
params,
|
410 |
+
group,
|
411 |
+
lr=lr,
|
412 |
+
wd=wd,
|
413 |
+
momentum=momentum,
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
self.base(
|
417 |
+
params,
|
418 |
+
group,
|
419 |
+
lr=lr,
|
420 |
+
wd=wd,
|
421 |
+
momentum=momentum,
|
422 |
+
)
|
423 |
+
|
424 |
+
############################
|
425 |
+
# AdamW backup #
|
426 |
+
############################
|
427 |
+
|
428 |
+
params = [p for p in group["params"] if not self.state[p]["use_muon"]]
|
429 |
+
lr = group["lr"]
|
430 |
+
beta1, beta2 = group["adamw_betas"]
|
431 |
+
eps = group["adamw_eps"]
|
432 |
+
weight_decay = group["wd"]
|
433 |
+
|
434 |
+
for p in params:
|
435 |
+
g = p.grad
|
436 |
+
if g is None:
|
437 |
+
continue
|
438 |
+
state = self.state[p]
|
439 |
+
if "step" not in state:
|
440 |
+
state["step"] = 0
|
441 |
+
state["moment1"] = torch.zeros_like(g)
|
442 |
+
state["moment2"] = torch.zeros_like(g)
|
443 |
+
state["step"] += 1
|
444 |
+
step = state["step"]
|
445 |
+
buf1 = state["moment1"]
|
446 |
+
buf2 = state["moment2"]
|
447 |
+
buf1.lerp_(g, 1 - beta1)
|
448 |
+
buf2.lerp_(g.square(), 1 - beta2)
|
449 |
+
|
450 |
+
g = buf1 / (eps + buf2.sqrt())
|
451 |
+
|
452 |
+
bias_correction1 = 1 - beta1**step
|
453 |
+
bias_correction2 = 1 - beta2**step
|
454 |
+
scale = bias_correction1 / bias_correction2**0.5
|
455 |
+
p.data.mul_(1 - lr * weight_decay)
|
456 |
+
p.data.add_(g, alpha=-lr / scale)
|
457 |
+
|
458 |
+
return loss
|
build/torch26-cxx11-rocm62-x86_64-linux/optimizer/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .muon import Muon
|
2 |
+
|
3 |
+
__all__ = [
|
4 |
+
"Muon",
|
5 |
+
]
|
build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _optimizer_b4b3752_dirty
|
3 |
+
ops = torch.ops._optimizer_b4b3752_dirty
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_optimizer_b4b3752_dirty::{op_name}"
|
build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_optimizer_20250614121529.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d2f60369ba2bd0a0f84e053d857d37496137ff476dc21561f211b1fa39651990
|
3 |
+
size 1749784
|
build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_optimizer_20250614123843.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f4d790535f99b7b362a966e802a547654f31749f5f28a0207493870927f1d8d2
|
3 |
+
size 1749784
|
build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:b440dd9a60711a498010068e91d0ad013cd0b8ac732c16b5d1d17e5d4ec0f9b4
|
3 |
+
size 1749784
|
build/torch26-cxx11-rocm62-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f50ea9cab62a5bd06d886516d3917e4490e65aa9addd1cbb84fc81c6f9a9d5b1
|
3 |
+
size 1749744
|
build/torch26-cxx11-rocm62-x86_64-linux/optimizer/muon.py
ADDED
@@ -0,0 +1,458 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.distributed as dist
|
6 |
+
from torch.distributed._tensor import DTensor
|
7 |
+
|
8 |
+
|
9 |
+
# TODO leave original url and consider LICENSE
|
10 |
+
# This code snippet is a modified version adapted from the following GitHub repository:
|
11 |
+
# https://github.com/KellerJordan/Muon/blob/master/muon.py
|
12 |
+
def _zeropower_via_newtonschulz5(G, steps):
|
13 |
+
"""
|
14 |
+
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
|
15 |
+
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
|
16 |
+
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
|
17 |
+
zero even beyond the point where the iteration no longer converges all the way to one everywhere
|
18 |
+
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
|
19 |
+
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
|
20 |
+
performance at all relative to UV^T, where USV^T = G is the SVD.
|
21 |
+
"""
|
22 |
+
assert len(G.shape) == 2
|
23 |
+
a, b, c = (3.4445, -4.7750, 2.0315)
|
24 |
+
X = G # no manual typecast
|
25 |
+
if G.size(0) > G.size(1):
|
26 |
+
X = X.T
|
27 |
+
# Ensure spectral norm is at most 1
|
28 |
+
X = X / (X.norm() + 1e-7)
|
29 |
+
X = X.bfloat16()
|
30 |
+
# Perform the NS iterations
|
31 |
+
for _ in range(steps):
|
32 |
+
A = X @ X.T
|
33 |
+
# B = (
|
34 |
+
# b * A + c * A @ A
|
35 |
+
# ) # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
|
36 |
+
B = torch.addmm(A, A, A, alpha=c, beta=b)
|
37 |
+
# X = a * X + B @ X
|
38 |
+
X = torch.addmm(X, B, X, alpha=1.0, beta=a)
|
39 |
+
|
40 |
+
if G.size(0) > G.size(1):
|
41 |
+
X = X.T
|
42 |
+
return X.to(G.dtype)
|
43 |
+
|
44 |
+
|
45 |
+
@dataclass
|
46 |
+
class _muon_state:
|
47 |
+
# TODO: use Optional
|
48 |
+
worker_rank: int | None = None
|
49 |
+
gathered_grad: torch.Tensor | None = None
|
50 |
+
computed_u: torch.Tensor | None = None
|
51 |
+
scattered_u: torch.Tensor | None = None
|
52 |
+
gather_event: torch.cuda.Event | None = None
|
53 |
+
compute_event: torch.cuda.Event | None = None
|
54 |
+
|
55 |
+
|
56 |
+
def _gather(p, state, rank, comm_stream):
|
57 |
+
g = p.grad
|
58 |
+
mesh = g.device_mesh
|
59 |
+
|
60 |
+
if rank == state.worker_rank:
|
61 |
+
gather_list = [torch.empty_like(g.to_local()) for _ in range(mesh.mesh.numel())]
|
62 |
+
else:
|
63 |
+
gather_list = None
|
64 |
+
|
65 |
+
with torch.cuda.stream(comm_stream):
|
66 |
+
torch.distributed.gather(
|
67 |
+
g.to_local(),
|
68 |
+
dst=state.worker_rank,
|
69 |
+
gather_list=gather_list,
|
70 |
+
group=mesh.get_group(),
|
71 |
+
)
|
72 |
+
if rank == state.worker_rank:
|
73 |
+
# TODO: Consider ,,,
|
74 |
+
if state.gathered_grad is not None:
|
75 |
+
raise RuntimeError(
|
76 |
+
"Gather event already exists, which should not happen."
|
77 |
+
)
|
78 |
+
state.gathered_grad = torch.cat(gather_list, dim=0)
|
79 |
+
state.gather_event = torch.cuda.Event()
|
80 |
+
state.gather_event.record()
|
81 |
+
else:
|
82 |
+
state.gathered_grad = None
|
83 |
+
state.gather_event = None
|
84 |
+
|
85 |
+
|
86 |
+
def _compute_u(state, steps, rank, compute_stream):
|
87 |
+
with torch.cuda.stream(compute_stream):
|
88 |
+
if rank == state.worker_rank:
|
89 |
+
if state.gather_event is None:
|
90 |
+
raise RuntimeError("Gather event must be set before compute.")
|
91 |
+
compute_stream.wait_event(state.gather_event)
|
92 |
+
u = _zeropower_via_newtonschulz5(state.gathered_grad, steps)
|
93 |
+
state.computed_u = u
|
94 |
+
state.compute_event = torch.cuda.Event()
|
95 |
+
state.compute_event.record()
|
96 |
+
else:
|
97 |
+
state.computed_u = None
|
98 |
+
state.compute_event = None
|
99 |
+
|
100 |
+
|
101 |
+
def _scatter(p, state, rank, comm_stream):
|
102 |
+
u = state.computed_u
|
103 |
+
mesh = p.device_mesh
|
104 |
+
|
105 |
+
with torch.cuda.stream(comm_stream):
|
106 |
+
if rank == state.worker_rank:
|
107 |
+
if state.compute_event is None:
|
108 |
+
raise RuntimeError("Compute event must be set before scatter.")
|
109 |
+
comm_stream.wait_event(state.compute_event)
|
110 |
+
scatter_list = list(torch.split(u, p.size(0) // mesh.mesh.numel(), dim=0))
|
111 |
+
else:
|
112 |
+
scatter_list = None
|
113 |
+
|
114 |
+
u = torch.empty_like(p.to_local())
|
115 |
+
torch.distributed.scatter(
|
116 |
+
u,
|
117 |
+
scatter_list=scatter_list,
|
118 |
+
src=state.worker_rank,
|
119 |
+
group=mesh.get_group(),
|
120 |
+
)
|
121 |
+
u = DTensor.from_local(
|
122 |
+
u,
|
123 |
+
placements=p.placements,
|
124 |
+
device_mesh=mesh,
|
125 |
+
)
|
126 |
+
|
127 |
+
state.scattered_u = u
|
128 |
+
|
129 |
+
|
130 |
+
class Muon(torch.optim.Optimizer):
|
131 |
+
"""
|
132 |
+
Muon - MomentUm Orthogonalized by Newton-schulz
|
133 |
+
|
134 |
+
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
|
135 |
+
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
|
136 |
+
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
|
137 |
+
the advantage that it can be stably run in bfloat16 on the GPU.
|
138 |
+
|
139 |
+
Some warnings:
|
140 |
+
- We believe this optimizer is unlikely to work well for training with small batch size.
|
141 |
+
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
|
142 |
+
|
143 |
+
Arguments:
|
144 |
+
muon_params: The parameters to be optimized by Muon.
|
145 |
+
lr: The learning rate. The updates will have spectral norm of `lr`. (0.02 is a good default)
|
146 |
+
momentum: The momentum used by the internal SGD. (0.95 is a good default)
|
147 |
+
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
|
148 |
+
ns_steps: The number of Newton-Schulz iterations to run. (6 is probably always enough)
|
149 |
+
adamw_params: The parameters to be optimized by AdamW. Any parameters in `muon_params` which are
|
150 |
+
{0, 1}-D or are detected as being the embed or lm_head will be optimized by AdamW as well.
|
151 |
+
adamw_lr: The learning rate for the internal AdamW.
|
152 |
+
adamw_betas: The betas for the internal AdamW.
|
153 |
+
adamw_eps: The epsilon for the internal AdamW.
|
154 |
+
adamw_wd: The weight decay for the internal AdamW.
|
155 |
+
"""
|
156 |
+
|
157 |
+
def __init__(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
is_muon_func,
|
161 |
+
lr=1e-3,
|
162 |
+
momentum=0.95,
|
163 |
+
nesterov=True,
|
164 |
+
ns_steps=5,
|
165 |
+
adamw_wd=0.1,
|
166 |
+
adamw_betas=(0.9, 0.95),
|
167 |
+
adamw_eps=1e-8,
|
168 |
+
debug=False,
|
169 |
+
):
|
170 |
+
defaults = dict(
|
171 |
+
lr=lr,
|
172 |
+
wd=adamw_wd,
|
173 |
+
momentum=momentum,
|
174 |
+
nesterov=nesterov,
|
175 |
+
ns_steps=ns_steps,
|
176 |
+
adamw_betas=adamw_betas,
|
177 |
+
adamw_eps=adamw_eps,
|
178 |
+
)
|
179 |
+
|
180 |
+
super().__init__(model.parameters(), defaults)
|
181 |
+
self.is_muon_func = is_muon_func
|
182 |
+
self.model = model
|
183 |
+
|
184 |
+
if not dist.is_initialized():
|
185 |
+
raise RuntimeError(
|
186 |
+
"Muon optimizer requires distributed training to be initialized."
|
187 |
+
)
|
188 |
+
|
189 |
+
self.rank = dist.get_rank()
|
190 |
+
|
191 |
+
self.comm_stream = torch.cuda.Stream()
|
192 |
+
self.compute_stream = torch.cuda.Stream()
|
193 |
+
self.debug = debug
|
194 |
+
|
195 |
+
def __setstate__(self, state):
|
196 |
+
# Sort parameters into those for which we will use Muon, and those for which we will not
|
197 |
+
super().__setstate__(state)
|
198 |
+
for name, p in self.model.named_parameters():
|
199 |
+
if self.is_muon_func(p, name):
|
200 |
+
# Use Muon for every parameter in muon_params which is >= 2D and doesn't look like an embedding or head layer
|
201 |
+
assert p.ndim == 2, p.ndim
|
202 |
+
self.state[p]["use_muon"] = True
|
203 |
+
self.state[p]["orig_shape"] = p.shape
|
204 |
+
else:
|
205 |
+
# Do not use Muon for parameters in adamw_params
|
206 |
+
self.state[p]["use_muon"] = False
|
207 |
+
|
208 |
+
def _calc_flops(self, G, steps):
|
209 |
+
assert len(G.shape) == 2
|
210 |
+
M, N = G.shape
|
211 |
+
if M > N:
|
212 |
+
M, N = N, M
|
213 |
+
|
214 |
+
return steps * ((M**3) * 2 + (M**2 * N) * 4 + M * N * 2 + M**2 * 3)
|
215 |
+
|
216 |
+
def adjust_lr_for_muon(self, lr, param_shape):
|
217 |
+
A, B = param_shape[:2]
|
218 |
+
# We adjust the learning rate and weight decay based on the size of the parameter matrix
|
219 |
+
# as describted in the paper
|
220 |
+
adjusted_ratio = 0.2 * math.sqrt(max(A, B))
|
221 |
+
adjusted_lr = lr * adjusted_ratio
|
222 |
+
return adjusted_lr
|
223 |
+
|
224 |
+
def init_state_and_assign_params(self, params, group):
|
225 |
+
param_to_state = {}
|
226 |
+
param_to_flops = {}
|
227 |
+
|
228 |
+
total_flops = 0
|
229 |
+
for p in params:
|
230 |
+
g = p.grad
|
231 |
+
if g is None:
|
232 |
+
continue
|
233 |
+
assert g.ndim == 2, "Muon only supports 2D parameters."
|
234 |
+
|
235 |
+
flops = self._calc_flops(g, group["ns_steps"])
|
236 |
+
param_to_flops[id(p)] = flops
|
237 |
+
total_flops += flops
|
238 |
+
|
239 |
+
if self.debug:
|
240 |
+
print(f"Total TFLOPs for Muon: {total_flops / 1e12:.2f} TFLOPs", flush=True)
|
241 |
+
|
242 |
+
ordered_params = sorted(
|
243 |
+
params, key=lambda p: param_to_flops[id(p)], reverse=True
|
244 |
+
)
|
245 |
+
|
246 |
+
round_robin = 0
|
247 |
+
mesh = None
|
248 |
+
for p in ordered_params:
|
249 |
+
if mesh is None:
|
250 |
+
mesh = p.device_mesh
|
251 |
+
if mesh.ndim != 1:
|
252 |
+
raise NotImplementedError(
|
253 |
+
"Muon requires a 1D mesh for distributed training yet."
|
254 |
+
)
|
255 |
+
elif mesh != p.device_mesh:
|
256 |
+
raise ValueError("All parameters must be on the same mesh.")
|
257 |
+
|
258 |
+
param_to_state[id(p)] = _muon_state()
|
259 |
+
param_to_state[id(p)].worker_rank = mesh.mesh[round_robin].item()
|
260 |
+
|
261 |
+
round_robin = (round_robin + 1) % mesh.mesh.numel()
|
262 |
+
|
263 |
+
return param_to_state, ordered_params
|
264 |
+
|
265 |
+
def base(self, params, group, lr, wd, momentum):
|
266 |
+
# generate weight updates in distributed fashion
|
267 |
+
for p in params:
|
268 |
+
g = p.grad
|
269 |
+
if g is None:
|
270 |
+
continue
|
271 |
+
if g.ndim > 2:
|
272 |
+
g = g.view(g.size(0), -1)
|
273 |
+
assert g is not None
|
274 |
+
|
275 |
+
# calc update
|
276 |
+
state = self.state[p]
|
277 |
+
if "momentum_buffer" not in state:
|
278 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
279 |
+
buf = state["momentum_buffer"]
|
280 |
+
buf.mul_(momentum).add_(g)
|
281 |
+
if group["nesterov"]:
|
282 |
+
g = g.add(buf, alpha=momentum)
|
283 |
+
else:
|
284 |
+
g = buf
|
285 |
+
|
286 |
+
u = _zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
|
287 |
+
|
288 |
+
# scale update
|
289 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
290 |
+
|
291 |
+
# apply weight decay
|
292 |
+
p.data.mul_(1 - lr * wd)
|
293 |
+
|
294 |
+
# apply update
|
295 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
296 |
+
|
297 |
+
def _update_g(self, p, g, group, momentum):
|
298 |
+
# calc update
|
299 |
+
state = self.state[p]
|
300 |
+
if "momentum_buffer" not in state:
|
301 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
302 |
+
buf = state["momentum_buffer"]
|
303 |
+
buf.mul_(momentum).add_(g)
|
304 |
+
if group["nesterov"]:
|
305 |
+
g = g.add(buf, alpha=momentum)
|
306 |
+
else:
|
307 |
+
g = buf
|
308 |
+
return g
|
309 |
+
|
310 |
+
def _update_p(self, p, u, lr, wd):
|
311 |
+
# scale update
|
312 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
313 |
+
# apply weight decay
|
314 |
+
p.data.mul_(1 - lr * wd)
|
315 |
+
# apply update
|
316 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
317 |
+
|
318 |
+
def parallel(self, params, group, lr, wd, momentum):
|
319 |
+
"""
|
320 |
+
Perform a parallel optimization step using Muon.
|
321 |
+
"""
|
322 |
+
|
323 |
+
for p in params:
|
324 |
+
g = p.grad
|
325 |
+
if g is None:
|
326 |
+
continue
|
327 |
+
if g.ndim > 2:
|
328 |
+
g = g.view(g.size(0), -1)
|
329 |
+
|
330 |
+
# Update g in the local rank
|
331 |
+
g = self._update_g(
|
332 |
+
p,
|
333 |
+
g,
|
334 |
+
group,
|
335 |
+
momentum=momentum,
|
336 |
+
)
|
337 |
+
p.grad = g
|
338 |
+
|
339 |
+
param_to_state, ordered_params = self.init_state_and_assign_params(
|
340 |
+
params, group
|
341 |
+
)
|
342 |
+
|
343 |
+
def enqueue_gathers(start_idx, chunk_size):
|
344 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
345 |
+
state = param_to_state[id(p)]
|
346 |
+
_gather(p, state, self.rank, self.comm_stream)
|
347 |
+
|
348 |
+
def enqueue_computes(start_idx, chunk_size):
|
349 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
350 |
+
state = param_to_state[id(p)]
|
351 |
+
_compute_u(state, group["ns_steps"], self.rank, self.compute_stream)
|
352 |
+
|
353 |
+
def enqueue_scatters(start_idx, chunk_size):
|
354 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
355 |
+
state = param_to_state[id(p)]
|
356 |
+
_scatter(p, state, self.rank, self.comm_stream)
|
357 |
+
|
358 |
+
chunk_size = params[0].device_mesh.mesh.numel()
|
359 |
+
|
360 |
+
# Wait grad update
|
361 |
+
self.comm_stream.wait_stream(torch.cuda.current_stream())
|
362 |
+
|
363 |
+
enqueue_gathers(0, chunk_size)
|
364 |
+
for i in range(0, len(params) + chunk_size - 1, chunk_size):
|
365 |
+
enqueue_computes(i, chunk_size)
|
366 |
+
enqueue_gathers(i + chunk_size, chunk_size)
|
367 |
+
enqueue_scatters(i, chunk_size)
|
368 |
+
|
369 |
+
torch.cuda.current_stream().wait_stream(self.comm_stream)
|
370 |
+
|
371 |
+
for p in params:
|
372 |
+
g = p.grad
|
373 |
+
if g is None:
|
374 |
+
continue
|
375 |
+
|
376 |
+
# Update p with sharded u
|
377 |
+
state = param_to_state[id(p)]
|
378 |
+
self._update_p(
|
379 |
+
p,
|
380 |
+
state.scattered_u,
|
381 |
+
lr=lr,
|
382 |
+
wd=wd,
|
383 |
+
)
|
384 |
+
|
385 |
+
def step(self, closure=None):
|
386 |
+
"""Perform a single optimization step.
|
387 |
+
|
388 |
+
Args:
|
389 |
+
closure (Callable, optional): A closure that reevaluates the model
|
390 |
+
and returns the loss.
|
391 |
+
"""
|
392 |
+
loss = None
|
393 |
+
if closure is not None:
|
394 |
+
with torch.enable_grad():
|
395 |
+
loss = closure()
|
396 |
+
|
397 |
+
for group in self.param_groups:
|
398 |
+
############################
|
399 |
+
# Muon #
|
400 |
+
############################
|
401 |
+
|
402 |
+
params = [p for p in group["params"] if self.state[p]["use_muon"]]
|
403 |
+
lr = group["lr"]
|
404 |
+
wd = group["wd"]
|
405 |
+
momentum = group["momentum"]
|
406 |
+
|
407 |
+
if isinstance(params[0].data, DTensor):
|
408 |
+
self.parallel(
|
409 |
+
params,
|
410 |
+
group,
|
411 |
+
lr=lr,
|
412 |
+
wd=wd,
|
413 |
+
momentum=momentum,
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
self.base(
|
417 |
+
params,
|
418 |
+
group,
|
419 |
+
lr=lr,
|
420 |
+
wd=wd,
|
421 |
+
momentum=momentum,
|
422 |
+
)
|
423 |
+
|
424 |
+
############################
|
425 |
+
# AdamW backup #
|
426 |
+
############################
|
427 |
+
|
428 |
+
params = [p for p in group["params"] if not self.state[p]["use_muon"]]
|
429 |
+
lr = group["lr"]
|
430 |
+
beta1, beta2 = group["adamw_betas"]
|
431 |
+
eps = group["adamw_eps"]
|
432 |
+
weight_decay = group["wd"]
|
433 |
+
|
434 |
+
for p in params:
|
435 |
+
g = p.grad
|
436 |
+
if g is None:
|
437 |
+
continue
|
438 |
+
state = self.state[p]
|
439 |
+
if "step" not in state:
|
440 |
+
state["step"] = 0
|
441 |
+
state["moment1"] = torch.zeros_like(g)
|
442 |
+
state["moment2"] = torch.zeros_like(g)
|
443 |
+
state["step"] += 1
|
444 |
+
step = state["step"]
|
445 |
+
buf1 = state["moment1"]
|
446 |
+
buf2 = state["moment2"]
|
447 |
+
buf1.lerp_(g, 1 - beta1)
|
448 |
+
buf2.lerp_(g.square(), 1 - beta2)
|
449 |
+
|
450 |
+
g = buf1 / (eps + buf2.sqrt())
|
451 |
+
|
452 |
+
bias_correction1 = 1 - beta1**step
|
453 |
+
bias_correction2 = 1 - beta2**step
|
454 |
+
scale = bias_correction1 / bias_correction2**0.5
|
455 |
+
p.data.mul_(1 - lr * weight_decay)
|
456 |
+
p.data.add_(g, alpha=-lr / scale)
|
457 |
+
|
458 |
+
return loss
|
build/torch26-cxx98-cu118-x86_64-linux/optimizer/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .muon import Muon
|
2 |
+
|
3 |
+
__all__ = [
|
4 |
+
"Muon",
|
5 |
+
]
|
build/torch26-cxx98-cu118-x86_64-linux/optimizer/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _optimizer_b4b3752_dirty
|
3 |
+
ops = torch.ops._optimizer_b4b3752_dirty
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_optimizer_b4b3752_dirty::{op_name}"
|
build/torch26-cxx98-cu118-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a8f8e7d78ed9a095b882cf764fd9c80a0b0810fb961ba9e8545656fc4cb0b0d7
|
3 |
+
size 1787200
|
build/torch26-cxx98-cu118-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:002dab6441bcad54ab4e7c064b5806acfd45170eb33cfa059745ba6e0c349607
|
3 |
+
size 1787192
|
build/torch26-cxx98-cu118-x86_64-linux/optimizer/muon.py
ADDED
@@ -0,0 +1,458 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.distributed as dist
|
6 |
+
from torch.distributed._tensor import DTensor
|
7 |
+
|
8 |
+
|
9 |
+
# TODO leave original url and consider LICENSE
|
10 |
+
# This code snippet is a modified version adapted from the following GitHub repository:
|
11 |
+
# https://github.com/KellerJordan/Muon/blob/master/muon.py
|
12 |
+
def _zeropower_via_newtonschulz5(G, steps):
|
13 |
+
"""
|
14 |
+
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
|
15 |
+
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
|
16 |
+
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
|
17 |
+
zero even beyond the point where the iteration no longer converges all the way to one everywhere
|
18 |
+
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
|
19 |
+
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
|
20 |
+
performance at all relative to UV^T, where USV^T = G is the SVD.
|
21 |
+
"""
|
22 |
+
assert len(G.shape) == 2
|
23 |
+
a, b, c = (3.4445, -4.7750, 2.0315)
|
24 |
+
X = G # no manual typecast
|
25 |
+
if G.size(0) > G.size(1):
|
26 |
+
X = X.T
|
27 |
+
# Ensure spectral norm is at most 1
|
28 |
+
X = X / (X.norm() + 1e-7)
|
29 |
+
X = X.bfloat16()
|
30 |
+
# Perform the NS iterations
|
31 |
+
for _ in range(steps):
|
32 |
+
A = X @ X.T
|
33 |
+
# B = (
|
34 |
+
# b * A + c * A @ A
|
35 |
+
# ) # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
|
36 |
+
B = torch.addmm(A, A, A, alpha=c, beta=b)
|
37 |
+
# X = a * X + B @ X
|
38 |
+
X = torch.addmm(X, B, X, alpha=1.0, beta=a)
|
39 |
+
|
40 |
+
if G.size(0) > G.size(1):
|
41 |
+
X = X.T
|
42 |
+
return X.to(G.dtype)
|
43 |
+
|
44 |
+
|
45 |
+
@dataclass
|
46 |
+
class _muon_state:
|
47 |
+
# TODO: use Optional
|
48 |
+
worker_rank: int | None = None
|
49 |
+
gathered_grad: torch.Tensor | None = None
|
50 |
+
computed_u: torch.Tensor | None = None
|
51 |
+
scattered_u: torch.Tensor | None = None
|
52 |
+
gather_event: torch.cuda.Event | None = None
|
53 |
+
compute_event: torch.cuda.Event | None = None
|
54 |
+
|
55 |
+
|
56 |
+
def _gather(p, state, rank, comm_stream):
|
57 |
+
g = p.grad
|
58 |
+
mesh = g.device_mesh
|
59 |
+
|
60 |
+
if rank == state.worker_rank:
|
61 |
+
gather_list = [torch.empty_like(g.to_local()) for _ in range(mesh.mesh.numel())]
|
62 |
+
else:
|
63 |
+
gather_list = None
|
64 |
+
|
65 |
+
with torch.cuda.stream(comm_stream):
|
66 |
+
torch.distributed.gather(
|
67 |
+
g.to_local(),
|
68 |
+
dst=state.worker_rank,
|
69 |
+
gather_list=gather_list,
|
70 |
+
group=mesh.get_group(),
|
71 |
+
)
|
72 |
+
if rank == state.worker_rank:
|
73 |
+
# TODO: Consider ,,,
|
74 |
+
if state.gathered_grad is not None:
|
75 |
+
raise RuntimeError(
|
76 |
+
"Gather event already exists, which should not happen."
|
77 |
+
)
|
78 |
+
state.gathered_grad = torch.cat(gather_list, dim=0)
|
79 |
+
state.gather_event = torch.cuda.Event()
|
80 |
+
state.gather_event.record()
|
81 |
+
else:
|
82 |
+
state.gathered_grad = None
|
83 |
+
state.gather_event = None
|
84 |
+
|
85 |
+
|
86 |
+
def _compute_u(state, steps, rank, compute_stream):
|
87 |
+
with torch.cuda.stream(compute_stream):
|
88 |
+
if rank == state.worker_rank:
|
89 |
+
if state.gather_event is None:
|
90 |
+
raise RuntimeError("Gather event must be set before compute.")
|
91 |
+
compute_stream.wait_event(state.gather_event)
|
92 |
+
u = _zeropower_via_newtonschulz5(state.gathered_grad, steps)
|
93 |
+
state.computed_u = u
|
94 |
+
state.compute_event = torch.cuda.Event()
|
95 |
+
state.compute_event.record()
|
96 |
+
else:
|
97 |
+
state.computed_u = None
|
98 |
+
state.compute_event = None
|
99 |
+
|
100 |
+
|
101 |
+
def _scatter(p, state, rank, comm_stream):
|
102 |
+
u = state.computed_u
|
103 |
+
mesh = p.device_mesh
|
104 |
+
|
105 |
+
with torch.cuda.stream(comm_stream):
|
106 |
+
if rank == state.worker_rank:
|
107 |
+
if state.compute_event is None:
|
108 |
+
raise RuntimeError("Compute event must be set before scatter.")
|
109 |
+
comm_stream.wait_event(state.compute_event)
|
110 |
+
scatter_list = list(torch.split(u, p.size(0) // mesh.mesh.numel(), dim=0))
|
111 |
+
else:
|
112 |
+
scatter_list = None
|
113 |
+
|
114 |
+
u = torch.empty_like(p.to_local())
|
115 |
+
torch.distributed.scatter(
|
116 |
+
u,
|
117 |
+
scatter_list=scatter_list,
|
118 |
+
src=state.worker_rank,
|
119 |
+
group=mesh.get_group(),
|
120 |
+
)
|
121 |
+
u = DTensor.from_local(
|
122 |
+
u,
|
123 |
+
placements=p.placements,
|
124 |
+
device_mesh=mesh,
|
125 |
+
)
|
126 |
+
|
127 |
+
state.scattered_u = u
|
128 |
+
|
129 |
+
|
130 |
+
class Muon(torch.optim.Optimizer):
|
131 |
+
"""
|
132 |
+
Muon - MomentUm Orthogonalized by Newton-schulz
|
133 |
+
|
134 |
+
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
|
135 |
+
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
|
136 |
+
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
|
137 |
+
the advantage that it can be stably run in bfloat16 on the GPU.
|
138 |
+
|
139 |
+
Some warnings:
|
140 |
+
- We believe this optimizer is unlikely to work well for training with small batch size.
|
141 |
+
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
|
142 |
+
|
143 |
+
Arguments:
|
144 |
+
muon_params: The parameters to be optimized by Muon.
|
145 |
+
lr: The learning rate. The updates will have spectral norm of `lr`. (0.02 is a good default)
|
146 |
+
momentum: The momentum used by the internal SGD. (0.95 is a good default)
|
147 |
+
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
|
148 |
+
ns_steps: The number of Newton-Schulz iterations to run. (6 is probably always enough)
|
149 |
+
adamw_params: The parameters to be optimized by AdamW. Any parameters in `muon_params` which are
|
150 |
+
{0, 1}-D or are detected as being the embed or lm_head will be optimized by AdamW as well.
|
151 |
+
adamw_lr: The learning rate for the internal AdamW.
|
152 |
+
adamw_betas: The betas for the internal AdamW.
|
153 |
+
adamw_eps: The epsilon for the internal AdamW.
|
154 |
+
adamw_wd: The weight decay for the internal AdamW.
|
155 |
+
"""
|
156 |
+
|
157 |
+
def __init__(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
is_muon_func,
|
161 |
+
lr=1e-3,
|
162 |
+
momentum=0.95,
|
163 |
+
nesterov=True,
|
164 |
+
ns_steps=5,
|
165 |
+
adamw_wd=0.1,
|
166 |
+
adamw_betas=(0.9, 0.95),
|
167 |
+
adamw_eps=1e-8,
|
168 |
+
debug=False,
|
169 |
+
):
|
170 |
+
defaults = dict(
|
171 |
+
lr=lr,
|
172 |
+
wd=adamw_wd,
|
173 |
+
momentum=momentum,
|
174 |
+
nesterov=nesterov,
|
175 |
+
ns_steps=ns_steps,
|
176 |
+
adamw_betas=adamw_betas,
|
177 |
+
adamw_eps=adamw_eps,
|
178 |
+
)
|
179 |
+
|
180 |
+
super().__init__(model.parameters(), defaults)
|
181 |
+
self.is_muon_func = is_muon_func
|
182 |
+
self.model = model
|
183 |
+
|
184 |
+
if not dist.is_initialized():
|
185 |
+
raise RuntimeError(
|
186 |
+
"Muon optimizer requires distributed training to be initialized."
|
187 |
+
)
|
188 |
+
|
189 |
+
self.rank = dist.get_rank()
|
190 |
+
|
191 |
+
self.comm_stream = torch.cuda.Stream()
|
192 |
+
self.compute_stream = torch.cuda.Stream()
|
193 |
+
self.debug = debug
|
194 |
+
|
195 |
+
def __setstate__(self, state):
|
196 |
+
# Sort parameters into those for which we will use Muon, and those for which we will not
|
197 |
+
super().__setstate__(state)
|
198 |
+
for name, p in self.model.named_parameters():
|
199 |
+
if self.is_muon_func(p, name):
|
200 |
+
# Use Muon for every parameter in muon_params which is >= 2D and doesn't look like an embedding or head layer
|
201 |
+
assert p.ndim == 2, p.ndim
|
202 |
+
self.state[p]["use_muon"] = True
|
203 |
+
self.state[p]["orig_shape"] = p.shape
|
204 |
+
else:
|
205 |
+
# Do not use Muon for parameters in adamw_params
|
206 |
+
self.state[p]["use_muon"] = False
|
207 |
+
|
208 |
+
def _calc_flops(self, G, steps):
|
209 |
+
assert len(G.shape) == 2
|
210 |
+
M, N = G.shape
|
211 |
+
if M > N:
|
212 |
+
M, N = N, M
|
213 |
+
|
214 |
+
return steps * ((M**3) * 2 + (M**2 * N) * 4 + M * N * 2 + M**2 * 3)
|
215 |
+
|
216 |
+
def adjust_lr_for_muon(self, lr, param_shape):
|
217 |
+
A, B = param_shape[:2]
|
218 |
+
# We adjust the learning rate and weight decay based on the size of the parameter matrix
|
219 |
+
# as describted in the paper
|
220 |
+
adjusted_ratio = 0.2 * math.sqrt(max(A, B))
|
221 |
+
adjusted_lr = lr * adjusted_ratio
|
222 |
+
return adjusted_lr
|
223 |
+
|
224 |
+
def init_state_and_assign_params(self, params, group):
|
225 |
+
param_to_state = {}
|
226 |
+
param_to_flops = {}
|
227 |
+
|
228 |
+
total_flops = 0
|
229 |
+
for p in params:
|
230 |
+
g = p.grad
|
231 |
+
if g is None:
|
232 |
+
continue
|
233 |
+
assert g.ndim == 2, "Muon only supports 2D parameters."
|
234 |
+
|
235 |
+
flops = self._calc_flops(g, group["ns_steps"])
|
236 |
+
param_to_flops[id(p)] = flops
|
237 |
+
total_flops += flops
|
238 |
+
|
239 |
+
if self.debug:
|
240 |
+
print(f"Total TFLOPs for Muon: {total_flops / 1e12:.2f} TFLOPs", flush=True)
|
241 |
+
|
242 |
+
ordered_params = sorted(
|
243 |
+
params, key=lambda p: param_to_flops[id(p)], reverse=True
|
244 |
+
)
|
245 |
+
|
246 |
+
round_robin = 0
|
247 |
+
mesh = None
|
248 |
+
for p in ordered_params:
|
249 |
+
if mesh is None:
|
250 |
+
mesh = p.device_mesh
|
251 |
+
if mesh.ndim != 1:
|
252 |
+
raise NotImplementedError(
|
253 |
+
"Muon requires a 1D mesh for distributed training yet."
|
254 |
+
)
|
255 |
+
elif mesh != p.device_mesh:
|
256 |
+
raise ValueError("All parameters must be on the same mesh.")
|
257 |
+
|
258 |
+
param_to_state[id(p)] = _muon_state()
|
259 |
+
param_to_state[id(p)].worker_rank = mesh.mesh[round_robin].item()
|
260 |
+
|
261 |
+
round_robin = (round_robin + 1) % mesh.mesh.numel()
|
262 |
+
|
263 |
+
return param_to_state, ordered_params
|
264 |
+
|
265 |
+
def base(self, params, group, lr, wd, momentum):
|
266 |
+
# generate weight updates in distributed fashion
|
267 |
+
for p in params:
|
268 |
+
g = p.grad
|
269 |
+
if g is None:
|
270 |
+
continue
|
271 |
+
if g.ndim > 2:
|
272 |
+
g = g.view(g.size(0), -1)
|
273 |
+
assert g is not None
|
274 |
+
|
275 |
+
# calc update
|
276 |
+
state = self.state[p]
|
277 |
+
if "momentum_buffer" not in state:
|
278 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
279 |
+
buf = state["momentum_buffer"]
|
280 |
+
buf.mul_(momentum).add_(g)
|
281 |
+
if group["nesterov"]:
|
282 |
+
g = g.add(buf, alpha=momentum)
|
283 |
+
else:
|
284 |
+
g = buf
|
285 |
+
|
286 |
+
u = _zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
|
287 |
+
|
288 |
+
# scale update
|
289 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
290 |
+
|
291 |
+
# apply weight decay
|
292 |
+
p.data.mul_(1 - lr * wd)
|
293 |
+
|
294 |
+
# apply update
|
295 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
296 |
+
|
297 |
+
def _update_g(self, p, g, group, momentum):
|
298 |
+
# calc update
|
299 |
+
state = self.state[p]
|
300 |
+
if "momentum_buffer" not in state:
|
301 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
302 |
+
buf = state["momentum_buffer"]
|
303 |
+
buf.mul_(momentum).add_(g)
|
304 |
+
if group["nesterov"]:
|
305 |
+
g = g.add(buf, alpha=momentum)
|
306 |
+
else:
|
307 |
+
g = buf
|
308 |
+
return g
|
309 |
+
|
310 |
+
def _update_p(self, p, u, lr, wd):
|
311 |
+
# scale update
|
312 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
313 |
+
# apply weight decay
|
314 |
+
p.data.mul_(1 - lr * wd)
|
315 |
+
# apply update
|
316 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
317 |
+
|
318 |
+
def parallel(self, params, group, lr, wd, momentum):
|
319 |
+
"""
|
320 |
+
Perform a parallel optimization step using Muon.
|
321 |
+
"""
|
322 |
+
|
323 |
+
for p in params:
|
324 |
+
g = p.grad
|
325 |
+
if g is None:
|
326 |
+
continue
|
327 |
+
if g.ndim > 2:
|
328 |
+
g = g.view(g.size(0), -1)
|
329 |
+
|
330 |
+
# Update g in the local rank
|
331 |
+
g = self._update_g(
|
332 |
+
p,
|
333 |
+
g,
|
334 |
+
group,
|
335 |
+
momentum=momentum,
|
336 |
+
)
|
337 |
+
p.grad = g
|
338 |
+
|
339 |
+
param_to_state, ordered_params = self.init_state_and_assign_params(
|
340 |
+
params, group
|
341 |
+
)
|
342 |
+
|
343 |
+
def enqueue_gathers(start_idx, chunk_size):
|
344 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
345 |
+
state = param_to_state[id(p)]
|
346 |
+
_gather(p, state, self.rank, self.comm_stream)
|
347 |
+
|
348 |
+
def enqueue_computes(start_idx, chunk_size):
|
349 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
350 |
+
state = param_to_state[id(p)]
|
351 |
+
_compute_u(state, group["ns_steps"], self.rank, self.compute_stream)
|
352 |
+
|
353 |
+
def enqueue_scatters(start_idx, chunk_size):
|
354 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
355 |
+
state = param_to_state[id(p)]
|
356 |
+
_scatter(p, state, self.rank, self.comm_stream)
|
357 |
+
|
358 |
+
chunk_size = params[0].device_mesh.mesh.numel()
|
359 |
+
|
360 |
+
# Wait grad update
|
361 |
+
self.comm_stream.wait_stream(torch.cuda.current_stream())
|
362 |
+
|
363 |
+
enqueue_gathers(0, chunk_size)
|
364 |
+
for i in range(0, len(params) + chunk_size - 1, chunk_size):
|
365 |
+
enqueue_computes(i, chunk_size)
|
366 |
+
enqueue_gathers(i + chunk_size, chunk_size)
|
367 |
+
enqueue_scatters(i, chunk_size)
|
368 |
+
|
369 |
+
torch.cuda.current_stream().wait_stream(self.comm_stream)
|
370 |
+
|
371 |
+
for p in params:
|
372 |
+
g = p.grad
|
373 |
+
if g is None:
|
374 |
+
continue
|
375 |
+
|
376 |
+
# Update p with sharded u
|
377 |
+
state = param_to_state[id(p)]
|
378 |
+
self._update_p(
|
379 |
+
p,
|
380 |
+
state.scattered_u,
|
381 |
+
lr=lr,
|
382 |
+
wd=wd,
|
383 |
+
)
|
384 |
+
|
385 |
+
def step(self, closure=None):
|
386 |
+
"""Perform a single optimization step.
|
387 |
+
|
388 |
+
Args:
|
389 |
+
closure (Callable, optional): A closure that reevaluates the model
|
390 |
+
and returns the loss.
|
391 |
+
"""
|
392 |
+
loss = None
|
393 |
+
if closure is not None:
|
394 |
+
with torch.enable_grad():
|
395 |
+
loss = closure()
|
396 |
+
|
397 |
+
for group in self.param_groups:
|
398 |
+
############################
|
399 |
+
# Muon #
|
400 |
+
############################
|
401 |
+
|
402 |
+
params = [p for p in group["params"] if self.state[p]["use_muon"]]
|
403 |
+
lr = group["lr"]
|
404 |
+
wd = group["wd"]
|
405 |
+
momentum = group["momentum"]
|
406 |
+
|
407 |
+
if isinstance(params[0].data, DTensor):
|
408 |
+
self.parallel(
|
409 |
+
params,
|
410 |
+
group,
|
411 |
+
lr=lr,
|
412 |
+
wd=wd,
|
413 |
+
momentum=momentum,
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
self.base(
|
417 |
+
params,
|
418 |
+
group,
|
419 |
+
lr=lr,
|
420 |
+
wd=wd,
|
421 |
+
momentum=momentum,
|
422 |
+
)
|
423 |
+
|
424 |
+
############################
|
425 |
+
# AdamW backup #
|
426 |
+
############################
|
427 |
+
|
428 |
+
params = [p for p in group["params"] if not self.state[p]["use_muon"]]
|
429 |
+
lr = group["lr"]
|
430 |
+
beta1, beta2 = group["adamw_betas"]
|
431 |
+
eps = group["adamw_eps"]
|
432 |
+
weight_decay = group["wd"]
|
433 |
+
|
434 |
+
for p in params:
|
435 |
+
g = p.grad
|
436 |
+
if g is None:
|
437 |
+
continue
|
438 |
+
state = self.state[p]
|
439 |
+
if "step" not in state:
|
440 |
+
state["step"] = 0
|
441 |
+
state["moment1"] = torch.zeros_like(g)
|
442 |
+
state["moment2"] = torch.zeros_like(g)
|
443 |
+
state["step"] += 1
|
444 |
+
step = state["step"]
|
445 |
+
buf1 = state["moment1"]
|
446 |
+
buf2 = state["moment2"]
|
447 |
+
buf1.lerp_(g, 1 - beta1)
|
448 |
+
buf2.lerp_(g.square(), 1 - beta2)
|
449 |
+
|
450 |
+
g = buf1 / (eps + buf2.sqrt())
|
451 |
+
|
452 |
+
bias_correction1 = 1 - beta1**step
|
453 |
+
bias_correction2 = 1 - beta2**step
|
454 |
+
scale = bias_correction1 / bias_correction2**0.5
|
455 |
+
p.data.mul_(1 - lr * weight_decay)
|
456 |
+
p.data.add_(g, alpha=-lr / scale)
|
457 |
+
|
458 |
+
return loss
|
build/torch26-cxx98-cu124-x86_64-linux/optimizer/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .muon import Muon
|
2 |
+
|
3 |
+
__all__ = [
|
4 |
+
"Muon",
|
5 |
+
]
|
build/torch26-cxx98-cu124-x86_64-linux/optimizer/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _optimizer_b4b3752_dirty
|
3 |
+
ops = torch.ops._optimizer_b4b3752_dirty
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_optimizer_b4b3752_dirty::{op_name}"
|
build/torch26-cxx98-cu124-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:ab2379d932e40d10bee55f032bd16d2e4d9c1920bc5500628006f8a0eb8abd39
|
3 |
+
size 1824192
|
build/torch26-cxx98-cu124-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:7f499350bb19eca6c3da1bb72e46023834b8411ce00730854273b588b2cd9206
|
3 |
+
size 1824184
|
build/torch26-cxx98-cu124-x86_64-linux/optimizer/muon.py
ADDED
@@ -0,0 +1,458 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.distributed as dist
|
6 |
+
from torch.distributed._tensor import DTensor
|
7 |
+
|
8 |
+
|
9 |
+
# TODO leave original url and consider LICENSE
|
10 |
+
# This code snippet is a modified version adapted from the following GitHub repository:
|
11 |
+
# https://github.com/KellerJordan/Muon/blob/master/muon.py
|
12 |
+
def _zeropower_via_newtonschulz5(G, steps):
|
13 |
+
"""
|
14 |
+
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
|
15 |
+
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
|
16 |
+
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
|
17 |
+
zero even beyond the point where the iteration no longer converges all the way to one everywhere
|
18 |
+
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
|
19 |
+
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
|
20 |
+
performance at all relative to UV^T, where USV^T = G is the SVD.
|
21 |
+
"""
|
22 |
+
assert len(G.shape) == 2
|
23 |
+
a, b, c = (3.4445, -4.7750, 2.0315)
|
24 |
+
X = G # no manual typecast
|
25 |
+
if G.size(0) > G.size(1):
|
26 |
+
X = X.T
|
27 |
+
# Ensure spectral norm is at most 1
|
28 |
+
X = X / (X.norm() + 1e-7)
|
29 |
+
X = X.bfloat16()
|
30 |
+
# Perform the NS iterations
|
31 |
+
for _ in range(steps):
|
32 |
+
A = X @ X.T
|
33 |
+
# B = (
|
34 |
+
# b * A + c * A @ A
|
35 |
+
# ) # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
|
36 |
+
B = torch.addmm(A, A, A, alpha=c, beta=b)
|
37 |
+
# X = a * X + B @ X
|
38 |
+
X = torch.addmm(X, B, X, alpha=1.0, beta=a)
|
39 |
+
|
40 |
+
if G.size(0) > G.size(1):
|
41 |
+
X = X.T
|
42 |
+
return X.to(G.dtype)
|
43 |
+
|
44 |
+
|
45 |
+
@dataclass
|
46 |
+
class _muon_state:
|
47 |
+
# TODO: use Optional
|
48 |
+
worker_rank: int | None = None
|
49 |
+
gathered_grad: torch.Tensor | None = None
|
50 |
+
computed_u: torch.Tensor | None = None
|
51 |
+
scattered_u: torch.Tensor | None = None
|
52 |
+
gather_event: torch.cuda.Event | None = None
|
53 |
+
compute_event: torch.cuda.Event | None = None
|
54 |
+
|
55 |
+
|
56 |
+
def _gather(p, state, rank, comm_stream):
|
57 |
+
g = p.grad
|
58 |
+
mesh = g.device_mesh
|
59 |
+
|
60 |
+
if rank == state.worker_rank:
|
61 |
+
gather_list = [torch.empty_like(g.to_local()) for _ in range(mesh.mesh.numel())]
|
62 |
+
else:
|
63 |
+
gather_list = None
|
64 |
+
|
65 |
+
with torch.cuda.stream(comm_stream):
|
66 |
+
torch.distributed.gather(
|
67 |
+
g.to_local(),
|
68 |
+
dst=state.worker_rank,
|
69 |
+
gather_list=gather_list,
|
70 |
+
group=mesh.get_group(),
|
71 |
+
)
|
72 |
+
if rank == state.worker_rank:
|
73 |
+
# TODO: Consider ,,,
|
74 |
+
if state.gathered_grad is not None:
|
75 |
+
raise RuntimeError(
|
76 |
+
"Gather event already exists, which should not happen."
|
77 |
+
)
|
78 |
+
state.gathered_grad = torch.cat(gather_list, dim=0)
|
79 |
+
state.gather_event = torch.cuda.Event()
|
80 |
+
state.gather_event.record()
|
81 |
+
else:
|
82 |
+
state.gathered_grad = None
|
83 |
+
state.gather_event = None
|
84 |
+
|
85 |
+
|
86 |
+
def _compute_u(state, steps, rank, compute_stream):
|
87 |
+
with torch.cuda.stream(compute_stream):
|
88 |
+
if rank == state.worker_rank:
|
89 |
+
if state.gather_event is None:
|
90 |
+
raise RuntimeError("Gather event must be set before compute.")
|
91 |
+
compute_stream.wait_event(state.gather_event)
|
92 |
+
u = _zeropower_via_newtonschulz5(state.gathered_grad, steps)
|
93 |
+
state.computed_u = u
|
94 |
+
state.compute_event = torch.cuda.Event()
|
95 |
+
state.compute_event.record()
|
96 |
+
else:
|
97 |
+
state.computed_u = None
|
98 |
+
state.compute_event = None
|
99 |
+
|
100 |
+
|
101 |
+
def _scatter(p, state, rank, comm_stream):
|
102 |
+
u = state.computed_u
|
103 |
+
mesh = p.device_mesh
|
104 |
+
|
105 |
+
with torch.cuda.stream(comm_stream):
|
106 |
+
if rank == state.worker_rank:
|
107 |
+
if state.compute_event is None:
|
108 |
+
raise RuntimeError("Compute event must be set before scatter.")
|
109 |
+
comm_stream.wait_event(state.compute_event)
|
110 |
+
scatter_list = list(torch.split(u, p.size(0) // mesh.mesh.numel(), dim=0))
|
111 |
+
else:
|
112 |
+
scatter_list = None
|
113 |
+
|
114 |
+
u = torch.empty_like(p.to_local())
|
115 |
+
torch.distributed.scatter(
|
116 |
+
u,
|
117 |
+
scatter_list=scatter_list,
|
118 |
+
src=state.worker_rank,
|
119 |
+
group=mesh.get_group(),
|
120 |
+
)
|
121 |
+
u = DTensor.from_local(
|
122 |
+
u,
|
123 |
+
placements=p.placements,
|
124 |
+
device_mesh=mesh,
|
125 |
+
)
|
126 |
+
|
127 |
+
state.scattered_u = u
|
128 |
+
|
129 |
+
|
130 |
+
class Muon(torch.optim.Optimizer):
|
131 |
+
"""
|
132 |
+
Muon - MomentUm Orthogonalized by Newton-schulz
|
133 |
+
|
134 |
+
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
|
135 |
+
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
|
136 |
+
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
|
137 |
+
the advantage that it can be stably run in bfloat16 on the GPU.
|
138 |
+
|
139 |
+
Some warnings:
|
140 |
+
- We believe this optimizer is unlikely to work well for training with small batch size.
|
141 |
+
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
|
142 |
+
|
143 |
+
Arguments:
|
144 |
+
muon_params: The parameters to be optimized by Muon.
|
145 |
+
lr: The learning rate. The updates will have spectral norm of `lr`. (0.02 is a good default)
|
146 |
+
momentum: The momentum used by the internal SGD. (0.95 is a good default)
|
147 |
+
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
|
148 |
+
ns_steps: The number of Newton-Schulz iterations to run. (6 is probably always enough)
|
149 |
+
adamw_params: The parameters to be optimized by AdamW. Any parameters in `muon_params` which are
|
150 |
+
{0, 1}-D or are detected as being the embed or lm_head will be optimized by AdamW as well.
|
151 |
+
adamw_lr: The learning rate for the internal AdamW.
|
152 |
+
adamw_betas: The betas for the internal AdamW.
|
153 |
+
adamw_eps: The epsilon for the internal AdamW.
|
154 |
+
adamw_wd: The weight decay for the internal AdamW.
|
155 |
+
"""
|
156 |
+
|
157 |
+
def __init__(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
is_muon_func,
|
161 |
+
lr=1e-3,
|
162 |
+
momentum=0.95,
|
163 |
+
nesterov=True,
|
164 |
+
ns_steps=5,
|
165 |
+
adamw_wd=0.1,
|
166 |
+
adamw_betas=(0.9, 0.95),
|
167 |
+
adamw_eps=1e-8,
|
168 |
+
debug=False,
|
169 |
+
):
|
170 |
+
defaults = dict(
|
171 |
+
lr=lr,
|
172 |
+
wd=adamw_wd,
|
173 |
+
momentum=momentum,
|
174 |
+
nesterov=nesterov,
|
175 |
+
ns_steps=ns_steps,
|
176 |
+
adamw_betas=adamw_betas,
|
177 |
+
adamw_eps=adamw_eps,
|
178 |
+
)
|
179 |
+
|
180 |
+
super().__init__(model.parameters(), defaults)
|
181 |
+
self.is_muon_func = is_muon_func
|
182 |
+
self.model = model
|
183 |
+
|
184 |
+
if not dist.is_initialized():
|
185 |
+
raise RuntimeError(
|
186 |
+
"Muon optimizer requires distributed training to be initialized."
|
187 |
+
)
|
188 |
+
|
189 |
+
self.rank = dist.get_rank()
|
190 |
+
|
191 |
+
self.comm_stream = torch.cuda.Stream()
|
192 |
+
self.compute_stream = torch.cuda.Stream()
|
193 |
+
self.debug = debug
|
194 |
+
|
195 |
+
def __setstate__(self, state):
|
196 |
+
# Sort parameters into those for which we will use Muon, and those for which we will not
|
197 |
+
super().__setstate__(state)
|
198 |
+
for name, p in self.model.named_parameters():
|
199 |
+
if self.is_muon_func(p, name):
|
200 |
+
# Use Muon for every parameter in muon_params which is >= 2D and doesn't look like an embedding or head layer
|
201 |
+
assert p.ndim == 2, p.ndim
|
202 |
+
self.state[p]["use_muon"] = True
|
203 |
+
self.state[p]["orig_shape"] = p.shape
|
204 |
+
else:
|
205 |
+
# Do not use Muon for parameters in adamw_params
|
206 |
+
self.state[p]["use_muon"] = False
|
207 |
+
|
208 |
+
def _calc_flops(self, G, steps):
|
209 |
+
assert len(G.shape) == 2
|
210 |
+
M, N = G.shape
|
211 |
+
if M > N:
|
212 |
+
M, N = N, M
|
213 |
+
|
214 |
+
return steps * ((M**3) * 2 + (M**2 * N) * 4 + M * N * 2 + M**2 * 3)
|
215 |
+
|
216 |
+
def adjust_lr_for_muon(self, lr, param_shape):
|
217 |
+
A, B = param_shape[:2]
|
218 |
+
# We adjust the learning rate and weight decay based on the size of the parameter matrix
|
219 |
+
# as describted in the paper
|
220 |
+
adjusted_ratio = 0.2 * math.sqrt(max(A, B))
|
221 |
+
adjusted_lr = lr * adjusted_ratio
|
222 |
+
return adjusted_lr
|
223 |
+
|
224 |
+
def init_state_and_assign_params(self, params, group):
|
225 |
+
param_to_state = {}
|
226 |
+
param_to_flops = {}
|
227 |
+
|
228 |
+
total_flops = 0
|
229 |
+
for p in params:
|
230 |
+
g = p.grad
|
231 |
+
if g is None:
|
232 |
+
continue
|
233 |
+
assert g.ndim == 2, "Muon only supports 2D parameters."
|
234 |
+
|
235 |
+
flops = self._calc_flops(g, group["ns_steps"])
|
236 |
+
param_to_flops[id(p)] = flops
|
237 |
+
total_flops += flops
|
238 |
+
|
239 |
+
if self.debug:
|
240 |
+
print(f"Total TFLOPs for Muon: {total_flops / 1e12:.2f} TFLOPs", flush=True)
|
241 |
+
|
242 |
+
ordered_params = sorted(
|
243 |
+
params, key=lambda p: param_to_flops[id(p)], reverse=True
|
244 |
+
)
|
245 |
+
|
246 |
+
round_robin = 0
|
247 |
+
mesh = None
|
248 |
+
for p in ordered_params:
|
249 |
+
if mesh is None:
|
250 |
+
mesh = p.device_mesh
|
251 |
+
if mesh.ndim != 1:
|
252 |
+
raise NotImplementedError(
|
253 |
+
"Muon requires a 1D mesh for distributed training yet."
|
254 |
+
)
|
255 |
+
elif mesh != p.device_mesh:
|
256 |
+
raise ValueError("All parameters must be on the same mesh.")
|
257 |
+
|
258 |
+
param_to_state[id(p)] = _muon_state()
|
259 |
+
param_to_state[id(p)].worker_rank = mesh.mesh[round_robin].item()
|
260 |
+
|
261 |
+
round_robin = (round_robin + 1) % mesh.mesh.numel()
|
262 |
+
|
263 |
+
return param_to_state, ordered_params
|
264 |
+
|
265 |
+
def base(self, params, group, lr, wd, momentum):
|
266 |
+
# generate weight updates in distributed fashion
|
267 |
+
for p in params:
|
268 |
+
g = p.grad
|
269 |
+
if g is None:
|
270 |
+
continue
|
271 |
+
if g.ndim > 2:
|
272 |
+
g = g.view(g.size(0), -1)
|
273 |
+
assert g is not None
|
274 |
+
|
275 |
+
# calc update
|
276 |
+
state = self.state[p]
|
277 |
+
if "momentum_buffer" not in state:
|
278 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
279 |
+
buf = state["momentum_buffer"]
|
280 |
+
buf.mul_(momentum).add_(g)
|
281 |
+
if group["nesterov"]:
|
282 |
+
g = g.add(buf, alpha=momentum)
|
283 |
+
else:
|
284 |
+
g = buf
|
285 |
+
|
286 |
+
u = _zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
|
287 |
+
|
288 |
+
# scale update
|
289 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
290 |
+
|
291 |
+
# apply weight decay
|
292 |
+
p.data.mul_(1 - lr * wd)
|
293 |
+
|
294 |
+
# apply update
|
295 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
296 |
+
|
297 |
+
def _update_g(self, p, g, group, momentum):
|
298 |
+
# calc update
|
299 |
+
state = self.state[p]
|
300 |
+
if "momentum_buffer" not in state:
|
301 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
302 |
+
buf = state["momentum_buffer"]
|
303 |
+
buf.mul_(momentum).add_(g)
|
304 |
+
if group["nesterov"]:
|
305 |
+
g = g.add(buf, alpha=momentum)
|
306 |
+
else:
|
307 |
+
g = buf
|
308 |
+
return g
|
309 |
+
|
310 |
+
def _update_p(self, p, u, lr, wd):
|
311 |
+
# scale update
|
312 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
313 |
+
# apply weight decay
|
314 |
+
p.data.mul_(1 - lr * wd)
|
315 |
+
# apply update
|
316 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
317 |
+
|
318 |
+
def parallel(self, params, group, lr, wd, momentum):
|
319 |
+
"""
|
320 |
+
Perform a parallel optimization step using Muon.
|
321 |
+
"""
|
322 |
+
|
323 |
+
for p in params:
|
324 |
+
g = p.grad
|
325 |
+
if g is None:
|
326 |
+
continue
|
327 |
+
if g.ndim > 2:
|
328 |
+
g = g.view(g.size(0), -1)
|
329 |
+
|
330 |
+
# Update g in the local rank
|
331 |
+
g = self._update_g(
|
332 |
+
p,
|
333 |
+
g,
|
334 |
+
group,
|
335 |
+
momentum=momentum,
|
336 |
+
)
|
337 |
+
p.grad = g
|
338 |
+
|
339 |
+
param_to_state, ordered_params = self.init_state_and_assign_params(
|
340 |
+
params, group
|
341 |
+
)
|
342 |
+
|
343 |
+
def enqueue_gathers(start_idx, chunk_size):
|
344 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
345 |
+
state = param_to_state[id(p)]
|
346 |
+
_gather(p, state, self.rank, self.comm_stream)
|
347 |
+
|
348 |
+
def enqueue_computes(start_idx, chunk_size):
|
349 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
350 |
+
state = param_to_state[id(p)]
|
351 |
+
_compute_u(state, group["ns_steps"], self.rank, self.compute_stream)
|
352 |
+
|
353 |
+
def enqueue_scatters(start_idx, chunk_size):
|
354 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
355 |
+
state = param_to_state[id(p)]
|
356 |
+
_scatter(p, state, self.rank, self.comm_stream)
|
357 |
+
|
358 |
+
chunk_size = params[0].device_mesh.mesh.numel()
|
359 |
+
|
360 |
+
# Wait grad update
|
361 |
+
self.comm_stream.wait_stream(torch.cuda.current_stream())
|
362 |
+
|
363 |
+
enqueue_gathers(0, chunk_size)
|
364 |
+
for i in range(0, len(params) + chunk_size - 1, chunk_size):
|
365 |
+
enqueue_computes(i, chunk_size)
|
366 |
+
enqueue_gathers(i + chunk_size, chunk_size)
|
367 |
+
enqueue_scatters(i, chunk_size)
|
368 |
+
|
369 |
+
torch.cuda.current_stream().wait_stream(self.comm_stream)
|
370 |
+
|
371 |
+
for p in params:
|
372 |
+
g = p.grad
|
373 |
+
if g is None:
|
374 |
+
continue
|
375 |
+
|
376 |
+
# Update p with sharded u
|
377 |
+
state = param_to_state[id(p)]
|
378 |
+
self._update_p(
|
379 |
+
p,
|
380 |
+
state.scattered_u,
|
381 |
+
lr=lr,
|
382 |
+
wd=wd,
|
383 |
+
)
|
384 |
+
|
385 |
+
def step(self, closure=None):
|
386 |
+
"""Perform a single optimization step.
|
387 |
+
|
388 |
+
Args:
|
389 |
+
closure (Callable, optional): A closure that reevaluates the model
|
390 |
+
and returns the loss.
|
391 |
+
"""
|
392 |
+
loss = None
|
393 |
+
if closure is not None:
|
394 |
+
with torch.enable_grad():
|
395 |
+
loss = closure()
|
396 |
+
|
397 |
+
for group in self.param_groups:
|
398 |
+
############################
|
399 |
+
# Muon #
|
400 |
+
############################
|
401 |
+
|
402 |
+
params = [p for p in group["params"] if self.state[p]["use_muon"]]
|
403 |
+
lr = group["lr"]
|
404 |
+
wd = group["wd"]
|
405 |
+
momentum = group["momentum"]
|
406 |
+
|
407 |
+
if isinstance(params[0].data, DTensor):
|
408 |
+
self.parallel(
|
409 |
+
params,
|
410 |
+
group,
|
411 |
+
lr=lr,
|
412 |
+
wd=wd,
|
413 |
+
momentum=momentum,
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
self.base(
|
417 |
+
params,
|
418 |
+
group,
|
419 |
+
lr=lr,
|
420 |
+
wd=wd,
|
421 |
+
momentum=momentum,
|
422 |
+
)
|
423 |
+
|
424 |
+
############################
|
425 |
+
# AdamW backup #
|
426 |
+
############################
|
427 |
+
|
428 |
+
params = [p for p in group["params"] if not self.state[p]["use_muon"]]
|
429 |
+
lr = group["lr"]
|
430 |
+
beta1, beta2 = group["adamw_betas"]
|
431 |
+
eps = group["adamw_eps"]
|
432 |
+
weight_decay = group["wd"]
|
433 |
+
|
434 |
+
for p in params:
|
435 |
+
g = p.grad
|
436 |
+
if g is None:
|
437 |
+
continue
|
438 |
+
state = self.state[p]
|
439 |
+
if "step" not in state:
|
440 |
+
state["step"] = 0
|
441 |
+
state["moment1"] = torch.zeros_like(g)
|
442 |
+
state["moment2"] = torch.zeros_like(g)
|
443 |
+
state["step"] += 1
|
444 |
+
step = state["step"]
|
445 |
+
buf1 = state["moment1"]
|
446 |
+
buf2 = state["moment2"]
|
447 |
+
buf1.lerp_(g, 1 - beta1)
|
448 |
+
buf2.lerp_(g.square(), 1 - beta2)
|
449 |
+
|
450 |
+
g = buf1 / (eps + buf2.sqrt())
|
451 |
+
|
452 |
+
bias_correction1 = 1 - beta1**step
|
453 |
+
bias_correction2 = 1 - beta2**step
|
454 |
+
scale = bias_correction1 / bias_correction2**0.5
|
455 |
+
p.data.mul_(1 - lr * weight_decay)
|
456 |
+
p.data.add_(g, alpha=-lr / scale)
|
457 |
+
|
458 |
+
return loss
|
build/torch26-cxx98-cu126-x86_64-linux/optimizer/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .muon import Muon
|
2 |
+
|
3 |
+
__all__ = [
|
4 |
+
"Muon",
|
5 |
+
]
|
build/torch26-cxx98-cu126-x86_64-linux/optimizer/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _optimizer_b4b3752_dirty
|
3 |
+
ops = torch.ops._optimizer_b4b3752_dirty
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_optimizer_b4b3752_dirty::{op_name}"
|
build/torch26-cxx98-cu126-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3c3282a321487a6faa532afe43bc1298731983c50e2a1acdff5480ff6e4df34e
|
3 |
+
size 1824192
|
build/torch26-cxx98-cu126-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a5b49ed642e1c320da3932377033ad90031124f4ec24b2d1c95fd976ff28346c
|
3 |
+
size 1824184
|
build/torch26-cxx98-cu126-x86_64-linux/optimizer/muon.py
ADDED
@@ -0,0 +1,458 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.distributed as dist
|
6 |
+
from torch.distributed._tensor import DTensor
|
7 |
+
|
8 |
+
|
9 |
+
# TODO leave original url and consider LICENSE
|
10 |
+
# This code snippet is a modified version adapted from the following GitHub repository:
|
11 |
+
# https://github.com/KellerJordan/Muon/blob/master/muon.py
|
12 |
+
def _zeropower_via_newtonschulz5(G, steps):
|
13 |
+
"""
|
14 |
+
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
|
15 |
+
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
|
16 |
+
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
|
17 |
+
zero even beyond the point where the iteration no longer converges all the way to one everywhere
|
18 |
+
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
|
19 |
+
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
|
20 |
+
performance at all relative to UV^T, where USV^T = G is the SVD.
|
21 |
+
"""
|
22 |
+
assert len(G.shape) == 2
|
23 |
+
a, b, c = (3.4445, -4.7750, 2.0315)
|
24 |
+
X = G # no manual typecast
|
25 |
+
if G.size(0) > G.size(1):
|
26 |
+
X = X.T
|
27 |
+
# Ensure spectral norm is at most 1
|
28 |
+
X = X / (X.norm() + 1e-7)
|
29 |
+
X = X.bfloat16()
|
30 |
+
# Perform the NS iterations
|
31 |
+
for _ in range(steps):
|
32 |
+
A = X @ X.T
|
33 |
+
# B = (
|
34 |
+
# b * A + c * A @ A
|
35 |
+
# ) # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
|
36 |
+
B = torch.addmm(A, A, A, alpha=c, beta=b)
|
37 |
+
# X = a * X + B @ X
|
38 |
+
X = torch.addmm(X, B, X, alpha=1.0, beta=a)
|
39 |
+
|
40 |
+
if G.size(0) > G.size(1):
|
41 |
+
X = X.T
|
42 |
+
return X.to(G.dtype)
|
43 |
+
|
44 |
+
|
45 |
+
@dataclass
|
46 |
+
class _muon_state:
|
47 |
+
# TODO: use Optional
|
48 |
+
worker_rank: int | None = None
|
49 |
+
gathered_grad: torch.Tensor | None = None
|
50 |
+
computed_u: torch.Tensor | None = None
|
51 |
+
scattered_u: torch.Tensor | None = None
|
52 |
+
gather_event: torch.cuda.Event | None = None
|
53 |
+
compute_event: torch.cuda.Event | None = None
|
54 |
+
|
55 |
+
|
56 |
+
def _gather(p, state, rank, comm_stream):
|
57 |
+
g = p.grad
|
58 |
+
mesh = g.device_mesh
|
59 |
+
|
60 |
+
if rank == state.worker_rank:
|
61 |
+
gather_list = [torch.empty_like(g.to_local()) for _ in range(mesh.mesh.numel())]
|
62 |
+
else:
|
63 |
+
gather_list = None
|
64 |
+
|
65 |
+
with torch.cuda.stream(comm_stream):
|
66 |
+
torch.distributed.gather(
|
67 |
+
g.to_local(),
|
68 |
+
dst=state.worker_rank,
|
69 |
+
gather_list=gather_list,
|
70 |
+
group=mesh.get_group(),
|
71 |
+
)
|
72 |
+
if rank == state.worker_rank:
|
73 |
+
# TODO: Consider ,,,
|
74 |
+
if state.gathered_grad is not None:
|
75 |
+
raise RuntimeError(
|
76 |
+
"Gather event already exists, which should not happen."
|
77 |
+
)
|
78 |
+
state.gathered_grad = torch.cat(gather_list, dim=0)
|
79 |
+
state.gather_event = torch.cuda.Event()
|
80 |
+
state.gather_event.record()
|
81 |
+
else:
|
82 |
+
state.gathered_grad = None
|
83 |
+
state.gather_event = None
|
84 |
+
|
85 |
+
|
86 |
+
def _compute_u(state, steps, rank, compute_stream):
|
87 |
+
with torch.cuda.stream(compute_stream):
|
88 |
+
if rank == state.worker_rank:
|
89 |
+
if state.gather_event is None:
|
90 |
+
raise RuntimeError("Gather event must be set before compute.")
|
91 |
+
compute_stream.wait_event(state.gather_event)
|
92 |
+
u = _zeropower_via_newtonschulz5(state.gathered_grad, steps)
|
93 |
+
state.computed_u = u
|
94 |
+
state.compute_event = torch.cuda.Event()
|
95 |
+
state.compute_event.record()
|
96 |
+
else:
|
97 |
+
state.computed_u = None
|
98 |
+
state.compute_event = None
|
99 |
+
|
100 |
+
|
101 |
+
def _scatter(p, state, rank, comm_stream):
|
102 |
+
u = state.computed_u
|
103 |
+
mesh = p.device_mesh
|
104 |
+
|
105 |
+
with torch.cuda.stream(comm_stream):
|
106 |
+
if rank == state.worker_rank:
|
107 |
+
if state.compute_event is None:
|
108 |
+
raise RuntimeError("Compute event must be set before scatter.")
|
109 |
+
comm_stream.wait_event(state.compute_event)
|
110 |
+
scatter_list = list(torch.split(u, p.size(0) // mesh.mesh.numel(), dim=0))
|
111 |
+
else:
|
112 |
+
scatter_list = None
|
113 |
+
|
114 |
+
u = torch.empty_like(p.to_local())
|
115 |
+
torch.distributed.scatter(
|
116 |
+
u,
|
117 |
+
scatter_list=scatter_list,
|
118 |
+
src=state.worker_rank,
|
119 |
+
group=mesh.get_group(),
|
120 |
+
)
|
121 |
+
u = DTensor.from_local(
|
122 |
+
u,
|
123 |
+
placements=p.placements,
|
124 |
+
device_mesh=mesh,
|
125 |
+
)
|
126 |
+
|
127 |
+
state.scattered_u = u
|
128 |
+
|
129 |
+
|
130 |
+
class Muon(torch.optim.Optimizer):
|
131 |
+
"""
|
132 |
+
Muon - MomentUm Orthogonalized by Newton-schulz
|
133 |
+
|
134 |
+
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
|
135 |
+
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
|
136 |
+
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
|
137 |
+
the advantage that it can be stably run in bfloat16 on the GPU.
|
138 |
+
|
139 |
+
Some warnings:
|
140 |
+
- We believe this optimizer is unlikely to work well for training with small batch size.
|
141 |
+
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
|
142 |
+
|
143 |
+
Arguments:
|
144 |
+
muon_params: The parameters to be optimized by Muon.
|
145 |
+
lr: The learning rate. The updates will have spectral norm of `lr`. (0.02 is a good default)
|
146 |
+
momentum: The momentum used by the internal SGD. (0.95 is a good default)
|
147 |
+
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
|
148 |
+
ns_steps: The number of Newton-Schulz iterations to run. (6 is probably always enough)
|
149 |
+
adamw_params: The parameters to be optimized by AdamW. Any parameters in `muon_params` which are
|
150 |
+
{0, 1}-D or are detected as being the embed or lm_head will be optimized by AdamW as well.
|
151 |
+
adamw_lr: The learning rate for the internal AdamW.
|
152 |
+
adamw_betas: The betas for the internal AdamW.
|
153 |
+
adamw_eps: The epsilon for the internal AdamW.
|
154 |
+
adamw_wd: The weight decay for the internal AdamW.
|
155 |
+
"""
|
156 |
+
|
157 |
+
def __init__(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
is_muon_func,
|
161 |
+
lr=1e-3,
|
162 |
+
momentum=0.95,
|
163 |
+
nesterov=True,
|
164 |
+
ns_steps=5,
|
165 |
+
adamw_wd=0.1,
|
166 |
+
adamw_betas=(0.9, 0.95),
|
167 |
+
adamw_eps=1e-8,
|
168 |
+
debug=False,
|
169 |
+
):
|
170 |
+
defaults = dict(
|
171 |
+
lr=lr,
|
172 |
+
wd=adamw_wd,
|
173 |
+
momentum=momentum,
|
174 |
+
nesterov=nesterov,
|
175 |
+
ns_steps=ns_steps,
|
176 |
+
adamw_betas=adamw_betas,
|
177 |
+
adamw_eps=adamw_eps,
|
178 |
+
)
|
179 |
+
|
180 |
+
super().__init__(model.parameters(), defaults)
|
181 |
+
self.is_muon_func = is_muon_func
|
182 |
+
self.model = model
|
183 |
+
|
184 |
+
if not dist.is_initialized():
|
185 |
+
raise RuntimeError(
|
186 |
+
"Muon optimizer requires distributed training to be initialized."
|
187 |
+
)
|
188 |
+
|
189 |
+
self.rank = dist.get_rank()
|
190 |
+
|
191 |
+
self.comm_stream = torch.cuda.Stream()
|
192 |
+
self.compute_stream = torch.cuda.Stream()
|
193 |
+
self.debug = debug
|
194 |
+
|
195 |
+
def __setstate__(self, state):
|
196 |
+
# Sort parameters into those for which we will use Muon, and those for which we will not
|
197 |
+
super().__setstate__(state)
|
198 |
+
for name, p in self.model.named_parameters():
|
199 |
+
if self.is_muon_func(p, name):
|
200 |
+
# Use Muon for every parameter in muon_params which is >= 2D and doesn't look like an embedding or head layer
|
201 |
+
assert p.ndim == 2, p.ndim
|
202 |
+
self.state[p]["use_muon"] = True
|
203 |
+
self.state[p]["orig_shape"] = p.shape
|
204 |
+
else:
|
205 |
+
# Do not use Muon for parameters in adamw_params
|
206 |
+
self.state[p]["use_muon"] = False
|
207 |
+
|
208 |
+
def _calc_flops(self, G, steps):
|
209 |
+
assert len(G.shape) == 2
|
210 |
+
M, N = G.shape
|
211 |
+
if M > N:
|
212 |
+
M, N = N, M
|
213 |
+
|
214 |
+
return steps * ((M**3) * 2 + (M**2 * N) * 4 + M * N * 2 + M**2 * 3)
|
215 |
+
|
216 |
+
def adjust_lr_for_muon(self, lr, param_shape):
|
217 |
+
A, B = param_shape[:2]
|
218 |
+
# We adjust the learning rate and weight decay based on the size of the parameter matrix
|
219 |
+
# as describted in the paper
|
220 |
+
adjusted_ratio = 0.2 * math.sqrt(max(A, B))
|
221 |
+
adjusted_lr = lr * adjusted_ratio
|
222 |
+
return adjusted_lr
|
223 |
+
|
224 |
+
def init_state_and_assign_params(self, params, group):
|
225 |
+
param_to_state = {}
|
226 |
+
param_to_flops = {}
|
227 |
+
|
228 |
+
total_flops = 0
|
229 |
+
for p in params:
|
230 |
+
g = p.grad
|
231 |
+
if g is None:
|
232 |
+
continue
|
233 |
+
assert g.ndim == 2, "Muon only supports 2D parameters."
|
234 |
+
|
235 |
+
flops = self._calc_flops(g, group["ns_steps"])
|
236 |
+
param_to_flops[id(p)] = flops
|
237 |
+
total_flops += flops
|
238 |
+
|
239 |
+
if self.debug:
|
240 |
+
print(f"Total TFLOPs for Muon: {total_flops / 1e12:.2f} TFLOPs", flush=True)
|
241 |
+
|
242 |
+
ordered_params = sorted(
|
243 |
+
params, key=lambda p: param_to_flops[id(p)], reverse=True
|
244 |
+
)
|
245 |
+
|
246 |
+
round_robin = 0
|
247 |
+
mesh = None
|
248 |
+
for p in ordered_params:
|
249 |
+
if mesh is None:
|
250 |
+
mesh = p.device_mesh
|
251 |
+
if mesh.ndim != 1:
|
252 |
+
raise NotImplementedError(
|
253 |
+
"Muon requires a 1D mesh for distributed training yet."
|
254 |
+
)
|
255 |
+
elif mesh != p.device_mesh:
|
256 |
+
raise ValueError("All parameters must be on the same mesh.")
|
257 |
+
|
258 |
+
param_to_state[id(p)] = _muon_state()
|
259 |
+
param_to_state[id(p)].worker_rank = mesh.mesh[round_robin].item()
|
260 |
+
|
261 |
+
round_robin = (round_robin + 1) % mesh.mesh.numel()
|
262 |
+
|
263 |
+
return param_to_state, ordered_params
|
264 |
+
|
265 |
+
def base(self, params, group, lr, wd, momentum):
|
266 |
+
# generate weight updates in distributed fashion
|
267 |
+
for p in params:
|
268 |
+
g = p.grad
|
269 |
+
if g is None:
|
270 |
+
continue
|
271 |
+
if g.ndim > 2:
|
272 |
+
g = g.view(g.size(0), -1)
|
273 |
+
assert g is not None
|
274 |
+
|
275 |
+
# calc update
|
276 |
+
state = self.state[p]
|
277 |
+
if "momentum_buffer" not in state:
|
278 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
279 |
+
buf = state["momentum_buffer"]
|
280 |
+
buf.mul_(momentum).add_(g)
|
281 |
+
if group["nesterov"]:
|
282 |
+
g = g.add(buf, alpha=momentum)
|
283 |
+
else:
|
284 |
+
g = buf
|
285 |
+
|
286 |
+
u = _zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
|
287 |
+
|
288 |
+
# scale update
|
289 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
290 |
+
|
291 |
+
# apply weight decay
|
292 |
+
p.data.mul_(1 - lr * wd)
|
293 |
+
|
294 |
+
# apply update
|
295 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
296 |
+
|
297 |
+
def _update_g(self, p, g, group, momentum):
|
298 |
+
# calc update
|
299 |
+
state = self.state[p]
|
300 |
+
if "momentum_buffer" not in state:
|
301 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
302 |
+
buf = state["momentum_buffer"]
|
303 |
+
buf.mul_(momentum).add_(g)
|
304 |
+
if group["nesterov"]:
|
305 |
+
g = g.add(buf, alpha=momentum)
|
306 |
+
else:
|
307 |
+
g = buf
|
308 |
+
return g
|
309 |
+
|
310 |
+
def _update_p(self, p, u, lr, wd):
|
311 |
+
# scale update
|
312 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
313 |
+
# apply weight decay
|
314 |
+
p.data.mul_(1 - lr * wd)
|
315 |
+
# apply update
|
316 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
317 |
+
|
318 |
+
def parallel(self, params, group, lr, wd, momentum):
|
319 |
+
"""
|
320 |
+
Perform a parallel optimization step using Muon.
|
321 |
+
"""
|
322 |
+
|
323 |
+
for p in params:
|
324 |
+
g = p.grad
|
325 |
+
if g is None:
|
326 |
+
continue
|
327 |
+
if g.ndim > 2:
|
328 |
+
g = g.view(g.size(0), -1)
|
329 |
+
|
330 |
+
# Update g in the local rank
|
331 |
+
g = self._update_g(
|
332 |
+
p,
|
333 |
+
g,
|
334 |
+
group,
|
335 |
+
momentum=momentum,
|
336 |
+
)
|
337 |
+
p.grad = g
|
338 |
+
|
339 |
+
param_to_state, ordered_params = self.init_state_and_assign_params(
|
340 |
+
params, group
|
341 |
+
)
|
342 |
+
|
343 |
+
def enqueue_gathers(start_idx, chunk_size):
|
344 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
345 |
+
state = param_to_state[id(p)]
|
346 |
+
_gather(p, state, self.rank, self.comm_stream)
|
347 |
+
|
348 |
+
def enqueue_computes(start_idx, chunk_size):
|
349 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
350 |
+
state = param_to_state[id(p)]
|
351 |
+
_compute_u(state, group["ns_steps"], self.rank, self.compute_stream)
|
352 |
+
|
353 |
+
def enqueue_scatters(start_idx, chunk_size):
|
354 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
355 |
+
state = param_to_state[id(p)]
|
356 |
+
_scatter(p, state, self.rank, self.comm_stream)
|
357 |
+
|
358 |
+
chunk_size = params[0].device_mesh.mesh.numel()
|
359 |
+
|
360 |
+
# Wait grad update
|
361 |
+
self.comm_stream.wait_stream(torch.cuda.current_stream())
|
362 |
+
|
363 |
+
enqueue_gathers(0, chunk_size)
|
364 |
+
for i in range(0, len(params) + chunk_size - 1, chunk_size):
|
365 |
+
enqueue_computes(i, chunk_size)
|
366 |
+
enqueue_gathers(i + chunk_size, chunk_size)
|
367 |
+
enqueue_scatters(i, chunk_size)
|
368 |
+
|
369 |
+
torch.cuda.current_stream().wait_stream(self.comm_stream)
|
370 |
+
|
371 |
+
for p in params:
|
372 |
+
g = p.grad
|
373 |
+
if g is None:
|
374 |
+
continue
|
375 |
+
|
376 |
+
# Update p with sharded u
|
377 |
+
state = param_to_state[id(p)]
|
378 |
+
self._update_p(
|
379 |
+
p,
|
380 |
+
state.scattered_u,
|
381 |
+
lr=lr,
|
382 |
+
wd=wd,
|
383 |
+
)
|
384 |
+
|
385 |
+
def step(self, closure=None):
|
386 |
+
"""Perform a single optimization step.
|
387 |
+
|
388 |
+
Args:
|
389 |
+
closure (Callable, optional): A closure that reevaluates the model
|
390 |
+
and returns the loss.
|
391 |
+
"""
|
392 |
+
loss = None
|
393 |
+
if closure is not None:
|
394 |
+
with torch.enable_grad():
|
395 |
+
loss = closure()
|
396 |
+
|
397 |
+
for group in self.param_groups:
|
398 |
+
############################
|
399 |
+
# Muon #
|
400 |
+
############################
|
401 |
+
|
402 |
+
params = [p for p in group["params"] if self.state[p]["use_muon"]]
|
403 |
+
lr = group["lr"]
|
404 |
+
wd = group["wd"]
|
405 |
+
momentum = group["momentum"]
|
406 |
+
|
407 |
+
if isinstance(params[0].data, DTensor):
|
408 |
+
self.parallel(
|
409 |
+
params,
|
410 |
+
group,
|
411 |
+
lr=lr,
|
412 |
+
wd=wd,
|
413 |
+
momentum=momentum,
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
self.base(
|
417 |
+
params,
|
418 |
+
group,
|
419 |
+
lr=lr,
|
420 |
+
wd=wd,
|
421 |
+
momentum=momentum,
|
422 |
+
)
|
423 |
+
|
424 |
+
############################
|
425 |
+
# AdamW backup #
|
426 |
+
############################
|
427 |
+
|
428 |
+
params = [p for p in group["params"] if not self.state[p]["use_muon"]]
|
429 |
+
lr = group["lr"]
|
430 |
+
beta1, beta2 = group["adamw_betas"]
|
431 |
+
eps = group["adamw_eps"]
|
432 |
+
weight_decay = group["wd"]
|
433 |
+
|
434 |
+
for p in params:
|
435 |
+
g = p.grad
|
436 |
+
if g is None:
|
437 |
+
continue
|
438 |
+
state = self.state[p]
|
439 |
+
if "step" not in state:
|
440 |
+
state["step"] = 0
|
441 |
+
state["moment1"] = torch.zeros_like(g)
|
442 |
+
state["moment2"] = torch.zeros_like(g)
|
443 |
+
state["step"] += 1
|
444 |
+
step = state["step"]
|
445 |
+
buf1 = state["moment1"]
|
446 |
+
buf2 = state["moment2"]
|
447 |
+
buf1.lerp_(g, 1 - beta1)
|
448 |
+
buf2.lerp_(g.square(), 1 - beta2)
|
449 |
+
|
450 |
+
g = buf1 / (eps + buf2.sqrt())
|
451 |
+
|
452 |
+
bias_correction1 = 1 - beta1**step
|
453 |
+
bias_correction2 = 1 - beta2**step
|
454 |
+
scale = bias_correction1 / bias_correction2**0.5
|
455 |
+
p.data.mul_(1 - lr * weight_decay)
|
456 |
+
p.data.add_(g, alpha=-lr / scale)
|
457 |
+
|
458 |
+
return loss
|
build/torch27-cxx11-cu118-x86_64-linux/optimizer/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .muon import Muon
|
2 |
+
|
3 |
+
__all__ = [
|
4 |
+
"Muon",
|
5 |
+
]
|
build/torch27-cxx11-cu118-x86_64-linux/optimizer/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _optimizer_b4b3752_dirty
|
3 |
+
ops = torch.ops._optimizer_b4b3752_dirty
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_optimizer_b4b3752_dirty::{op_name}"
|
build/torch27-cxx11-cu118-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:de82486a39ded94bfe7eeaa862459944a93e284fd0d919329979bb67db3c367f
|
3 |
+
size 1787376
|
build/torch27-cxx11-cu118-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8ac9027c4a93801e9f19f1e9e94a9ed33b27e92c72797053c3de55e2a6fbb41d
|
3 |
+
size 1787368
|
build/torch27-cxx11-cu118-x86_64-linux/optimizer/muon.py
ADDED
@@ -0,0 +1,458 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.distributed as dist
|
6 |
+
from torch.distributed._tensor import DTensor
|
7 |
+
|
8 |
+
|
9 |
+
# TODO leave original url and consider LICENSE
|
10 |
+
# This code snippet is a modified version adapted from the following GitHub repository:
|
11 |
+
# https://github.com/KellerJordan/Muon/blob/master/muon.py
|
12 |
+
def _zeropower_via_newtonschulz5(G, steps):
|
13 |
+
"""
|
14 |
+
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
|
15 |
+
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
|
16 |
+
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
|
17 |
+
zero even beyond the point where the iteration no longer converges all the way to one everywhere
|
18 |
+
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
|
19 |
+
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
|
20 |
+
performance at all relative to UV^T, where USV^T = G is the SVD.
|
21 |
+
"""
|
22 |
+
assert len(G.shape) == 2
|
23 |
+
a, b, c = (3.4445, -4.7750, 2.0315)
|
24 |
+
X = G # no manual typecast
|
25 |
+
if G.size(0) > G.size(1):
|
26 |
+
X = X.T
|
27 |
+
# Ensure spectral norm is at most 1
|
28 |
+
X = X / (X.norm() + 1e-7)
|
29 |
+
X = X.bfloat16()
|
30 |
+
# Perform the NS iterations
|
31 |
+
for _ in range(steps):
|
32 |
+
A = X @ X.T
|
33 |
+
# B = (
|
34 |
+
# b * A + c * A @ A
|
35 |
+
# ) # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
|
36 |
+
B = torch.addmm(A, A, A, alpha=c, beta=b)
|
37 |
+
# X = a * X + B @ X
|
38 |
+
X = torch.addmm(X, B, X, alpha=1.0, beta=a)
|
39 |
+
|
40 |
+
if G.size(0) > G.size(1):
|
41 |
+
X = X.T
|
42 |
+
return X.to(G.dtype)
|
43 |
+
|
44 |
+
|
45 |
+
@dataclass
|
46 |
+
class _muon_state:
|
47 |
+
# TODO: use Optional
|
48 |
+
worker_rank: int | None = None
|
49 |
+
gathered_grad: torch.Tensor | None = None
|
50 |
+
computed_u: torch.Tensor | None = None
|
51 |
+
scattered_u: torch.Tensor | None = None
|
52 |
+
gather_event: torch.cuda.Event | None = None
|
53 |
+
compute_event: torch.cuda.Event | None = None
|
54 |
+
|
55 |
+
|
56 |
+
def _gather(p, state, rank, comm_stream):
|
57 |
+
g = p.grad
|
58 |
+
mesh = g.device_mesh
|
59 |
+
|
60 |
+
if rank == state.worker_rank:
|
61 |
+
gather_list = [torch.empty_like(g.to_local()) for _ in range(mesh.mesh.numel())]
|
62 |
+
else:
|
63 |
+
gather_list = None
|
64 |
+
|
65 |
+
with torch.cuda.stream(comm_stream):
|
66 |
+
torch.distributed.gather(
|
67 |
+
g.to_local(),
|
68 |
+
dst=state.worker_rank,
|
69 |
+
gather_list=gather_list,
|
70 |
+
group=mesh.get_group(),
|
71 |
+
)
|
72 |
+
if rank == state.worker_rank:
|
73 |
+
# TODO: Consider ,,,
|
74 |
+
if state.gathered_grad is not None:
|
75 |
+
raise RuntimeError(
|
76 |
+
"Gather event already exists, which should not happen."
|
77 |
+
)
|
78 |
+
state.gathered_grad = torch.cat(gather_list, dim=0)
|
79 |
+
state.gather_event = torch.cuda.Event()
|
80 |
+
state.gather_event.record()
|
81 |
+
else:
|
82 |
+
state.gathered_grad = None
|
83 |
+
state.gather_event = None
|
84 |
+
|
85 |
+
|
86 |
+
def _compute_u(state, steps, rank, compute_stream):
|
87 |
+
with torch.cuda.stream(compute_stream):
|
88 |
+
if rank == state.worker_rank:
|
89 |
+
if state.gather_event is None:
|
90 |
+
raise RuntimeError("Gather event must be set before compute.")
|
91 |
+
compute_stream.wait_event(state.gather_event)
|
92 |
+
u = _zeropower_via_newtonschulz5(state.gathered_grad, steps)
|
93 |
+
state.computed_u = u
|
94 |
+
state.compute_event = torch.cuda.Event()
|
95 |
+
state.compute_event.record()
|
96 |
+
else:
|
97 |
+
state.computed_u = None
|
98 |
+
state.compute_event = None
|
99 |
+
|
100 |
+
|
101 |
+
def _scatter(p, state, rank, comm_stream):
|
102 |
+
u = state.computed_u
|
103 |
+
mesh = p.device_mesh
|
104 |
+
|
105 |
+
with torch.cuda.stream(comm_stream):
|
106 |
+
if rank == state.worker_rank:
|
107 |
+
if state.compute_event is None:
|
108 |
+
raise RuntimeError("Compute event must be set before scatter.")
|
109 |
+
comm_stream.wait_event(state.compute_event)
|
110 |
+
scatter_list = list(torch.split(u, p.size(0) // mesh.mesh.numel(), dim=0))
|
111 |
+
else:
|
112 |
+
scatter_list = None
|
113 |
+
|
114 |
+
u = torch.empty_like(p.to_local())
|
115 |
+
torch.distributed.scatter(
|
116 |
+
u,
|
117 |
+
scatter_list=scatter_list,
|
118 |
+
src=state.worker_rank,
|
119 |
+
group=mesh.get_group(),
|
120 |
+
)
|
121 |
+
u = DTensor.from_local(
|
122 |
+
u,
|
123 |
+
placements=p.placements,
|
124 |
+
device_mesh=mesh,
|
125 |
+
)
|
126 |
+
|
127 |
+
state.scattered_u = u
|
128 |
+
|
129 |
+
|
130 |
+
class Muon(torch.optim.Optimizer):
|
131 |
+
"""
|
132 |
+
Muon - MomentUm Orthogonalized by Newton-schulz
|
133 |
+
|
134 |
+
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
|
135 |
+
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
|
136 |
+
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
|
137 |
+
the advantage that it can be stably run in bfloat16 on the GPU.
|
138 |
+
|
139 |
+
Some warnings:
|
140 |
+
- We believe this optimizer is unlikely to work well for training with small batch size.
|
141 |
+
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
|
142 |
+
|
143 |
+
Arguments:
|
144 |
+
muon_params: The parameters to be optimized by Muon.
|
145 |
+
lr: The learning rate. The updates will have spectral norm of `lr`. (0.02 is a good default)
|
146 |
+
momentum: The momentum used by the internal SGD. (0.95 is a good default)
|
147 |
+
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
|
148 |
+
ns_steps: The number of Newton-Schulz iterations to run. (6 is probably always enough)
|
149 |
+
adamw_params: The parameters to be optimized by AdamW. Any parameters in `muon_params` which are
|
150 |
+
{0, 1}-D or are detected as being the embed or lm_head will be optimized by AdamW as well.
|
151 |
+
adamw_lr: The learning rate for the internal AdamW.
|
152 |
+
adamw_betas: The betas for the internal AdamW.
|
153 |
+
adamw_eps: The epsilon for the internal AdamW.
|
154 |
+
adamw_wd: The weight decay for the internal AdamW.
|
155 |
+
"""
|
156 |
+
|
157 |
+
def __init__(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
is_muon_func,
|
161 |
+
lr=1e-3,
|
162 |
+
momentum=0.95,
|
163 |
+
nesterov=True,
|
164 |
+
ns_steps=5,
|
165 |
+
adamw_wd=0.1,
|
166 |
+
adamw_betas=(0.9, 0.95),
|
167 |
+
adamw_eps=1e-8,
|
168 |
+
debug=False,
|
169 |
+
):
|
170 |
+
defaults = dict(
|
171 |
+
lr=lr,
|
172 |
+
wd=adamw_wd,
|
173 |
+
momentum=momentum,
|
174 |
+
nesterov=nesterov,
|
175 |
+
ns_steps=ns_steps,
|
176 |
+
adamw_betas=adamw_betas,
|
177 |
+
adamw_eps=adamw_eps,
|
178 |
+
)
|
179 |
+
|
180 |
+
super().__init__(model.parameters(), defaults)
|
181 |
+
self.is_muon_func = is_muon_func
|
182 |
+
self.model = model
|
183 |
+
|
184 |
+
if not dist.is_initialized():
|
185 |
+
raise RuntimeError(
|
186 |
+
"Muon optimizer requires distributed training to be initialized."
|
187 |
+
)
|
188 |
+
|
189 |
+
self.rank = dist.get_rank()
|
190 |
+
|
191 |
+
self.comm_stream = torch.cuda.Stream()
|
192 |
+
self.compute_stream = torch.cuda.Stream()
|
193 |
+
self.debug = debug
|
194 |
+
|
195 |
+
def __setstate__(self, state):
|
196 |
+
# Sort parameters into those for which we will use Muon, and those for which we will not
|
197 |
+
super().__setstate__(state)
|
198 |
+
for name, p in self.model.named_parameters():
|
199 |
+
if self.is_muon_func(p, name):
|
200 |
+
# Use Muon for every parameter in muon_params which is >= 2D and doesn't look like an embedding or head layer
|
201 |
+
assert p.ndim == 2, p.ndim
|
202 |
+
self.state[p]["use_muon"] = True
|
203 |
+
self.state[p]["orig_shape"] = p.shape
|
204 |
+
else:
|
205 |
+
# Do not use Muon for parameters in adamw_params
|
206 |
+
self.state[p]["use_muon"] = False
|
207 |
+
|
208 |
+
def _calc_flops(self, G, steps):
|
209 |
+
assert len(G.shape) == 2
|
210 |
+
M, N = G.shape
|
211 |
+
if M > N:
|
212 |
+
M, N = N, M
|
213 |
+
|
214 |
+
return steps * ((M**3) * 2 + (M**2 * N) * 4 + M * N * 2 + M**2 * 3)
|
215 |
+
|
216 |
+
def adjust_lr_for_muon(self, lr, param_shape):
|
217 |
+
A, B = param_shape[:2]
|
218 |
+
# We adjust the learning rate and weight decay based on the size of the parameter matrix
|
219 |
+
# as describted in the paper
|
220 |
+
adjusted_ratio = 0.2 * math.sqrt(max(A, B))
|
221 |
+
adjusted_lr = lr * adjusted_ratio
|
222 |
+
return adjusted_lr
|
223 |
+
|
224 |
+
def init_state_and_assign_params(self, params, group):
|
225 |
+
param_to_state = {}
|
226 |
+
param_to_flops = {}
|
227 |
+
|
228 |
+
total_flops = 0
|
229 |
+
for p in params:
|
230 |
+
g = p.grad
|
231 |
+
if g is None:
|
232 |
+
continue
|
233 |
+
assert g.ndim == 2, "Muon only supports 2D parameters."
|
234 |
+
|
235 |
+
flops = self._calc_flops(g, group["ns_steps"])
|
236 |
+
param_to_flops[id(p)] = flops
|
237 |
+
total_flops += flops
|
238 |
+
|
239 |
+
if self.debug:
|
240 |
+
print(f"Total TFLOPs for Muon: {total_flops / 1e12:.2f} TFLOPs", flush=True)
|
241 |
+
|
242 |
+
ordered_params = sorted(
|
243 |
+
params, key=lambda p: param_to_flops[id(p)], reverse=True
|
244 |
+
)
|
245 |
+
|
246 |
+
round_robin = 0
|
247 |
+
mesh = None
|
248 |
+
for p in ordered_params:
|
249 |
+
if mesh is None:
|
250 |
+
mesh = p.device_mesh
|
251 |
+
if mesh.ndim != 1:
|
252 |
+
raise NotImplementedError(
|
253 |
+
"Muon requires a 1D mesh for distributed training yet."
|
254 |
+
)
|
255 |
+
elif mesh != p.device_mesh:
|
256 |
+
raise ValueError("All parameters must be on the same mesh.")
|
257 |
+
|
258 |
+
param_to_state[id(p)] = _muon_state()
|
259 |
+
param_to_state[id(p)].worker_rank = mesh.mesh[round_robin].item()
|
260 |
+
|
261 |
+
round_robin = (round_robin + 1) % mesh.mesh.numel()
|
262 |
+
|
263 |
+
return param_to_state, ordered_params
|
264 |
+
|
265 |
+
def base(self, params, group, lr, wd, momentum):
|
266 |
+
# generate weight updates in distributed fashion
|
267 |
+
for p in params:
|
268 |
+
g = p.grad
|
269 |
+
if g is None:
|
270 |
+
continue
|
271 |
+
if g.ndim > 2:
|
272 |
+
g = g.view(g.size(0), -1)
|
273 |
+
assert g is not None
|
274 |
+
|
275 |
+
# calc update
|
276 |
+
state = self.state[p]
|
277 |
+
if "momentum_buffer" not in state:
|
278 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
279 |
+
buf = state["momentum_buffer"]
|
280 |
+
buf.mul_(momentum).add_(g)
|
281 |
+
if group["nesterov"]:
|
282 |
+
g = g.add(buf, alpha=momentum)
|
283 |
+
else:
|
284 |
+
g = buf
|
285 |
+
|
286 |
+
u = _zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
|
287 |
+
|
288 |
+
# scale update
|
289 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
290 |
+
|
291 |
+
# apply weight decay
|
292 |
+
p.data.mul_(1 - lr * wd)
|
293 |
+
|
294 |
+
# apply update
|
295 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
296 |
+
|
297 |
+
def _update_g(self, p, g, group, momentum):
|
298 |
+
# calc update
|
299 |
+
state = self.state[p]
|
300 |
+
if "momentum_buffer" not in state:
|
301 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
302 |
+
buf = state["momentum_buffer"]
|
303 |
+
buf.mul_(momentum).add_(g)
|
304 |
+
if group["nesterov"]:
|
305 |
+
g = g.add(buf, alpha=momentum)
|
306 |
+
else:
|
307 |
+
g = buf
|
308 |
+
return g
|
309 |
+
|
310 |
+
def _update_p(self, p, u, lr, wd):
|
311 |
+
# scale update
|
312 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
313 |
+
# apply weight decay
|
314 |
+
p.data.mul_(1 - lr * wd)
|
315 |
+
# apply update
|
316 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
317 |
+
|
318 |
+
def parallel(self, params, group, lr, wd, momentum):
|
319 |
+
"""
|
320 |
+
Perform a parallel optimization step using Muon.
|
321 |
+
"""
|
322 |
+
|
323 |
+
for p in params:
|
324 |
+
g = p.grad
|
325 |
+
if g is None:
|
326 |
+
continue
|
327 |
+
if g.ndim > 2:
|
328 |
+
g = g.view(g.size(0), -1)
|
329 |
+
|
330 |
+
# Update g in the local rank
|
331 |
+
g = self._update_g(
|
332 |
+
p,
|
333 |
+
g,
|
334 |
+
group,
|
335 |
+
momentum=momentum,
|
336 |
+
)
|
337 |
+
p.grad = g
|
338 |
+
|
339 |
+
param_to_state, ordered_params = self.init_state_and_assign_params(
|
340 |
+
params, group
|
341 |
+
)
|
342 |
+
|
343 |
+
def enqueue_gathers(start_idx, chunk_size):
|
344 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
345 |
+
state = param_to_state[id(p)]
|
346 |
+
_gather(p, state, self.rank, self.comm_stream)
|
347 |
+
|
348 |
+
def enqueue_computes(start_idx, chunk_size):
|
349 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
350 |
+
state = param_to_state[id(p)]
|
351 |
+
_compute_u(state, group["ns_steps"], self.rank, self.compute_stream)
|
352 |
+
|
353 |
+
def enqueue_scatters(start_idx, chunk_size):
|
354 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
355 |
+
state = param_to_state[id(p)]
|
356 |
+
_scatter(p, state, self.rank, self.comm_stream)
|
357 |
+
|
358 |
+
chunk_size = params[0].device_mesh.mesh.numel()
|
359 |
+
|
360 |
+
# Wait grad update
|
361 |
+
self.comm_stream.wait_stream(torch.cuda.current_stream())
|
362 |
+
|
363 |
+
enqueue_gathers(0, chunk_size)
|
364 |
+
for i in range(0, len(params) + chunk_size - 1, chunk_size):
|
365 |
+
enqueue_computes(i, chunk_size)
|
366 |
+
enqueue_gathers(i + chunk_size, chunk_size)
|
367 |
+
enqueue_scatters(i, chunk_size)
|
368 |
+
|
369 |
+
torch.cuda.current_stream().wait_stream(self.comm_stream)
|
370 |
+
|
371 |
+
for p in params:
|
372 |
+
g = p.grad
|
373 |
+
if g is None:
|
374 |
+
continue
|
375 |
+
|
376 |
+
# Update p with sharded u
|
377 |
+
state = param_to_state[id(p)]
|
378 |
+
self._update_p(
|
379 |
+
p,
|
380 |
+
state.scattered_u,
|
381 |
+
lr=lr,
|
382 |
+
wd=wd,
|
383 |
+
)
|
384 |
+
|
385 |
+
def step(self, closure=None):
|
386 |
+
"""Perform a single optimization step.
|
387 |
+
|
388 |
+
Args:
|
389 |
+
closure (Callable, optional): A closure that reevaluates the model
|
390 |
+
and returns the loss.
|
391 |
+
"""
|
392 |
+
loss = None
|
393 |
+
if closure is not None:
|
394 |
+
with torch.enable_grad():
|
395 |
+
loss = closure()
|
396 |
+
|
397 |
+
for group in self.param_groups:
|
398 |
+
############################
|
399 |
+
# Muon #
|
400 |
+
############################
|
401 |
+
|
402 |
+
params = [p for p in group["params"] if self.state[p]["use_muon"]]
|
403 |
+
lr = group["lr"]
|
404 |
+
wd = group["wd"]
|
405 |
+
momentum = group["momentum"]
|
406 |
+
|
407 |
+
if isinstance(params[0].data, DTensor):
|
408 |
+
self.parallel(
|
409 |
+
params,
|
410 |
+
group,
|
411 |
+
lr=lr,
|
412 |
+
wd=wd,
|
413 |
+
momentum=momentum,
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
self.base(
|
417 |
+
params,
|
418 |
+
group,
|
419 |
+
lr=lr,
|
420 |
+
wd=wd,
|
421 |
+
momentum=momentum,
|
422 |
+
)
|
423 |
+
|
424 |
+
############################
|
425 |
+
# AdamW backup #
|
426 |
+
############################
|
427 |
+
|
428 |
+
params = [p for p in group["params"] if not self.state[p]["use_muon"]]
|
429 |
+
lr = group["lr"]
|
430 |
+
beta1, beta2 = group["adamw_betas"]
|
431 |
+
eps = group["adamw_eps"]
|
432 |
+
weight_decay = group["wd"]
|
433 |
+
|
434 |
+
for p in params:
|
435 |
+
g = p.grad
|
436 |
+
if g is None:
|
437 |
+
continue
|
438 |
+
state = self.state[p]
|
439 |
+
if "step" not in state:
|
440 |
+
state["step"] = 0
|
441 |
+
state["moment1"] = torch.zeros_like(g)
|
442 |
+
state["moment2"] = torch.zeros_like(g)
|
443 |
+
state["step"] += 1
|
444 |
+
step = state["step"]
|
445 |
+
buf1 = state["moment1"]
|
446 |
+
buf2 = state["moment2"]
|
447 |
+
buf1.lerp_(g, 1 - beta1)
|
448 |
+
buf2.lerp_(g.square(), 1 - beta2)
|
449 |
+
|
450 |
+
g = buf1 / (eps + buf2.sqrt())
|
451 |
+
|
452 |
+
bias_correction1 = 1 - beta1**step
|
453 |
+
bias_correction2 = 1 - beta2**step
|
454 |
+
scale = bias_correction1 / bias_correction2**0.5
|
455 |
+
p.data.mul_(1 - lr * weight_decay)
|
456 |
+
p.data.add_(g, alpha=-lr / scale)
|
457 |
+
|
458 |
+
return loss
|
build/torch27-cxx11-cu126-x86_64-linux/optimizer/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .muon import Muon
|
2 |
+
|
3 |
+
__all__ = [
|
4 |
+
"Muon",
|
5 |
+
]
|
build/torch27-cxx11-cu126-x86_64-linux/optimizer/_ops.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from . import _optimizer_b4b3752_dirty
|
3 |
+
ops = torch.ops._optimizer_b4b3752_dirty
|
4 |
+
|
5 |
+
def add_op_namespace_prefix(op_name: str):
|
6 |
+
"""
|
7 |
+
Prefix op by namespace.
|
8 |
+
"""
|
9 |
+
return f"_optimizer_b4b3752_dirty::{op_name}"
|
build/torch27-cxx11-cu126-x86_64-linux/optimizer/_optimizer_20250614125054.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:cb02d3818a89c819a5a12d066ce56da0ebc4f3da491cb045ae380c5b9319e592
|
3 |
+
size 1824256
|
build/torch27-cxx11-cu126-x86_64-linux/optimizer/_optimizer_b4b3752_dirty.abi3.so
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:b425a7fd854402508da5af17fa88f305753a09474686d6ec7afe540b3c5c082e
|
3 |
+
size 1824256
|
build/torch27-cxx11-cu126-x86_64-linux/optimizer/muon.py
ADDED
@@ -0,0 +1,458 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.distributed as dist
|
6 |
+
from torch.distributed._tensor import DTensor
|
7 |
+
|
8 |
+
|
9 |
+
# TODO leave original url and consider LICENSE
|
10 |
+
# This code snippet is a modified version adapted from the following GitHub repository:
|
11 |
+
# https://github.com/KellerJordan/Muon/blob/master/muon.py
|
12 |
+
def _zeropower_via_newtonschulz5(G, steps):
|
13 |
+
"""
|
14 |
+
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
|
15 |
+
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
|
16 |
+
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
|
17 |
+
zero even beyond the point where the iteration no longer converges all the way to one everywhere
|
18 |
+
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
|
19 |
+
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
|
20 |
+
performance at all relative to UV^T, where USV^T = G is the SVD.
|
21 |
+
"""
|
22 |
+
assert len(G.shape) == 2
|
23 |
+
a, b, c = (3.4445, -4.7750, 2.0315)
|
24 |
+
X = G # no manual typecast
|
25 |
+
if G.size(0) > G.size(1):
|
26 |
+
X = X.T
|
27 |
+
# Ensure spectral norm is at most 1
|
28 |
+
X = X / (X.norm() + 1e-7)
|
29 |
+
X = X.bfloat16()
|
30 |
+
# Perform the NS iterations
|
31 |
+
for _ in range(steps):
|
32 |
+
A = X @ X.T
|
33 |
+
# B = (
|
34 |
+
# b * A + c * A @ A
|
35 |
+
# ) # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
|
36 |
+
B = torch.addmm(A, A, A, alpha=c, beta=b)
|
37 |
+
# X = a * X + B @ X
|
38 |
+
X = torch.addmm(X, B, X, alpha=1.0, beta=a)
|
39 |
+
|
40 |
+
if G.size(0) > G.size(1):
|
41 |
+
X = X.T
|
42 |
+
return X.to(G.dtype)
|
43 |
+
|
44 |
+
|
45 |
+
@dataclass
|
46 |
+
class _muon_state:
|
47 |
+
# TODO: use Optional
|
48 |
+
worker_rank: int | None = None
|
49 |
+
gathered_grad: torch.Tensor | None = None
|
50 |
+
computed_u: torch.Tensor | None = None
|
51 |
+
scattered_u: torch.Tensor | None = None
|
52 |
+
gather_event: torch.cuda.Event | None = None
|
53 |
+
compute_event: torch.cuda.Event | None = None
|
54 |
+
|
55 |
+
|
56 |
+
def _gather(p, state, rank, comm_stream):
|
57 |
+
g = p.grad
|
58 |
+
mesh = g.device_mesh
|
59 |
+
|
60 |
+
if rank == state.worker_rank:
|
61 |
+
gather_list = [torch.empty_like(g.to_local()) for _ in range(mesh.mesh.numel())]
|
62 |
+
else:
|
63 |
+
gather_list = None
|
64 |
+
|
65 |
+
with torch.cuda.stream(comm_stream):
|
66 |
+
torch.distributed.gather(
|
67 |
+
g.to_local(),
|
68 |
+
dst=state.worker_rank,
|
69 |
+
gather_list=gather_list,
|
70 |
+
group=mesh.get_group(),
|
71 |
+
)
|
72 |
+
if rank == state.worker_rank:
|
73 |
+
# TODO: Consider ,,,
|
74 |
+
if state.gathered_grad is not None:
|
75 |
+
raise RuntimeError(
|
76 |
+
"Gather event already exists, which should not happen."
|
77 |
+
)
|
78 |
+
state.gathered_grad = torch.cat(gather_list, dim=0)
|
79 |
+
state.gather_event = torch.cuda.Event()
|
80 |
+
state.gather_event.record()
|
81 |
+
else:
|
82 |
+
state.gathered_grad = None
|
83 |
+
state.gather_event = None
|
84 |
+
|
85 |
+
|
86 |
+
def _compute_u(state, steps, rank, compute_stream):
|
87 |
+
with torch.cuda.stream(compute_stream):
|
88 |
+
if rank == state.worker_rank:
|
89 |
+
if state.gather_event is None:
|
90 |
+
raise RuntimeError("Gather event must be set before compute.")
|
91 |
+
compute_stream.wait_event(state.gather_event)
|
92 |
+
u = _zeropower_via_newtonschulz5(state.gathered_grad, steps)
|
93 |
+
state.computed_u = u
|
94 |
+
state.compute_event = torch.cuda.Event()
|
95 |
+
state.compute_event.record()
|
96 |
+
else:
|
97 |
+
state.computed_u = None
|
98 |
+
state.compute_event = None
|
99 |
+
|
100 |
+
|
101 |
+
def _scatter(p, state, rank, comm_stream):
|
102 |
+
u = state.computed_u
|
103 |
+
mesh = p.device_mesh
|
104 |
+
|
105 |
+
with torch.cuda.stream(comm_stream):
|
106 |
+
if rank == state.worker_rank:
|
107 |
+
if state.compute_event is None:
|
108 |
+
raise RuntimeError("Compute event must be set before scatter.")
|
109 |
+
comm_stream.wait_event(state.compute_event)
|
110 |
+
scatter_list = list(torch.split(u, p.size(0) // mesh.mesh.numel(), dim=0))
|
111 |
+
else:
|
112 |
+
scatter_list = None
|
113 |
+
|
114 |
+
u = torch.empty_like(p.to_local())
|
115 |
+
torch.distributed.scatter(
|
116 |
+
u,
|
117 |
+
scatter_list=scatter_list,
|
118 |
+
src=state.worker_rank,
|
119 |
+
group=mesh.get_group(),
|
120 |
+
)
|
121 |
+
u = DTensor.from_local(
|
122 |
+
u,
|
123 |
+
placements=p.placements,
|
124 |
+
device_mesh=mesh,
|
125 |
+
)
|
126 |
+
|
127 |
+
state.scattered_u = u
|
128 |
+
|
129 |
+
|
130 |
+
class Muon(torch.optim.Optimizer):
|
131 |
+
"""
|
132 |
+
Muon - MomentUm Orthogonalized by Newton-schulz
|
133 |
+
|
134 |
+
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
|
135 |
+
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
|
136 |
+
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
|
137 |
+
the advantage that it can be stably run in bfloat16 on the GPU.
|
138 |
+
|
139 |
+
Some warnings:
|
140 |
+
- We believe this optimizer is unlikely to work well for training with small batch size.
|
141 |
+
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
|
142 |
+
|
143 |
+
Arguments:
|
144 |
+
muon_params: The parameters to be optimized by Muon.
|
145 |
+
lr: The learning rate. The updates will have spectral norm of `lr`. (0.02 is a good default)
|
146 |
+
momentum: The momentum used by the internal SGD. (0.95 is a good default)
|
147 |
+
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
|
148 |
+
ns_steps: The number of Newton-Schulz iterations to run. (6 is probably always enough)
|
149 |
+
adamw_params: The parameters to be optimized by AdamW. Any parameters in `muon_params` which are
|
150 |
+
{0, 1}-D or are detected as being the embed or lm_head will be optimized by AdamW as well.
|
151 |
+
adamw_lr: The learning rate for the internal AdamW.
|
152 |
+
adamw_betas: The betas for the internal AdamW.
|
153 |
+
adamw_eps: The epsilon for the internal AdamW.
|
154 |
+
adamw_wd: The weight decay for the internal AdamW.
|
155 |
+
"""
|
156 |
+
|
157 |
+
def __init__(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
is_muon_func,
|
161 |
+
lr=1e-3,
|
162 |
+
momentum=0.95,
|
163 |
+
nesterov=True,
|
164 |
+
ns_steps=5,
|
165 |
+
adamw_wd=0.1,
|
166 |
+
adamw_betas=(0.9, 0.95),
|
167 |
+
adamw_eps=1e-8,
|
168 |
+
debug=False,
|
169 |
+
):
|
170 |
+
defaults = dict(
|
171 |
+
lr=lr,
|
172 |
+
wd=adamw_wd,
|
173 |
+
momentum=momentum,
|
174 |
+
nesterov=nesterov,
|
175 |
+
ns_steps=ns_steps,
|
176 |
+
adamw_betas=adamw_betas,
|
177 |
+
adamw_eps=adamw_eps,
|
178 |
+
)
|
179 |
+
|
180 |
+
super().__init__(model.parameters(), defaults)
|
181 |
+
self.is_muon_func = is_muon_func
|
182 |
+
self.model = model
|
183 |
+
|
184 |
+
if not dist.is_initialized():
|
185 |
+
raise RuntimeError(
|
186 |
+
"Muon optimizer requires distributed training to be initialized."
|
187 |
+
)
|
188 |
+
|
189 |
+
self.rank = dist.get_rank()
|
190 |
+
|
191 |
+
self.comm_stream = torch.cuda.Stream()
|
192 |
+
self.compute_stream = torch.cuda.Stream()
|
193 |
+
self.debug = debug
|
194 |
+
|
195 |
+
def __setstate__(self, state):
|
196 |
+
# Sort parameters into those for which we will use Muon, and those for which we will not
|
197 |
+
super().__setstate__(state)
|
198 |
+
for name, p in self.model.named_parameters():
|
199 |
+
if self.is_muon_func(p, name):
|
200 |
+
# Use Muon for every parameter in muon_params which is >= 2D and doesn't look like an embedding or head layer
|
201 |
+
assert p.ndim == 2, p.ndim
|
202 |
+
self.state[p]["use_muon"] = True
|
203 |
+
self.state[p]["orig_shape"] = p.shape
|
204 |
+
else:
|
205 |
+
# Do not use Muon for parameters in adamw_params
|
206 |
+
self.state[p]["use_muon"] = False
|
207 |
+
|
208 |
+
def _calc_flops(self, G, steps):
|
209 |
+
assert len(G.shape) == 2
|
210 |
+
M, N = G.shape
|
211 |
+
if M > N:
|
212 |
+
M, N = N, M
|
213 |
+
|
214 |
+
return steps * ((M**3) * 2 + (M**2 * N) * 4 + M * N * 2 + M**2 * 3)
|
215 |
+
|
216 |
+
def adjust_lr_for_muon(self, lr, param_shape):
|
217 |
+
A, B = param_shape[:2]
|
218 |
+
# We adjust the learning rate and weight decay based on the size of the parameter matrix
|
219 |
+
# as describted in the paper
|
220 |
+
adjusted_ratio = 0.2 * math.sqrt(max(A, B))
|
221 |
+
adjusted_lr = lr * adjusted_ratio
|
222 |
+
return adjusted_lr
|
223 |
+
|
224 |
+
def init_state_and_assign_params(self, params, group):
|
225 |
+
param_to_state = {}
|
226 |
+
param_to_flops = {}
|
227 |
+
|
228 |
+
total_flops = 0
|
229 |
+
for p in params:
|
230 |
+
g = p.grad
|
231 |
+
if g is None:
|
232 |
+
continue
|
233 |
+
assert g.ndim == 2, "Muon only supports 2D parameters."
|
234 |
+
|
235 |
+
flops = self._calc_flops(g, group["ns_steps"])
|
236 |
+
param_to_flops[id(p)] = flops
|
237 |
+
total_flops += flops
|
238 |
+
|
239 |
+
if self.debug:
|
240 |
+
print(f"Total TFLOPs for Muon: {total_flops / 1e12:.2f} TFLOPs", flush=True)
|
241 |
+
|
242 |
+
ordered_params = sorted(
|
243 |
+
params, key=lambda p: param_to_flops[id(p)], reverse=True
|
244 |
+
)
|
245 |
+
|
246 |
+
round_robin = 0
|
247 |
+
mesh = None
|
248 |
+
for p in ordered_params:
|
249 |
+
if mesh is None:
|
250 |
+
mesh = p.device_mesh
|
251 |
+
if mesh.ndim != 1:
|
252 |
+
raise NotImplementedError(
|
253 |
+
"Muon requires a 1D mesh for distributed training yet."
|
254 |
+
)
|
255 |
+
elif mesh != p.device_mesh:
|
256 |
+
raise ValueError("All parameters must be on the same mesh.")
|
257 |
+
|
258 |
+
param_to_state[id(p)] = _muon_state()
|
259 |
+
param_to_state[id(p)].worker_rank = mesh.mesh[round_robin].item()
|
260 |
+
|
261 |
+
round_robin = (round_robin + 1) % mesh.mesh.numel()
|
262 |
+
|
263 |
+
return param_to_state, ordered_params
|
264 |
+
|
265 |
+
def base(self, params, group, lr, wd, momentum):
|
266 |
+
# generate weight updates in distributed fashion
|
267 |
+
for p in params:
|
268 |
+
g = p.grad
|
269 |
+
if g is None:
|
270 |
+
continue
|
271 |
+
if g.ndim > 2:
|
272 |
+
g = g.view(g.size(0), -1)
|
273 |
+
assert g is not None
|
274 |
+
|
275 |
+
# calc update
|
276 |
+
state = self.state[p]
|
277 |
+
if "momentum_buffer" not in state:
|
278 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
279 |
+
buf = state["momentum_buffer"]
|
280 |
+
buf.mul_(momentum).add_(g)
|
281 |
+
if group["nesterov"]:
|
282 |
+
g = g.add(buf, alpha=momentum)
|
283 |
+
else:
|
284 |
+
g = buf
|
285 |
+
|
286 |
+
u = _zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
|
287 |
+
|
288 |
+
# scale update
|
289 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
290 |
+
|
291 |
+
# apply weight decay
|
292 |
+
p.data.mul_(1 - lr * wd)
|
293 |
+
|
294 |
+
# apply update
|
295 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
296 |
+
|
297 |
+
def _update_g(self, p, g, group, momentum):
|
298 |
+
# calc update
|
299 |
+
state = self.state[p]
|
300 |
+
if "momentum_buffer" not in state:
|
301 |
+
state["momentum_buffer"] = torch.zeros_like(g)
|
302 |
+
buf = state["momentum_buffer"]
|
303 |
+
buf.mul_(momentum).add_(g)
|
304 |
+
if group["nesterov"]:
|
305 |
+
g = g.add(buf, alpha=momentum)
|
306 |
+
else:
|
307 |
+
g = buf
|
308 |
+
return g
|
309 |
+
|
310 |
+
def _update_p(self, p, u, lr, wd):
|
311 |
+
# scale update
|
312 |
+
adjusted_lr = self.adjust_lr_for_muon(lr, p.shape)
|
313 |
+
# apply weight decay
|
314 |
+
p.data.mul_(1 - lr * wd)
|
315 |
+
# apply update
|
316 |
+
p.data.add_(u, alpha=-adjusted_lr)
|
317 |
+
|
318 |
+
def parallel(self, params, group, lr, wd, momentum):
|
319 |
+
"""
|
320 |
+
Perform a parallel optimization step using Muon.
|
321 |
+
"""
|
322 |
+
|
323 |
+
for p in params:
|
324 |
+
g = p.grad
|
325 |
+
if g is None:
|
326 |
+
continue
|
327 |
+
if g.ndim > 2:
|
328 |
+
g = g.view(g.size(0), -1)
|
329 |
+
|
330 |
+
# Update g in the local rank
|
331 |
+
g = self._update_g(
|
332 |
+
p,
|
333 |
+
g,
|
334 |
+
group,
|
335 |
+
momentum=momentum,
|
336 |
+
)
|
337 |
+
p.grad = g
|
338 |
+
|
339 |
+
param_to_state, ordered_params = self.init_state_and_assign_params(
|
340 |
+
params, group
|
341 |
+
)
|
342 |
+
|
343 |
+
def enqueue_gathers(start_idx, chunk_size):
|
344 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
345 |
+
state = param_to_state[id(p)]
|
346 |
+
_gather(p, state, self.rank, self.comm_stream)
|
347 |
+
|
348 |
+
def enqueue_computes(start_idx, chunk_size):
|
349 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
350 |
+
state = param_to_state[id(p)]
|
351 |
+
_compute_u(state, group["ns_steps"], self.rank, self.compute_stream)
|
352 |
+
|
353 |
+
def enqueue_scatters(start_idx, chunk_size):
|
354 |
+
for p in ordered_params[start_idx : start_idx + chunk_size]:
|
355 |
+
state = param_to_state[id(p)]
|
356 |
+
_scatter(p, state, self.rank, self.comm_stream)
|
357 |
+
|
358 |
+
chunk_size = params[0].device_mesh.mesh.numel()
|
359 |
+
|
360 |
+
# Wait grad update
|
361 |
+
self.comm_stream.wait_stream(torch.cuda.current_stream())
|
362 |
+
|
363 |
+
enqueue_gathers(0, chunk_size)
|
364 |
+
for i in range(0, len(params) + chunk_size - 1, chunk_size):
|
365 |
+
enqueue_computes(i, chunk_size)
|
366 |
+
enqueue_gathers(i + chunk_size, chunk_size)
|
367 |
+
enqueue_scatters(i, chunk_size)
|
368 |
+
|
369 |
+
torch.cuda.current_stream().wait_stream(self.comm_stream)
|
370 |
+
|
371 |
+
for p in params:
|
372 |
+
g = p.grad
|
373 |
+
if g is None:
|
374 |
+
continue
|
375 |
+
|
376 |
+
# Update p with sharded u
|
377 |
+
state = param_to_state[id(p)]
|
378 |
+
self._update_p(
|
379 |
+
p,
|
380 |
+
state.scattered_u,
|
381 |
+
lr=lr,
|
382 |
+
wd=wd,
|
383 |
+
)
|
384 |
+
|
385 |
+
def step(self, closure=None):
|
386 |
+
"""Perform a single optimization step.
|
387 |
+
|
388 |
+
Args:
|
389 |
+
closure (Callable, optional): A closure that reevaluates the model
|
390 |
+
and returns the loss.
|
391 |
+
"""
|
392 |
+
loss = None
|
393 |
+
if closure is not None:
|
394 |
+
with torch.enable_grad():
|
395 |
+
loss = closure()
|
396 |
+
|
397 |
+
for group in self.param_groups:
|
398 |
+
############################
|
399 |
+
# Muon #
|
400 |
+
############################
|
401 |
+
|
402 |
+
params = [p for p in group["params"] if self.state[p]["use_muon"]]
|
403 |
+
lr = group["lr"]
|
404 |
+
wd = group["wd"]
|
405 |
+
momentum = group["momentum"]
|
406 |
+
|
407 |
+
if isinstance(params[0].data, DTensor):
|
408 |
+
self.parallel(
|
409 |
+
params,
|
410 |
+
group,
|
411 |
+
lr=lr,
|
412 |
+
wd=wd,
|
413 |
+
momentum=momentum,
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
self.base(
|
417 |
+
params,
|
418 |
+
group,
|
419 |
+
lr=lr,
|
420 |
+
wd=wd,
|
421 |
+
momentum=momentum,
|
422 |
+
)
|
423 |
+
|
424 |
+
############################
|
425 |
+
# AdamW backup #
|
426 |
+
############################
|
427 |
+
|
428 |
+
params = [p for p in group["params"] if not self.state[p]["use_muon"]]
|
429 |
+
lr = group["lr"]
|
430 |
+
beta1, beta2 = group["adamw_betas"]
|
431 |
+
eps = group["adamw_eps"]
|
432 |
+
weight_decay = group["wd"]
|
433 |
+
|
434 |
+
for p in params:
|
435 |
+
g = p.grad
|
436 |
+
if g is None:
|
437 |
+
continue
|
438 |
+
state = self.state[p]
|
439 |
+
if "step" not in state:
|
440 |
+
state["step"] = 0
|
441 |
+
state["moment1"] = torch.zeros_like(g)
|
442 |
+
state["moment2"] = torch.zeros_like(g)
|
443 |
+
state["step"] += 1
|
444 |
+
step = state["step"]
|
445 |
+
buf1 = state["moment1"]
|
446 |
+
buf2 = state["moment2"]
|
447 |
+
buf1.lerp_(g, 1 - beta1)
|
448 |
+
buf2.lerp_(g.square(), 1 - beta2)
|
449 |
+
|
450 |
+
g = buf1 / (eps + buf2.sqrt())
|
451 |
+
|
452 |
+
bias_correction1 = 1 - beta1**step
|
453 |
+
bias_correction2 = 1 - beta2**step
|
454 |
+
scale = bias_correction1 / bias_correction2**0.5
|
455 |
+
p.data.mul_(1 - lr * weight_decay)
|
456 |
+
p.data.add_(g, alpha=-lr / scale)
|
457 |
+
|
458 |
+
return loss
|