forked from arcee-ai/mergekit
-
Notifications
You must be signed in to change notification settings - Fork 0
/
graph.py
298 lines (237 loc) · 8.52 KB
/
graph.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
from abc import ABC, abstractmethod
from typing import Any, Dict, Iterator, List, Optional, Set, Tuple, Union
import networkx
import safetensors.torch
import torch
import tqdm
from pydantic import BaseModel
from typing_extensions import Protocol, TypeAlias
from common import ModelReference
from lazy_tensors import LazyTensorLoader, TensorWriter
class TensorReference(BaseModel):
model: Optional[ModelReference]
key: str
def __str__(self) -> str:
if self.model is not None:
ns = str(self.model)
else:
ns = "_"
return ns + ":" + self.key
class Config:
frozen = True
class InputShard(BaseModel):
path: str
class Config:
frozen = True
ModelComponent: TypeAlias = Union[TensorReference, InputShard]
class Operation(BaseModel):
function: str
inputs: List[ModelComponent]
kwargs: Optional[Dict[str, Any]] = None
class Config:
frozen = True
class ProceduralRule(ABC):
@abstractmethod
def can_generate_rule(self, component: ModelComponent) -> bool:
...
@abstractmethod
def generate_rule(self, component: ModelComponent) -> Optional[Operation]:
...
class LoadTensorRule(ProceduralRule):
model: ModelReference
tensor_paths: Dict[str, str]
def __init__(
self,
model: ModelReference,
tensor_paths: Dict[str, str],
dtype: Optional[str],
cuda: bool,
):
self.model = model
self.tensor_paths = tensor_paths
self.dtype = dtype
self.cuda = cuda
def can_generate_rule(self, component: ModelComponent) -> bool:
return (
isinstance(component, TensorReference)
and component.model == self.model
and component.key in self.tensor_paths
)
def generate_rule(self, component: ModelComponent) -> Operation:
if not self.can_generate_rule(component):
return None
return Operation(
function="load_tensor",
inputs=[],
kwargs={
"model": component.model,
"key": component.key,
"dtype": self.dtype,
"cuda": self.cuda,
},
)
class ShardNoopRule(ProceduralRule):
def can_generate_rule(self, component: ModelComponent) -> bool:
return isinstance(component, InputShard)
def generate_rule(self, component: ModelComponent) -> Operation:
return Operation(function="noop")
class RuleSet:
static: Dict[ModelComponent, Operation]
procedural: List[ProceduralRule]
def __init__(
self,
static: Optional[Dict[ModelComponent, Operation]] = None,
procedural: Optional[List[ProceduralRule]] = None,
):
if not static:
static = {}
if not procedural:
procedural = []
self.static = static
self.procedural = procedural
def get(self, component: ModelComponent) -> Optional[Operation]:
if component in self.static:
return self.static[component]
for p in self.procedural:
if p.can_generate_rule(component):
op = p.generate_rule(component)
if op:
return op
return None
class OperationProtocol(Protocol):
def __call__(
self, tensors: Dict[TensorReference, torch.Tensor], **kwargs
) -> Optional[torch.Tensor]:
...
class Executor:
rules: RuleSet
loaders: Dict[ModelReference, LazyTensorLoader]
targets: List[ModelComponent]
operations: Dict[str, OperationProtocol]
gpu_shard_buffer: bool = False
def compare_component_key(self, c: ModelComponent):
if isinstance(c, InputShard):
return (c.path, None)
if c.model:
shard_path = self.loaders[c.model].index.tensor_paths[c.key]
else:
shard_path = ""
return (shard_path, c.key)
def __init__(
self,
models: List[ModelReference],
targets: List[ModelComponent],
rules: RuleSet,
operations: Optional[Dict[str, OperationProtocol]] = None,
cache_dir: Optional[str] = None,
dtype: Optional[str] = None,
cuda: bool = False,
gpu_shard_buffer: bool = False,
):
self.targets = targets
self.loaders = {
ref: LazyTensorLoader(ref.tensor_index(cache_dir=cache_dir))
for ref in models
}
for model, loader in self.loaders.items():
rules.procedural.append(
LoadTensorRule(model, loader.index.tensor_paths, dtype=dtype, cuda=cuda)
)
rules.procedural.append(ShardNoopRule())
if operations is None:
operations = {}
self.operations = operations
self.rules = rules
self.gpu_shard_buffer = gpu_shard_buffer
def run(self, out_path: str):
writer = TensorWriter(out_path)
for ref, tensor in tqdm.tqdm(self.generate_tensors(), total=len(self.targets)):
if not self.gpu_shard_buffer:
tensor = tensor.cpu()
writer.save_tensor(ref.key, tensor)
writer.finalize()
def generate_tensors(self) -> Iterator[Tuple[TensorReference, torch.Tensor]]:
schedule = self._schedule_ops()
last_use = {}
for idx, (component, op) in enumerate(schedule):
for j in range(len(schedule) - 1, idx, -1):
if component in schedule[j][1].inputs:
break
last_use[component] = j
tensors: Dict[ModelComponent, torch.Tensor] = {}
for idx, (component, op) in enumerate(schedule):
tensor_args = {}
for ref in op.inputs:
if isinstance(ref, InputShard):
continue
tensor_args[ref] = tensors[ref]
res = self._perform_operation(op, tensor_args)
if res is not None:
tensors[component] = res
if component in self.targets:
yield (component, res)
# expire unreferenced tensors
expired = []
for key in tensors:
if idx > last_use[key]:
expired.append(key)
for key in expired:
del tensors[key]
def _perform_operation(
self, operation: Operation, tensor_args: Dict[TensorReference, torch.Tensor]
) -> Optional[torch.Tensor]:
if operation.function == "load_tensor":
res = self.loaders[operation.kwargs["model"]].get_tensor(
operation.kwargs["key"]
)
if operation.kwargs["dtype"]:
res = res.to(dtype=operation.kwargs["dtype"])
if operation.kwargs["cuda"]:
res = res.cuda()
return res
elif operation.function == "pack_shard":
safetensors.torch.save_file(
{key.key: value for (key, value) in tensor_args.items()},
operation.kwargs["path"],
metadata={"format": "pt"},
)
elif operation.function == "noop":
print("noop!")
return None
elif operation.function in self.operations:
return self.operations[operation.function](
input_tensors=tensor_args, **operation.kwargs
)
else:
raise RuntimeError(f"Unimplemented function {operation.function}")
def _schedule_ops(self):
dependencies, ops = self._build_dependencies()
edge_tups = []
for a in dependencies:
for b in dependencies[a]:
edge_tups.append((b, a))
graph = networkx.DiGraph(edge_tups)
res = list(
networkx.lexicographical_topological_sort(
graph, key=self.compare_component_key
)
)
return [(r, ops[r]) for r in res]
def _build_dependencies(self):
dependencies: Dict[ModelComponent, Set[ModelComponent]] = {}
ops: Dict[ModelComponent, Operation] = {}
def _visit(node: ModelComponent):
if node in ops:
return
operation = self.rules.get(node)
if not operation:
raise RuntimeError(f"No rule to produce {node}")
ops[node] = operation
dependencies[node] = set()
for dependency in operation.inputs:
dependencies[node].add(dependency)
for dependency in operation.inputs:
_visit(dependency)
for t in self.targets:
_visit(t)
return dependencies, ops