karl commited on
Commit
5d3e4b5
·
1 Parent(s): 49049b7

added more input/output datas and drafted deduplicating tensor hash code. may not run have run into issues detering testing agian.

Browse files
Files changed (3) hide show
  1. _bighash.py +91 -0
  2. _safetensors.py +40 -18
  3. run_test.py +24 -12
_bighash.py ADDED
@@ -0,0 +1,91 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ # xxhash primes converted to int64
4
+ XXPRIME_1 = -(11400714785074694791 ^ 0xffffffffffffffff) - 1
5
+ XXPRIME_2 = -(14029467366897019727 ^ 0xffffffffffffffff) - 1
6
+ XXPRIME_5 = 2870177450012600261
7
+ # python tuple value
8
+ XXPRIME_5_3527539 = XXPRIME_5 ^ 3527539
9
+
10
+ def tensor_python_tuple_hash(items, out_or_in_place):
11
+ # https://github.com/python/cpython/blob/v3.13.2/Objects/tupleobject.c#L321
12
+ # this is apparently a simplified & modified form of xxhash
13
+ # learning xxhash could likely improve the code in this file
14
+
15
+ len_ = len(items)
16
+ # first iteration pulled out to provide storage placement
17
+ if out_or_in_place is None:
18
+ # in place
19
+ acc = torch.add(XXPRIME_5, items[0], alpha=XXPRIME_2, out=items[0])
20
+ else:
21
+ # place in out_or_in_place
22
+ acc = torch.add(XXPRIME_5, items[0], alpha=XXPRIME_2, out=out_or_in_place)
23
+ # bitwise rotation
24
+ upshift = acc << 31
25
+ acc >>= 33
26
+ acc &= 0x7fffffff # mask int64 sign extension
27
+ acc |= upshift
28
+
29
+ acc *= XXPRIME_1
30
+
31
+ for i in range(1, len_):
32
+ # acc += x * prime2
33
+ acc.add_(items[i], alpha=XXPRIME_2)
34
+
35
+ # bitwise rotation
36
+ upshift = acc << 31
37
+ acc >>= 33
38
+ acc &= 0x7fffffff # mask int64 sign extension
39
+ acc |= upshift
40
+
41
+ acc *= XXPRIME_1
42
+
43
+ acc += (len_ ^ XXPRIME_5_3527539)
44
+ return acc
45
+
46
+ def hash(buffer, *incoming_unhashed_ints):
47
+ if len(buffer) < 16:
48
+ return buffer.__hash__()
49
+
50
+ # first pass
51
+ # - allocate storage
52
+ # - place unhashed ints
53
+
54
+ words = len(buffer) // 8
55
+ dwords = words // 2
56
+ incoming_data = torch.frombuffer(buffer, count=words, dtype=torch.int64)
57
+ incoming_unhashed_ints = [int.from_bytes(buffer[words*8:],'little')] + list(incoming_unhashed_ints)
58
+
59
+ incoming_hashable_length = words & ~1
60
+ incoming_unhashed_length = words & 1
61
+ incoming_unhashed_int_length = len(incoming_unhashed_ints)
62
+ incoming_hashable_data = incoming_data[:incoming_hashable_length].view(2,-1)
63
+ incoming_unhashed_data = incoming_data[incoming_hashable_length:]
64
+
65
+ storage = torch.empty([dwords + incoming_unhashed_length + incoming_unhashed_int_length], dtype=torch.int64)
66
+ hashed_data = tensor_python_tuple_hash(incoming_hashable_data, out_or_in_place=storage[:dwords])
67
+ storage[dwords:-incoming_unhashed_int_length] = incoming_unhashed_data
68
+ for idx in range(incoming_unhashed_int_length):
69
+ storage[dwords+idx] = incoming_unhashed_ints[idx]
70
+
71
+ incoming_data = storage
72
+ words = len(incoming_data)
73
+ dwords = words // 2
74
+
75
+ # iterative passes
76
+ while words > 1:
77
+ incoming_hashable_length = words & ~1
78
+ incoming_unhashed_length = words & 1
79
+ incoming_hashable_data = incoming_data[:incoming_hashable_length].view(2,-1)
80
+ incoming_unhashed_tensor_data = incoming_data[incoming_hashable_length:]
81
+
82
+ hashed_data = tensor_python_tuple_hash(incoming_hashable_data, out_or_in_place=None)
83
+ words = dwords + incoming_unhashed_length
84
+ storage[dwords:words] = incoming_unhashed_tensor_data
85
+ incoming_data = storage[:words]
86
+ dwords = words // 2
87
+
88
+ return incoming_data[0].item()
89
+
90
+ if __name__ == '__main__':
91
+ print(hash(bytearray(b'the quick brown fox jumped over the lazy dog')))
_safetensors.py CHANGED
@@ -3,6 +3,8 @@ import contextlib, json, mmap, os
3
 
4
  import torch
5
 
 
 
6
  class WritingSafeTensors:
7
  def __init__(self, name, file_size=16*1024*1024*1024, **metadata):
8
  self.name = name.removesuffix('.safetensors')
@@ -10,36 +12,41 @@ class WritingSafeTensors:
10
  self.file = self.File(self.name + '.safetensors')
11
  self.files = {self.file.filename:self.file}
12
  self.file_size = file_size
13
- self.total_size = 0
14
  self.weight_map = {}
 
15
  def add(self, name, tensor):
16
  print(name, '...')
17
- self.weight_map[name] = self.file
18
- self.file.add(name, tensor)
19
- if self.file.size >= self.file_size:
 
 
 
 
20
  self.file.undo(name, tensor)
21
  ct = len(self.files)
22
  if len(self.files) == 1:
23
  self.file.rename(f'{self.name}-{ct:05}.safetensors')
24
  self.file.set_metadata(index = str(ct))
25
  self.files = {self.file.filename:self.file}
26
- self.file.finalize()
27
- self.total_size += self.file.size
28
  ct += 1
29
  self.file = self.File(f'{self.name}-{ct:05}.safetensors', index = ct)
30
  self.files[self.file.filename] = self.file
31
  self.file.add(name, tensor)
32
  self.weight_map[name] = self.file
33
  def finalize(self):
34
- self.file.finalize()
35
- if len(self.files) > 1:
36
- self.total_size += self.file.size
37
  index_name = self.name + '.safetensors.index.json'
38
  print(index_name, '...')
 
39
  tot = len(self.files)
40
  for ct, file in enumerate(self.files.values()):
41
  ct += 1
42
  file.rename(f'{self.name}-{ct:05}-of-{tot:06}.safetensors')
 
 
43
  with open(index_name, 'w') as fh:
44
  json.dump(
45
  {
@@ -82,6 +89,7 @@ class WritingSafeTensors:
82
 
83
  class File:
84
  def __init__(self, filename, **metadata):
 
85
  self.filename = filename
86
  self.fd = os.open(self.filename, os.O_RDWR | os.O_CREAT)
87
  self.size = 0
@@ -97,28 +105,43 @@ class WritingSafeTensors:
97
  os.truncate(self.filename, new_capacity)
98
  self.mmapview = memoryview(mmap.mmap(self.fd, new_capacity))
99
  self.capacity = new_capacity
100
- def add(self, name, tensor):
101
  length = tensor.numel() * tensor.dtype.itemsize
102
- self._reserve(length)
103
- end = self.size + length
104
- torch.frombuffer(
105
- self.mmapview[self.size:end],
106
- dtype=tensor.dtype, count=tensor.numel(),
107
- ).view(tensor.shape)[:] = tensor
 
 
 
 
 
 
 
 
 
 
 
 
 
108
  self.header[name] = {
109
  'dtype':
110
  str(tensor.dtype).rsplit('.',1)[-1]
111
  .replace('float','F')
112
  .replace('uint','U')
113
  .replace('int','I')
 
114
  .removesuffix('fn')
115
  .upper(),
116
  'shape':
117
  list(tensor.shape),
118
  'data_offsets':
119
- [self.size, end],
120
  }
121
  self.size = end
 
122
  def undo(self, name, tensor):
123
  length = tensor.numel() * tensor.dtype.itemsize
124
  del self.header[name]
@@ -153,4 +176,3 @@ class WritingSafeTensors:
153
  os.close(self.fd)
154
  self.finalized = True
155
  os.unlink(self.filename)
156
-
 
3
 
4
  import torch
5
 
6
+ from _bighash import hash
7
+
8
  class WritingSafeTensors:
9
  def __init__(self, name, file_size=16*1024*1024*1024, **metadata):
10
  self.name = name.removesuffix('.safetensors')
 
12
  self.file = self.File(self.name + '.safetensors')
13
  self.files = {self.file.filename:self.file}
14
  self.file_size = file_size
 
15
  self.weight_map = {}
16
+ self.hash_map = {}
17
  def add(self, name, tensor):
18
  print(name, '...')
19
+ tensor_hash = self.file.add(name, tensor)
20
+ image_of = self.hash_map.get(tensor_hash)
21
+ if image_of is not None:
22
+ self.file.undo(name, tensor)
23
+ imaged_hash = self.weight_map[image_of].add(name, tensor, image_of)
24
+ assert imaged_hash == tensor_hash
25
+ elif self.file.size >= self.file_size:
26
  self.file.undo(name, tensor)
27
  ct = len(self.files)
28
  if len(self.files) == 1:
29
  self.file.rename(f'{self.name}-{ct:05}.safetensors')
30
  self.file.set_metadata(index = str(ct))
31
  self.files = {self.file.filename:self.file}
 
 
32
  ct += 1
33
  self.file = self.File(f'{self.name}-{ct:05}.safetensors', index = ct)
34
  self.files[self.file.filename] = self.file
35
  self.file.add(name, tensor)
36
  self.weight_map[name] = self.file
37
  def finalize(self):
38
+ if len(self.files) == 1:
39
+ self.file.finalize()
40
+ else:
41
  index_name = self.name + '.safetensors.index.json'
42
  print(index_name, '...')
43
+ total_size = 0
44
  tot = len(self.files)
45
  for ct, file in enumerate(self.files.values()):
46
  ct += 1
47
  file.rename(f'{self.name}-{ct:05}-of-{tot:06}.safetensors')
48
+ file.finalize()
49
+ total_size += file.size
50
  with open(index_name, 'w') as fh:
51
  json.dump(
52
  {
 
89
 
90
  class File:
91
  def __init__(self, filename, **metadata):
92
+ print(filename, '...')
93
  self.filename = filename
94
  self.fd = os.open(self.filename, os.O_RDWR | os.O_CREAT)
95
  self.size = 0
 
105
  os.truncate(self.filename, new_capacity)
106
  self.mmapview = memoryview(mmap.mmap(self.fd, new_capacity))
107
  self.capacity = new_capacity
108
+ def add(self, name, tensor, image_of=None):
109
  length = tensor.numel() * tensor.dtype.itemsize
110
+ if image_of is None:
111
+ self._reserve(length)
112
+ start, end = self.size, self.size + length
113
+ torch.frombuffer(
114
+ self.mmapview[start : end],
115
+ dtype=tensor.dtype, count=tensor.numel(),
116
+ ).view(tensor.shape or [1])[:] = tensor
117
+ else:
118
+ image = self.header[image_of]
119
+ start, end = image['data_offsets']
120
+ assert end - start == length
121
+ assert (tensor == torch.frombuffer(
122
+ self.mmapview[start : end],
123
+ dtype=tensor.dtype, count=tensor.numel(),
124
+ ).view(tensor.shape or [1])[:]).all()
125
+
126
+ tensor.flatten()
127
+ tensor_hash = hash(self.mmapview[start : end])
128
+
129
  self.header[name] = {
130
  'dtype':
131
  str(tensor.dtype).rsplit('.',1)[-1]
132
  .replace('float','F')
133
  .replace('uint','U')
134
  .replace('int','I')
135
+ .removesuffix('uz')
136
  .removesuffix('fn')
137
  .upper(),
138
  'shape':
139
  list(tensor.shape),
140
  'data_offsets':
141
+ [start, end],
142
  }
143
  self.size = end
144
+ return tensor_hash
145
  def undo(self, name, tensor):
146
  length = tensor.numel() * tensor.dtype.itemsize
147
  del self.header[name]
 
176
  os.close(self.fd)
177
  self.finalized = True
178
  os.unlink(self.filename)
 
run_test.py CHANGED
@@ -67,33 +67,45 @@ SafeTensors = WritingSafeTensors(
67
  IDX = 0 # IDX is unused
68
  module_prefixes = {mod : name + '.' if name else '' for name, mod in pipe.model.named_modules()}
69
  tensors = {}
 
 
 
 
 
 
 
 
 
 
 
 
70
  def hook(module, inputs, kwinputs, outputs):
71
  global IDX
72
  prefix = module_prefixes[module]
 
 
73
  HAS_HF_HOOK = hasattr(module, '_hf_hook')
74
  if HAS_HF_HOOK:
75
  inputs, kwinputs = module._hf_hook.pre_forward(module, *inputs, **kwinputs)
76
  for idx, input in enumerate(inputs):
77
- if isinstance(input, torch.Tensor):
78
- SafeTensors.add(f'{prefix}input.{idx}', input);
79
  for key, input in kwinputs.items():
80
- if isinstance(input, torch.Tensor):
81
- SafeTensors.add(f'{prefix}input.{key}', input);
82
- if STORE_WEIGHTS and not list(module.children()):
83
- for wtname, wt in list(module.named_parameters()) + list(module.named_buffers()):
84
- SafeTensors.add(f'{prefix}{wtname}', wt)
 
85
  if HAS_HF_HOOK:
86
  outputs = module._hf_hook.post_forward(module, outputs)
87
  if isinstance(outputs, torch.Tensor):
88
- SafeTensors.add(f'{prefix}output', outputs);
89
  elif isinstance(outputs, dict):
90
  for key, output in outputs.items():
91
- if isinstance(output, torch.Tensor):
92
- SafeTensors.add(f'{prefix}output.{key}', output);
93
  else:
94
  for idx, output in enumerate(outputs):
95
- if isinstance(output, torch.Tensor):
96
- SafeTensors.add(f'{prefix}output.{idx}', output);
97
  IDX += 1
98
 
99
  for module in pipe.model.modules():
 
67
  IDX = 0 # IDX is unused
68
  module_prefixes = {mod : name + '.' if name else '' for name, mod in pipe.model.named_modules()}
69
  tensors = {}
70
+ def add_if_tensor(name, tensor):
71
+ if not isinstance(tensor, torch.Tensor):
72
+ try:
73
+ tensor = torch.tensor(tensor)
74
+ except:
75
+ try:
76
+ for idx, subtensor in enumerate(tensor):
77
+ add_if_tensor(f'{name}.{idx}', subtensor)
78
+ except:
79
+ pass
80
+ return
81
+ SafeTensors.add(name, tensor);
82
  def hook(module, inputs, kwinputs, outputs):
83
  global IDX
84
  prefix = module_prefixes[module]
85
+ if not prefix:
86
+ import pdb; pdb.set_trace()
87
  HAS_HF_HOOK = hasattr(module, '_hf_hook')
88
  if HAS_HF_HOOK:
89
  inputs, kwinputs = module._hf_hook.pre_forward(module, *inputs, **kwinputs)
90
  for idx, input in enumerate(inputs):
91
+ add_if_tensor(f'{prefix}input.{idx}', input);
 
92
  for key, input in kwinputs.items():
93
+ add_if_tensor(f'{prefix}input.{key}', input);
94
+ if STORE_WEIGHTS:
95
+ for wtname, wt in module.named_buffers(recurse=False):
96
+ add_if_tensor(f'{prefix}{wtname}', wt)
97
+ for wtname, wt in module.named_parameters(recurse=False):
98
+ add_if_tensor(f'{prefix}{wtname}', wt)
99
  if HAS_HF_HOOK:
100
  outputs = module._hf_hook.post_forward(module, outputs)
101
  if isinstance(outputs, torch.Tensor):
102
+ add_if_tensor(f'{prefix}output', outputs);
103
  elif isinstance(outputs, dict):
104
  for key, output in outputs.items():
105
+ add_if_tensor(f'{prefix}output.{key}', output);
 
106
  else:
107
  for idx, output in enumerate(outputs):
108
+ add_if_tensor(f'{prefix}output.{idx}', output);
 
109
  IDX += 1
110
 
111
  for module in pipe.model.modules():