zxdu20 commited on
Commit
2e1be30
1 Parent(s): c949d03

Add support for loading quantized model

Browse files
Files changed (3) hide show
  1. configuration_chatglm.py +3 -0
  2. modeling_chatglm.py +30 -4
  3. quantization.py +44 -30
configuration_chatglm.py CHANGED
@@ -70,6 +70,7 @@ class ChatGLMConfig(PretrainedConfig):
70
  max_sequence_length=2048,
71
  inner_hidden_size=16384,
72
  position_encoding_2d=True,
 
73
  pre_seq_len=None,
74
  prefix_projection=False,
75
  **kwargs
@@ -86,8 +87,10 @@ class ChatGLMConfig(PretrainedConfig):
86
  self.eos_token_id = eos_token_id
87
  self.pad_token_id = pad_token_id
88
  self.position_encoding_2d = position_encoding_2d
 
89
  self.pre_seq_len = pre_seq_len
90
  self.prefix_projection = prefix_projection
 
91
  super().__init__(
92
  pad_token_id=pad_token_id,
93
  bos_token_id=bos_token_id,
 
70
  max_sequence_length=2048,
71
  inner_hidden_size=16384,
72
  position_encoding_2d=True,
73
+ quantization_bit=0,
74
  pre_seq_len=None,
75
  prefix_projection=False,
76
  **kwargs
 
87
  self.eos_token_id = eos_token_id
88
  self.pad_token_id = pad_token_id
89
  self.position_encoding_2d = position_encoding_2d
90
+ self.quantization_bit = quantization_bit
91
  self.pre_seq_len = pre_seq_len
92
  self.prefix_projection = prefix_projection
93
+
94
  super().__init__(
95
  pad_token_id=pad_token_id,
96
  bos_token_id=bos_token_id,
modeling_chatglm.py CHANGED
@@ -139,6 +139,7 @@ class PrefixEncoder(torch.nn.Module):
139
  Input shape: (batch-size, prefix-length)
140
  Output shape: (batch-size, prefix-length, 2*layers*hidden)
141
  """
 
142
  def __init__(self, config):
143
  super().__init__()
144
  self.prefix_projection = config.prefix_projection
@@ -216,6 +217,13 @@ class RotaryEmbedding(torch.nn.Module):
216
  self.cos_cached, self.sin_cached = cos_cached, sin_cached
217
  return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]
218
 
 
 
 
 
 
 
 
219
 
220
  def rotate_half(x):
221
  x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:]
@@ -931,7 +939,6 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
931
  gmask=use_gmask
932
  )
933
 
934
-
935
  # [seq_len, batch, hidden_size]
936
  hidden_states = inputs_embeds.transpose(0, 1)
937
 
@@ -999,7 +1006,7 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
999
 
1000
 
1001
  class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
1002
- def __init__(self, config):
1003
  super().__init__(config)
1004
 
1005
  # self.hidden_size = config.hidden_size
@@ -1019,6 +1026,13 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
1019
  dtype=torch.half
1020
  )
1021
 
 
 
 
 
 
 
 
1022
  def get_output_embeddings(self):
1023
  return self.lm_head
1024
 
@@ -1351,7 +1365,19 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
1351
  break
1352
  yield input_ids
1353
 
1354
- def quantize(self, bits: int):
 
 
 
1355
  from .quantization import quantize
1356
- self.transformer = quantize(self.transformer, bits)
 
 
 
 
 
 
 
 
 
1357
  return self
 
139
  Input shape: (batch-size, prefix-length)
140
  Output shape: (batch-size, prefix-length, 2*layers*hidden)
141
  """
142
+
143
  def __init__(self, config):
144
  super().__init__()
145
  self.prefix_projection = config.prefix_projection
 
217
  self.cos_cached, self.sin_cached = cos_cached, sin_cached
218
  return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]
219
 
220
+ def _apply(self, fn):
221
+ if self.cos_cached is not None:
222
+ self.cos_cached = fn(self.cos_cached)
223
+ if self.sin_cached is not None:
224
+ self.sin_cached = fn(self.sin_cached)
225
+ return super()._apply(fn)
226
+
227
 
228
  def rotate_half(x):
229
  x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:]
 
939
  gmask=use_gmask
940
  )
941
 
 
942
  # [seq_len, batch, hidden_size]
943
  hidden_states = inputs_embeds.transpose(0, 1)
944
 
 
1006
 
1007
 
1008
  class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
1009
+ def __init__(self, config: ChatGLMConfig):
1010
  super().__init__(config)
1011
 
1012
  # self.hidden_size = config.hidden_size
 
1026
  dtype=torch.half
1027
  )
1028
 
1029
+ self.config = config
1030
+
1031
+ self.quantized = False
1032
+
1033
+ if self.config.quantization_bit:
1034
+ self.quantize(self.config.quantization_bit, empty_init=True)
1035
+
1036
  def get_output_embeddings(self):
1037
  return self.lm_head
1038
 
 
1365
  break
1366
  yield input_ids
1367
 
1368
+ def quantize(self, bits: int, empty_init=False, **kwargs):
1369
+ if bits == 0:
1370
+ return
1371
+
1372
  from .quantization import quantize
1373
+
1374
+ if self.quantized:
1375
+ logger.info("Already quantized.")
1376
+ return self
1377
+
1378
+ self.quantized = True
1379
+
1380
+ self.config.quantization_bit = bits
1381
+
1382
+ self.transformer = quantize(self.transformer, bits, empty_init=empty_init, **kwargs)
1383
  return self
quantization.py CHANGED
@@ -5,9 +5,40 @@ import bz2
5
  import torch
6
  import base64
7
  import ctypes
 
8
 
9
  from typing import List
10
- from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11
 
12
 
13
  class W8A16Linear(torch.autograd.Function):
@@ -33,30 +64,6 @@ class W8A16Linear(torch.autograd.Function):
33
  return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None, None
34
 
35
 
36
- class Kernel:
37
- def __init__(self, code: bytes, function_names: List[str]):
38
- self.code = code
39
- self._function_names = function_names
40
- self._cmodule = LazyKernelCModule(self.code)
41
-
42
- for name in self._function_names:
43
- setattr(self, name, KernelFunction(self._cmodule, name))
44
-
45
-
46
- quantization_code = "$QlpoOTFBWSZTWU9yuJUAQHN//////////f/n/8/n///n//bt4dTidcVx8X3V9FV/92/v4B7/AD5FBQFAAAChSgKpFCFAFVSigUAAAEKhSgUUqgFBKigqVREQAABQBQIANDTTIGI00BkZBkNGE0A0BkBkGQGRkaNAaAGQNBoGgDIAAYIGTI0DQAQAaGmmQMRpoDIyDIaMJoBoDIDIMgMjI0aA0AMgaDQNAGQAAwQMmRoGgAgA0NNMgYjTQGRkGQ0YTQDQGQGQZAZGRo0BoAZA0GgaAMgABggZMjQNABABoaaZAxGmgMjIMhowmgGgMgMgyAyMjRoDQAyBoNA0AZAADBAyZGgaAAmqU1NEgJqnptU/Sn4jRR6J6epk2pqb1Q/SgAPUGgyNNGjQ2SBpoAZAAGg0NB6mgDIAAAAA2oaApSREBNAARhGiYEaEwU8pvImlP0k2aam1GaGqbFNM1MHpTwmkepmyU9R6nqPKekHqNNPUxNGhp6n6p6QaZ6o9TG1GMqcoV9ly6nRanHlq6zPNbnGZNi6HSug+2nPiZ13XcnFYZW+45W11CumhzYhchOJ2GLLV1OBjBjGf4TptOddTSOcVxhqYZMYwZXZZY00zI1paX5X9J+b+f4e+x43RXSxXPOdquiGpduatGyXneN696M9t4HU2eR5XX/kPhP261NTx3JO1Ow7LyuDmeo9a7d351T1ZxnvnrvYnrXv/hXxPCeuYx2XsNmO003eg9J3Z6U7b23meJ4ri01OdzTk9BNO96brz+qT5nuvvH3ds/G+m/JcG/F2XYuhXlvO+jP7U3XgrzPN/lr8Sf1n6j4j7jZs+s/T0tNaNNYzTs12rxjwztHlnire3Nzc3N1wuBwOBwXBvZfoHpD7rFmR99V5vj3aXza3xdBbXMalubTg/jIv5dfAi54Pdc75j4z412n3Npj3Ld/ENm7a3b/Cod6h/ret1/5vn/C+l+gdslMvgPSLJ8d8q+U66fevYn/tW1chleEtNTGlcHCbLRlq0tHzF5tsbbZZfHjjLgZu42XCuC3NrdjTasZGNzgxPIrGqp7r3p7L2p5XjnpPSmTd5XtzqnB6U87zzg1Ol0zd0zsLszxR6lkxp35u6/teL0L0W922cR7Lu1lpL9CsHirzuM2T+BgsyViT6LHcm0/Vr6U/7LGGyJeqTEjt0PHWhF5mCT7R9mtlDwriYv0Tyr/OxYt6qp5r0mPVT0608TqnqMZaarU2nFwrTzzlrs1ed7z1ux60wyr4ydCaTi3enW8x68x0zU7tXSlcmPSW1mGpWJMg4zmPC2lK96tp0OE80y4MfEvnZj8zGluR6b22ki1Ou9V2nCd9xovcPvcYMZYy0lvN60ScZ45vN6yeCeeXFb1lVjnnCar5fwXwE2bzJ4HI1XVPXfXZMm44GUsMpYsmLB65TuVdm0cl0b+i/wGNN66XjeV7zuPpHcnK/juhhjdfId5jMdE5nN0dGmmm2zZs2cexD5n9p/dY352XsvXHaZNWWsmmS1atjR452nYudzvqv2HMRyvNNnlMcDl3R2+yx2uVrBubTW9icHDVtbNXlZm7jma1rM4VurZZd2y6nUau7ZXZ7bVU+mnoOVxZGMrVmvX60605JwmzGZhhhjTWtaaaMaaGTGmNMZasY0iX8VMUl8eepaIrzGSpemWOQyZORk2bNpjUybMmxqYmknCGCFynutfksaZpjTNMaaatM0xsxcGR0sociNqxNSmhhR1ZJPbsn8qyF0t2qH6iYBclclalbtTTcHTDsPaX6rlnElph2Jyumumtynv2Kk8GI7rsvXbIcJgHJOSaSXnnGaI3m87RtVXJOZ/YtgdTE6Wpha6ZlE8ayXkef1fh602r2WwvfMXtMdLlkfnLFdYYwYso+bWqm7yJqHXZGw2nrS5ZanSYnWlxBxMF1V940K2wdrI7R6OYf7DGGamMmTSbRhlS45xmVOumF1EyPCmHrrN8wwZOOrdNtLeMtzFzDlWnfTBxMk2NaXIZHBYxYLD4w8yju0ao65Vz1OIXoS9dLanwCe1PWrYuWMqf1if1z2k2yYfKJ741PDgno1ZQ8DRqvUny3mNoWTzGO6m1DkrJI8JiR5cSd+vZdGOO8nrMoc5+NDUFsMSXaZJeNlMmGLtJsovOsUp7I9S5VojKxF6bTVEelXqlfJobQr3LozSh2Jk7VcrVMfhXqszGWMzNqGhqZY0OadxkyyMssKugZR0KNFXBHlqwmJgTE/BNVMk6ItJXZMR0H47GpXv/DMOvNkmVuaV1PRfEdxuqc7Hcd+ZV/zTLaRxWk0nl9CdCeM6mn5rstHIBcpiuwmUZXeq81DacHI2rmrZ5SuE5mOZd6LQrZg9mx32TprA8BMo5jKN6yLTCi3WzQaZSuhzTtM1fUTGVpG8Tw+KXI0tjEpiWxtLYynOlktSbVlaI5kxP8TDH8kx50xoxi5KcA4pcja8KWLRlO/Ks6q06ergnvm1ca3Tq8Uw7LTUsmWyctXPWmpitl/uvGcWTGXGuAXDfhqazGmjkxcJW5hMMMMpYsXl2TZYtVOddG3XCarUt6Ptq9CZXSNzyuRzqRZOjsxdBbFVz6OA5HI43r1jityVlVpVkxmOsyaYWE1NTGq1sOVh36mHMcxtSvcy70edG0ZGR3I1Go1GRlV7mWWo1G0ZGRqlvH40l7o4m5xMWLLLYyNjnqc8556mdPqLJ31n/1nWOncxzG1tizrHs/Z+d2vP/B/l8wdJ6rHUn2nbbDq4p6htFtYzMMMTaZis1K5GKzGNmxhmUx2DDlZ/qNnIx41xnaMfCZWYaZWtNLTNW8ND4Fw1MyZOCdM428suKG1ehW8TesOydg7J+YYcD4cYR+8dFK6M4E3HM9ZfRNNL+Sn6rsl4DsrDl2HpPCnfxjGXtbZtYys1ttlyJ4T+BvexjGWRjMszK4Jpc77D3GyuVD7q0+G8m9G+2+rGm7cOR2y7FdtY2XUYx/oNlfRYxhMYyYZkyyg55enna9Kt/FFi6GMMwYwdwxWgxGMLKYmUyGExTKMZkMFhkymKuh0NOBNnBu+23LdwDoZYYzGGMxtORaTU1pjTGWTTGGtMrNWUsyyTTLLG1qy2ZjbK2DBllWqxMtBMaYZQmcE7zvvRcTkclUwdkxTaSdyySt/7fpL+T1v516Ji97fwr5JbLu305zMn5+GMTTZ9F+y7ExwmGVfG44yxn3dLv6l5i+Wth1jCrDq21nW9LqvvDzz3Vf3LLH/O/32TJ/erx3bXftO4eF+G956D952K/An4NfvOpjFjExjevP/UmE0fIoZXx6/w6lX/no3D0bLt+ixjieBM6ksRd0yB4Lt2SwYNE+gd1detlZWUnpiZfGfFaK+4PyCa/v18V8X75pe9fLXzp7l3VjF76vWZmHwGz1IZNWT7b8yddJ4q5kyrVdfru6atWc7bVYztL9Jf4GXvT+Y8m9/YsXP6H018a8D4XVOqvfzqeR+6yZOD8dPv0+U7/q5Pl+2dNb0MjzGVH5p6MNQ7cOWvw62U9aHE8DprDek+McLyvDz+te+9Zhq5+YTruufMcWMabqysTmZVWjKPfnK0wyVcrsuhjZRdLkHNvD72b9abriOSGIxiLixMOoalNPXzy+wT/tf+U6HHONfsz+xe8ufHBdQWWGWLA9if0rsnmrxK5LvRZQeWsTCsrmOYy8VteVfuRfcVTtDLItLIsMYxZLdU/DbtSemxF6Z6Zo5WBXE4tFdCyVMMXMTEMZXVlS6Xec2T4e0tHsRcEuWshcJ2YsNF5rUx1E8ifCq6Z+ZP7qdCeu/aTwFd53l16/o0NOw6O3dLavP4Hbi4RdmuDk6DoYaninC0+o4uZjbJ7Rxeu0/FbuFg+q7DVS6fQe0rZ6NDGUNNU6DEqOaLTicKnYZMnBWruljQxoaS3dZhocDge0bSTyOvdAbG5hxe2xji7E/L55xX13wWNDi6HCekcFxfCPGxY0MXC+s7afWaMdDyjyr+o8Rudm/NabOZvdl274zH4f5XK9z6On1Pe/K5TdPAslg77BjuO6Y3eO7GqvOPG/stknp1leyvLL0Z7bl9I4noMvLkzytLhWYzrOZzLXCORe028rORzOg4N/L0HlMOQ3Pgmnbb6KczlabORpu980q37TBqRu0/p3PO6234Bl03Ynuz+9W7gnsEcmvYaYY3aMYY0wx3pYd+ujsXauWdaY5Xkbtl23fPzFHiDB/QMo0yFjBllYxTQYYyxkrwn7JufwJ/PfgJ+C83X69ni6zvXcnyXabv0ncbLwsceS+RNlyN2mnneJtX0ngYO0+e+0+UnA+Wch3ji8hj5an4h+i6XBySU4n+R0roVcbw5yvHrmr4Yw8Y7x6c+9POPYHI5HI5HI5HI5HGXGww4nE4nrVyOR8XeqPEO7PLOiukYa3Novk5hV4cdtYZLI93e+uxff2jRo0aNGjRo0aNG1bVtW1dy3m83m8+tQ5ZzHw3nObwOu8La9Rc1dtkdS8A3eTk823tnktXWlxN6Oixe06zrN70Isd9jiOgZFq9yfkPqP/SLhN2Myl8jDM43bl1nbcb4cO57jlh8Jow6pzXZdL4dyODTuuhu77FyO27DdwdRxmvO+O+3N2+BdqyTwLHVczDVY4UPE4O66/ZO2cx1LFzVdSXtF7G4HMbrauOHRw6c8FdZ5m9fHZHYZXfTlZquyynSyTTKke6vcffSD9pzPA/G7n7jxPmuhc1DHMynPMrGL6AdewYmwu5ko+UUyTwrMv27rPH1v1nGqd87+p6N6LU8k3NEng53xXyHS97+44OSg/sy/hn+Se6yfYNjW0/uTgP+PvWYzLMmjhcLB/gGpri6H83/84eUXWT6T9Hsv7785z/7z4icpW+zfXypuR7rx/gMdZb1/wC678pcs8/2a3mDitGHxl9mfPlll5MafWWqxk/eYuTDgcNMzDGWLWvsuglNxs53GtN6uWpktlW1tZZYcuinMMWmnNnJydze3b2Y1McBxrBkXw799izLMZZYyy0TkbsGM4p03S2uVu5s/XXUdSdec6smVxZYYGpVmT8A+8ajuEyV5FatkvVru2x6uxGXXbH4A+jvgP4GMYy3iPLXzq/6z65+E005ey+cwMZD3fZcqc6xpjTFjQ0P3U+e++cPYmTIwj0nrK5NPTfl3WvpfLtXDcb2HQMudYOxFXQBor4L4T6vrOauFctYXJQ++NUWmJe5bmx1jDiZS1dTqWxo4GR8jm3fttpmPHppk9PEyv4/y8/sO07XacOmcqc0x2Vi9BvNJvN5oW8x4mOsydpidRxMYJPx06m1bqPzq9KtK8sxXNXFodD/+MYYaJTLwOhc9brCsV18oOR1i4tXChyTkq4lf4y1Ke+9axjDHqs1mfBbMXuP4Hzi+X7t8vzv7bHerrUPgPCxhjre4fXdfLNtNM+Jd+Zdh8xd8wP87uNPoPgv4W7/5P2BuxfsMabNnMnza+54Pdi5U671GPZY8CehX8Voeoo7FHpkeEc6715FwHZrIrUrHaviPUbPZHND+IhczrP6FcYvhOZ0Di/ETt0OI+YwNWR9r7tpf6WDeZKZDB1+z2IthOl1mPyb5FluvEx9h9d0NnM0Y1XPFkWIsk1WotJ0PBMmkvjvQTd0e71tfeV+8r8lQ/tpzpsmxJ+InrI/dj2UajUajVTUajatRqNRtGo1Go1Go4wjeMpZFMVV9CHbofPraLsJ3JpWV2XOoanCuFky4y3PPNxucK2uKC1Lbdb1eo+m5XomN6HfeZsabHLHRX/K+offtNGGmHWctcVcG44MdSqsOLY9VzX+Zxfxn2HPdWTpzWvkrtJ8M5zorrKcquRytJ5N5DZmcaW02l76nWO+BqPXm1A2Ry/0q71dH/mqrqeFjkYxjEXtsX8qubTk67rGycyqsdm4tZx5D6D5hhi0waaWmiaMP81Yjii5qxPlPuU/GfTL1Y5E6Jyfiq63qTa39A4J0sOGDgO9WF9bOXl0XfPRbsY2bPNKPy1YrFYrFYmRhhlTIyMjJWJYZHXuCXI8OoXsvfljGLFicNifpp2XunoPiG1wtx3p1Tah+/DD66OnVtVXP9rKbVxOnL0tR/rHtqB5UDErUVcl11D4qqvjpOcxX7armUNJB3LpW6bxVvD08e8h3odKKvyCFZBdSh2FVcST9xV3n3T8t1j7Kr9qgrqXg+13Pt5U7JCvFXVIV1YG5lRhkVYZJYYDDD4KOIMoHCp26WS8GB7uBh2zIdgq/PKyInjV2STShuoapUdCpX1yTwqq/z1VvET7Kh5nVPkO8YyxjLt2MaaMmWTLQvx3qnzltnXW0p2jxgbEtSny/Osv8Y9pLMXYoHVPAhkVdWVeODhR6q9/Sxe2liwwZWMVvFXfRkeIDxAePUPIrdJ4ey6yquzH+PD/bUOWAu05qVHtFd8rrKHSoeNIOUqrYr3FXyToqfYJgwmJdKpXXOwYYegNNGMzfZPp/t3t/DVs4zjNTN61rRqaWaa4NYbRjTa0tWwy2Y2tGN8ZO8ofNKq4j9SL7I+cSm4/6ovLV5HNXLI0jJidwrtk6ynCaP6Z++GjRlWS3tLeW129Mi9evxU9mtz6s5J3Z7M2ngTgnKvmpomxpaLCzPfmx0JWE+m3NLDDGOX47RctdYYNK5jakdqLkRlI39n590T5zctGSwwZZDJj6kW8XSi6ot2MmWWJ0DUT3nuvebBudScjZ79g8cWJ8av0k+/bE5WKd5MdbFpbDVMxu1DVMmtNZGJvq1mtRbn6M+g/kP0FwDwr7quZs7xosNGpbscyxhhd9TyJyFwbLcxlTasg75vW7TsV5K7ji44XPMMrdoj+Y3rT0Hie62nlYV/pwczzOmdLqLhYkzGMzCZWGMQzGMSsZYY6Di1t4nlJ+Em63mJxrVLxPbYxNEdgc1dU2iOKyoYYWjNrEeHTYybVk0atSa7ehuwsWMWTqn1TrnS6hYsi71d1+s+k+ic70e20fzE/VaTdxT9ZtU4GIXdeNx3X77guYYfpHeTQjaMX6brOu4OY4K7Y2d9mbHarI5ox3p4GpJ2Vd/Tst60f7j999pppjR+Q/Qf8J/VaORs3cji7FfFuN61+ui9s8hix1OCh5KGVV23BPXvZfz3CLyHpix+exi8z/KnCnosY2eunor+cxyPO/xJ0vKey9OvE9VjqaYu0x3Z3jd6o2b1T12D+F8l232lwaaacD5LE8LBxu7WTlbWraWpew8Xexjel3E+wWD4APITdNqR8F3R3T0lunCQ4GaE9R37DxeCYfcHi4xci5ovKfxVs55y2hf+65E/Xdp6jR5nrebTmi5incpkyOjs50JvrZwstbbW6kfuuQw+2mykf/EXNFzxfKTrxew929TR6bWnGL//F3JFOFCQT3K4lQ"
47
-
48
- kernels = Kernel(
49
- bz2.decompress(base64.b64decode(quantization_code)),
50
- [
51
- "int4WeightCompression",
52
- "int4WeightExtractionFloat",
53
- "int4WeightExtractionHalf",
54
- "int8WeightExtractionFloat",
55
- "int8WeightExtractionHalf",
56
- ],
57
- )
58
-
59
-
60
  def compress_int4_weight(weight: torch.Tensor): # (n, m)
61
  with torch.cuda.device(weight.device):
62
  n, m = weight.size(0), weight.size(1)
@@ -111,18 +118,18 @@ def extract_weight_to_half(weight: torch.Tensor, scale_list: torch.Tensor, sourc
111
 
112
 
113
  class QuantizedLinear(Linear):
114
- def __init__(self, weight_bit_width: int, weight_tensor=None, bias_tensor=None, *args, **kwargs):
115
  super(QuantizedLinear, self).__init__(*args, **kwargs)
116
  self.weight_bit_width = weight_bit_width
117
 
118
  shape = self.weight.shape
119
  del self.weight
120
 
121
- if weight_tensor is None:
122
  self.weight = torch.empty(
123
  shape[0], shape[1] * weight_bit_width // 8, dtype=torch.int8, device=kwargs["device"]
124
  )
125
- self.weight_scale = torch.empty(shape[0], dtype=kwargs["params_dtype"], device=kwargs["device"])
126
  else:
127
  self.weight_scale = (weight_tensor.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)).half()
128
  self.weight = torch.round(weight_tensor / self.weight_scale[:, None]).to(torch.int8)
@@ -131,7 +138,10 @@ class QuantizedLinear(Linear):
131
 
132
  self.weight = Parameter(self.weight.to(kwargs["device"]), requires_grad=False)
133
  self.weight_scale = Parameter(self.weight_scale.to(kwargs["device"]), requires_grad=False)
134
- self.bias = Parameter(bias_tensor.to(kwargs["device"]), requires_grad=False)
 
 
 
135
 
136
  def forward(self, input):
137
  output = W8A16Linear.apply(input, self.weight, self.weight_scale, self.weight_bit_width)
@@ -140,7 +150,7 @@ class QuantizedLinear(Linear):
140
  return output
141
 
142
 
143
- def quantize(model, weight_bit_width):
144
  """Replace fp16 linear with quantized linear"""
145
 
146
  for layer in model.layers:
@@ -153,6 +163,7 @@ def quantize(model, weight_bit_width):
153
  bias=True,
154
  dtype=torch.half,
155
  device=layer.attention.query_key_value.weight.device,
 
156
  )
157
  layer.attention.dense = QuantizedLinear(
158
  weight_bit_width=weight_bit_width,
@@ -163,6 +174,7 @@ def quantize(model, weight_bit_width):
163
  bias=True,
164
  dtype=torch.half,
165
  device=layer.attention.dense.weight.device,
 
166
  )
167
  layer.mlp.dense_h_to_4h = QuantizedLinear(
168
  weight_bit_width=weight_bit_width,
@@ -173,6 +185,7 @@ def quantize(model, weight_bit_width):
173
  bias=True,
174
  dtype=torch.half,
175
  device=layer.mlp.dense_h_to_4h.weight.device,
 
176
  )
177
  layer.mlp.dense_4h_to_h = QuantizedLinear(
178
  weight_bit_width=weight_bit_width,
@@ -183,5 +196,6 @@ def quantize(model, weight_bit_width):
183
  bias=True,
184
  dtype=torch.half,
185
  device=layer.mlp.dense_4h_to_h.weight.device,
 
186
  )
187
  return model
 
5
  import torch
6
  import base64
7
  import ctypes
8
+ from transformers.utils import logging
9
 
10
  from typing import List
11
+ from functools import partial
12
+
13
+ logger = logging.get_logger(__name__)
14
+
15
+ try:
16
+ from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up
17
+
18
+ class Kernel:
19
+ def __init__(self, code: bytes, function_names: List[str]):
20
+ self.code = code
21
+ self._function_names = function_names
22
+ self._cmodule = LazyKernelCModule(self.code)
23
+
24
+ for name in self._function_names:
25
+ setattr(self, name, KernelFunction(self._cmodule, name))
26
+
27
+ quantization_code = "$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"
28
+
29
+ kernels = Kernel(
30
+ bz2.decompress(base64.b64decode(quantization_code)),
31
+ [
32
+ "int4WeightCompression",
33
+ "int4WeightExtractionFloat",
34
+ "int4WeightExtractionHalf",
35
+ "int8WeightExtractionFloat",
36
+ "int8WeightExtractionHalf",
37
+ ],
38
+ )
39
+ except Exception as exception:
40
+ kernels = None
41
+ logger.warning("Failed to load cpm_kernels:" + str(exception))
42
 
43
 
44
  class W8A16Linear(torch.autograd.Function):
 
64
  return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None, None
65
 
66
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
67
  def compress_int4_weight(weight: torch.Tensor): # (n, m)
68
  with torch.cuda.device(weight.device):
69
  n, m = weight.size(0), weight.size(1)
 
118
 
119
 
120
  class QuantizedLinear(Linear):
121
+ def __init__(self, weight_bit_width: int, weight_tensor=None, bias_tensor=None, empty_init=False, *args, **kwargs):
122
  super(QuantizedLinear, self).__init__(*args, **kwargs)
123
  self.weight_bit_width = weight_bit_width
124
 
125
  shape = self.weight.shape
126
  del self.weight
127
 
128
+ if weight_tensor is None or empty_init:
129
  self.weight = torch.empty(
130
  shape[0], shape[1] * weight_bit_width // 8, dtype=torch.int8, device=kwargs["device"]
131
  )
132
+ self.weight_scale = torch.empty(shape[0], dtype=kwargs["dtype"], device=kwargs["device"])
133
  else:
134
  self.weight_scale = (weight_tensor.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)).half()
135
  self.weight = torch.round(weight_tensor / self.weight_scale[:, None]).to(torch.int8)
 
138
 
139
  self.weight = Parameter(self.weight.to(kwargs["device"]), requires_grad=False)
140
  self.weight_scale = Parameter(self.weight_scale.to(kwargs["device"]), requires_grad=False)
141
+ if bias_tensor is not None:
142
+ self.bias = Parameter(bias_tensor.to(kwargs["device"]), requires_grad=False)
143
+ else:
144
+ self.bias = None
145
 
146
  def forward(self, input):
147
  output = W8A16Linear.apply(input, self.weight, self.weight_scale, self.weight_bit_width)
 
150
  return output
151
 
152
 
153
+ def quantize(model, weight_bit_width, empty_init=False, **kwargs):
154
  """Replace fp16 linear with quantized linear"""
155
 
156
  for layer in model.layers:
 
163
  bias=True,
164
  dtype=torch.half,
165
  device=layer.attention.query_key_value.weight.device,
166
+ empty_init=empty_init
167
  )
168
  layer.attention.dense = QuantizedLinear(
169
  weight_bit_width=weight_bit_width,
 
174
  bias=True,
175
  dtype=torch.half,
176
  device=layer.attention.dense.weight.device,
177
+ empty_init=empty_init
178
  )
179
  layer.mlp.dense_h_to_4h = QuantizedLinear(
180
  weight_bit_width=weight_bit_width,
 
185
  bias=True,
186
  dtype=torch.half,
187
  device=layer.mlp.dense_h_to_4h.weight.device,
188
+ empty_init=empty_init
189
  )
190
  layer.mlp.dense_4h_to_h = QuantizedLinear(
191
  weight_bit_width=weight_bit_width,
 
196
  bias=True,
197
  dtype=torch.half,
198
  device=layer.mlp.dense_4h_to_h.weight.device,
199
+ empty_init=empty_init
200
  )
201
  return model