-
Notifications
You must be signed in to change notification settings - Fork 1
/
KV_process.py
404 lines (341 loc) · 14.1 KB
/
KV_process.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
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
import math
import torch
import torch.nn as nn
import operator
import skvq_quant
from typing import Literal, TypeAlias
from transformers import PreTrainedModel, LlamaForCausalLM, MistralForCausalLM
RodMeta: TypeAlias = tuple[dict[str, torch.Tensor], dict[str, torch.Tensor]]
class SKVQuantProcessor(nn.Module):
def __init__(
self,
K_target_bitwidth,
V_target_bitwidth,
gsize: int,
hidden: int,
clipping: list[float] = None,
reorder_meta: RodMeta = None,
smooth_scale: RodMeta = None,
KIVI_mode: bool = False,
fp8:bool = False,
fake_quant: bool = True,
# model: PreTrainedModel = None,
) -> None:
"""
reorder_meta: [reorder_indices, cluster_st_inds]
"""
super().__init__()
self.K_target_bitwidth = K_target_bitwidth
self.V_target_bitwidth = V_target_bitwidth
self.clipping = clipping
if reorder_meta:
reorder_meta[0]["k"] = reorder_meta[0]["k"].to(torch.int16)
reorder_meta[0]["v"] = reorder_meta[0]["v"].to(torch.int16)
reorder_meta[1]["k"] = reorder_meta[1]["k"].to(torch.int16)
reorder_meta[1]["v"] = reorder_meta[1]["v"].to(torch.int16)
self.reorder_idx = reorder_meta[0] if reorder_meta else None
self.group_st_idx = reorder_meta[1] if reorder_meta else None
self.smooth_scale = smooth_scale
self.gsize = gsize
self.fake_quant = fake_quant
self.KIVI_mode = KIVI_mode
self.fp8 = fp8
self.pack_sim = True
self.hidden = hidden
self.layer_idx:int = None
@torch.no_grad()
def reorder_weight(
self, model: PreTrainedModel, reorder_indices: list[torch.Tensor]
):
proj_map = {
LlamaForCausalLM: {
"q": operator.attrgetter("self_attn.q_proj"),
"k": operator.attrgetter("self_attn.k_proj"),
"v": operator.attrgetter("self_attn.v_proj"),
"o": operator.attrgetter("self_attn.o_proj"),
},
MistralForCausalLM: {
"q": operator.attrgetter("self_attn.q_proj"),
"k": operator.attrgetter("self_attn.k_proj"),
"v": operator.attrgetter("self_attn.v_proj"),
"o": operator.attrgetter("self_attn.o_proj"),
},
}
wname_map: dict[str, str] = proj_map.get(model.__class__, None)
assert wname_map is not None, f"Not supported for {model.__class__}"
for idx, layer in enumerate(model.model.layers):
for wtype, wgetter in wname_map.items():
weight: nn.Parameter = wgetter(layer)
dev = weight.weight.device
# if wtype in ["q", "k"] and idx == 0:
if wtype in ["k"]:
reorder_index = reorder_indices["k"][idx].to(dev)
weight.weight.data = weight.weight.data[reorder_index, :]
if wtype == "v":
reorder_index = reorder_indices["v"][idx].to(dev)
weight.weight.data = weight.weight.data[reorder_index, :]
# elif wtype == "o":
# reorder_index = reorder_indices["v"][idx].to(dev)
# weight.weight.data = weight.weight.data[:, reorder_index]
def quant_pytorch(
self,
ttype: Literal["k", "v"],
tensor: torch.Tensor,
)->tuple[torch.Tensor, None, None]:
bs, num_heads, seqlen, head_dim = tensor.shape
assert num_heads * head_dim == self.hidden
if seqlen == 0:
return tensor, None, None
dtype = tensor.dtype
qbits = self.K_target_bitwidth if ttype == "k" else self.V_target_bitwidth
per_channel = self.KIVI_mode and (ttype == "k")
if round(qbits) == qbits:
max_int = (1 << int(qbits)) - 1
else:
assert qbits == 1.5
max_int = 2
def pack_tensor(t: torch.Tensor, store_type=torch.uint8):
'''
pack along the last dim
'''
store_bits = torch.tensor([], dtype=torch.uint8).element_size() * 8
val_bits = int(math.ceil(qbits))
pack_num = store_bits // val_bits
pack_gst_lis = [0]
if self.group_st_idx is None:
gst_lis = torch.arange(0, num_heads*head_dim+1, self.gsize)
else:
gst_lis = self.group_st_idx["k"]
pgst = 0
res = []
for i in range(len(gst_lis)-1):
gst, ged = gst_lis[i], gst_lis[i+1]
gsize = ged - gst
pgst += int(math.ceil(gsize / pack_num))
pack_gst_lis.append(pgst)
pack_group = []
for j in range(gst, ged, pack_num):
pack_val = torch.zeros((bs, seqlen, 1), dtype=store_type).to(t.device)
for k in range(pack_num):
if j + k < ged:
pack_val += t[:,:, j + k:j+k+1].to(store_type) << ((pack_num - k - 1) * val_bits)
pack_group.append(pack_val)
res.append(torch.cat(pack_group, dim=-1))
res = torch.cat(res, dim=-1)
return res
def quant(t: torch.Tensor):
"""
Asymmetric Dynamic Quantiztion
Assume the last dim is reduction dim (group dim)
return: (quant_tensor, scale, zp), shape of quant param: [bs, seqlen, 1]
"""
if qbits == 16:
return t, None, None
gmin, gmax = t.aminmax(dim=-1, keepdim=True)
clip_scale = self.clipping[self.layer_idx]
gmin = gmin * clip_scale
gmax = gmax * clip_scale
zp = gmin
scale = torch.clamp((gmax - gmin) / max_int, min=1e-5)
# quant
res = (
t.sub(zp)
.div(scale)
.clamp(0, max_int)
.round()
)
if self.fp8:
scale = scale.to(torch.float8_e4m3fn).to(dtype)
zp = zp.to(torch.float8_e4m3fn).to(dtype)
if self.fake_quant:
# dequant
res = (
res.to(dtype)
.mul(scale)
.add(zp)
)
return res, None, None
else:
return res, scale, zp
def back_to_original(t: torch.Tensor, reorder_indices: torch.Tensor):
# https://stackoverflow.com/questions/52127723/pytorch-better-way-to-get-back-original-tensor-order-after-torch-sort
return t.gather(-1, reorder_indices.argsort(-1).expand(t.shape))
if per_channel:
assert seqlen % self.gsize == 0
assert self.fake_quant
# [bs, num_heads, head_dim, seqlen]
t_reshape = tensor.transpose(2, 3)
else:
t_reshape = tensor.transpose(1, 2).reshape(bs, seqlen, num_heads * head_dim)
if self.smooth_scale is not None:
assert not per_channel
smooth_scale = self.smooth_scale[ttype].to(tensor.dtype).to(tensor.device)
assert smooth_scale.shape[0] == self.hidden
t_reshape = t_reshape.mul(smooth_scale)
if self.reorder_idx is not None:
assert not per_channel
# reorder KV
reorder_idx = self.reorder_idx[ttype].long().to(tensor.device)
assert reorder_idx.shape[0] == self.hidden
reordered_kv = t_reshape[..., reorder_idx]
res = torch.empty_like(reordered_kv)
gst = self.group_st_idx[ttype].long()
if not self.fake_quant:
gscale = []
gzp = []
# quant reordered KV
for i in range(len(gst) - 1):
qdata, scale, zp = quant(reordered_kv[..., gst[i] : gst[i + 1]])
res[..., gst[i] : gst[i + 1]] = qdata
if not self.fake_quant:
gscale.append(scale)
gzp.append(zp)
if not self.fake_quant:
scale = torch.cat(gscale, dim = -1)
zp = torch.cat(gscale, dim = -1)
if self.fake_quant:
# back to original(ordered KV)
res = back_to_original(res, reorder_idx)
if self.smooth_scale is not None:
res = res.div(smooth_scale)
kv_quant = res.reshape(bs, seqlen, num_heads, head_dim).transpose(1, 2)
# if not self.fake_quant and self.pack_sim:
# kv_quant = pack_tensor(kv_quant.transpose(1,2).reshape(bs,seqlen, -1))
return kv_quant, scale, zp
else:
t_reshape = t_reshape.reshape(bs, seqlen, num_heads * head_dim // self.gsize, self.gsize)
qdata, scale, zp = quant(t_reshape)
if not self.fake_quant:
scale, zp = scale.squeeze(-1), zp.squeeze(-1)
if self.fake_quant and self.smooth_scale is not None:
qdata = qdata.reshape(bs, seqlen, -1).div(smooth_scale)
if per_channel:
kv_quant = qdata.reshape(bs, num_heads, head_dim, seqlen).transpose(2,3)
else:
kv_quant = qdata.reshape(bs, seqlen, num_heads, head_dim).transpose(1,2)
# if not self.fake_quant and self.pack_sim:
# kv_quant = pack_tensor(kv_quant.transpose(1,2).reshape(bs,seqlen, -1))
return kv_quant, scale, zp
def quant_cuda(
self,
ttype: Literal["k", "v"],
tensor: torch.Tensor,
)->tuple[torch.Tensor,torch.Tensor,torch.Tensor]:
qbits = self.K_target_bitwidth if ttype == "k" else self.V_target_bitwidth
bs, num_heads, seqlen, head_dim = tensor.shape
if seqlen == 0:
return tensor, None, None
qbits = self.K_target_bitwidth if ttype == "k" else self.V_target_bitwidth
assert not (self.KIVI_mode and (ttype == "k"))
if qbits == 16:
return tensor, None, None
# [bs, seqlen, num_heads, head_dim]
t_reshape = tensor.transpose(1, 2).contiguous()
smooth_scale = None
reorder_idx = None
gst_idx = None
if self.smooth_scale is not None:
smooth_scale = self.smooth_scale[ttype].to(tensor.dtype).to(tensor.device).contiguous()
assert smooth_scale.shape[0] == num_heads * head_dim
if self.reorder_idx is not None:
reorder_idx = self.reorder_idx[ttype].to(torch.int16).to(tensor.device)
# print(f"cache device: {tensor.device}")
gst_idx = self.group_st_idx[ttype].to(torch.int16).to(tensor.device)
assert reorder_idx.shape[0] == num_heads * head_dim
if self.fake_quant:
# [bs, seq_len, num_heads, head_dim]
fake_quant = skvq_quant.skvq_quant_fake(
t_reshape,
gst_idx,
reorder_idx,
smooth_scale,
qbits, self.gsize, self.hidden, self.fp8, self.clipping[self.layer_idx],
)
# [bs, num_heads, seq_len, head_dim]
fake_quant = fake_quant.transpose(1,2).contiguous()
return fake_quant, None, None
# [bs, seq_len, pack_hidden], [bs, seq_len, num_groups]
pack, scale, zp = skvq_quant.skvq_quant_pack(
t_reshape,
gst_idx,
reorder_idx,
smooth_scale,
qbits, self.gsize, self.hidden, self.fp8, self.clipping[self.layer_idx],
)
return pack, scale, zp
def dequant_pytorch(
self,
ttype: Literal["k", "v"],
tensor: torch.Tensor,
scale: torch.Tensor,
zp: torch.Tensor,
):
raise NotImplementedError()
def dequant_cuda(
self,
ttype: Literal["k", "v"],
pack: torch.Tensor,
scale: torch.Tensor,
zp: torch.Tensor,
):
# [bs, seq_len, pack_hidden]
assert len(pack.shape) == 3
qbits = self.K_target_bitwidth if ttype == "k" else self.V_target_bitwidth
smooth_scale = None
reorder_idx = None
gst_idx = None
if self.smooth_scale is not None:
smooth_scale = self.smooth_scale[ttype].to(scale.dtype).to(scale.device)
if self.reorder_idx is not None:
reorder_idx = self.reorder_idx[ttype].to(scale.device)
gst_idx = self.group_st_idx[ttype].to(scale.device)
dequant = skvq_quant.skvq_dequant_unpack(
pack, scale, zp,
gst_idx,
reorder_idx,
smooth_scale,
qbits, self.gsize, self.hidden, self.fp8,
)
return dequant
def quantization(
self,
ttype: Literal["k", "v"],
tensor: torch.Tensor,
impl: Literal["py", "cuda", "triton"] = "cuda",
):
assert ttype in ["k", "v"]
assert impl in ["py", "cuda"]
if self.KIVI_mode:
impl = "py"
if tensor is None:
return None, None, None
if impl == "py":
return self.quant_pytorch(ttype, tensor)
elif impl == "cuda":
return self.quant_cuda(ttype, tensor)
elif impl == "triton":
raise ValueError(f"{impl} not supported")
else:
raise ValueError(f"{impl} not supported")
def forward(self, K, V):
quantized_K = self.quantization("k", K)
quantized_V = self.quantization("v", V)
return quantized_K, quantized_V
def dequant(
self,
ttype: Literal["k", "v"],
quant_data: torch.Tensor,
scale: torch.Tensor,
zp: torch.Tensor,
impl: Literal["py", "cuda", "triton"] = "cuda",
):
assert ttype in ["k", "v"]
assert impl in ["py", "cuda", "triton"]
if impl == "py":
return self.dequant_pytorch(ttype, quant_data, scale, zp)
elif impl == "cuda":
return self.dequant_cuda(ttype, quant_data, scale, zp)
elif impl == "triton":
raise ValueError(f"{impl} not supported")
else:
raise ValueError(f"{impl} not supported")