quantizer.py 17.3 KB
Newer Older
Ji Lin's avatar
Ji Lin committed
1
import torch
2
import inspect
Casper's avatar
Casper committed
3
4
5
6
import logging
import functools
import torch.nn as nn
from tqdm import tqdm
Vik Paruchuri's avatar
Vik Paruchuri committed
7
from typing import Dict, List
Casper's avatar
Casper committed
8
9
10
from collections import defaultdict
from awq.utils.utils import clear_memory
from awq.utils.calib_data import get_calib_dataset
Casper Hansen's avatar
Casper Hansen committed
11
from awq.quantize.scale import apply_scale, apply_clip
Casper's avatar
Casper committed
12
from awq.modules.linear import WQLinear_GEMM, WQLinear_GEMV
13
14
15
16
17
18
19
from awq.utils.module import (
    append_str_prefix,
    get_op_name,
    get_named_linears,
    set_op_by_name,
    exclude_layers_to_not_quantize
)
Casper's avatar
Casper committed
20
21
22


class AwqQuantizer:
23
    def __init__(self, awq_model, model, tokenizer, w_bit, group_size, version, 
24
25
                       calib_data, split, text_column, duo_scaling, modules_to_not_convert=None,
                       export_compatible=False) -> None:
Casper Hansen's avatar
Casper Hansen committed
26
        self.awq_model = awq_model
Casper's avatar
Casper committed
27
28
29
30
31
32
33
34
        self.model = model
        self.tokenizer = tokenizer
        self.w_bit = w_bit
        self.group_size = group_size
        self.version = version
        self.calib_data = calib_data
        self.split = split
        self.text_column = text_column
35
        self.duo_scaling = duo_scaling
36
        self.export_compatible = export_compatible
37
        self.modules_to_not_convert = modules_to_not_convert if modules_to_not_convert is not None else []
Casper Hansen's avatar
Casper Hansen committed
38
        self.modules, self.module_kwargs, self.inps = self.init_quant()
39
    
Casper's avatar
Casper committed
40
41
42
43
44
45
46
47
    def pseudo_quantize_tensor(self, w: torch.Tensor, get_scale_zp=False):
        org_w_shape = w.shape
        if self.group_size > 0:
            assert org_w_shape[-1] % self.group_size == 0
            w = w.reshape(-1, self.group_size)
        assert w.dim() == 2

        # zero point quantization
Ji Lin's avatar
Ji Lin committed
48
49
        max_val = w.amax(dim=1, keepdim=True)
        min_val = w.amin(dim=1, keepdim=True)
Casper's avatar
Casper committed
50
        max_int = 2 ** self.w_bit - 1
Ji Lin's avatar
Ji Lin committed
51
52
53
        min_int = 0
        scales = (max_val - min_val).clamp(min=1e-5) / max_int
        zeros = (-torch.round(min_val / scales)).clamp_(min_int, max_int)
Casper's avatar
Casper committed
54
55
56
57
58
59
60
61
62
63
64
65
66
67

        assert torch.isnan(scales).sum() == 0
        assert torch.isnan(w).sum() == 0

        w = (torch.clamp(torch.round(w / scales) + zeros, min_int, max_int) - zeros) * scales
        assert torch.isnan(w).sum() == 0

        w = w.reshape(org_w_shape)

        if get_scale_zp:
            return w, scales.view(w.shape[0], -1), zeros.view(w.shape[0], -1)
        else:
            return w
    
Casper's avatar
Casper committed
68
69
70
71
72
73
74
75
76
77
78
79
80
    def pseudo_dequantize_tensor(self, w: nn.Linear, scales: torch.Tensor, zeros: torch.Tensor):
        # get repeated count
        repeat_count = w.weight.data.shape[-1] // zeros.shape[-1]

        # get zeros and scales in correct shape
        zeros = zeros.repeat(1, repeat_count).reshape(w.weight.data.shape)
        scales = scales.repeat(1, repeat_count).reshape(w.weight.data.shape)

        # dequantize
        w = (w.weight.data - zeros) * scales

        return w
    
Casper Hansen's avatar
Casper Hansen committed
81
82
    def quantize(self):
        for i in tqdm(range(len(self.modules)), desc="AWQ"):
Casper's avatar
Casper committed
83
84
85
            # Move module and inputs to correct device
            common_device = next(self.modules[i].parameters()).device
            if common_device is None or str(common_device) == "cpu":
86
                self.modules[i] = self.modules[i].cuda("cuda:" + str(i % torch.cuda.device_count()))
Casper's avatar
Casper committed
87
                common_device = next(self.modules[i].parameters()).device
88
89
90
91
92
93
94

            if self.module_kwargs.get("position_ids") is not None:
                self.module_kwargs["position_ids"] = self.module_kwargs["position_ids"].to(common_device)

            if self.module_kwargs.get("attention_mask") is not None:
                self.module_kwargs["attention_mask"] = self.module_kwargs["attention_mask"].to(common_device)

Casper's avatar
Casper committed
95
96
            self.inps = self.inps.to(common_device)

Casper's avatar
Casper committed
97
98
            # [STEP 1]: Get layer, extract linear modules, extract input features
            named_linears = get_named_linears(self.modules[i])
99
100

            # Filter out the linear layers we don't want to exclude
101
            named_linears = exclude_layers_to_not_quantize(named_linears, self.modules_to_not_convert)
102

Casper's avatar
Casper committed
103
104
105
106
            input_feat = self._get_input_feat(self.modules[i], named_linears)
            clear_memory()

            # [STEP 2]: Compute and apply scale list
Vik Paruchuri's avatar
Vik Paruchuri committed
107
            module_config: List[Dict] = self.awq_model.get_layers_for_scaling(
Casper's avatar
Casper committed
108
109
                self.modules[i], input_feat, self.module_kwargs
            )
Casper Hansen's avatar
Casper Hansen committed
110
            scales_list = [self._search_best_scale(self.modules[i], **layer) for layer in module_config]
Casper's avatar
Casper committed
111
112
113
114
            apply_scale(self.modules[i], scales_list, input_feat_dict=input_feat)
            scales_list = append_str_prefix(scales_list, get_op_name(self.model, self.modules[i]) + ".")

            # [STEP 3]: Compute and apply clipping list
Casper Hansen's avatar
Casper Hansen committed
115
116
117
            clip_list = self._search_best_clip(self.modules[i], named_linears, input_feat)
            apply_clip(self.modules[i], clip_list)
            clip_list = append_str_prefix(clip_list, get_op_name(self.model, self.modules[i]) + ".")
Casper's avatar
Casper committed
118
119

            # [STEP 4]: Quantize weights
120
121
122
123
124
125
126
127
128
            if not self.export_compatible:
                self._apply_quant(self.modules[i], named_linears)
            
            clear_memory()
    
    def pack(self):
        for i in tqdm(range(len(self.modules)), desc="Packing"):
            named_linears = get_named_linears(self.modules[i])
            named_linears = exclude_layers_to_not_quantize(named_linears, self.modules_to_not_convert)
129
130
131
            self._apply_quant(self.modules[i], named_linears)
            clear_memory()
    
Vik Paruchuri's avatar
Vik Paruchuri committed
132
    def _apply_quant(self, module, named_linears: Dict[str, nn.Linear]):
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
        for name, linear_layer in named_linears.items():
            # NOTE: small regression in perplexity if linear layer uses .cpu().float()
            linear_layer = linear_layer.cuda().half()

            linear_layer.weight.data, scales, zeros = self.pseudo_quantize_tensor(
                linear_layer.weight.data, 
                get_scale_zp=True
            )

            if self.version == 'GEMM':
                scales = scales.t().contiguous()
                zeros = zeros.t().contiguous()
                q_linear_module = WQLinear_GEMM

            elif self.version  == 'GEMV':
                q_linear_module = WQLinear_GEMV
Casper's avatar
Casper committed
149
            
150
151
152
153
154
155
156
157
158
159
160
161
            q_linear = q_linear_module.from_linear(
                linear=linear_layer,
                w_bit=self.w_bit,
                group_size=self.group_size,
                init_only=False,
                scales=scales,
                zeros=zeros
            )

            linear_layer.cpu()
            q_linear.to(next(module.parameters()).device)
            set_op_by_name(module, name, q_linear)
Casper's avatar
Casper committed
162
163
164
            clear_memory()

    @torch.no_grad()
Vik Paruchuri's avatar
Vik Paruchuri committed
165
    def _search_best_scale(self, module, prev_op, layers: List[nn.Linear], inp: torch.Tensor, module2inspect=None, kwargs={}):
Casper Hansen's avatar
Casper Hansen committed
166
167
168
169
170
171
172
        if module2inspect is None:
            assert len(layers) == 1
            module2inspect = layers[0]
        
        if "use_cache" in kwargs:
            kwargs.pop("use_cache")
        
Casper's avatar
Casper committed
173
        # Put x on the right device
Casper Hansen's avatar
Casper Hansen committed
174
        inp = inp.to(next(module2inspect.parameters()).device)
Casper's avatar
Casper committed
175
176

        # [STEP 1]: Compute maximum of weight
Casper Hansen's avatar
Casper Hansen committed
177
178
        weight = torch.cat([_m.weight for _m in layers], dim=0)
        org_shape = weight.shape
Casper's avatar
Casper committed
179
        weight = weight.view(-1, self.group_size)
Casper Hansen's avatar
Casper Hansen committed
180
181
182
        w_scale = weight.abs() / weight.abs().amax(dim=1, keepdim=True)
        w_scale = w_scale.view(org_shape)
        w_max = w_scale.mean(0)
Casper's avatar
Casper committed
183
184
185
        clear_memory(weight)

        # [STEP 2]: Compute maximum of x
Casper Hansen's avatar
Casper Hansen committed
186
        x_max = inp.abs().view(-1, inp.shape[-1]).mean(0)
Casper's avatar
Casper committed
187

Casper Hansen's avatar
Casper Hansen committed
188
        # [STEP 3]: Compute output of module
Casper's avatar
Casper committed
189
        with torch.no_grad():
190
191
192
            module_kwargs = self._sanitize_kwargs(kwargs, module2inspect)

            fp16_output = module2inspect(inp, **module_kwargs)
193
194
            if isinstance(fp16_output, tuple):
                fp16_output = fp16_output[0]
Casper's avatar
Casper committed
195
196
197
        
        # [STEP 4]: Compute loss
        best_scales = self._compute_best_scale(
Casper Hansen's avatar
Casper Hansen committed
198
            inp, w_max, x_max, module2inspect,
199
            layers, fp16_output, module_kwargs
Casper's avatar
Casper committed
200
201
        )
        
Casper Hansen's avatar
Casper Hansen committed
202
        return (get_op_name(module, prev_op), tuple([get_op_name(module, m) for m in layers]), best_scales)
Casper's avatar
Casper committed
203

Vik Paruchuri's avatar
Vik Paruchuri committed
204
    def _compute_best_scale(self, x, w_max, x_max, module2inspect, linears2scale: List[nn.Linear],
205
                                  fp16_output, kwargs={}):
Casper's avatar
Casper committed
206
207
208
        """
        Compute loss and select best scales

Casper's avatar
Casper committed
209
        L(s) = || Q(W * s) (s^-1 * X) - W * X ||
Casper's avatar
Casper committed
210
211
212
213
214
215
216
217
218
219
220
        Q: weight quantization function | pseudo_quantize_tensor(W * s)
        X: inputs from calib dataset    | X
        W: original weights in FP16     | layer
        s: per channel scaling factor   | s^-1 * X
        """
        n_grid = 20
        history = []
        best_ratio = -1
        best_scales = None
        best_error = float('inf')

Casper Hansen's avatar
Casper Hansen committed
221
        org_sd = {k: v.cpu() for k, v in module2inspect.state_dict().items()}
Casper's avatar
Casper committed
222
223
224
225
226
        
        device = x.device
        x_max = x_max.view(-1).to(device)
        w_max = w_max.view(-1).to(device)
        
Casper's avatar
Casper committed
227
228
        for ratio in range(n_grid):
            # create new scales
Casper's avatar
Casper committed
229
            ratio = ratio / n_grid
230

Casper Hansen's avatar
Casper Hansen committed
231
            # NOTE: s^-1 * x is fused here, according to paper
232
233
234
235
            if self.duo_scaling:
                scales = (x_max.pow(ratio) / w_max.pow(1-ratio)).clamp(min=1e-4)
            else:
                scales = x_max.pow(ratio).clamp(min=1e-4).view(-1)
Casper's avatar
Casper committed
236
            scales = scales / (scales.max() * scales.min()).sqrt()
Casper's avatar
Casper committed
237
            scales_view = scales.view(1, -1).to(device)
238

Casper Hansen's avatar
Casper Hansen committed
239
            # Q(W * s)
Casper's avatar
Casper committed
240
            for fc in linears2scale:
Casper's avatar
Casper committed
241
242
                fc.weight.mul_(scales_view)
                fc.weight.data = self.pseudo_quantize_tensor(fc.weight.data) / scales_view
Casper's avatar
Casper committed
243

244
245
246
247
248
            # W * X
            int_w_output = module2inspect(x, **kwargs)
            if isinstance(int_w_output, tuple):
                int_w_output = int_w_output[0]
            
Casper Hansen's avatar
Casper Hansen committed
249
250
            # compute mean squared error (L2 norm)
            loss = (fp16_output - int_w_output).float().pow(2).mean().item() # NOTE: float prevents overflow
Casper's avatar
Casper committed
251
252

            history.append(loss)
Casper's avatar
Casper committed
253
            if loss < best_error:
Casper's avatar
Casper committed
254
255
                best_error = loss
                best_ratio = ratio
Casper's avatar
Casper committed
256
                best_scales = scales.clone()
Casper Hansen's avatar
Casper Hansen committed
257
            module2inspect.load_state_dict(org_sd)
Casper's avatar
Casper committed
258

Casper's avatar
Casper committed
259
260
261
262
263
264
        if best_ratio == -1:
            logging.debug(history)
            raise Exception

        assert torch.isnan(best_scales).sum() == 0, best_scales

Casper Hansen's avatar
Casper Hansen committed
265
        return best_scales.detach().cpu()
Casper's avatar
Casper committed
266

Casper Hansen's avatar
Casper Hansen committed
267
268
269
270
    @torch.no_grad()
    def _search_best_clip(self, layer, named_linears, input_feat):
        clip_list = []
        avoid_clipping = ["q_", "k_", "query", "key", "Wqkv"]
Casper's avatar
Casper committed
271

Casper Hansen's avatar
Casper Hansen committed
272
273
274
275
276
277
278
279
280
281
        for name in named_linears:
            # due to qk bmm, it is hard to clip precisely
            if any([_ in name for _ in avoid_clipping]):
                continue

            named_linears[name].cuda()
            max_val = self._compute_best_clip(named_linears[name].weight, input_feat[name])
            clip_list.append((name, max_val))

            named_linears[name].cpu()
Casper Hansen's avatar
Casper Hansen committed
282
283
        
        return clip_list
Casper Hansen's avatar
Casper Hansen committed
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300

    @torch.no_grad()
    def _compute_best_clip(self, w: torch.Tensor, input_feat: torch.Tensor, n_grid=20, max_shrink=0.5, n_sample_token=512):
        assert w.dim() == 2
        org_w_shape = w.shape
        # w           [co, ci]      -> [co, 1, n_group, group size]
        # input_feat  [n_token, ci] -> [1, n_token, n_group, group size]
        group_size = self.group_size if self.group_size > 0 else w.shape[1]
        input_feat = input_feat.view(-1, input_feat.shape[-1])
        input_feat = input_feat.reshape(1, input_feat.shape[0], -1, group_size)
        input_feat = input_feat[:, 0::input_feat.shape[1] // n_sample_token]
        w = w.reshape(w.shape[0], 1, -1, group_size)

        oc_batch_size = 256 if w.shape[0] % 256 == 0 else 64  # prevent OOM
        assert w.shape[0] % oc_batch_size == 0
        w_all = w
        best_max_val_all = []
Casper's avatar
Casper committed
301

Casper Hansen's avatar
Casper Hansen committed
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
        for i_b in range(w.shape[0] // oc_batch_size):
            w = w_all[i_b * oc_batch_size: (i_b + 1) * oc_batch_size]

            org_max_val = w.abs().amax(dim=-1, keepdim=True)  # co, 1, n_group, 1

            best_max_val = org_max_val.clone()
            min_errs = torch.ones_like(org_max_val) * 1e9
            input_feat = input_feat.to(w.device)
            org_out = (input_feat * w).sum(dim=-1)  # co, n_token, n_group

            for i_s in range(int(max_shrink * n_grid)):
                max_val = org_max_val * (1 - i_s / n_grid)
                min_val = - max_val
                cur_w = torch.clamp(w, min_val, max_val)
                q_w = self.pseudo_quantize_tensor(cur_w)
                cur_out = (input_feat * q_w).sum(dim=-1)

                # co, 1, n_group, 1
                err = (cur_out - org_out).pow(2).mean(dim=1).view(min_errs.shape)
                del cur_w
                del cur_out
                cur_best_idx = err < min_errs
                min_errs[cur_best_idx] = err[cur_best_idx]
                best_max_val[cur_best_idx] = max_val[cur_best_idx]
            best_max_val_all.append(best_max_val)

        best_max_val = torch.cat(best_max_val_all, dim=0)

        clear_memory(input_feat)
        clear_memory(org_out)

        return best_max_val.squeeze(1)

    def init_quant(self, n_samples=128, seqlen=512):
        modules = self.awq_model.get_model_layers(self.model)
Casper's avatar
Casper committed
337
338
339
340
341
342
343
344
345
        samples = get_calib_dataset(
            data=self.calib_data, tokenizer=self.tokenizer, n_samples=n_samples, block_size=seqlen,
            split=self.split, text_column=self.text_column
        )
        samples = torch.cat(samples, dim=0)

        inps = []
        layer_kwargs = {}

Casper Hansen's avatar
Casper Hansen committed
346
347
        modules[0] = modules[0].cuda()
        self.awq_model.move_embed(self.model, "cuda")
Casper's avatar
Casper committed
348
349
350
351
352
353
354
355
356
        
        # get input and kwargs to layer 0
        # with_kwargs is only supported in PyTorch 2.0
        # use this Catcher hack for now
        class Catcher(nn.Module):
            def __init__(self, module):
                super().__init__()
                self.module = module

357
358
359
360
361
362
363
364
365
366
            def forward(self, *args, **kwargs):
                # assume first input to forward is hidden states
                if len(args) > 0:
                    hidden_states = args[0]
                    del args
                else:
                    first_key = list(kwargs.keys())[0]
                    hidden_states = kwargs.pop(first_key)

                inps.append(hidden_states)
Casper's avatar
Casper committed
367
368
369
370
                layer_kwargs.update(kwargs)
                raise ValueError  # early exit to break later inference

        # patch layer 0 to catch input and kwargs
Casper Hansen's avatar
Casper Hansen committed
371
        modules[0] = Catcher(modules[0])
Casper's avatar
Casper committed
372
373
374
375
        try:
            self.model(samples.to(next(self.model.parameters()).device))
        except ValueError:  # work with early exit
            pass
376
377
378
379
380
381
382
        
        # Update the layer kwargs with `prepare_inputs_for_generation` method
        # that takes care of everything to avoid unexpected errors.
        layer_kwargs = self.model.prepare_inputs_for_generation(samples, **layer_kwargs)
        # Pop the input_ids as they are not needed at all.
        layer_kwargs.pop("input_ids")

Casper's avatar
Casper committed
383
        del samples
Casper Hansen's avatar
Casper Hansen committed
384
        modules[0] = modules[0].module  # restore
Casper's avatar
Casper committed
385
386
        inps = inps[0]

Casper Hansen's avatar
Casper Hansen committed
387
388
        modules[0] = modules[0].cpu()
        self.awq_model.move_embed(self.model, "cpu")
Casper's avatar
Casper committed
389
390
        
        clear_memory()
391
        
Casper's avatar
Casper committed
392
        if layer_kwargs.get("attention_mask") is not None:
393
            layer_kwargs["attention_mask"] = layer_kwargs["attention_mask"].to("cuda")
Casper's avatar
Casper committed
394

Casper Hansen's avatar
Casper Hansen committed
395
        return modules, layer_kwargs, inps
Casper's avatar
Casper committed
396
397
398
399
400
401
402
403
404
405
    
    def _get_input_feat(self, layer, named_linears):
        # firstly, get input features of all linear layers
        def cache_input_hook(m, x, y, name, feat_dict):
            x = x[0]
            x = x.detach().cpu()
            feat_dict[name].append(x)

        input_feat = defaultdict(list)
        handles = []
406
407
408
409
410

        # FIXME: Workaround for Mixtral to use block_sparse_moe input features
        if self.awq_model.model_type == "mixtral":
            named_linears = {**named_linears, "block_sparse_moe": layer.block_sparse_moe}

Casper's avatar
Casper committed
411
412
413
414
        for name in named_linears:
            handles.append(named_linears[name].register_forward_hook(
                functools.partial(cache_input_hook, name=name,
                                feat_dict=input_feat)))
Casper Hansen's avatar
Casper Hansen committed
415
        self.inps = self.inps.to(next(layer.parameters()).device)  # in case multi-gpu
Casper's avatar
Casper committed
416
        # get output as next layer's input
417
418
419
420
421
422
423
        
        # Sanitize the kwargs in case we use transformers version that contains
        # kwargs that are not handled by the module.
        # Useful for trust_remote_code models.
        module_kwargs = self._sanitize_kwargs(self.module_kwargs, layer)

        self.inps = layer(self.inps, **module_kwargs)[0]
Casper's avatar
Casper committed
424
425
426
427
428
429
        for h in handles:
            h.remove()
        # now solve for scaling and clipping
        input_feat = {k: torch.cat(v, dim=0) for k, v in input_feat.items()}
        
        return input_feat
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449


    def _sanitize_kwargs(self, inputs_kwargs, module):
        """
        Remove the arguments that are not supported in the module's
        forward pass to avoid breaking behaviour between different versions
        of transformers. 

        Args:
            inputs_kwargs (`dict`):
                The input dictionary to pass to the model layer
            module (`torch.nn.Module`):
                Target module to quantize.
        """
        module_signature = inspect.signature(module.forward).parameters
        sanitized_kwargs = {}
        for k, v in  inputs_kwargs.items():
            if k in module_signature:
                sanitized_kwargs[k] = v
        return sanitized_kwargs