File size: 14,025 Bytes
dc9bb20
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from dataclasses import dataclass
from math import prod
from typing import Optional
import warnings
from warnings import warn

import torch

import bitsandbytes.functional as F

# The inverse transformation for the colTuring and colAmpere format were contributed by Alex Borzunov:
# https://github.com/bigscience-workshop/petals/blob/main/src/petals/utils/linear8bitlt_patch.py


"""
    This class pools outlier dimensions across layers.
    This is particularly important for small models where outlier features
    are less systematic and occur with low frequency.
"""


class GlobalOutlierPooler:
    _instance = None

    def __init__(self):
        raise RuntimeError("Call get_instance() instead")

    def initialize(self):
        self.outliers = set()
        self.model_dim = None

    @classmethod
    def get_instance(cls):
        if cls._instance is None:
            cls._instance = cls.__new__(cls)
            cls._instance.initialize()
        return cls._instance

    def add_outliers(self, outlier_idx, feature_dim):
        if self.model_dim is None:
            self.model_dim = feature_dim
        if feature_dim != self.model_dim:
            return  # we do not encode outliers for the 2nd FFN layer

        self.outliers.update(outlier_idx.tolist())

    def get_current_outlier_idx(self):
        return torch.Tensor(list(self.outliers)).to(torch.int64)


_is_compiling = torch.compiler.is_compiling


@dataclass
class MatmulLtState:
    _tile_indices: Optional[torch.Tensor] = None  # TODO: remove

    force_no_igemmlt: bool = False

    CB: Optional[torch.Tensor] = None
    CxB: Optional[torch.Tensor] = None  # TODO: Deprecate/remove
    SB: Optional[torch.Tensor] = None
    SCB: Optional[torch.Tensor] = None

    CxBt: Optional[torch.Tensor] = None  # TODO: Deprecate/remove
    SBt: Optional[torch.Tensor] = None
    CBt: Optional[torch.Tensor] = None

    subB: Optional[torch.Tensor] = None

    outlier_pool: Optional[GlobalOutlierPooler] = None
    has_accumulated_gradients = False
    threshold = 0.0
    idx: Optional[torch.Tensor] = None
    is_training = True
    has_fp16_weights = True
    use_pool = False
    formatB = "row"  # TODO: Deprecate/remove

    def reset_grads(self):
        self.CB = None
        self.CxB = None
        self.SB = None
        self.SCB = None

        self.CxBt = None
        self.SBt = None
        self.CBt = None

    @property
    def tile_indices(self):
        raise ValueError("tile_indices is no longer supported.")


class MatMul8bitLt(torch.autograd.Function):
    @staticmethod
    def forward(
        ctx: torch.autograd.function.FunctionCtx,
        A: torch.Tensor,
        B: torch.Tensor,
        out: Optional[torch.Tensor] = None,
        bias: Optional[torch.Tensor] = None,
        state: Optional[MatmulLtState] = None,
    ):
        state = state or MatmulLtState()

        # default of pytorch behavior if inputs are empty
        ctx.is_empty = False
        if prod(A.shape) == 0:
            ctx.is_empty = True
            ctx.A = A
            ctx.B = B
            ctx.bias = bias
            if A.shape[-1] == B.shape[0]:
                return torch.empty(A.shape[:-1] + B.shape[1:], dtype=A.dtype, device=A.device)
            else:
                return torch.empty(A.shape[:-1] + B.shape[:1], dtype=A.dtype, device=A.device)

        input_shape = A.shape

        # Cast A to fp16
        if A.dtype != torch.float16 and not _is_compiling():
            warnings.warn(f"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization")

        if len(A.shape) == 3:
            A = A.reshape(-1, A.shape[-1])

        # 1. Quantize A. Note that as a side-effect, outliers are suppressed in CA/CAt.
        if ctx.needs_input_grad[1]:
            # Slower path
            CA, CAt, SCA, SCAt, outlier_cols = F.int8_double_quant(A.to(torch.float16), threshold=state.threshold)
        else:
            # Fast path
            CA, SCA, outlier_cols = F.int8_vectorwise_quant(A.to(torch.float16), threshold=state.threshold)
            CAt = SCAt = None

        has_grad = False

        if state.has_fp16_weights or state.CB is None:
            has_grad = getattr(B, "grad", None) is not None
            is_transposed = not B.is_contiguous() and B.shape[0] == B.stride(1)
            if is_transposed:
                B = B.contiguous()

            if (state.is_training and not has_grad) or state.CB is None or state.SCB is None:
                state.reset_grads()

                # 2. Quantize B
                state.CB, state.SCB, _ = F.int8_vectorwise_quant(B.to(torch.float16))

        # Handle sparse decomposition
        if state.threshold > 0.0:
            state.idx = outlier_cols

            # Mixed Int8 Matmul + Dequant + Bias
            output, subA = torch.ops.bitsandbytes.int8_mixed_scaled_mm(
                A,
                CA,
                state.CB,
                SCA,
                state.SCB,
                outlier_cols,
                bias,
            )

        else:
            # Int8 Matmul + Dequant + Bias
            output = torch.ops.bitsandbytes.int8_scaled_mm.default(
                CA, state.CB, SCA, state.SCB, bias=bias, dtype=A.dtype
            )
            subA = None

        # 5. Save state
        ctx.state = state

        ctx.grad_shape = input_shape
        ctx.dtype_A = A.dtype
        ctx.dtype_bias = None if bias is None else bias.dtype

        if any(ctx.needs_input_grad[:2]):
            ctx.tensors = (CAt, subA, A)
            ctx.tensor_states = (SCAt, state.idx)
        else:
            ctx.tensors = [None, None, None]
            ctx.tensor_states = (None, None)
            ctx.save_for_backward(None, None)

        output_shape = (*input_shape[:-1], state.CB.shape[0])

        if len(input_shape) == 3:
            return output.reshape(output_shape)

        return output

    @staticmethod
    def backward(ctx: torch.autograd.function.FunctionCtx, grad_output: torch.Tensor):
        if ctx.is_empty:
            bias_grad = None if ctx.bias is None else torch.zeros_like(ctx.bias)
            return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None

        req_gradA, req_gradB, _, req_gradBias, _ = ctx.needs_input_grad
        CAt, subA, _A = ctx.tensors
        SCAt, idx = ctx.tensor_states
        state: MatmulLtState = ctx.state
        grad_A = grad_B = grad_bias = None

        if req_gradBias:
            # compute grad_bias first before changing grad_output dtype
            grad_bias = grad_output.sum(0, dtype=ctx.dtype_bias)

        # Cast grad_output to fp16
        if len(grad_output.shape) == 3:
            grad_output = grad_output.reshape(-1, grad_output.shape[-1]).contiguous()

        if req_gradB:
            Cgrad, _, _, SCgradt, _ = F.int8_double_quant(grad_output.to(torch.float16))

            grad_B = torch.ops.bitsandbytes.int8_scaled_mm.default(
                Cgrad.t().contiguous(),
                CAt.t(),
                SCgradt,
                SCAt,
                dtype=torch.float16,
            )

            if state.threshold > 0.0 and subA is not None and subA.numel() > 0:
                grad_B[:, idx] += torch.matmul(grad_output.t(), subA)

        if req_gradA:
            if state.CB is not None:
                CB = state.CB.to(ctx.dtype_A, copy=True).mul_(state.SCB.unsqueeze(1).mul(1.0 / 127.0))
                grad_A = torch.matmul(grad_output.to(ctx.dtype_A), CB).view(ctx.grad_shape)
            else:
                raise Exception("State must contain CB matrix for backward")

        return grad_A, grad_B, None, grad_bias, None


class MatMul8bitFp(torch.autograd.Function):
    # For Intel CPU and XPU MatMul8bitFp is much faster (~3x) than MatMul8bitLt in finetune.
    # Because the MatMul8bitLt has more mechanisms in computing grad.
    # We don't have fast kernel for quant/dequant 8bit in CPU/XPU, so it's very slow.
    # We'd like to use dequant + matmul to run finetune with good performance.

    @staticmethod
    def forward(ctx, A, B, out=None, bias=None, state=MatmulLtState):
        if state.has_fp16_weights or state.CB is None:
            has_grad = getattr(B, "grad", None) is not None
            is_transposed = not B.is_contiguous() and B.shape[0] == B.stride(1)
            if is_transposed:
                B = B.contiguous()

            if (state.is_training and not has_grad) or state.CB is None or state.SCB is None:
                state.reset_grads()
                state.CB, state.SCB, _ = F.int8_vectorwise_quant(B.to(torch.float16))
                B = state.CB

        CB = state.CB.data.to(A.dtype).mul_(state.SCB.unsqueeze(1).mul(1.0 / 127.0))
        output = torch.nn.functional.linear(A, CB, bias)
        ctx.state = state
        ctx.dtype_A = A.dtype
        ctx.grad_shape = A.shape
        ctx.A = A
        ctx.dtype_bias = None if bias is None else bias.dtype
        return output

    @staticmethod
    def backward(ctx, grad_output):
        req_gradA, req_gradB, _, req_gradBias, _ = ctx.needs_input_grad
        A = ctx.A
        state = ctx.state
        grad_A = grad_B = grad_bias = None
        if req_gradBias:
            # compute grad_bias first before changing grad_output dtype
            grad_bias = grad_output.sum(0, dtype=ctx.dtype_bias)

        # Cast grad_output to fp16
        if len(grad_output.shape) == 3:
            grad_output = grad_output.reshape(-1, grad_output.shape[-1]).contiguous()

        if req_gradB:
            grad_B = torch.matmul(A.t(), grad_output).t()

        if req_gradA:
            if state.CB is not None:
                CB = state.CB.to(ctx.dtype_A, copy=True).mul_(state.SCB.unsqueeze(1).mul(1.0 / 127.0))
                grad_A = torch.matmul(grad_output.to(ctx.dtype_A), CB).view(ctx.grad_shape)
            else:
                raise Exception("State must contain CB matrix for backward")

        return grad_A, grad_B, None, grad_bias, None


class MatMul4Bit(torch.autograd.Function):
    # forward is the same, but we added the fallback for pre-turing GPUs
    # backward is mostly the same, but adds one extra clause (see "elif state.CxB is not None")

    @staticmethod
    def forward(ctx, A, B, out=None, bias=None, quant_state: Optional[F.QuantState] = None):
        # default of pytorch behavior if inputs are empty
        ctx.is_empty = False
        if prod(A.shape) == 0:
            ctx.is_empty = True
            ctx.A = A
            ctx.B = B
            ctx.bias = bias
            B_shape = quant_state.shape
            if A.shape[-1] == B_shape[0]:
                return torch.empty(A.shape[:-1] + B_shape[1:], dtype=A.dtype, device=A.device)
            else:
                return torch.empty(A.shape[:-1] + B_shape[:1], dtype=A.dtype, device=A.device)

        # 1. Dequantize
        # 2. MatmulnN
        output = torch.nn.functional.linear(A, F.dequantize_4bit(B, quant_state).to(A.dtype).t(), bias)

        # 3. Save state
        ctx.state = quant_state
        ctx.dtype_A, ctx.dtype_B, ctx.dtype_bias = A.dtype, B.dtype, None if bias is None else bias.dtype

        if any(ctx.needs_input_grad[:2]):
            ctx.tensors = (None, B)
        else:
            ctx.tensors = (None, None)

        return output

    @staticmethod
    def backward(ctx, grad_output):
        if ctx.is_empty:
            bias_grad = None if ctx.bias is None else torch.zeros_like(ctx.bias)
            return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None

        req_gradA, _, _, req_gradBias, _ = ctx.needs_input_grad
        _, B = ctx.tensors

        grad_A, grad_B, grad_bias = None, None, None

        if req_gradBias:
            # compute grad_bias first before changing grad_output dtype
            grad_bias = grad_output.sum(0, dtype=ctx.dtype_bias)

        # not supported by PyTorch. TODO: create work-around
        # if req_gradB: grad_B = torch.matmul(grad_output.t(), A)
        if req_gradA:
            grad_A = torch.matmul(grad_output, F.dequantize_4bit(B, ctx.state).to(grad_output.dtype).t())

        return grad_A, grad_B, None, grad_bias, None


def matmul(
    A: torch.Tensor,
    B: torch.Tensor,
    out: Optional[torch.Tensor] = None,
    state: Optional[MatmulLtState] = None,
    threshold=0.0,
    bias: Optional[torch.Tensor] = None,
):
    state = state or MatmulLtState()
    if threshold > 0.0:
        state.threshold = threshold
    # MatMul8bitLt is slower because no fast kernel for quant/dequant 8bit in CPU/XPU
    if state.is_training:
        if A.device.type in ("cpu", "xpu"):
            return MatMul8bitFp.apply(A, B, out, bias, state)
    return MatMul8bitLt.apply(A, B, out, bias, state)


def matmul_4bit(
    A: torch.Tensor,
    B: torch.Tensor,
    quant_state: F.QuantState,
    out: Optional[torch.Tensor] = None,
    bias: Optional[torch.Tensor] = None,
):
    assert quant_state is not None
    # Change dtype to input dtype on CPU
    if A.device.type == "cpu":
        quant_state.dtype = A.dtype

        if getattr(quant_state, "packing_format_for_cpu", False):
            out = F.gemv_4bit(A, B, out, state=quant_state)
            if bias is not None:
                out += bias
            return out
        else:
            return MatMul4Bit.apply(A, B, out, bias, quant_state)

    if A.numel() == A.shape[-1] and A.requires_grad == False and A.device.type != "hpu":
        if A.shape[-1] % quant_state.blocksize != 0:
            warn(
                f"Some matrices hidden dimension is not a multiple of {quant_state.blocksize} and efficient inference kernels are not supported for these (slow). Matrix input size found: {A.shape}",
            )
            return MatMul4Bit.apply(A, B, out, bias, quant_state)
        else:
            out = F.gemv_4bit(A, B.t(), out, state=quant_state)
            if bias is not None:
                out += bias
            return out
    else:
        return MatMul4Bit.apply(A, B, out, bias, quant_state)