File size: 13,379 Bytes
f7eb3fa
 
 
 
 
 
57c2394
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f7eb3fa
 
 
 
 
 
 
 
 
 
 
 
57c2394
f7eb3fa
57c2394
f7eb3fa
 
 
 
 
 
 
 
 
 
 
 
57c2394
 
f7eb3fa
 
 
 
 
57c2394
 
f7eb3fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
57c2394
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import torch

from ._ops import ops

__all__ = [
    "matmul_packed_w4a4_int8",
    "matmul_packed_adaln_int4_cpu",
    "matmul_packed_weight",
    "quantize_activations_cpu",
    "quantize_activations_int8",
    "quantize_activations_packed_w4",
    "supports_cpu_activation",
    "supports_cpu_adaln",
    "supports_device",
]


def supports_device(device_type: str) -> bool:
    dispatch_keys = {
        "cpu": "CPU",
        "cuda": "CUDA",
        "mps": "MPS",
    }
    try:
        dispatch_key = dispatch_keys[device_type]
    except KeyError as exc:
        raise ValueError(f"unknown device type {device_type!r}") from exc
    return bool(
        torch._C._dispatch_has_kernel_for_dispatch_key(  # noqa: SLF001
            ops.matmul_packed_weight._qualified_op_name,  # noqa: SLF001
            dispatch_key,
        )
    )


def supports_cpu_activation() -> bool:
    try:
        operation = ops.quantize_activations_cpu
        qualified_name = operation._qualified_op_name  # noqa: SLF001
    except (AttributeError, RuntimeError):
        return False
    return bool(
        torch._C._dispatch_has_kernel_for_dispatch_key(  # noqa: SLF001
            qualified_name,
            "CPU",
        )
    )


def supports_cpu_adaln() -> bool:
    try:
        operation = ops.matmul_packed_adaln_int4_cpu
        qualified_name = operation._qualified_op_name  # noqa: SLF001
    except (AttributeError, RuntimeError):
        return False
    return bool(
        torch._C._dispatch_has_kernel_for_dispatch_key(  # noqa: SLF001
            qualified_name,
            "CPU",
        )
    )


def quantize_activations_cpu(
    x: torch.Tensor,
    permutation: torch.Tensor,
    signs: torch.Tensor,
    centroids: torch.Tensor,
    boundaries: torch.Tensor,
    *,
    eps: float,
    inv_sqrt_block: float,
    block_size: int,
) -> torch.Tensor:
    if x.device.type != "cpu":
        raise RuntimeError("native CPU activation quantization requires CPU tensors")
    if x.dtype not in {torch.float32, torch.float16, torch.bfloat16}:
        raise ValueError("x must be float32, float16, or bfloat16")
    if block_size <= 0 or block_size & (block_size - 1):
        raise ValueError("block_size must be a positive power of two")
    dim = x.shape[-1]
    if dim % block_size != 0:
        raise ValueError("block_size must divide the input dimension")

    original_shape = x.shape
    values = x.contiguous().reshape(-1, dim)
    out = torch.empty_like(values)
    permutation_dtype = torch.int32 if permutation.dtype == torch.int32 else torch.int64
    permutation_values = permutation.to(device="cpu", dtype=permutation_dtype).contiguous()
    sign_values = signs.to(device="cpu", dtype=torch.int8).contiguous()
    centroid_values = centroids.to(device="cpu", dtype=torch.float32).contiguous()
    boundary_values = boundaries.to(device="cpu", dtype=torch.float32).contiguous()
    ops.quantize_activations_cpu(
        out,
        values,
        permutation_values,
        sign_values,
        centroid_values,
        boundary_values,
        eps,
        inv_sqrt_block,
        block_size,
    )
    return out.reshape(original_shape)


def matmul_packed_weight(
    x: torch.Tensor,
    packed_weight_indices: torch.Tensor,
    row_norms: torch.Tensor,
    centroids: torch.Tensor,
    *,
    bits: int,
    out_features: int,
    in_features: int,
    bias: torch.Tensor | None = None,
    block_m: int = 64,
    block_n: int = 64,
    block_k: int = 128,
) -> torch.Tensor:
    if bits <= 0 or bits > 8:
        raise ValueError("bits must be in [1, 8]")
    if x.shape[-1] != in_features:
        raise ValueError(f"expected input last dimension {in_features}, got {x.shape[-1]}")
    if block_m <= 0 or block_n <= 0 or block_k <= 0:
        raise ValueError("packed matmul tile sizes must be positive")

    original_shape = x.shape
    x_2d = x.contiguous().reshape(-1, in_features)
    out = torch.empty((x_2d.shape[0], out_features), device=x.device, dtype=x.dtype)
    packed = packed_weight_indices.to(device=x.device, dtype=torch.uint8).contiguous()
    auxiliary_dtype = torch.bfloat16 if x.device.type == "cuda" else torch.float32
    norms = row_norms.to(device=x.device, dtype=auxiliary_dtype).contiguous()
    centroid_values = centroids.to(device=x.device, dtype=torch.float32).contiguous()
    if bias is None:
        bias_values = torch.empty((1,), device=x.device, dtype=torch.float32)
        has_bias = False
    else:
        bias_dtype = x.dtype if x.device.type == "cuda" else torch.float32
        bias_values = bias.to(device=x.device, dtype=bias_dtype).contiguous()
        has_bias = True

    ops.matmul_packed_weight(
        out,
        x_2d,
        packed,
        norms,
        centroid_values,
        bias_values,
        has_bias,
        bits,
        out_features,
        in_features,
        block_m,
        block_n,
        block_k,
    )
    return out.reshape(*original_shape[:-1], out_features)


def matmul_packed_adaln_int4_cpu(
    x: torch.Tensor,
    packed_weight: torch.Tensor,
    scales: torch.Tensor,
    *,
    out_features: int,
    in_features: int,
    group_size: int,
    bias: torch.Tensor | None = None,
) -> torch.Tensor:
    if x.device.type != "cpu":
        raise RuntimeError("native packed AdaLN requires CPU tensors")
    if x.shape[-1] != in_features:
        raise ValueError(f"expected input last dimension {in_features}, got {x.shape[-1]}")
    if group_size <= 0:
        raise ValueError("group_size must be positive")

    original_shape = x.shape
    x_2d = x.to(dtype=torch.bfloat16).contiguous().reshape(-1, in_features)
    out = torch.empty((x_2d.shape[0], out_features), dtype=torch.bfloat16)
    packed = packed_weight.to(device="cpu", dtype=torch.uint8).contiguous()
    scale_values = scales.to(device="cpu", dtype=torch.float32).contiguous()
    if bias is None:
        bias_values = torch.empty((1,), dtype=torch.float32)
        has_bias = False
    else:
        bias_values = (
            bias.to(device="cpu", dtype=torch.bfloat16).to(dtype=torch.float32).contiguous()
        )
        has_bias = True
    ops.matmul_packed_adaln_int4_cpu(
        out,
        x_2d,
        packed,
        scale_values,
        bias_values,
        has_bias,
        out_features,
        in_features,
        group_size,
    )
    return out.reshape(*original_shape[:-1], out_features)


def matmul_packed_w4a4_int8(
    packed_activations: torch.Tensor,
    packed_weight_indices: torch.Tensor,
    token_norms: torch.Tensor,
    row_norms: torch.Tensor,
    activation_codes: torch.Tensor,
    weight_codes: torch.Tensor,
    *,
    activation_scale: float,
    weight_scale: float,
    out_features: int,
    in_features: int,
    bias: torch.Tensor | None = None,
    output_dtype: torch.dtype = torch.bfloat16,
    tile_m: int = 128,
    tile_n: int = 128,
    async_packed: bool = False,
    weight_k_major: bool = False,
) -> torch.Tensor:
    if not packed_activations.is_cuda:
        raise RuntimeError("packed W4A4 INT8 matmul requires CUDA tensors")
    if in_features <= 0 or in_features % 64 != 0:
        raise ValueError("in_features must be positive and divisible by 64")
    if packed_activations.shape[-1] != in_features // 2:
        raise ValueError(
            f"expected packed activation last dimension {in_features // 2}, "
            f"got {packed_activations.shape[-1]}"
        )
    if output_dtype not in {torch.float16, torch.bfloat16}:
        raise ValueError("output_dtype must be float16 or bfloat16")
    if (tile_m, tile_n) not in {(128, 128), (256, 128), (128, 256)}:
        raise ValueError("tile must be 128x128, 256x128, or 128x256")

    original_shape = packed_activations.shape
    activations = (
        packed_activations.to(dtype=torch.uint8).contiguous().reshape(-1, in_features // 2)
    )
    weights = packed_weight_indices.to(device=activations.device, dtype=torch.uint8).contiguous()
    norms = token_norms.to(device=activations.device, dtype=torch.float32).contiguous()
    weight_norms = row_norms.to(device=activations.device, dtype=torch.bfloat16).contiguous()
    activation_code_values = activation_codes.to(
        device=activations.device, dtype=torch.int8
    ).contiguous()
    weight_code_values = weight_codes.to(device=activations.device, dtype=torch.int8).contiguous()
    out = torch.empty(
        (activations.shape[0], out_features),
        device=activations.device,
        dtype=output_dtype,
    )
    if bias is None:
        bias_values = out
        has_bias = False
    else:
        bias_values = bias.to(device=activations.device, dtype=output_dtype).contiguous()
        has_bias = True

    ops.matmul_packed_w4a4_int8(
        out,
        activations,
        weights,
        norms,
        weight_norms,
        activation_code_values,
        weight_code_values,
        bias_values,
        has_bias,
        activation_scale,
        weight_scale,
        out_features,
        in_features,
        tile_m,
        tile_n,
        async_packed,
        weight_k_major,
    )
    return out.reshape(*original_shape[:-1], out_features)


def quantize_activations_packed_w4(
    x: torch.Tensor,
    permutation: torch.Tensor,
    signs: torch.Tensor,
    boundaries: torch.Tensor,
    *,
    eps: float = 1e-12,
    inv_sqrt_block: float,
    threads: int = 256,
) -> tuple[torch.Tensor, torch.Tensor]:
    if not x.is_cuda:
        raise RuntimeError("native packed W4 activation quantization requires CUDA tensors")
    if x.dtype not in {torch.float16, torch.bfloat16}:
        raise ValueError("x must be float16 or bfloat16")
    dim = x.shape[-1]
    if dim not in {512, 1024, 2048, 4096, 8192, 16384}:
        raise ValueError(
            "native packed W4 activation quantization supports dimensions "
            "512, 1024, 2048, 4096, 8192, and 16384"
        )
    if threads not in {128, 256, 512}:
        raise ValueError("threads must be 128, 256, or 512")
    if permutation.numel() != dim or signs.numel() != dim:
        raise ValueError("permutation and signs must match the input dimension")
    if boundaries.numel() != 15:
        raise ValueError("boundaries must contain 15 values")

    original_shape = x.shape
    values = x.contiguous().reshape(-1, dim)
    packed = torch.empty((values.shape[0], dim // 2), device=x.device, dtype=torch.uint8)
    norms = torch.empty(values.shape[0], device=x.device, dtype=torch.float32)
    permutation_dtype = torch.int32 if permutation.dtype == torch.int32 else torch.int64
    permutation_values = permutation.to(device=x.device, dtype=permutation_dtype).contiguous()
    sign_values = signs.to(device=x.device, dtype=torch.int8).contiguous()
    boundary_values = boundaries.to(device=x.device, dtype=torch.float32).contiguous()
    ops.quantize_activations_packed_w4(
        packed,
        norms,
        values,
        permutation_values,
        sign_values,
        boundary_values,
        eps,
        inv_sqrt_block,
        threads,
    )
    return (
        packed.reshape(*original_shape[:-1], dim // 2),
        norms.reshape(original_shape[:-1]),
    )


def quantize_activations_int8(
    x: torch.Tensor,
    permutation: torch.Tensor,
    signs: torch.Tensor,
    boundaries: torch.Tensor,
    codes: torch.Tensor,
    *,
    eps: float = 1e-12,
    inv_sqrt_block: float,
    threads: int = 256,
) -> tuple[torch.Tensor, torch.Tensor]:
    if not x.is_cuda:
        raise RuntimeError("native INT8 activation quantization requires CUDA tensors")
    if x.dtype not in {torch.float16, torch.bfloat16}:
        raise ValueError("x must be float16 or bfloat16")
    dim = x.shape[-1]
    if dim not in {512, 1024, 2048, 4096, 8192, 12288, 16384}:
        raise ValueError(
            "native INT8 activation quantization supports dimensions "
            "512, 1024, 2048, 4096, 8192, 12288, and 16384"
        )
    if threads not in {128, 256, 512}:
        raise ValueError("threads must be 128, 256, or 512")
    if permutation.numel() != dim or signs.numel() != dim:
        raise ValueError("permutation and signs must match the input dimension")
    if boundaries.numel() != 15:
        raise ValueError("boundaries must contain 15 values")
    if codes.numel() != 16:
        raise ValueError("codes must contain 16 values")

    original_shape = x.shape
    values = x.contiguous().reshape(-1, dim)
    quantized = torch.empty(values.shape, device=x.device, dtype=torch.int8)
    norms = torch.empty(values.shape[0], device=x.device, dtype=torch.float32)
    permutation_dtype = torch.int32 if permutation.dtype == torch.int32 else torch.int64
    permutation_values = permutation.to(device=x.device, dtype=permutation_dtype).contiguous()
    sign_values = signs.to(device=x.device, dtype=torch.int8).contiguous()
    boundary_values = boundaries.to(device=x.device, dtype=torch.float32).contiguous()
    code_values = codes.to(device=x.device, dtype=torch.int8).contiguous()
    ops.quantize_activations_int8(
        quantized,
        norms,
        values,
        permutation_values,
        sign_values,
        boundary_values,
        code_values,
        eps,
        inv_sqrt_block,
        threads,
    )
    return (
        quantized.reshape(*original_shape[:-1], dim),
        norms.reshape(original_shape[:-1]),
    )