File size: 4,177 Bytes
e04ed2d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c453fcd
 
 
 
 
 
e04ed2d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Flattened from lib/codebook/__init__.py.

Provides `kdict` (imported by bitshift.py) and registers the quip_lib torch.library
ops for each compiled kernel shape. The registration is pure torch.library metadata;
the actual CUDA impl only runs when qtip_kernels is importable. In manifest
('train-fixW') inference the fused-kernel path is never taken, so this whole module is
effectively just supplying an (empty) `kdict` symbol — but we keep the op registration
so the eval/kernel path still works verbatim if the compiled .so happens to be present.
"""
import torch  # must precede qtip_kernels so libc10 is loaded (the .so links it)

try:
    import qtip_kernels
except Exception as _qtip_err:  # ABI mismatch (torch upgrade) / not built / no CUDA
    qtip_kernels = None
    import warnings
    warnings.warn(
        f"qtip_kernels unavailable ({_qtip_err}); CUDA trellis kernels disabled - "
        f"falling back to torch decode. has_kernel() is forced False, so inference "
        f"runs the (pure-torch) manifest path.")

kernels = [
    (5120, 1, 5120, 4),
    (5120, 1, 5120, 3),
    (5120, 1, 5120, 2),
    (1024, 1, 5120, 4),
    (1024, 1, 5120, 3),
    (1024, 1, 5120, 2),
    (8192, 1, 5120, 4),
    (8192, 1, 5120, 3),
    (8192, 1, 5120, 2),
    (5120, 1, 8192, 4),
    (5120, 1, 8192, 3),
    (5120, 1, 8192, 2),
    (1024, 1, 3072, 4),
    (8192, 1, 3072, 4),
    (3072, 1, 8192, 4),
    (3072, 1, 3072, 4),
    (53248, 1, 16384, 2),
    (53248, 1, 16384, 3),
    (53248, 1, 16384, 4),
    (16384, 1, 53248, 2),
    (16384, 1, 53248, 3),
    (16384, 1, 53248, 4),
    (1024, 1, 16384, 2),
    (1024, 1, 16384, 3),
    (1024, 1, 16384, 4),
    (16384, 1, 16384, 2),
    (16384, 1, 16384, 3),
    (16384, 1, 16384, 4),
    (4096, 1, 14336, 2),
    (4096, 1, 14336, 3),
    (4096, 1, 14336, 4),
    (14336, 1, 4096, 2),
    (14336, 1, 4096, 3),
    (14336, 1, 4096, 4),
    (1024, 1, 4096, 2),
    (1024, 1, 4096, 3),
    (1024, 1, 4096, 4),
    (4096, 1, 4096, 2),
    (4096, 1, 11008, 2),
    (4096, 1, 12288, 2),
    (11008, 1, 4096, 2),
    (12288, 1, 4096, 2),
    (22016, 1, 4096, 2),
    (8192, 1, 8192, 2),
    (10240, 1, 8192, 2),
    (10240, 1, 8192, 3),
    (10240, 1, 8192, 4),
    (57344, 1, 8192, 2),
    (57344, 1, 8192, 3),
    (57344, 1, 8192, 4),
    (8192, 1, 1024, 2),
    (8192, 1, 28672, 2),
    (28672, 1, 8192, 2),
    (1024, 1, 8192, 2),
    (4096, 1, 4096, 3),
    (4096, 1, 11008, 3),
    (4096, 1, 12288, 3),
    (11008, 1, 4096, 3),
    (12288, 1, 4096, 3),
    (22016, 1, 4096, 3),
    (8192, 1, 8192, 3),
    (8192, 1, 1024, 3),
    (8192, 1, 28672, 3),
    (28672, 1, 8192, 3),
    (1024, 1, 8192, 3),
    (4096, 1, 4096, 4),
    (4096, 1, 11008, 4),
    (4096, 1, 12288, 4),
    (11008, 1, 4096, 4),
    (12288, 1, 4096, 4),
    (22016, 1, 4096, 4),
    (8192, 1, 8192, 4),
    (8192, 1, 1024, 4),
    (8192, 1, 28672, 4),
    (28672, 1, 8192, 4),
    (1024, 1, 8192, 4),
    (27648, 1, 5120, 2),
    (27648, 1, 5120, 3),
    (27648, 1, 5120, 4),
    (5120, 1, 27648, 2),
    (5120, 1, 27648, 3),
    (5120, 1, 27648, 4),
]

kdict = {}

for m, n, k, bitrate in kernels:
    name = f"decompress_matvec_qtip_{m}_{n}_{k}_{bitrate}"
    kernel_name = f"qtip_kernels.decompress_matvec_16_9_{bitrate}_1_{m}_{n}_{k}"
    try:
        torch.library.define(
            f"quip_lib::{name}",
            "(Tensor compressed, Tensor x, Tensor codebook) -> Tensor")
    except RuntimeError:
        continue   # op already registered in this process (loading another checkpoint in one kernel)
    exec(f"""\
@torch.library.register_fake("quip_lib::{name}")
def {name}_abstract(
        compressed: torch.Tensor,
        x: torch.Tensor,
        codebook: torch.Tensor) -> torch.Tensor:
    return torch.zeros(1, {m}, dtype=torch.float32, device=x.device)

@torch.library.impl("quip_lib::{name}", "cuda")
def {name}_cuda(
        compressed: torch.Tensor,
        x: torch.Tensor,
        codebook: torch.Tensor) -> torch.Tensor:
    out = torch.zeros(({m}, 1), dtype=torch.float32, device=x.device)
    {kernel_name}(out, compressed.reshape(-1).view(torch.int32), x.to(torch.float16).T, codebook.reshape(-1))
    return out.T
    """)