File size: 6,026 Bytes
5a76465
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""RTX benchmark for grouped W4A4 expert projection launch aggregation."""

from __future__ import annotations

import argparse
import importlib
import sys
from pathlib import Path

import torch

PACKAGE = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(PACKAGE / "tests"))
from test_grouped_moe_gemv import load_source_ops, sfb_bytes  # noqa: E402


def elapsed_us(fn, warmup: int = 20, repeats: int = 100) -> float:
    for _ in range(warmup):
        fn()
    torch.cuda.synchronize()
    start = torch.cuda.Event(enable_timing=True)
    end = torch.cuda.Event(enable_timing=True)
    start.record()
    for _ in range(repeats):
        fn()
    end.record()
    end.synchronize()
    return start.elapsed_time(end) * 1000.0 / repeats


def load_ops(backend: str, artifact: str | None):
    if backend == "source":
        return load_source_ops()
    if artifact:
        sys.path.insert(0, artifact)
    return importlib.import_module("grouped_moe_gemv")


def run_case(ops, label: str, m: int, top_k: int, n: int, k: int) -> dict:
    experts = 8
    x = torch.randn((m, k), device="cuda", dtype=torch.bfloat16) * 0.2
    packed = torch.empty((m, k // 2), device="cuda", dtype=torch.uint8)
    sfa = torch.empty((sfb_bytes(m, k),), device="cuda", dtype=torch.uint8)
    weights = torch.full(
        (experts, n, k // 2), 0x11, device="cuda", dtype=torch.uint8
    )
    sfb = torch.full(
        (experts, sfb_bytes(n, k)), 0x38, device="cuda", dtype=torch.uint8
    )
    alpha = torch.ones((experts,), device="cuda", dtype=torch.float32)
    idx = (
        torch.arange(m * top_k, device="cuda", dtype=torch.int32)
        .reshape(m, top_k)
        .remainder(experts)
        .contiguous()
    )
    out = torch.empty((m, top_k, n), device="cuda", dtype=torch.bfloat16)
    routed_x = x[:, None, :].expand(m, top_k, k).reshape(m * top_k, k).contiguous()
    w4a16_out = torch.empty((m * top_k, n), device="cuda", dtype=torch.bfloat16)

    def grouped_region():
        ops.quantize_activations_nvfp4_bf16(x, packed=packed, sfa=sfa)
        ops.grouped_w4a4_gemv_bf16(
            packed, weights, sfa, sfb, alpha, idx, out=out
        )

    def grouped_kernel():
        ops.grouped_w4a4_gemv_bf16(
            packed, weights, sfa, sfb, alpha, idx, out=out
        )

    route_out = torch.empty((m, 1, n), device="cuda", dtype=torch.bfloat16)
    route_idx = [idx[:, route : route + 1].contiguous() for route in range(top_k)]

    def route_loop():
        ops.quantize_activations_nvfp4_bf16(x, packed=packed, sfa=sfa)
        for route in range(top_k):
            ops.grouped_w4a4_gemv_bf16(
                packed,
                weights,
                sfa,
                sfb,
                alpha,
                route_idx[route],
                out=route_out,
            )

    pair_packed = torch.empty((1, k // 2), device="cuda", dtype=torch.uint8)
    pair_sfa = torch.empty((sfb_bytes(1, k),), device="cuda", dtype=torch.uint8)
    pair_out = torch.empty((1, 1, n), device="cuda", dtype=torch.bfloat16)
    pair_idx = [
        idx[token : token + 1, route : route + 1].contiguous()
        for token in range(m)
        for route in range(top_k)
    ]

    def pair_loop():
        pair = 0
        for token in range(m):
            for _route in range(top_k):
                ops.quantize_activations_nvfp4_bf16(
                    x[token : token + 1], packed=pair_packed, sfa=pair_sfa
                )
                ops.grouped_w4a4_gemv_bf16(
                    pair_packed,
                    weights,
                    pair_sfa,
                    sfb,
                    alpha,
                    pair_idx[pair],
                    out=pair_out,
                )
                pair += 1

    grouped_us = elapsed_us(grouped_region)
    grouped_kernel_us = elapsed_us(grouped_kernel)
    def w4a16_region():
        ops.grouped_w4a16_gemv_bf16(
            routed_x,
            weights,
            sfb,
            alpha,
            idx.reshape(-1),
            w_stride=n * k // 2,
            sfb_stride=sfb.shape[1],
            n=n,
            out=w4a16_out,
        )

    w4a16_us = elapsed_us(w4a16_region)
    route_us = elapsed_us(route_loop) if top_k > 1 else grouped_us
    pair_repeats = 30 if m * top_k >= 32 else 100
    pair_us = elapsed_us(pair_loop, warmup=5, repeats=pair_repeats)
    return {
        "label": label,
        "M": m,
        "top_k": top_k,
        "pairs": m * top_k,
        "N": n,
        "K": k,
        "grouped_us": grouped_us,
        "grouped_kernel_us": grouped_kernel_us,
        "route_loop_us": route_us,
        "pair_loop_us": pair_us,
        "w4a16_us": w4a16_us,
        "vs_route_loop": route_us / grouped_us,
        "vs_pair_loop": pair_us / grouped_us,
        "vs_w4a16": w4a16_us / grouped_us,
    }


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("--backend", choices=["source", "installed"], default="source")
    parser.add_argument("--artifact")
    args = parser.parse_args()
    ops = load_ops(args.backend, args.artifact)
    cases = [
        ("gate_up_decode", 1, 8, 1024, 2048),
        ("gate_up_verify", 7, 8, 1024, 2048),
        ("down_decode", 8, 1, 2048, 512),
        ("down_verify", 56, 1, 2048, 512),
    ]
    print("label,M,top_k,pairs,N,K,w4a4_region_us,w4a4_kernel_us,w4a16_us,route_loop_us,pair_loop_us,vs_w4a16,vs_route,vs_pair")
    for case in cases:
        result = run_case(ops, *case)
        print(
            f"{result['label']},{result['M']},{result['top_k']},{result['pairs']},"
            f"{result['N']},{result['K']},{result['grouped_us']:.3f},"
            f"{result['grouped_kernel_us']:.3f},{result['w4a16_us']:.3f},"
            f"{result['route_loop_us']:.3f},"
            f"{result['pair_loop_us']:.3f},{result['vs_w4a16']:.2f}x,"
            f"{result['vs_route_loop']:.2f}x,{result['vs_pair_loop']:.2f}x"
        )
    return 0


if __name__ == "__main__":
    raise SystemExit(main())