File size: 5,060 Bytes
1db1507
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Compare staged-bias v1 and fused-bias v2 FP8 GELU MLP entries."""

from __future__ import annotations

import argparse
import importlib
import importlib.util
import json
import os
import statistics
import sys
from pathlib import Path

import torch


ROOT = Path(__file__).resolve().parents[2]
TEST_FILE = ROOT / "flashrt-fp8-ffn" / "tests" / "test_fp8_ffn.py"
SHAPES = {
    "groot_vit": (128, 1024, 4096, 1024),
    "groot_deepstack": (128, 4096, 4096, 2048),
    "groot_action_dit": (41, 1536, 6144, 1536),
    "pi05_decoder": (10, 1024, 4096, 1024),
}


def helpers_module():
    spec = importlib.util.spec_from_file_location("fp8_ffn_helpers", TEST_FILE)
    if spec is None or spec.loader is None:
        raise RuntimeError(f"cannot load {TEST_FILE}")
    module = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(module)
    return module


def time_us(fn, warmup: int, iterations: int, rounds: int = 5) -> float:
    samples = []
    for _ in range(rounds):
        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(iterations):
            fn()
        end.record()
        end.synchronize()
        samples.append(start.elapsed_time(end) * 1000.0 / iterations)
    return float(statistics.median(samples))


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("--backend", choices=("source", "installed"), default="source")
    parser.add_argument("--artifact")
    parser.add_argument("--warmup", type=int, default=20)
    parser.add_argument("--iterations", type=int, default=100)
    parser.add_argument("--json-out")
    args = parser.parse_args()
    helpers = helpers_module()
    if args.backend == "source":
        ops = helpers.load_source_ops()
    else:
        if args.artifact:
            sys.path.insert(0, args.artifact)
        try:
            ops = importlib.import_module("flashrt_fp8_ffn")
        finally:
            if args.artifact:
                sys.path.remove(args.artifact)

    rows = []
    for name, shape in SHAPES.items():
        tensors = helpers.make_case(*shape)
        m, _, h, n = shape
        hidden_v1 = torch.empty((m, h), device="cuda", dtype=torch.bfloat16)
        hidden_fp8_v1 = torch.empty_like(hidden_v1, dtype=helpers.fp8_dtype())
        out_v1_buf = torch.empty((m, n), device="cuda", dtype=torch.bfloat16)
        hidden_v2 = torch.empty_like(hidden_v1)
        hidden_fp8_v2 = torch.empty_like(hidden_fp8_v1)
        out_v2_buf = torch.empty_like(out_v1_buf)
        if args.backend == "source":
            v1 = lambda: ops.fp8_gelu_mlp_bf16(
                *tensors, hidden=hidden_v1, hidden_fp8=hidden_fp8_v1,
                out=out_v1_buf,
            )
            v2 = lambda: ops.fp8_gelu_mlp_v2_bf16(
                *tensors, hidden=hidden_v2, hidden_fp8=hidden_fp8_v2,
                out=out_v2_buf,
            )
        else:
            v1 = lambda: ops.fp8_gelu_mlp_bf16(
                *tensors, hidden_bf16=hidden_v1, hidden_fp8=hidden_fp8_v1,
                out=out_v1_buf,
            )
            v2 = lambda: ops.fp8_gelu_mlp_v2_bf16(
                *tensors, hidden_bf16=hidden_v2, hidden_fp8=hidden_fp8_v2,
                out=out_v2_buf,
            )
        out_v1 = v1()
        out_v2 = v2()
        torch.cuda.synchronize()
        old = os.environ.get("FLASHRT_FP8_FFN_REQUIRE_BIAS_EPILOGUE")
        os.environ["FLASHRT_FP8_FFN_REQUIRE_BIAS_EPILOGUE"] = "1"
        try:
            v2()
            torch.cuda.synchronize()
            fused_hit = True
        except RuntimeError as error:
            if "BIAS epilogue was required" not in str(error):
                raise
            fused_hit = False
        finally:
            if old is None:
                os.environ.pop("FLASHRT_FP8_FFN_REQUIRE_BIAS_EPILOGUE", None)
            else:
                os.environ["FLASHRT_FP8_FFN_REQUIRE_BIAS_EPILOGUE"] = old
        metrics = helpers.distribution_metrics(out_v2, out_v1)
        v1_us = time_us(v1, args.warmup, args.iterations)
        v2_us = time_us(v2, args.warmup, args.iterations)
        row = {
            "shape": name,
            "M": shape[0],
            "K": shape[1],
            "H": shape[2],
            "N": shape[3],
            "fused_bias_epilogue": fused_hit,
            "v1_us": v1_us,
            "v2_us": v2_us,
            "speedup": v1_us / v2_us,
            **metrics,
        }
        rows.append(row)
        print(json.dumps(row, sort_keys=True), flush=True)

    result = {
        "device": torch.cuda.get_device_name(),
        "capability": list(torch.cuda.get_device_capability()),
        "torch": torch.__version__,
        "cuda": torch.version.cuda,
        "rows": rows,
    }
    if args.json_out:
        Path(args.json_out).write_text(json.dumps(result, indent=2) + "\n")
    return 0


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