entangle/examples/verify_aws_llama.py at main · ByteDance-Seed/entangle · GitHub

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
# Copyright (c) 2025 ByteDance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
from typing import Callable, Sequence

import numpy as np
from entangle.convert.convert import *
from entangle.convert.mappings.vllm._custom_op import *
from entangle.sgraph import sexpr
from entangle.sgraph.egraph import SExprECondition
from entangle.sgraph.sskeleton import *
from entangle.tools.config import ExplorativeConfig


class MyConfig(ExplorativeConfig):
    def __init__(self, args: Sequence[str]):
        super().__init__(args)
        parser = argparse.ArgumentParser()
        parser.add_argument("--num_layers", type=int, default=1, choices=[1, 2, 4, 8, 12, 16])
        parser.add_argument("--tp", type=int, default=2, choices=[2, 4, 8])
        args = parser.parse_args(args)
        self.num_layers = args.num_layers
        self.tp = args.tp
        self.world_size = self.tp

    def build_preconditions(self) -> list[SExprECondition]:
        placeholder: Callable[[str], SExpr] = sexpr.get_placeholder_maker(self.get_sexpr)
        num_layers = self.num_layers
        world_size = self.world_size
        # fmt: off
        q_range = [i//world_size for i in [0, 2048]]
        k_range = [i//world_size for i in [2048, 2560]]
        v_range = [i//world_size for i in [2560, 3072]]
        base = 6 + num_layers * 2
        bc_base1 = 77 + num_layers*9
        bc_base2 = 84 + num_layers*9
        bc_s_gap = 102
        bc_t_gap = 114
        layers = [
            [
                # rms_norm weight
                SExprECondition.all_eq(f"Sn__r0__p{base+7*l}.{base+1+7*l}", [f"Dn__r{i}__p{base+7*l}.{base+1+7*l}" for i in range(world_size)], placeholder),
                # qkv weight
                SExprECondition(
                    inputs=[
                        s:=placeholder(f"Sn__r0__p{base+1+7*l}.{base+2+7*l}"),
                        *(d:=[placeholder(f"Dn__r{i}__p{base+1+7*l}.{base+2+7*l}") for i in range(world_size)])],
                    eclasses=[
                        [
                            s,
                            sexpr.concat(
                                [
                                    sexpr.concat([sexpr.slice(di, 1, *q_range) for di in d], dim=1),
                                    sexpr.concat([sexpr.slice(di, 1, *k_range) for di in d], dim=1),
                                    sexpr.concat([sexpr.slice(di, 1, *v_range) for di in d], dim=1),
                                ],
                                dim=1,
                            ),
                        ]
                    ]
                ),
                # A broadcasted constant
                SExprECondition.concat_target(f"Sn__r0__broadcast.{bc_base1+bc_s_gap*l}", [f"Dn__r{i}__broadcast.{bc_base1+bc_t_gap*l}" for i in range(world_size)], 2, placeholder),
                # A broadcasted constant: -30000
                SExprECondition.concat_target(f"Sn__r0__broadcast.{bc_base2+bc_s_gap*l}", [f"Dn__r{i}__broadcast.{bc_base2+bc_t_gap*l}" for i in range(world_size)], 1, placeholder),
                # Output projection weight at src/transformers_neuronx/layers/attention.py:886
                SExprECondition.concat_target(f"Sn__r0__p{base+2+7*l}.{base+3+7*l}", [f"Dn__r{i}__p{base+2+7*l}.{base+3+7*l}" for i in range(world_size)], 1, placeholder),
                # rms_norm weight
                SExprECondition.all_eq(f"Sn__r0__p{base+3+7*l}.{base+4+7*l}", [f"Dn__r{i}__p{base+3+7*l}.{base+4+7*l}" for i in range(world_size)], placeholder),
                # gated mlp weight 1: src/transformers_neuronx/llama/hlo.py:312
                SExprECondition.concat_target(f"Sn__r0__p{base+4+7*l}.{base+5+7*l}", [f"Dn__r{i}__p{base+4+7*l}.{base+5+7*l}" for i in range(world_size)], 1, placeholder),
                # gated mlp weight 2: src/transformers_neuronx/llama/hlo.py:312
                SExprECondition.concat_target(f"Sn__r0__p{base+5+7*l}.{base+6+7*l}", [f"Dn__r{i}__p{base+5+7*l}.{base+6+7*l}" for i in range(world_size)], 1, placeholder),
                # gated mlp weight 3: src/transformers_neuronx/llama/hlo.py:312
                SExprECondition.concat_target(f"Sn__r0__p{base+6+7*l}.{base+7+7*l}", [f"Dn__r{i}__p{base+6+7*l}.{base+7+7*l}" for i in range(world_size)], 1, placeholder),
                # rms_norm weight
                SExprECondition.all_eq(f"Sn__r0__p{base+7+7*l}.{base+8+7*l}", [f"Dn__r{i}__p{base+7+7*l}.{base+8+7*l}" for i in range(world_size)], placeholder),
                # Weight for computing logits
                SExprECondition.concat_target(f"Sn__r0__p{base+8+7*l}.{base+9+7*l}", [f"Dn__r{i}__p{base+8+7*l}.{base+9+7*l}" for i in range(world_size)], 1, placeholder),
            ]
            for l in range(num_layers)
        ]

        return [
            # Hidden
            SExprECondition.all_eq(f"Sn__r0__p0.1", [f"Dn__r{i}__p0.1" for i in range(world_size)], placeholder),
            # inv_freq
            SExprECondition.all_eq(f"Sn__r0__constant.{13+9*num_layers}", [f"Dn__r{i}__constant.{13+9*num_layers}" for i in range(world_size)], placeholder),
            # cache id? src/transformers_neuronx/layers/transformer.py:86
            SExprECondition.all_eq(f"Sn__r0__p1.2", [f"Dn__r{i}__p1.2" for i in range(world_size)], placeholder),
            # `start_ids` at src/transformers_neuronx/layers/transformer.py:97
            SExprECondition.all_eq(f"Sn__r0__p2.3", [f"Dn__r{i}__p2.3" for i in range(world_size)], placeholder),
            SExprECondition.just_map(f"Sn__r0__iota.{20+9*num_layers}", [f"Dn__r{i}__iota.{20+9*num_layers}" for i in range(world_size)], placeholder),
            # index select when done: src/transformers_neuronx/layers/transformer.py:331
            SExprECondition.all_eq(f"Sn__r0__p3.4", [f"Dn__r{i}__p3.4" for i in range(world_size)], placeholder),
        ] + list(chain(*layers))
        # fmt: on

    def oracle_group_id_to_size(self, group_id: str | int) -> int:
        return self.tp

    def build_expected(self) -> list[SExprECondition]:
        placeholder: Callable[[str], SExpr] = sexpr.get_placeholder_maker(self.get_sexpr)
        world_size = self.world_size
        num_layers = self.num_layers
        return [
            # fmt: off
            SExprECondition(
                inputs=[
                    s:=placeholder(f"Sn__r0__reshape.{38+111*num_layers}"),
                    *(d := [placeholder(f"Dn__r{i}__reshape.{38+123*num_layers}") for i in range(world_size)])
                ],
                eclasses=[
                    [s, sexpr.concat(d, dim=0).clone_with(name="EXPECTED_Sn__r0__reshape.149")],
                ],
            ),
            # fmt: on
        ]