Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
rocm_moe_utils.py 4.45 KB
Newer Older
maxiao's avatar
adapt to sglang v0.5.2rc1 on dcu  
maxiao committed
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 
# Adapted from https://github.com/vllm-project/vllm/blob/v0.9.1rc2/vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from enum import IntEnum
from functools import cache
from typing import Optional

import torch

from sglang.srt.utils import direct_register_custom_op, get_bool_env_var, is_hip

_is_hip = is_hip()
_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip


class ActivationMethod(IntEnum):
    # This allows interfacing with AITER ActivationType enum
    # without importing the ActivationType enum from AITER globally.
    SILU = 0
    GELU = 1


def rocm_aiter_asm_moe_tkw1_impl(
    hidden_states: torch.Tensor,
    w1: torch.Tensor,
    w2: torch.Tensor,
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    fc1_scale: Optional[torch.Tensor] = None,
    fc2_scale: Optional[torch.Tensor] = None,
    fc1_smooth_scale: Optional[torch.Tensor] = None,
    fc2_smooth_scale: Optional[torch.Tensor] = None,
    a16: bool = False,
    per_tensor_quant_scale: Optional[torch.Tensor] = None,
    expert_mask: Optional[torch.Tensor] = None,
    activation_method: int = ActivationMethod.SILU.value,
) -> torch.Tensor:

    from aiter import ActivationType
    from aiter.fused_moe_bf16_asm import asm_moe_tkw1

    activation = ActivationType(activation_method)

    return asm_moe_tkw1(
        hidden_states,
        w1,
        w2,
        topk_weights,
        topk_ids,
        fc1_scale=fc1_scale,
        fc2_scale=fc2_scale,
        fc1_smooth_scale=fc1_smooth_scale,
        fc2_smooth_scale=fc2_smooth_scale,
        a16=a16,
        per_tensor_quant_scale=per_tensor_quant_scale,
        expert_mask=expert_mask,
        activation=activation,
    )


def rocm_aiter_asm_moe_tkw1_fake(
    hidden_states: torch.Tensor,
    w1: torch.Tensor,
    w2: torch.Tensor,
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    fc1_scale: Optional[torch.Tensor] = None,
    fc2_scale: Optional[torch.Tensor] = None,
    fc1_smooth_scale: Optional[torch.Tensor] = None,
    fc2_smooth_scale: Optional[torch.Tensor] = None,
    a16: bool = False,
    per_tensor_quant_scale: Optional[torch.Tensor] = None,
    expert_mask: Optional[torch.Tensor] = None,
    activation_method: int = ActivationMethod.SILU.value,
) -> torch.Tensor:
    return torch.empty_like(hidden_states)


if _use_aiter:

    direct_register_custom_op(
        op_name="rocm_aiter_asm_moe_tkw1",
        op_func=rocm_aiter_asm_moe_tkw1_impl,
        mutates_args=[],
        fake_impl=rocm_aiter_asm_moe_tkw1_fake,
    )


def rocm_fused_experts_tkw1(
    hidden_states: torch.Tensor,
    w1: torch.Tensor,
    w2: torch.Tensor,
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    activation: str = "silu",
    apply_router_weight_on_input: bool = False,
    use_fp8_w8a8: bool = False,
    per_channel_quant: bool = False,
    w1_scale: Optional[torch.Tensor] = None,
    w2_scale: Optional[torch.Tensor] = None,
    a1_scale: Optional[torch.Tensor] = None,
    a2_scale: Optional[torch.Tensor] = None,
    block_shape: Optional[list[int]] = None,
) -> torch.Tensor:

    activation_method = (
        ActivationMethod.SILU if activation == "silu" else ActivationMethod.GELU
    )
    # All AITER Fused MoE kernels are expecting the following datatypes
    topk_weights = topk_weights.to(torch.float32)
    topk_ids = topk_ids.to(torch.int32)

    # w8a8 per-channel quantization
    if per_channel_quant and apply_router_weight_on_input and use_fp8_w8a8:
        # AITER tkw1 kernel for FP8 models with `apply_router_weight_on_input`
        # This applies topk_weights on the GEMM output of the first FC layer
        #  rather than the second FC.
        assert (
            topk_weights.dim() == 2
        ), "`topk_weights` should be in shape (num_tokens, topk)"
        assert topk_weights.shape[-1] == 1, (
            "Only support topk=1 when" " `apply_router_weight_on_input` is True"
        )

        return torch.ops.sglang.rocm_aiter_asm_moe_tkw1(
            hidden_states,
            w1,
            w2,
            topk_weights,
            topk_ids,
            fc1_scale=w1_scale,
            fc2_scale=w2_scale,
            fc1_smooth_scale=None,
            fc2_smooth_scale=None,
            a16=False,
            per_tensor_quant_scale=None,
            expert_mask=None,
            activation_method=activation_method,
        )
    else:
        assert False, "This should not be called."