Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_tp1_quant.py 4.2 KB
Newer Older
Luka Govedič's avatar
[CI][torch.compile] Reduce e2e fusion test time (#33293)  
Luka Govedič 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 
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Callable

import pytest

from vllm.config import PassConfig

from .common import (
    INDUCTOR_GRAPH_PARTITION,
    AttentionBackendCase,
    Matches,
    custom_ops_combos,
    is_blackwell,
)
from .models import (
    FLASHINFER_ATTN,
    TRITON_ATTN,
    llama3_8b_fp4,
    llama3_8b_fp8,
    llama4_scout_fp4,
    llama4_scout_fp8,
    qwen3_a3b_fp8,
)


@pytest.mark.parametrize(
    "model_name, matches_fn, model_kwargs, hf_overrides, use_deepgemm",
    [
        (*llama3_8b_fp8, False),
        (*llama4_scout_fp8, False),
        (*qwen3_a3b_fp8, False),
        (*qwen3_a3b_fp8, True),
    ],
)
@pytest.mark.parametrize("attn_backend", [TRITON_ATTN, FLASHINFER_ATTN])
@pytest.mark.parametrize("n_layers", [6])
@pytest.mark.parametrize("custom_ops", custom_ops_combos("quant_fp8", "rms_norm"))
@pytest.mark.parametrize("inductor_graph_partition", INDUCTOR_GRAPH_PARTITION)
def test_tp1_fp8_fusions(
    model_name: str,
    matches_fn: Callable[[int], Matches],
    model_kwargs: dict,
    hf_overrides: Callable[[int], dict],
    attn_backend: AttentionBackendCase,
    n_layers: int,
    custom_ops: str,
    inductor_graph_partition: bool,
    use_deepgemm: bool,
    run_e2e_fusion_test,
    monkeypatch,
):
ElizaWszola's avatar
[Bugfix] Fix quant RMS norm fusion for quantization with TMA-aligned scales (#33255)  
ElizaWszola committed
53
54 
    if use_deepgemm and is_blackwell():
        # TODO(luka) DeepGEMM uses different quants, matching not supported
Luka Govedič's avatar
[CI][torch.compile] Reduce e2e fusion test time (#33293)  
Luka Govedič committed
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 
        #  - on Blackwell, uses a special quant fp8, currently not supported
        pytest.skip("DeepGEMM & quant matching not currently supported")

    matches = matches_fn(n_layers)

    if "qwen" in model_name.lower() and "-quant_fp8" in custom_ops:
        # This is why config forces +quant_fp8 by default
        pytest.skip("native QuantFP8 matching not supported for group quant")

    # Reduce size of model and skip weight loading time
    model_kwargs["hf_overrides"] = hf_overrides(n_layers)
    model_kwargs["load_format"] = "dummy"
    model_kwargs["max_model_len"] = 1024
    compilation_config = dict(
        use_inductor_graph_partition=inductor_graph_partition,
        custom_ops=custom_ops.split(","),
        pass_config=PassConfig(
            fuse_norm_quant=True,
            fuse_act_quant=True,
            fuse_attn_quant=True,
            enable_qk_norm_rope_fusion=True,
        ),
    )

    matches_check = [
        "rms_quant_fusion",
        "act_quant_fusion",
        "norm_rope_fusion",
        "attn_quant_fusion",
    ]

    run_e2e_fusion_test(
        model_name,
        matches,
        model_kwargs,
        attn_backend,
        compilation_config,
        matches_check,
        use_deepgemm=use_deepgemm,
    )


@pytest.mark.parametrize(
    "model_name, matches_fn, model_kwargs, hf_overrides",
    [llama3_8b_fp4, llama4_scout_fp4],
)
@pytest.mark.parametrize("attn_backend", [FLASHINFER_ATTN])
@pytest.mark.parametrize("n_layers", [6])
@pytest.mark.parametrize("custom_ops", custom_ops_combos("rms_norm"))
@pytest.mark.parametrize("inductor_graph_partition", INDUCTOR_GRAPH_PARTITION)
@pytest.mark.skipif(not is_blackwell(), reason="Blackwell required for fp4")
def test_tp1_fp4_fusions(
    model_name: str,
    matches_fn: Callable[[int], Matches],
    model_kwargs: dict,
    hf_overrides: Callable[[int], dict],
    attn_backend: AttentionBackendCase,
    n_layers: int,
    custom_ops: str,
    inductor_graph_partition: bool,
    run_e2e_fusion_test,
):
    matches = matches_fn(n_layers)

    # Reduce size of model and skip weight loading time
    model_kwargs["hf_overrides"] = hf_overrides(n_layers)
    model_kwargs["load_format"] = "dummy"
    model_kwargs["max_model_len"] = 1024

    compilation_config = dict(
        use_inductor_graph_partition=inductor_graph_partition,
        custom_ops=custom_ops.split(","),
        pass_config=PassConfig(
            fuse_norm_quant=True,
            fuse_act_quant=True,
            fuse_attn_quant=True,
            enable_qk_norm_rope_fusion=True,
        ),
    )

    matches_check = ["act_quant_fusion", "attn_quant_fusion", "norm_rope_fusion"]

    run_e2e_fusion_test(
        model_name,
        matches,
        model_kwargs,
        attn_backend,
        compilation_config,
        matches_check,
    )