File size: 6,238 Bytes
0558aa4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
#
# 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.

from os.path import basename, splitext

import nemo_run as run

from nemo.collections.llm.recipes.llama4_e16 import pretrain_recipe
from nemo.collections.nlp.modules.common.tokenizer_utils import get_nmt_tokenizer
from nemo.lightning.run.plugins import MemoryProfilePlugin, NsysPlugin

from ..argument_parser import parse_additional_slurm_params, parse_cli_args
from ..executors import slurm_executor
from ..helpers import (
    args_sanity_check,
    build_perf_env_plugin,
    get_user_configs,
    set_exp_logging_configs,
    set_primary_perf_configs,
)
from ..utils import dump_config_diff_from_base_recipe, hf_tokenizer


def override_recipe_configs(
    args: str,
    num_nodes: int,
    mbs: int,
    gbs: int,
    tp_size: int,
    pp_size: int,
    cp_size: int,
    vp_size: int,
    ep_size: int,
    etp_size: int,
    enable_cuda_graphs: bool,
):
    """
    llama4 e16 pre-train recipe aimed at achieving best possible performance and faster
    overall runtime.

    NOTE: Use fp8 precision training with caution. It might not give desirable results.
    """
    recipe = pretrain_recipe(performance_mode=True)
    recipe = set_primary_perf_configs(
        recipe,
        "pre_train",
        num_nodes,
        args.gpus_per_node,
        mbs,
        gbs,
        args.max_steps,
        tp_size,
        pp_size,
        cp_size,
        vp_size,
        ep_size,
        etp_size,
        enable_cuda_graphs=enable_cuda_graphs,
        use_mcore_fsdp=args.use_mcore_fsdp,
        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
        use_user_buffer_registration=args.use_user_buffer_registration,
        use_sharp=args.use_sharp,
        compute_dtype=args.compute_dtype,
        fp8_recipe=args.fp8_recipe,
        use_te_act_func=args.use_te_act_func,
        act_func_fp8_input_store=args.act_func_fp8_input_store,
    )
    recipe = set_exp_logging_configs(
        recipe, "pre_train", "llm", "llama4", args.tensorboard, args.wandb, args.wandb_prj_name, args.wandb_job_name
    )

    # data module configs
    if args.use_hf_tokenizer:
        recipe.data.tokenizer = hf_tokenizer('meta-llama/Llama-4-Scout-17B-16E-Instruct')
    else:
        recipe.data.tokenizer = run.Config(
            get_nmt_tokenizer, library="null", model_name="NullTokenizer", vocab_size=200000
        )
        recipe.model.tokenizer = recipe.data.tokenizer

    recipe.model.config.cross_entropy_fusion_impl = "te"
    recipe.model.config.cross_entropy_loss_fusion = True
    recipe.model.config.apply_rope_fusion = True
    recipe.model.config.moe_permute_fusion = True

    return recipe


if __name__ == "__main__":
    args = parse_cli_args().parse_args()
    args_sanity_check(args)
    # Parse additional SLURM parameters if provided
    additional_slurm_params = None
    if hasattr(args, 'additional_slurm_params') and args.additional_slurm_params:
        additional_slurm_params = parse_additional_slurm_params(args.additional_slurm_params)

    kwargs = get_user_configs(args.gpu.lower(), "pre_train", "llama4", "e16", args)
    num_nodes, mbs, gbs, tp_size, pp_size, cp_size, vp_size, ep_size, etp_size, enable_cuda_graphs, _, _, _ = kwargs[
        0:13
    ]

    recipe = override_recipe_configs(
        args, num_nodes, mbs, gbs, tp_size, pp_size, cp_size, vp_size, ep_size, etp_size, enable_cuda_graphs
    )

    exp_config = (
        f"{num_nodes}nodes_tp{tp_size}_pp{pp_size}_cp{cp_size}_vp{vp_size}_ep{ep_size}_etp{etp_size}_{mbs}mbs_{gbs}gbs"
    )
    exp_name = f"{splitext(basename(__file__))[0]}_{args.compute_dtype}_{exp_config}"

    # Workaround for CUDA graph illegal memory access error
    if not enable_cuda_graphs:
        custom_env_vars = {"PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True"}
    else:
        custom_env_vars = {}

    executor = slurm_executor(
        args.gpu.lower(),
        args.account,
        args.partition,
        args.log_dir,
        num_nodes,
        args.gpus_per_node,
        args.time_limit,
        args.container_image,
        custom_mounts=args.custom_mounts,
        custom_env_vars=custom_env_vars,
        hf_token=args.hf_token,
        nemo_home=args.nemo_home,
        wandb_key=args.wandb_key,
        network='sharp' if args.use_sharp else None,
        additional_slurm_params=additional_slurm_params,
    )

    plugins = [build_perf_env_plugin(args, pp_size=pp_size)]
    if args.enable_nsys:
        plugins.append(NsysPlugin(start_step=15, end_step=16, gen_shape=True))
    if args.enable_memory_profile:
        assert args.memory_profile_out_path is not None
        plugins.append(MemoryProfilePlugin(dir=args.memory_profile_out_path))

    with run.Experiment(exp_name) as exp:
        exp.add(
            recipe,
            executor=executor,
            name=exp_name,
            plugins=plugins,
        )

        if not args.dryrun:
            exp.run(sequential=True, detach=args.detach)
        else:
            exp.dryrun()

    if args.dump_config_diff_from_base_recipe:
        output_dir = exp.jobs[0].executor.job_dir
        # dump difference from base recipe
        base_recipe = pretrain_recipe(performance_mode=False)
        file_name = f"diff_from_base_recipe_{args.compute_dtype}.diff"
        dump_config_diff_from_base_recipe(base_recipe, recipe, output_dir, file_name=file_name)
        # dump difference from default perf recipe
        default_perf_recipe = pretrain_recipe(performance_mode=True)
        file_name = f"diff_from_default_perf_recipe_{args.compute_dtype}.diff"
        dump_config_diff_from_base_recipe(default_perf_recipe, recipe, output_dir, file_name=file_name)