File size: 6,890 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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
# 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_e128 import finetune_recipe, model
from nemo.collections.llm.recipes.precision.mixed_precision import bf16_with_fp8_mixed
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 (
    args_sanity_check,
    get_user_configs,
    hf_tokenizer,
    import_ckpt_experiment,
    prepare_squad_dataset_experiment,
    set_exp_logging_configs,
    set_primary_perf_configs,
    slurm_executor,
)

HF_MODEL_URI = "meta-llama/Llama-4-Maverick-17B-128E-Instruct"

# Set this to True if checkpoint is available at 'NEMO_HOME'. If set to False,
# extra Slurm job will be scheduled. In this case, if checkpoint is available
# at 'NEMO_HOME', fine-tuning job will use this checkpoint, else, it will be
# downloaded from HuggingFace
SKIP_IMPORT = False

# Set this to True if dataset is already downloaded. If set to False,
# dataset will be downloaded from HuggingFace
SKIP_DATASET_DOWNLOAD = False


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,
    use_mcore_fsdp: bool,
    recompute_layers: int,
    activation_offload_layers: int,
):
    """
    Llama4 e128 fine-tuning recipe aimed at achieving best possible performance.

    NOTE: Use fp8 precision training with caution. It might not give desirable results.
    """
    finetuning_scheme = "none" if args.finetuning == "sft" else args.finetuning

    recipe = finetune_recipe(peft_scheme=finetuning_scheme, performance_mode=True, packed_sequence=True)

    recipe = set_primary_perf_configs(
        recipe,
        finetuning_scheme,
        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=use_mcore_fsdp,
        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
        use_user_buffer_registration=args.use_user_buffer_registration,
        recompute_layers=recompute_layers,
        activation_offload_layers=activation_offload_layers,
        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,
        finetuning_scheme,
        "llm",
        "llama4",
        args.tensorboard,
        args.wandb,
        args.wandb_prj_name,
        args.wandb_job_name,
    )

    # data module configs
    recipe.data.tokenizer = hf_tokenizer(HF_MODEL_URI)

    # Compute dtype configs
    if args.compute_dtype.lower() == "fp8":
        recipe.trainer.plugins = bf16_with_fp8_mixed()
        recipe.trainer.plugins.grad_reduce_in_fp32 = False

    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(), "sft", "llama4", "e128", args)
    (
        num_nodes,
        mbs,
        gbs,
        tp_size,
        pp_size,
        cp_size,
        vp_size,
        ep_size,
        etp_size,
        enable_cuda_graphs,
        use_mcore_fsdp,
        recompute_layers,
        activation_offload_layers,
    ) = 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,
        use_mcore_fsdp,
        recompute_layers,
        activation_offload_layers,
    )
    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}"

    plugins = [build_perf_env_plugin(args, pp_size=pp_size)]

    if args.enable_nsys:
        plugins.append(NsysPlugin(start_step=5, end_step=6))
    if args.enable_memory_profile:
        assert args.memory_profile_out_path is not None
        plugins.append(MemoryProfilePlugin(dir=args.memory_profile_out_path))

    executor = slurm_executor(
        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={},
        hf_token=args.hf_token,
        nemo_home=args.nemo_home,
        wandb_key=args.wandb_key,
        additional_slurm_params=additional_slurm_params,
    )

    with run.Experiment(exp_name) as exp:
        if not SKIP_IMPORT:
            assert args.hf_token is not None, "HF token is required for importing checkpoint from HuggingFace"
            exp.add(*import_ckpt_experiment(executor, model(), source=f"hf://{HF_MODEL_URI}"))
        if not SKIP_DATASET_DOWNLOAD:
            exp.add(
                *prepare_squad_dataset_experiment(executor, HF_MODEL_URI, seq_length=4096, nemo_home=args.nemo_home)
            )
        exp.add(
            recipe,
            executor=executor,
            name=exp_name,
            plugins=plugins,
        )
        if not args.dryrun:
            exp.run(sequential=True, detach=args.detach)
        else:
            exp.dryrun()