Cannot finetune using NeMo in AWS: TypeError: finetune() got an unexpected keyword argument 'tokenizer'

Inizia da nemo_run/core/runners/fdl_runner.py e dal suo punto di ingresso fdl_direct_run, quindi segui come finetune_recipe arriva a finetune() e perché viene passato tokenizer. Riproduci il malfunzionamento usando il container NeMo fornito e la configurazione AWS/EKS. Il lavoro è completo quando la configurazione di finetuning segnalata non genera più l’errore imprevisto della parola chiave tokenizer.

Set up-
Running NeMo in EKS Cluster with 2 g6e.12xl instances (8 L402 gpus):

`import nemo_run as run
import json
import argparse
import os
from datetime import datetime
from functools import partial
from typing import Any, Optional
import lightning.pytorch as pl
from nemo.collections import llm
from nemo.utils import logging
from typing import TYPE_CHECKING, Any, Dict, List, Optional 
from custom_data_module import CustomDataModule
from multiprocessing import Process, set_start_method
from nemo.collections.llm import api

def configure_finetune_recipe(
    exp_name=None,
    work_dir=None,
    peft_scheme=None,
    lora_enabled=None,
    max_steps=None,
    nodes=None,
    gpu_devices=None,
    seq_length=None,
    micro_batch_size=None,
    global_batch_size=None,
    ):
   finetune_recipe = llm.llama31_8b.finetune_recipe(
       num_nodes=nodes,
       name=exp_name,
       dir=work_dir,
       peft_scheme=peft_scheme
    )
   finetune_recipe.resume.restore_config.path = "/data/fsx/nemo/cache/meta-llama/Llama-3.1-8B"
   finetune_recipe.trainer.devices = gpu_devices
   finetune_recipe.trainer.num_sanity_val_steps = 0
   finetune_recipe.trainer.max_steps = max_steps
   finetune_recipe.trainer.strategy.context_parallel_size = 1
   finetune_recipe.trainer.val_check_interval = 10
   finetune_recipe.data = custom_dataset(
       seq_length=seq_length,
       micro_batch_size=micro_batch_size,
       global_batch_size=global_batch_size
   )

   if lora_enabled:
       finetune_recipe.trainer.strategy.ddp = "megatron"
   
   return finetune_recipe
def skypilot_executor(
   nodes: int,
   pvc_mount: str,
   gpu_devices: int,
   gpus: str = "L40S",
   efa_devices: Optional[int] = None,
   custom_mounts: Optional[dict[str, str]] = None,
   container_image: str = "nvcr.io/nvidia/nemo:25.04.01",
   env_vars_file: str = "env_vars.json",
   pvc_name: str = "fsx-claim",
   lora_enabled: bool = False,
   memory: int = 180
) -> run.SkypilotExecutor:

   mounts = {}
   # Custom mounts are defined here.
   if custom_mounts:
       for k, v in custom_mounts.items():
           mounts[k] = v
   # Env vars for jobs are configured here
   with open(env_vars_file, 'r') as f:
       env_vars = json.load(f)

   packager = run.GitArchivePackager()

   shared_pod_config = {
        "kubernetes": {
            "pod_config": {
                "spec": {
                    "serviceAccountName": "default",
                    "securityContext": {
                        "runAsUser": 0,  
                        "fsGroup": 0    
                    },
                    "affinity": {
                        "nodeAffinity": {
                            "requiredDuringSchedulingIgnoredDuringExecution": {
                                "nodeSelectorTerms": [{
                                    "matchExpressions": [{
                                        "key": "nvidia.com/gpu.product",
                                        "operator": "In",
                                        "values": ["NVIDIA-L40S"]
                                    }]
                                }]
                            }
                        }
                    },
                    "containers": [{ 
                        "volumeMounts": [
                            {"name": "nemo-runs", "mountPath": pvc_mount},
                        ]
                    }],
                    "volumes": [{
                        "name": "nemo-runs",
                        "persistentVolumeClaim": {"claimName": pvc_name}
                    }]
                }
            }
        }
    }

   if efa_devices is not None:
        shared_pod_config["kubernetes"]["pod_config"]["spec"]["containers"][0]["resources"] = {
            "requests": {
                "vpc.amazonaws.com/efa": efa_devices
            },
            "limits": {
                "vpc.amazonaws.com/efa": efa_devices
            }
        }
  
   # This defines the skypilot executor.
   executor = run.SkypilotExecutor(
       cloud="kubernetes",
       gpus=gpus,
       gpus_per_node=gpu_devices,
       num_nodes=nodes,
       packager=packager,
       cluster_config_overrides=shared_pod_config,
       memory=memory
   )

   executor.container_image = container_image
   executor.file_mounts = mounts
   executor.env_vars = env_vars
   executor.env_vars["NEMORUN_HOME"] = pvc_mount
   executor.env_vars["NEMO_HOME"] = f"{pvc_mount}/fsx/nemo"
   executor.env_vars["NEMO_MODELS_CACHE"] = f"{pvc_mount}/fsx/nemo/cache"
   executor.env_vars["HF_HOME"] = f"{pvc_mount}/fsx/huggingface"
   executor.env_vars["HF_HUB_CACHE"] = f"{pvc_mount}/fsx/huggingface/hub"
   executor.env_vars["TRANSFORMERS_CACHE"] = f"{pvc_mount}/fsx/huggingface/transformers"
   executor.env_vars["NEMO_DISABLE_LOCK"] = "1"

   return executor
if __name__ == "__main__":
    pvc_mount = "/data"

    # LoRA is enabled by default, disabled if --disable_lora flag is present
    lora_enabled = True

    if lora_enabled:
        peft_scheme = "lora"
        exp_name = f"aws-nemo2-lora-finetune-{datetime.now().strftime('%Y%m%d-%H%M%S')}"
    else:
        peft_scheme = None
        exp_name = f"aws-nemo2-finetune-{datetime.now().strftime('%Y%m%d-%H%M%S')}"
    s3_out_dir = f"{pvc_mount}/experiments/{exp_name}"
    import_ckpt = configure_checkpoint_conversion()
    finetune_recipe = configure_finetune_recipe(
        exp_name=exp_name,
        work_dir=pvc_mount,
        peft_scheme=peft_scheme,
        lora_enabled=lora_enabled,
        max_steps=200,
        nodes=2,
        gpu_devices=4,
        seq_length=2048,
        micro_batch_size=1,
        global_batch_size=8
    )
    executor = skypilot_executor(
        nodes=2,
        gpus="l40s",
        gpu_devices=4,
        efa_devices=1,
        container_image="########.dkr.ecr.us-east-1.amazonaws.com/aws-nemo:25.04.01",
        custom_mounts={
            "/root/nemo": "."
        },
        env_vars_file="env_vars.json",
        pvc_name="fsx-claim",
        pvc_mount=pvc_mount,
        lora_enabled=lora_enabled,
        memory=180
    )
    # Set launcher based on number of nodes
    # if args.nodes > 1:
    executor.launcher = "torchrun"

    # executor.env_vars.update({
    #     "FI_PROVIDER": "efa",
    #     "NCCL_PROTO": "simple",
    #     "NCCL_DEBUG": "INFO"
    # })
    executor.env_vars.update({
        "NCCL_DEBUG": "INFO",
        "NCCL_SOCKET_IFNAME": "eth0",
        "NCCL_P2P_DISABLE": "0",
        "NCCL_NET_GDR_LEVEL": "0",
        "FI_PROVIDER": "efa",
    })

    # Set up the executor for the checkpoint conversion
    import_executor = executor.clone()

    with run.Experiment(exp_name, log_level="INFO") as exp:
        # exp.add(import_ckpt, executor=import_executor, name="checkpoint_conversion")
        exp.add(
            finetune_recipe, 
            executor=executor, 
            tail_logs=True, 
            name="finetuning"
        )

        local_executor = run.LocalExecutor()
        # Run the experiment
        exp.run(sequential=True, detach=True)
`

Error-

(head, rank=0, pid=3139) [default0]:Traceback (most recent call last):
(head, rank=0, pid=3139) [default0]:  File "<frozen runpy>", line 198, in _run_module_as_main
(head, rank=0, pid=3139) [default0]:  File "<frozen runpy>", line 88, in _run_code
(head, rank=0, pid=3139) [default0]:  File "/opt/NeMo-Run/nemo_run/core/runners/fdl_runner.py", line 72, in <module>
(head, rank=0, pid=3139) [default0]:    fdl_runner_app()
(head, rank=0, pid=3139) [default0]:  File "/usr/local/lib/python3.12/dist-packages/typer/main.py", line 339, in __call__
(head, rank=0, pid=3139) [default0]:    raise e
(head, rank=0, pid=3139) [default0]:  File "/usr/local/lib/python3.12/dist-packages/typer/main.py", line 322, in __call__
(head, rank=0, pid=3139) [default0]:    return get_command(self)(*args, **kwargs)
(head, rank=0, pid=3139) [default0]:           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(head, rank=0, pid=3139) [default0]:  File "/usr/local/lib/python3.12/dist-packages/click/core.py", line 1161, in __call__
(head, rank=0, pid=3139) [default0]:    return self.main(*args, **kwargs)
(head, rank=0, pid=3139) [default0]:           ^^^^^^^^^^^^^^^^^^^^^^^^^^
(head, rank=0, pid=3139) [default0]:  File "/usr/local/lib/python3.12/dist-packages/typer/core.py", line 677, in main
(head, rank=0, pid=3139) [default0]:    return _main(
(head, rank=0, pid=3139) [default0]:           ^^^^^^
(head, rank=0, pid=3139) [default0]:  File "/usr/local/lib/python3.12/dist-packages/typer/core.py", line 195, in _main
(head, rank=0, pid=3139) [default0]:    rv = self.invoke(ctx)
(head, rank=0, pid=3139) [default0]:         ^^^^^^^^^^^^^^^^
(head, rank=0, pid=3139) [default0]:  File "/usr/local/lib/python3.12/dist-packages/click/core.py", line 1443, in invoke
(head, rank=0, pid=3139) [default0]:    return ctx.invoke(self.callback, **ctx.params)
(head, rank=0, pid=3139) [default0]:           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(head, rank=0, pid=3139) [default0]:  File "/usr/local/lib/python3.12/dist-packages/click/core.py", line 788, in invoke
(head, rank=0, pid=3139) [default0]:    return __callback(*args, **kwargs)
(head, rank=0, pid=3139) [default0]:           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
(head, rank=0, pid=3139) [default0]:  File "/usr/local/lib/python3.12/dist-packages/typer/main.py", line 697, in wrapper
(head, rank=0, pid=3139) [default0]:    return callback(**use_params)
(head, rank=0, pid=3139) [default0]:           ^^^^^^^^^^^^^^^^^^^^^^
(head, rank=0, pid=3139) [default0]:  File "/opt/NeMo-Run/nemo_run/core/runners/fdl_runner.py", line 68, in fdl_direct_run
(head, rank=0, pid=3139) [default0]:    fdl_fn()
(head, rank=0, pid=3139) [default0]:TypeError: finetune() got an unexpected keyword argument 'tokenizer'