Showing results for 
Search instead for 
Did you mean: 
Data Engineering
Join discussions on data engineering best practices, architectures, and optimization strategies within the Databricks Community. Exchange insights and solutions with fellow data engineers.
Showing results for 
Search instead for 
Did you mean: 

Error "Distributed package doesn't have nccl built in" with Transformers Library.

New Contributor

I am trying to run a simple training script using HF's transformers library and am running into the error `Distributed package doesn't have nccl built in` error.

Runtime: DBR 13.0 ML - SPark 3.4.0 - Scala 2.12

Driver: i3.xlarge - 4 cores

Note: This is a CPU instance

I am trying to fine-tune a transformers model for Sequence Classification - essentially following this tutorial:

When I try to initialize TrainingArguments (TrainingArguments(output_dir="test_trainer")), I get the error trace

RuntimeError                              Traceback (most recent call last)
File <command-1074749622305054>:3
      1 from transformers import  TrainingArguments
----> 3 TrainingArguments(output_dir="test_trainer")
File <string>:108, in __init__(self, output_dir, overwrite_output_dir, do_train, do_eval, do_predict, evaluation_strategy, prediction_loss_only, per_device_train_batch_size, per_device_eval_batch_size, per_gpu_train_batch_size, per_gpu_eval_batch_size, gradient_accumulation_steps, eval_accumulation_steps, eval_delay, learning_rate, adam_beta1, adam_beta2, adam_epsilon, max_grad_norm, num_train_epochs, max_steps, lr_scheduler_type, warmup_ratio, warmup_steps, log_level, log_level_replica, log_on_each_node, logging_dir, logging_strategy, logging_first_step, logging_steps, logging_nan_inf_filter, save_strategy, save_steps, save_total_limit, save_on_each_node, no_cuda, use_mps_device, seed, data_seed, jit_mode_eval, use_ipex, bf16, fp16, fp16_opt_level, half_precision_backend, bf16_full_eval, fp16_full_eval, tf32, local_rank, xpu_backend, tpu_num_cores, tpu_metrics_debug, debug, dataloader_drop_last, eval_steps, dataloader_num_workers, past_index, run_name, disable_tqdm, remove_unused_columns, label_names, load_best_model_at_end, metric_for_best_model, greater_is_better, ignore_data_skip, sharded_ddp, fsdp, fsdp_min_num_params, fsdp_transformer_layer_cls_to_wrap, deepspeed, label_smoothing_factor, optim, optim_args, adafactor, group_by_length, length_column_name, report_to, ddp_find_unused_parameters, ddp_bucket_cap_mb, dataloader_pin_memory, skip_memory_metrics, use_legacy_prediction_loop, push_to_hub, resume_from_checkpoint, hub_model_id, hub_strategy, hub_token, hub_private_repo, gradient_checkpointing, include_inputs_for_metrics, fp16_backend, push_to_hub_model_id, push_to_hub_organization, push_to_hub_token, mp_parameters, auto_find_batch_size, full_determinism, torchdynamo, ray_scope, ddp_timeout, torch_compile, torch_compile_backend, torch_compile_mode)
File /databricks/python/lib/python3.10/site-packages/transformers/, in TrainingArguments.__post_init__(self)
   1162     warnings.warn(
   1163         "`--adafactor` is deprecated and will be removed in version 5 of 🤗 Transformers. Use `--optim"
   1164         " adafactor` instead",
   1165         FutureWarning,
   1166     )
   1167     self.optim = OptimizerNames.ADAFACTOR
   1169 if (
   1170     self.framework == "pt"
   1171     and is_torch_available()
-> 1172     and (self.device.type != "cuda")
   1173     and (get_xla_device_type(self.device) != "GPU")
   1174     and (self.fp16 or self.fp16_full_eval)
   1175 ):
   1176     raise ValueError(
   1177         "FP16 Mixed precision training with AMP or APEX (`--fp16`) and FP16 half precision evaluation"
   1178         " (`--fp16_full_eval`) can only be used on CUDA devices."
   1179     )
   1181 if (
   1182     self.framework == "pt"
   1183     and is_torch_available()
   1188     and (self.bf16 or self.bf16_full_eval)
   1189 ):
File /databricks/python/lib/python3.10/site-packages/transformers/, in TrainingArguments.device(self)
   1552 """
   1553 The device used by this process.
   1554 """
   1555 requires_backends(self, ["torch"])
-> 1556 return self._setup_devices
File /databricks/python/lib/python3.10/site-packages/transformers/utils/, in cached_property.__get__(self, obj, objtype)
     55 cached = getattr(obj, attr, None)
     56 if cached is None:
---> 57     cached = self.fget(obj)
     58     setattr(obj, attr, cached)
     59 return cached
File /databricks/python/lib/python3.10/site-packages/transformers/, in TrainingArguments._setup_devices(self)
   1537 else:
   1538     # Here, we'll use torch.distributed.
   1539     # Initializes the distributed backend which will take care of synchronizing nodes/GPUs
   1540     if not torch.distributed.is_initialized():
-> 1541         torch.distributed.init_process_group(backend="nccl", timeout=self.ddp_timeout_delta)
   1542     device = torch.device("cuda", self.local_rank)
   1543     self._n_gpu = 1
File /databricks/python/lib/python3.10/site-packages/torch/distributed/, in init_process_group(backend, init_method, timeout, world_size, rank, store, group_name, pg_options)
    757         # Use a PrefixStore to avoid accidental overrides of keys used by
    758         # different systems (e.g. RPC) in case the store is multi-tenant.
    759         store = PrefixStore("default_pg", store)
--> 761     default_pg = _new_process_group_helper(
    762         world_size,
    763         rank,
    764         [],
    765         backend,
    766         store,
    767         pg_options=pg_options,
    768         group_name=group_name,
    769         timeout=timeout,
    770     )
    771     _update_default_pg(default_pg)
    773 _pg_group_ranks[GroupMember.WORLD] = {i: i for i in range(GroupMember.WORLD.size())}  # type: ignore[attr-defined, index]
File /databricks/python/lib/python3.10/site-packages/torch/distributed/, in _new_process_group_helper(group_size, group_rank, global_ranks_in_group, backend, store, pg_options, group_name, timeout)
    884 elif backend == Backend.NCCL:
    885     if not is_nccl_available():
--> 886         raise RuntimeError("Distributed package doesn't have NCCL " "built in")
    887     if pg_options is not None:
    888         assert isinstance(
    889             pg_options, ProcessGroupNCCL.Options
    890         ), "Expected pg_options argument to be of type ProcessGroupNCCL.Options"
RuntimeError: Distributed package doesn't have NCCL built in


I have tried the following fix with no effect.


import os

os.environ["PL_TORCH_DISTRIBUTED_BACKEND"] = "gloo"


I can not find any other pointers.

Can anyone please give suggestions on what may be going on?


Not applicable

Hi @Anastassia Kornilova​ 

Great to meet you, and thanks for your question!

Let's see if your peers in the community have an answer to your question. Thanks.

New Contributor II
New Contributor II

Hi @anastassia_kor1,

For CPU-only training, TrainingArguments has a no_cuda flag that should be set.

For transformers==4.26.1 (MLR 13.0) and transformers==4.28.1 (MLR 13.1), there's an additional xpu_backend argument that needs to be set as well. Try using:

training_args = TrainingArguments(output_dir="outputs", no_cuda=True, xpu_backend="gloo")

 For transformers==4.29.2 (MLR 13.2), try using:

training_args = TrainingArguments(output_dir="outputs", no_cuda=True)

 It may be necessary to restart the cluster in order for this argument to take effect.