ContextQMD
Back to Transformers

Tensor parallelism for training

docs/source/en/tensor_parallelism.md

5.16.14.6 KB
Original Source
<!--Copyright 2026 The HuggingFace Team. 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. ⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be rendered properly in your Markdown viewer. -->

Tensor parallelism for training

Tensor parallelism (TP) splits weight matrices column-wise or row-wise across GPUs. Each GPU holds a shard, computes a partial result, and synchronizes with an all-reduce to produce the full output.

TP relies on frequent cross-GPU communication. It works best on hardware with fast intra-node links such as NVLink.

text
    ┌─────────────────────────────┐
    │       X  (replicated)       │
    └────┬──────────┬─────────┬───┘
         │          │         │
    ┌────▼───┐ ┌────▼───┐ ┌───▼────┐
    │ ▓▓▓ W₀ │ │ ░░░ W₁ │ │ ███ W₂ │
    │  X@W₀  │ │  X@W₁  │ │  X@W₂  │
    └────┬───┘ └────┬───┘ └───┬────┘
         └──────────┼─────────┘
               Y₀+Y₁+Y₂
    ┌────────────────────────────┐
    │          Y (full)          │
    └────────────────────────────┘

Transformers supports TP for architectures whose config defines base_model_tp_plan. Check that field first to see whether a model supports native TP.

py
from transformers import AutoConfig

config = AutoConfig.from_pretrained("Qwen/Qwen3-0.6B")
print(config.base_model_tp_plan is not None)
print(config.base_model_tp_plan)

If a model supports TP, create a [DistributedConfig] with the number of devices in tp_size and pass it to [~PreTrainedModel.from_pretrained]. Transformers uses the model's predefined plan, initializes the device mesh, and shards the supported layers for you.

You can also set tp_plan="auto" in [DistributedConfig]. When tp_size is omitted, it is inferred from WORLD_SIZE. Passing tp_plan directly to [~PreTrainedModel.from_pretrained] is deprecated and will be removed in v5.18.

[!WARNING] Don't use device_map with distributed_config. The two conflict at the weight-loading level. device_map places whole modules on specific GPUs, while tensor parallelism shards those same parameters across all GPUs.

py
import torch

from transformers import AutoModelForCausalLM, DistributedConfig

distributed_config = DistributedConfig(tp_size=4)

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-0.6B",
    dtype=torch.bfloat16,
    distributed_config=distributed_config,
)

[Trainer] detects the tensor parallel plan, reads tp_size from the model, and creates a [~accelerate.parallelism_config.ParallelismConfig] automatically.

Launch training on one node with 4 GPUs.

shell
torchrun --nproc-per-node 4 train_tp.py

ParallelismConfig

Pass [~accelerate.parallelism_config.ParallelismConfig] explicitly when combining TP with other parallelism techniques like FSDP.

py
import torch

from accelerate import ParallelismConfig
from transformers import AutoModelForCausalLM, DistributedConfig, TrainingArguments

distributed_config = DistributedConfig(tp_size=4)

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-0.6B",
    dtype=torch.bfloat16,
    distributed_config=distributed_config,
)

parallelism_config = ParallelismConfig(tp_size=4)

args = TrainingArguments(
    ...,
    parallelism_config=parallelism_config,
)

Next steps