Everything a training run needs, in one console
Grounded in what TIR actually ships — not slideware.
Managed fine-tuning pipelines
Foundation Studio walks you from model selection to launch: pick a base model, connect your Hugging Face token, configure data and hyperparameters, choose a GPU, hit Launch. Guides cover Mistral-7B, Gemma-7B, Llama 3.2 Vision, Stable Diffusion and SDXL.
Dataset management built in
Bring custom .jsonl, .json or parquet data via EOS-backed TIR Datasets, or link any public/private Hugging Face dataset. Set a validation split ratio and map your fields with a prompt configuration template.
Hyperparameters you control
Training type, epochs, learning rate, max steps, batch size, gradient accumulation and quantization options — plus a debug mode for fast dry runs before you commit GPU hours.
WandB experiment tracking
Add your Weights & Biases integration at job creation for full run tracking. Every run also streams training logs in the TIR console for debugging failed steps.
Checkpoints → Model Repository
Completed jobs push every checkpoint and adapter to a TIR Model Repository on E2E Object Storage — deployable to an inference endpoint in a few clicks.
Slurm-native Training Clusters
For pre-training and large-scale runs: multi-node clusters with Slurm scheduling (via Slinky + Pyxis), pre-installed CUDA/NCCL images, PyTorch DDP, Lightning and OpenMPI support, and custom containers via Enroot.
How a fine-tuning job works
The actual workflow in TIR's Foundation Studio, start to finish.
Create the job
Open Foundation Studio → Fine-Tune Models → Create Fine-Tuning Job. Name it, pick your base model, and attach your Hugging Face token (needed for gated models like Llama).
Prepare the dataset
Choose CUSTOM (upload .jsonl/parquet to an EOS dataset) or HUGGING FACE. Set the task, validation split ratio, and prompt configuration mapping your input/output fields.
Set hyperparameters
Epochs, learning rate, max steps, batch size, gradient accumulation, quantization. Optionally wire up WandB and enable debug mode for a quick validation pass.
Pick a GPU & launch
Select your machine — H100, H200 or A100 configurations — and click LAUNCH. Watch live logs per run while the job trains.
Collect the model
The fine-tuned model lands in your Model Repository with all checkpoints and adapters preserved.
Deploy it
Hit Deploy on the model repo: choose GPU, replicas and endpoint name, and serve it as an autoscaling API.
Fine-Tuning Jobs vs Training Clusters
Two ways to train on TIR — pick by how much control you need.
| Foundation Studio Fine-Tuning | Training Cluster (Slurm) | |
|---|---|---|
| Best for | LoRA/adapter fine-tunes of open LLMs & diffusion models | Multi-node distributed training and pre-training |
| Setup | Guided console wizard, no code required | Cluster in minutes; submit jobs via sbatch/srun |
| Frameworks | Managed pipeline (HF Transformers under the hood) | PyTorch DDP, PyTorch Lightning, Slurm, OpenMPI |
| Images | Managed by TIR | Ubuntu Slurm pre-built (CUDA, NCCL, drivers) or any Docker/OCI image via Enroot |
| Data | EOS datasets or Hugging Face | Datasets, Shared File System (SFS), Parallel File System (PFS) |
| Scaling | Single-job GPU selection | Elastic node scaling on a live cluster, no teardown |
| Resilience | Checkpoints saved to Model Repository | Node-level fault isolation, auto recovery, XID error visibility with one-click worker restart |
| Observability | Run logs + WandB | Live squeue, per-node DCGM GPU metrics, controller & node logs |
| Team access | Project-based | Persistent Login User Management — per-user SSH keys and isolated home directories |
Submit a distributed job the way HPC teams already do
Training Cluster nodes arrive with drivers, CUDA, NCCL and PyTorch ready. Checkpoint to shared storage so runs resume after any node event.
# illustrative — adapt paths/partitions to your cluster
#!/bin/bash
#SBATCH --job-name=llama-ft
#SBATCH --nodes=4
#SBATCH --gres=gpu:8
#SBATCH --partition=slinky
#SBATCH --time=24:00:00
srun torchrun \
--nnodes=$SLURM_NNODES \
--nproc_per_node=8 \
--rdzv_backend=c10d \
train.py --data /my_pfs/tokenized \
--checkpoint-dir /my_sfs/checkpointsMonitor the run from the console: Jobs tab for queue state, Monitoring tab for per-node GPU utilisation, Logs tab for controller and node output.
Training is half the loop
Serve what you train, on the same platform and the same GPUs.
Frequently Asked Questions
Everything you need to know about fine-tuning and training on TIR.
Training & Fine-tuning
Foundation Studio has guided flows for popular open models — documented walkthroughs include Mistral-7B, Gemma-7B, Llama 3.2 11B Vision, Stable Diffusion (Dreambooth/Textual Inversion) and SDXL — and any public or gated Hugging Face model works with your HF token. Prefer code? Run LLaMA-Factory on a TIR node.
Custom datasets in .json, .jsonl or parquet uploaded to EOS-backed TIR Datasets, or any Hugging Face dataset linked directly. Incorrect formats fail the run, so the wizard shows the expected structure up front.
No. Training Clusters deliver full Slurm via Slinky (Kubernetes-native Slurm) with queueing, priorities and multi-node coordination managed for you. You submit jobs; the console shows queue, node health and GPU metrics.
Failures are isolated to the node; the cluster recovers automatically after restart. Checkpoint periodically to SFS/PFS and resume from the last checkpoint. The Nodes tab surfaces NVIDIA XID hardware errors in real time with one-click worker restart.
GPU time is billed hourly in INR — H100 from ₹362/hr*, H200 from ₹436/hr* (₹189.2/hr* committed) — with committed plans for longer runs. *Indicative — see the live rate card for current pricing.
Your first fine-tune can start today
Real GPUs, no waitlist.