Terradev-cli 6.2.15
Cross-cloud GPU orchestration CLI.
License: Apache 2.0 - Free and open source for commercial and personal use.
https://terradev.cloud/
Terradev is a cross-cloud compute control plane for AI workloads, not just a provisioning wrapper.
Combines quoting, provisioning, topology optimization, training orchestration, inference tuning, and cost analytics in one CLI, with an accelerated idempotent runtime underneath.
Continued focus on lower cost, faster provisioning, and topology-aware execution with local credential storage.
Model agnostic. Dataset agnostic. GPU agnostic. Provider agnostic. The only thing Terradev is not agnostic about is correctness: it enforces topology, idempotency, and sequencing.
NOTES ON v6.2.4
- Version bumped to v6.2.4.
- Provider list consolidated to 17 active GPU cloud and inference providers:
aws — Amazon Web Servicesazure — Microsoft Azurebaseten — Basetencrusoe — Crusoe Clouddigitalocean — DigitalOceane2enetworks — E2E Networksgcore — Gcoregcp — Google Cloud Platformhuggingface — Hugging Facehyperstack — Hyperstackinferx — InferXlatitude — Latitude.shrunpod — RunPodsiliconflow — SiliconFlowtensordock — TensorDockvastai — Vast.aiyottalabs — YottaLabs
- Added Gcore (
gcore) as a new VM provider with full lifecycle support, API token auth, and drift monitoring.
- Removed OVHcloud from the active provider list.
- Removed Jio Cloud and Oracle Cloud from the active provider list; removed Alibaba Cloud references from the CLI UI.
- Onboarding,
configure, setup, quote, credential_prompt, and MCP tool schemas now all reflect the 17-provider set.
NOTES ON Multi-Stage Training Pipeline (v6.1.1)
- New
terradev train subcommands for the full post-pretraining lifecycle:
terradev train sft --model <id> --data <path> --nodes <ips>terradev train dpo --base-checkpoint <sft-ckpt> --data <pairs> --algorithm <dpo|simpo|kto|orpo>terradev train grpo --base-checkpoint <dpo-ckpt> --data <prompts> --framework <unsloth|openrlhf|trl>terradev train pipeline --config examples/training_pipeline.yaml
terradev_cli/core/training_stages.py and terradev_cli/core/training_pipeline.py provide declarative SFT / DPO / GRPO stage configs, provider-aware quote selection, auto-provisioning, checkpoint handoff, and DAG sequencing via the Python DAGExecutor.TrainingOrchestrator now supports multi-node remote SSH launch and end-to-end completion tracking. Training scripts and embedded configs are staged to every node and the master process is polled until the job finishes.- Unsloth GRPO uses the native
unsloth.GRPOTrainer with a default rule-based reward instead of a TRL fallback.
- CLI-style frameworks (
axolotl, llama-factory, ms-swift, trl, openrlhf) are wrapped in self-contained Python scripts that write their embedded config files at runtime, so they are safe to copy to remote nodes.
- See
examples/training_pipeline.yaml for a SFT → DPO → GRPO pipeline sample.
NOTES ON 6.0.8
- New agent subcommands —
terradev agent sandbox, terradev agent mesh, and terradev agent mcp are now real, dependency-resolving commands instead of placeholders:
terradev agent sandbox runs untrusted payloads with hardware-isolated runtimes (bwrap, runsc, firecracker, and Linux Landlock LSM), all discoverable/downloadable via DependencyManager.terradev agent mesh creates a decentralized peer-to-peer agent mesh using real libp2p (go-libp2p-daemon + p2pclient), A2A HTTP, and WireGuard encrypted transports.terradev agent mcp is a dynamic Model Context Protocol bridge with stdio and HTTP transports and a multi-server bridge.
NOTES ON 6.0.0
-
Unsloth (terradev train unsloth): optimized local LLM training, serving, and coding agents with 70% lower VRAM usage and faster training via Triton kernels. Subcommands: run, start, stop.
terradev train unsloth run --model unsloth/Llama-3.1-8Bterradev train unsloth run --model unsloth/Qwen3.6-7B-GGUF:Q4_K_M --port 8080terradev train unsloth start claude --model unsloth/Llama-3.1-8Bterradev train unsloth stop
-
Weaviate (terradev database weaviate): vector database operations with local, embedded, cloud, and custom environments. Subcommands: up, list-collections, create-collection, delete-collection, insert, query, hybrid-search.
terradev database weaviate up --environment localterradev database weaviate create-collection --name Article --vector-size 384terradev database weaviate insert --collection Article --objects '[{"properties": {"title": "Hello"}, "vector": [0.1, ...]}]'terradev database weaviate query --collection Article --vector '[0.1, ...]' --top-k 5terradev database weaviate hybrid-search --collection Article --query "generative AI" --top-k 5
-
Letta (terradev agent letta): stateful agents with long-horizon memory across sessions. Subcommands: create, list, chat, status, delete, remember.
terradev agent letta create --name my-agent --model openai/gpt-4.1terradev agent letta listterradev agent letta chat --agent-id <id> --message "hello"terradev agent letta remember --agent-id <id> --text "Our staging cluster is on us-east-1" --label factterradev agent letta status --agent-id <id>terradev agent letta delete --agent-id <id>
-
Removed integrations: terradev ml databricks and terradev ml langsmith command groups, all related MCP tools, and Jaeger tracing endpoints in the Helm values have been removed.
-
Version bumped to 6.0.0 to reflect these breaking changes.
NOTES ON 5.7.10
- Local Ollama integration (
terradev ml ollama): list, pull, generate, chat, inspect, and check running models on a local Ollama server (list, pull, generate, chat, info, ps).
- DeepEval integration (
terradev ml deepeval): install, run test suites, list metrics, and evaluate single LLM outputs with metrics like AnswerRelevancyMetric, FaithfulnessMetric, and HallucinationMetric (install, init, run, metrics, evaluate).
- MCP tool surface now at 237 tools, including
ollama_* and deepeval_* tool definitions and handlers.
NOTES ON 5.7.7
-
Vault command (terradev vault): store, sync, and use cloud API secrets from environment variables or the encrypted local vault. Designed for GitHub Actions / CI/CD pipelines where secrets are provided as TERRADEV_<PROVIDER>_<KEY> env vars.
terradev vault sync imports supported TERRADEV_<PROVIDER>_<KEY> secrets into the encrypted ~/.terradev/credentials.json (use --all to also import custom keys).terradev vault run -- <command> injects vault secrets into a sub-process and zeroizes them afterwards.terradev now automatically falls back to TERRADEV_* environment variables when the local vault file is missing, so terradev up works directly from GitHub Secrets without a separate configure step.
-
Database command (terradev database): universal database and vector store operations with SQLite, PostgreSQL, Qdrant, and Redis adapters. Subcommands: terradev database up, database down, database crud, database search, database sql, and database qdrant with its own search, scroll, upsert, create-collection, and delete-collection operations.
NOTES ON 5.6.0
Added API Gateway for inference serving with OpenAI/Anthropic/custom API entry and exit points:
-
Gateway Service (core/gateway_service.py): FastAPI-based gateway that provides OpenAI-compatible, Anthropic-compatible, and custom workflow API endpoints for inference serving. Integrates with Terradev's inference router and KV cache management for intelligent routing.
-
Gateway CLI command (terradev gateway):
terradev gateway
terradev gateway --host 0.0.0.0 --port 8080
terradev gateway --no-anthropic --max-concurrent 50
terradev gateway --model meta-llama/Llama-3.1-8B-Instruct
-
OpenAI-compatible endpoints:
- POST /v1/chat/completions
- POST /v1/completions
-
Anthropic-compatible endpoints:
- POST /v1/messages
- POST /v1/messages/batches
-
Custom workflow endpoints:
- POST /v1/custom/entry/{workflow_id}
- POST /v1/custom/exit/{workflow_id}
-
Management endpoints:
- GET /health
- GET /v1/gateway/status
-
Features:
- Streaming response support
- Configurable CORS, concurrent requests, timeouts
- Integration with inference router for intelligent routing
- Request/response transformation and validation
NOTES ON 5.3.9
Added LoRAX (LoRA eXchange) integration and HuggingFace PEFT import for production-grade multi-LoRA inference serving:
-
LoRAX Service (ml_services/lorax_service.py): Async HTTP client for Predibase LoRAX multi-LoRA inference server that serves thousands of fine-tuned models on a single GPU with dynamic adapter loading, heterogeneous continuous batching, and adapter exchange scheduling.
-
LoRAX CLI commands (terradev lora lorax):
terradev lora lorax deploy -m mistralai/Mistral-7B-Instruct-v0.1 --docker
terradev lora lorax test --host localhost --port 8080
terradev lora lorax list-adapters
terradev lora lorax load-adapter -a vineetsharma/qlora-adapter-Mistral-7B-Instruct-v0.1-gsm8k
terradev lora lorax unload-adapter -a my-adapter
terradev lora lorax generate -p "What is 2+2?" -a my-adapter
terradev lora lorax sync-registry
-
PEFT Import Service (ml_services/peft_import_service.py): Download, validate, and prepare LoRA adapters from HuggingFace using the PEFT library with auto-detection of rank, alpha, and target modules.
-
PEFT CLI commands (terradev lora peft):
terradev lora peft import -a vineetsharma/qlora-adapter-Mistral-7B-Instruct-v0.1-gsm8k
terradev lora peft import -a username/adapter --local-name my-adapter --register --base-model mistralai/Mistral-7B-Instruct-v0.1
terradev lora peft list
terradev lora peft validate -p ~/.terradev/peft_adapters/username--adapter
terradev lora peft delete -a username/adapter
-
LoRAX Helm Template (clusters/lorax-template/helm/): Production-ready Kubernetes manifests with GPU resource limits, storage configuration.
-
Registry Integration: One-step import from HuggingFace and automatic registration in Terradev LoRA registry with version tracking, cross-replica sync, and cost attribution.
NOTES ON 5.3.3
Added provider registration and profiling system for intelligent quirk-aware routing across 17 cloud providers, and registration for custom providers from .yaml import:
-
ProviderProfile schema (providers/types.py): Encodes provider-specific behaviors including API style (REST/GraphQL/JSON:API), authentication type (Bearer/Basic/HMAC/X-Api-Key), rate limits, spot instance support, egress costs, fallback routing, capacity checks, container image pinning, and spot interruption handling.
-
Built-in profiles (providers/provider_profiles.py): Pre-configured profiles for all 23 providers (RunPod, Vast.ai, Lambda Labs, AWS, GCP, Azure, Oracle, Crusoe, CoreWeave, DigitalOcean, Yotta Labs, E2E Networks, FluidStack, Alibaba, Hetzner, SiliconFlow, TensorDock, Baseten, HuggingFace, Hyperstack, InferX, Latitude).
-
Dynamic registration: Users can register custom provider profiles programmatically or load from YAML/JSON files for internal clusters or proprietary cloud providers.
-
Profile-aware routing (providers/registry.py): ProviderRegistry.ranked_providers() now incorporates provider profiles into scoring, using egress costs, fallback routing preferences, and spot preemption rates for intelligent provider selection.
-
CLI commands (terradev providers): New command group for managing custom provider profiles:
terradev providers load-profiles ~/.terradev/custom_providers.yaml
terradev providers list-profiles
terradev providers show-profile runpod
terradev providers remove-profile my_custom_provider
terradev providers export-example -o ~/.terradev/custom_providers.yaml
NOTES ON 5.2.1
Added two new BYOAPI providers: Yotta Labs (Shakti Cloud) and E2E Networks — India's leading GPU clouds. Yotta Labs uses a pod-based compute model (similar to RunPod), and E2E Networks is a traditional VM-style hyperscaler that is NSE-listed and MeitY empanelled. Both are BYOAPI: your key, stored locally, never touches a Terradev server.
terradev configure --provider yottalabs
terradev configure --provider e2enetworks
NOTES ON 5.0.0
We removed the paywall, open-sourced Terradev, and added accelerators for safe and snappy delivery...
With the DAG orchestrator, the execution graph enforces correct sequencing and idempotency at the runtime level. You or the agent can issue commands freely... the orchestrator ensures they're safe to execute.
217 tools not including subcommand/flags require heavy context. The MCP orchestrator processes tool calls with minimal overhead: deserializing, routing, executing, and responding faster than pure-Python-based MCP servers by an order of magnitude. For an agent running a complex provisioning workflow across 17 cloud providers, that compounds across every tool call in the chain.
BYOAPI Configuration
Your API keys are stored locally at ~/.terradev/credentials.json and never sent to Terradev servers.
terradev configure --provider runpod
terradev configure --provider vastai
terradev configure --provider aws
terradev configure --provider gcp
- 2-8x throughput improvements with vLLM optimization
- 30-50% bandwidth penalty eliminated with NUMA topology
- 2-5x CUDA Graph speedup with optimal topology
- Up to 90% cost savings with automatic provider switching
- <2 minute spot recovery with KV cache checkpointing
- up to 3.6x faster cold starts with weight streaming
- Up to 50% cost savings with MLA-aware VRAM estimation
Complete Tutorial
Step 1: Install Terradev
For all cloud provider SDKs and ML integrations:
pip install terradev-cli[all]
Verify and list commands:
Terradev supports 17 GPU cloud providers. Start with one, RunPod is the fastest to set up:
terradev setup runpod --quick
This shows you where to get your API key. Then configure it:
terradev configure --provider runpod
Paste your API key when prompted. It's stored locally at ~/.terradev/credentials.json, never sent to a Terradev server. Add more providers later:
terradev configure --provider vastai
terradev configure --provider lambda_labs
terradev configure --provider aws
The more providers you configure, the better your price coverage.
Step 3: Get Real-Time GPU Prices
Check pricing across every provider you've configured:
Output is a table sorted cheapest-first: price/hour, provider, region, spot vs. on-demand. Try different GPUs:
terradev quote -g H100
terradev quote -g L40S
terradev quote -g RTX4090
Step 4: Provision
Most clouds hand you GPUs with suboptimal topology by default. Your GPU and NIC end up on different NUMA nodes, RDMA is disabled, and the kubelet Topology Manager is set to none. That's a 30-50% bandwidth penalty on every distributed operation and you'll never see it in nvidia-smi.
When you provision through Terradev, topology optimization is automatic:
terradev provision -g H100 -n 4 --parallel 6
What happens behind the scenes:
- NUMA alignment — GPU and NIC forced to the same NUMA node
- GPUDirect RDMA — nvidia_peermem loaded, zero-copy GPU-to-GPU transfers
- CPU pinning — static CPU manager policy, no core migration
- SR-IOV — virtual functions created per GPU for isolated RDMA paths
- NCCL tuning — InfiniBand enabled, GDR_LEVEL=PIX, GDR_READ=1
You don't configure any of this. It's applied automatically.
To preview the plan without launching:
terradev provision -g A100 -n 2 --dry-run
To set a price ceiling:
terradev provision -g A100 --max-price 2.50
Step 5: Run a Workload
Option A — Run a command on your provisioned instance:
terradev execute -i <instance-id> -c "nvidia-smi"
terradev execute -i <instance-id> -c "python train.py"
Option B — One command that provisions, deploys a container, and runs:
terradev run --gpu A100 --image pytorch/pytorch:latest -c "python train.py"
Option C — Keep an inference server alive:
terradev run --gpu H100 --image vllm/vllm-openai:latest --keep-alive --port 8000
Step 6: Manage Your Instances
terradev status --live
terradev manage -i <instance-id> -a stop
terradev manage -i <instance-id> -a start
terradev manage -i <instance-id> -a terminate
Step 7: Track Costs and Find Savings
terradev analytics --days 30
terradev optimize
Step 8: Distributed Training Pipeline
Now that your nodes have correct topology, distributed training actually runs at full bandwidth:
terradev preflight
terradev train --script train.py --from-provision latest
terradev monitor --job my-job
terradev train-status
terradev checkpoint list --job my-job
The --from-provision latest flag auto-resolves IPs from your last provision command. Supports torchrun, DeepSpeed, Accelerate, and Megatron.
Step 9: Optimize vLLM Inference (The 6 Knobs)
If you're serving a model with vLLM, there are 6 settings most teams leave at defaults — each one costs throughput:
| Knob | Default | Optimized | Impact |
|---|
| max-num-batched-tokens | 2048 | 16384 | 8x throughput |
| gpu-memory-utilization | 0.90 | 0.95 | 5% more VRAM |
| max-num-seqs | 256/1024 | 512-2048 | Prevent queuing |
| enable-prefix-caching | OFF | ON | Free throughput win |
| enable-chunked-prefill | OFF | ON | Better prefill |
| CPU Cores | 2 + #GPUs | Optimized | Prevent starvation |
Auto-tune all six from your workload profile:
terradev ml vllm auto-optimize -s workload.json -m meta-llama/Llama-2-7b-hf -g 4
Or analyze a running server:
terradev ml vllm analyze -e http://localhost:8000
Benchmark:
terradev ml vllm benchmark -e http://localhost:8000 -c 10
Step 10: Deploy a MoE Model with Auto-Applied Optimizations
For large Mixture-of-Experts models (GLM-5, Qwen 3.5, DeepSeek V4), Terradev's MoE templates include every optimization auto-applied — KV cache offloading, speculative decoding, sleep mode, expert load balancing:
terradev provision --task clusters/moe-template/task.yaml \
--set model_id=Qwen/Qwen3.5-397B-A17B
Or a smaller model:
terradev provision --task clusters/moe-template/task.yaml \
--set model_id=Qwen/Qwen3.5-122B-A10B --set tp_size=4 --set gpu_count=4
What's auto-applied (no flags needed):
- KV cache offloading — spills to CPU DRAM, up to 9x throughput
- MTP speculative decoding — up to 2.8x faster generation
- Sleep mode — idle models hibernate to CPU RAM, 18-200x faster than cold restart
- Expert load balancing — rebalances routing at runtime
- LMCache — distributes KV cache across instances via Redis
Step 11: Disaggregated Prefill/Decode (Advanced)
This separates inference into two GPU pools optimized for each phase:
- Prefill (compute-bound) — processes input prompt, wants high FLOPS
- Decode (memory-bound) — generates tokens, wants high HBM bandwidth
The KV cache transfers between them via NIXL — zero-copy GPU-to-GPU over RDMA. This is why getting the NUMA topology right in Step 4 matters: NIXL only runs at full speed when the GPU and NIC share a PCIe switch.
terradev ml ray --deploy-pd \
--model zai-org/GLM-5-FP8 \
--prefill-tp 8 --decode-tp 1 --decode-dp 24
Terradev's inference router automatically uses sticky routing. Once a prefill GPU hands off a KV cache to a decode GPU, future requests with the same prefix go to that same decode GPU, avoiding redundant transfers.
Step 12: Create a Kubernetes GPU Cluster
For production, create a topology-optimized K8s cluster:
terradev k8s create my-cluster --gpu H100 --count 8 --prefer-spot
This auto-configures Karpenter NodePools with NUMA-aligned kubelet Topology Manager, GPUDirect RDMA, and PCIe locality enforcement.
terradev k8s list
terradev k8s info my-cluster
terradev k8s destroy my-cluster
Why This Order Matters
Each step builds on the one before it:
- Step 4: NUMA / RDMA / SR-IOV topology ← foundation
- Step 8: Distributed training at full BW ← depends on topology
- Step 9: vLLM knob tuning ← depends on correct memory layout
- Step 10: KV cache offloading + sleep mode ← depends on CPU bus not saturated
- Step 11: Disaggregated P/D ← depends on RDMA for KV transfer
If the provisioning layer is wrong, every optimization above it underperforms. A disaggregated P/D setup with a cross-NUMA KV transfer is slower than a monolithic setup with correct topology.
Terradev handles the foundation automatically so the rest of the stack works the way it's supposed to.
Quick Reference
terradev configure
terradev quote -g H100
terradev provision -g H100 -n 4
terradev run --gpu A100 --image ...
terradev status --live
terradev train --from-provision latest
terradev ml vllm auto-optimize
terradev k8s create
terradev analytics --days 30
terradev optimize
Troubleshooting Training Workflows
NCCL Connectivity Problems
terradev preflight --detailed
terradev execute -i <node-id> -c "nccl_test -b 8G -e 8G -s 1073741824"
terradev provision -g H100 -n 4 --parallel 6 --ensure-numa-alignment
GPU Memory Issues
terradev monitor --job <job-id> --memory-usage
terradev execute -i <node-id> -c "nvidia-smi --query-gpu=memory.used,memory.total --format=csv"
terradev train --script train.py --from-provision latest --script-args "--batch-size 16 --gradient-checkpointing"
Dataset Staging Failures
terradev stage --status --dataset-id <dataset-id>
terradev stage --list-cached --region us-east-1
terradev stage -d ./my-dataset --target-regions us-east-1 --parallel-streams 64 --compression zstd
FlashOptim Compatibility Issues
terradev train-status --job <job-id> | grep flashoptim
terradev preflight --flashoptim-check
terradev train --script train.py --flashoptim off --from-provision latest
terradev train --script train.py --flashoptim on --flashoptim-optimizer adamw --flashoptim-master-weight-bits 8
Checkpoint Recovery Issues
terradev checkpoint list --job <job-id> --verify
terradev checkpoint validate --checkpoint <checkpoint-path>
terradev checkpoint save --job <job-id> --force
terradev checkpoint repair --checkpoint <checkpoint-path>
Performance Optimization
Slow Training Speed
terradev monitor --job <job-id> --bottleneck-analysis
terradev execute -i <node-id> -c "nvtop --interval 1"
terradev train --script train.py --script-args "--mixed-precision --fp16"
terradev stage --hf-dataset <dataset> --target-regions us-east-1 --preprocess "shuffle,cache"
terradev provision -g H100 -n 8 --parallel 12
Network Bottlenecks
terradev preflight --network-test
terradev execute -i <node-id> -c "ibstat -v"
terradev provision -g H100 -n 4 --ensure-rdma --enable-gpudirect
Contributing
We welcome contributions! Please see our Contributing Guide for details.
License
Apache 2.0.
Support