GPU Pro: pick the cards yourself
Access on request GPU Pro is opened per project.
If GET /v1/server-types does not list it for your project, your project has no access to
it yet - ask support to open it. Until then the type answers server_type_not_found
(404), the same as a type that does not exist.
GPU Pro is ordered through the same POST /v1/servers as any other server. Unlike the
packaged GPU Server, nothing comes as a set: you name the
cards, then size the cores, the memory and the disk yourself, and you may choose the physical host
the server lands on.
1. Find the type
curl -H "Authorization: Bearer $FBX_TOKEN" \
https://api.fiberax.com/v1/server-types
{ "data": [
{ "id": 8, "name": "GPU Pro", "pricing_mode": "sliders", "gpu": true,
"description": "Build your own GPU setup: pick one or more cards, then set CPU, RAM and disk." } ] }
Other types omitted for brevity.
"gpu": true marks a type with graphics cards, and "pricing_mode": "sliders"
says you size it yourself. The catalogue is the authority on what your project may order: the type
is there exactly when your project has access to it.
2. Read the cards and the ceilings
curl -H "Authorization: Bearer $FBX_TOKEN" \
https://api.fiberax.com/v1/server-types/8/pricing
{ "data": { "server_type_id": 8, "pricing_mode": "sliders",
"limits": { "cpu": 128, "ram_gb": 900, "disk_gb_local": 6500, "disk_gb_ceph": 5000, "gpus": 16 },
"gpu_models": [ { "key": "l40s", "name": "L40S", "memory_gb": 48, "price": 0 },
{ "key": "rtx6000bw", "name": "RTX PRO 6000 Blackwell", "memory_gb": 96, "price": 0 },
{ "key": "l4", "name": "L4", "memory_gb": 24, "price": 0 },
{ "key": "h200nvl", "name": "H200 NVL", "memory_gb": 141, "price": 0 } ] } }
Fields omitted for brevity.
gpu_models lists the cards of this type: the key is what you order by,
memory_gb is the memory of the card itself. A price of 0 means the card is
included in the price of the type. Read the list rather than copying it - the models on offer can
change.
limits are the ceilings a request is checked against on this type: cores, memory,
the system disk for each storage class, and gpus - how many cards one server may take.
Cores, memory or disk above them are refused with resource_limit_exceeded (422),
more cards than gpus with gpu_count_exceeded (422); the refusal names the
ceiling it hit. Within them the size is refused only below the minimum of the image you pick
(image_minimum_not_met, 422 - the image list in step 3 carries min_cores,
min_memory_gb and min_disk_gb); what can still stop the request is room on
the hosts (no_gpu_capacity, 409).
3. Pick an image
curl -H "Authorization: Bearer $FBX_TOKEN" \
https://api.fiberax.com/v1/server-types/8/images
{ "data": [ { "id": "fcf9f69c...", "name": "Ubuntu 24.04", "type": "os" },
{ "id": "b8909e50...", "name": "Ubuntu 24.04 + NVIDIA driver 595-open", "type": "os" },
{ "id": "eb0446da...", "name": "Windows Server 2025", "type": "os" },
{ "id": "uiso-40", "name": "your own ISO image", "type": "useriso" } ] }
Fields omitted for brevity.
The entry with the graphics driver already installed is usually the one you want on a GPU
machine. An entry with "type": "useriso" is your own uploaded ISO, and a server
cannot be CREATED from it - such a request is refused with image_not_supported
(422). Insert it into the drive of an existing server instead (PUT /v1/servers/{id}/iso).
4. Choose the host - or leave it to the platform
A server with cards lives on one physical host, and on this type you may choose which. Skip this step and the platform places the server itself. The list names the hosts of the type:
curl -H "Authorization: Bearer $FBX_TOKEN" \
https://api.fiberax.com/v1/server-types/8/hosts
{ "data": { "host_selection": true,
"hosts": [ { "name": "node-1", "accepting": true },
{ "name": "node-2", "accepting": true },
{ "name": "node-3", "accepting": true },
{ "name": "node-4", "accepting": true } ] } }
host_selection says whether the type takes a host on create; name is the
value to send; accepting says whether the host takes new servers at all. Hosts are
named by labels, never by their real names, and a label stays with its host.
Describe the server you want in the query, and each host says whether it would take it:
curl -H "Authorization: Bearer $FBX_TOKEN" \
"https://api.fiberax.com/v1/server-types/8/hosts?gpu_model=l40s&gpus=1&cores=8&memory_mb=65536&disk_gb=200&disk_type=local"
{ "data": { "host_selection": true,
"hosts": [ { "name": "node-1", "accepting": true, "fits": false, "reason": "cards" },
{ "name": "node-2", "accepting": true, "fits": true, "reason": null },
{ "name": "node-3", "accepting": true, "fits": false, "reason": "cards" },
{ "name": "node-4", "accepting": true, "fits": false, "reason": "cards" } ] } }
gpu_model and gpus go together; the rest is optional, and what you do
not name is not checked. When fits is false, reason says why in one word:
cards, memory (the reserved memory of the host), local_disk,
no_local_storage or not_accepting. The answer is a verdict on your request,
not a count of what is free - the list carries no numbers. A host shown as fitting will take the
request, unless someone else takes the room first.
5. Order it
curl -X POST -H "Authorization: Bearer $FBX_TOKEN" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: trainer-02-2026-09-29" \
-d '{
"name": "trainer-02",
"server_type_id": 8,
"gpus": [ { "model": "l40s", "count": 1 } ],
"cores": 8,
"memory_mb": 65536,
"disk_gb": 200,
"disk_type": "local",
"host": "node-2",
"image": "b8909e50bf3274376f38830c8ef5a747",
"public_ipv4": { "reuse": true },
"ssh_keys": [126]
}' \
https://api.fiberax.com/v1/servers
The answer is 202 with the server and a task, like any create: wait for the task,
then read the server. A server with cards takes longer to start than an ordinary one - allow for
it in your timeouts.
gpus- one entry per model, keys fromgpu_models; the same model named twice simply adds up. More cards in total thanlimits.gpusgivesgpu_count_exceeded(422); a model outside this type's catalogue givesgpu_model_not_found(404); leaving the list out givesgpus_required(422).cores,memory_mb,disk_gb- as on any slider type, within the ceilings from step 2;memory_mbin whole gigabytes.disk_type- the storage class of the system disk:local(the default) is the disk of the host the cards live in,cephis shared storage. Each class has its own ceiling inlimits. A GPU type may be declared with a local system disk only - none is at the moment - and therecephanswersdisk_type_not_available(422) - do not infer this from the ceilings.host- optional, a label from step 4. A label this type does not have answershost_not_found(404); a type without host choice answershost_not_supported(422). If the host list cannot be read for a moment, the answer isserver_unavailable(503) withRetry-After- repeat, do not drop the field.- Access and money. The project owner needs a payment card or credit on the account, and a confirmed e-mail:
without them creating any server answers 402 (
card_verification_required,email_verification_required).
Read the server back and it carries what you ordered: gpus, disk_type,
and host - the host it runs on right now. The operator may move a server to another
host; host follows the move.
6. When there is no room
Cards are not shared between servers, so the room is finite. When the request does not fit:
{ "status": 409, "code": "no_gpu_capacity",
"detail": "No host of this type has the requested cards free: ask for fewer cards or another
card model, or try again later - cards free up when servers that hold them are
deleted. The reason is worked out from the platform inventory as it reads at the
moment of this answer; at the moment of the refusal it may have been another one." }
There is deliberately no Retry-After here: for a shortage, repeating the same
request will not free a card. One code covers several causes, and detail names which one
it was:
- not enough free cards - ask for fewer, or for a different model;
- not enough reserved memory on a host that has the cards - ask for less memory;
- not enough local disk there - ask for a smaller disk, or put the system disk on shared storage
with
"disk_type": "ceph"; - no local storage on the hosts at all - put the system disk on shared storage with
"disk_type": "ceph"; - the host that could take the cards is not answering - repeat the same request in a moment;
- the cluster is not available, or no host is ready to take a server - repeat it later.
With a host chosen, the refusal is about that host: detail names it and what it
lacks, and suggests choosing another host or leaving the choice to the platform. A server never
moves to another host on its own - that is the point of choosing. The list from step 4 answers
the same question before you order.
A refused request costs you nothing. If it asked for a new public address and the server was not created, that address is released again. An address you named yourself stays yours, as it was before the request.
7. After it is running
- The cards are fixed.
gpusis not accepted in an update: a different set of cards is a different server. - Cores and memory change like on any slider type, within the ceilings and within the
reserved memory of the host the cards are on - when that runs out, the answer is
no_gpu_capacity(409) without adetail, and the server stays as it was. - Data disks are always placed on shared storage, whatever the class of the system disk. A
type declared with a local system disk only would take no data disks at all
(
disk_type_not_available, 422). - No image can be taken from it.
POST /v1/server-imagesanswersserver_type_not_supported: the cards are attached by the address of a device on one specific host, and an image carrying that would not run anywhere else. - An ISO can still be inserted into its drive - that does not touch the cards.
- Everything else is the same as for any server: networks, firewall, backups, addresses, and the clean-up described in A server from start to finish.
- In Terraform the same fields are on
fiberax_server-gpus,disk_typeandhost, with the current host incurrent_host- and the host list is the data sourcefiberax_server_type_hosts. See the Terraform reference.