machine-type

A machine type is a catalog resource that defines the compute shape of an agent VM. Each environment references one machine type by name via machine_type_ref. The platform provides built-in machine types for common VM shapes on GCP and AWS. Built-in machine types are immutable: you cannot edit one or delete it outright. murmur rm machine-type <name> disables a built-in, and murmur set machine-type <name> with no input re-enables it. A VM reads its compute rate from its machine type and its storage rate from its disk type, so price every region your placements use in both entries. See Pricing for every built-in machine type’s current list rate.

Fields

Name	Type	Required	Description
`name`	string	yes	Unique identifier. DNS label format: `[a-z][a-z0-9-]{0,62}`.
`description`	string	no	Human-readable description shown in the dashboard. Maximum 1024 bytes.
`vcpus`	int32	yes	Number of virtual CPUs. Must be positive.
`memory_gb`	int32	yes	Memory in GiB. Must be positive.
`architecture`	string	yes	CPU architecture. Values: `ARCHITECTURE_AMD64`, `ARCHITECTURE_ARM64`.
`has_gpu`	bool	no	Whether this machine type includes a GPU. Default: `false`. Set it only where the machine type itself includes one: see GPU support.
`gpu_kind`	string	conditional	GPU kind (e.g. `t4`, `l4`, `a100`). Required when `has_gpu` is `true`. Must be empty when `has_gpu` is `false`.
`preserves_memory_on_sleep`	bool	no	Whether sleeping an agent on this shape preserves its memory, so it wakes where it left off. Shapes that are very large, carry a GPU, or are bare metal cannot preserve memory, and setting it on one is rejected.
`gce`	object	conditional	GCE machine type details. Exactly one of `gce` or `aws` is required.
`aws`	object	conditional	AWS EC2 instance type details. Exactly one of `gce` or `aws` is required.

GCE spec fields

Set gce for GCP machine types:

Name	Type	Required	Description
`gce.machine_type`	string	yes	GCE machine type (e.g. `n2-standard-32`). An accelerator-optimized machine type (e.g. `a2-highgpu-1g`) includes its own GPUs and must be paired with `has_gpu`. Must match `^[a-z][a-z0-9]([a-z0-9-]*[a-z0-9])?$`.
`gce.accelerator_type`	string	no	Must be empty. Separately attached accelerators are not yet supported: see GPU support.
`gce.accelerator_count`	int32	no	Must be `0`. Separately attached accelerators are not yet supported: see GPU support.
`gce.regions`	map	yes	Per-region cost configuration. Lists the regions this entry is priced in: include every region your placements use. At least one entry required. Keys are region strings (e.g. `us-central1`).
`gce.supported_disk_types`	string[]	no	GCE disk types this machine type accepts (e.g. `hyperdisk-balanced`), each matching a disk type’s `gce.disk_type`. Leave empty to accept any disk type: see Disk compatibility.

Disk compatibility

An environment names a machine type and a disk type independently, and not every GCE machine family accepts every disk type: the newer families take Hyperdisk and refuse Persistent Disk. List the disk types a machine type accepts in gce.supported_disk_types, and an environment pairing it with a disk type outside that list is rejected when the environment is written, rather than when GCE refuses to create the VM. Compatibility is declared, not inferred. A machine type whose gce.supported_disk_types is empty declares no restriction and pairs with any disk type. Each entry is matched verbatim against a disk type’s gce.disk_type, so entries must be non-empty and distinct. Both records the environment names are held to the pairing for as long as the environment stands: a machine type whose gce.supported_disk_types no longer covers a disk type an environment pairs it with is rejected, as is a disk type repointed at a gce.disk_type its environment’s machine type does not accept. The error names the environment the write would break, so either change is made by writing the environment onto a compatible pair first.

name: c4a-standard-4
vcpus: 4
memory_gb: 16
architecture: ARCHITECTURE_ARM64
gce:
  machine_type: c4a-standard-4
  supported_disk_types:
    - hyperdisk-balanced
  regions:
    us-central1:
      cost_per_hour: 0.16

AWS spec fields

Set aws for AWS machine types:

Name	Type	Required	Description
`aws.instance_type`	string	yes	EC2 instance type (e.g. `m5.8xlarge`). Must match `^[a-z][a-z0-9-]*\.[a-z0-9]+$`.
`aws.regions`	map	yes	Per-region cost configuration. Lists the regions this entry is priced in: include every region your placements use. At least one entry required. Keys are region strings (e.g. `us-east-1`).

GPU support

A GPU reaches an agent VM through the machine type itself. On GCE, that is an accelerator-optimized machine type: the a2, a3, a4, a4x, g2, and g4 series each include a fixed GPU model and count, so machine_type: a2-highgpu-1g with has_gpu: true provisions the GPUs that machine type comes with. On AWS, an EC2 GPU instance type such as g5.2xlarge works the same way. Nothing attaches a GPU to a GCE machine type that lacks one, so on GCE has_gpu and the machine type must agree:

has_gpu: true requires an accelerator-optimized gce.machine_type. On any other machine type it claims a GPU the VM would not have, and is rejected.
An accelerator-optimized gce.machine_type requires has_gpu: true (with gpu_kind), so the machine type describes the GPUs its VMs get.
gce.accelerator_type and gce.accelerator_count request accelerators attached separately to an ordinary machine type. That is not yet supported: setting either is rejected.

A rejected declaration fails the write, and an environment referencing a machine type stored with one fails VM creation with UNSUPPORTED_GPU_SHAPE.

Region configuration

Each entry in a regions map is a RegionConfig object:

Name	Type	Required	Description
`cost_per_hour`	double	yes	On-demand cost per hour in USD. Must be greater than 0 and at most 1000000, the largest rate murmur can bill.
`spot_cost_per_hour`	double	no	Spot cost per hour in USD. Set it to offer this machine type on Spot in this region; unset means Spot is not offered here. Must be greater than 0 when set. An environment `provisioning_models` list uses it.

Examples

Listing machine types

murmur get machine-type

murmur-m5-8xlarge
murmur-n2-standard-32

The list prints one name per line, with no header. A disabled built-in prints with a (disabled — re-enable: murmur set machine-type <name>) marker.

Reading a single machine type

murmur get machine-type murmur-n2-standard-32

GCE machine type

name: my-gce-large
description: "GCE n2-standard-16"
vcpus: 16
memory_gb: 64
architecture: ARCHITECTURE_AMD64
has_gpu: false
gce:
  machine_type: n2-standard-16
  regions:
    us-central1:
      cost_per_hour: 0.78

GCE GPU machine type

An accelerator-optimized machine type includes its GPUs, so no accelerator fields are set:

name: my-gce-gpu
description: "GCE a2-highgpu-1g with one A100"
vcpus: 12
memory_gb: 85
architecture: ARCHITECTURE_AMD64
has_gpu: true
gpu_kind: a100
gce:
  machine_type: a2-highgpu-1g
  regions:
    us-central1:
      cost_per_hour: 3.67

AWS machine type

name: my-aws-large
description: "EC2 m5.8xlarge"
vcpus: 32
memory_gb: 128
architecture: ARCHITECTURE_AMD64
has_gpu: false
aws:
  instance_type: m5.8xlarge
  regions:
    us-east-1:
      cost_per_hour: 1.536
    us-west-2:
      cost_per_hour: 1.536

Creating a machine type

cat <<'EOF' | murmur set machine-type my-aws-large
name: my-aws-large
description: "EC2 m5.8xlarge"
vcpus: 32
memory_gb: 128
architecture: ARCHITECTURE_AMD64
has_gpu: false
aws:
  instance_type: m5.8xlarge
  regions:
    us-east-1:
      cost_per_hour: 1.536
EOF

Errors

Code	Meaning	What to do
`INVALID_ARGUMENT`	`name is required`	Provide a `name` field.
`INVALID_ARGUMENT`	`name must match [a-z][a-z0-9-]{0,62}`	Use a lowercase DNS label for the name.
`INVALID_ARGUMENT`	`description exceeds 1024 byte limit`	Shorten the `description` to 1024 bytes or fewer.
`INVALID_ARGUMENT`	`vcpus must be positive`	Set `vcpus` to a value greater than 0.
`INVALID_ARGUMENT`	`memory_gb must be positive`	Set `memory_gb` to a value greater than 0.
`INVALID_ARGUMENT`	`architecture is required`	Set `architecture` to `ARCHITECTURE_AMD64` or `ARCHITECTURE_ARM64`.
`INVALID_ARGUMENT`	`unknown architecture value <N>`	Use a valid `architecture` value.
`INVALID_ARGUMENT`	`gpu_kind is required when has_gpu is true`	Provide `gpu_kind` when `has_gpu` is `true`.
`INVALID_ARGUMENT`	`gpu_kind must be empty when has_gpu is false`	Remove `gpu_kind`, or set `has_gpu` to `true` on a machine type that includes GPUs.
`INVALID_ARGUMENT`	`spec is required ...`	Provide exactly one of `aws` or `gce`.
`INVALID_ARGUMENT`	`aws.instance_type "<value>" does not match EC2 format`	Use a valid EC2 instance type string (e.g. `m5.8xlarge`).
`INVALID_ARGUMENT`	`gce.machine_type "<value>" does not match GCE format`	Use a valid GCE machine type string (e.g. `n2-standard-32`).
`INVALID_ARGUMENT`	`gce.accelerator_type and gce.accelerator_count must be unset: separately attached accelerators are not yet supported on GCE. Use an accelerator-optimized machine type, which includes its GPUs`	Remove `gce.accelerator_type` and `gce.accelerator_count`, and name an accelerator-optimized `machine_type` such as `a2-highgpu-1g`: see GPU support.
`INVALID_ARGUMENT`	`has_gpu requires an accelerator-optimized gce.machine_type (a2, a3, a4, a4x, g2, g4 series); "<value>" includes no GPU`	Name an accelerator-optimized `gce.machine_type`, or remove `has_gpu` and `gpu_kind`: see GPU support.
`INVALID_ARGUMENT`	`gce.machine_type "<value>" includes GPUs: set has_gpu and gpu_kind to describe them`	Set `has_gpu: true` and `gpu_kind` on the machine type, so it describes the GPUs its VMs get.
`INVALID_ARGUMENT`	`aws.regions must be non-empty`	Add at least one region to `aws.regions`.
`INVALID_ARGUMENT`	`gce.regions must be non-empty`	Add at least one region to `gce.regions`.
`INVALID_ARGUMENT`	`gce.supported_disk_types[<index>] must be non-empty`	Remove the empty entry, or name the disk type’s `gce.disk_type` it was meant to hold.
`INVALID_ARGUMENT`	`gce.supported_disk_types[<index>] repeats "<value>"`	Remove the duplicate entry: listing a disk type twice allows nothing extra.
`INVALID_ARGUMENT`	`aws.regions["<region>"].cost_per_hour must be > 0`	Set `cost_per_hour` to a positive value for the given region.
`INVALID_ARGUMENT`	`gce.regions["<region>"].cost_per_hour must be > 0`	Set `cost_per_hour` to a positive value for the given region.
`INVALID_ARGUMENT`	`<spec>.regions["<region>"].cost_per_hour <value> is above the maximum billable rate of 1e+06 USD`	Use a rate of 1000000 or less.
`INVALID_ARGUMENT`	`<spec>.regions["<region>"].spot_cost_per_hour must be > 0, or omitted where the machine type is not offered on Spot`	Set a positive Spot rate, or omit `spot_cost_per_hour`.
`INVALID_ARGUMENT`	`preserves_memory_on_sleep: <reason>`	This shape cannot preserve memory across a sleep. Remove `preserves_memory_on_sleep`.
`INVALID_ARGUMENT`	`platform machine types are reserved for builtins and cannot be written via the catalog`	Do not set the `platform` field. Built-in machine types are managed by the platform.
`FAILED_PRECONDITION`	`cannot delete machine-type "<name>": referenced by environment: <names>`	Remove the `machine_type_ref` from the referencing environments before deleting the machine type.

Environments and placements: concept overview
Catalog: how catalog resources work
murmur set: CLI command for creating and updating catalog resources
murmur get: CLI command for reading catalog resources

Getting Started

Core Concepts

Guides

Fields

GCE spec fields

Disk compatibility

AWS spec fields

GPU support

Region configuration

Examples

Listing machine types

Reading a single machine type

GCE machine type

GCE GPU machine type

AWS machine type

Creating a machine type

Errors

​Fields

​GCE spec fields

​Disk compatibility

​AWS spec fields

​GPU support

​Region configuration

​Examples

​Listing machine types

​Reading a single machine type

​GCE machine type

​GCE GPU machine type

​AWS machine type

​Creating a machine type

​Errors

​Related

Fields

GCE spec fields

Disk compatibility

AWS spec fields

GPU support

Region configuration

Examples

Listing machine types

Reading a single machine type

GCE machine type

GCE GPU machine type

AWS machine type

Creating a machine type

Errors

Related