For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python module

max.driver

Device programming for MAX: allocate buffers, manage devices, and dispatch work through device queues.

Devices

`Accelerator`	Creates an accelerator device with the specified ID and memory limit.
`CompletionFlag`	An 8-byte completion flag in pinned host memory mapped into a device's address space.
`CPU`	Creates a CPU device.
`Device`	Represents a compute device available for tensor operations.
`DeviceEvent`	Provides access to an event object.
`DeviceQueue`	Provides access to a queue of execution on a device.
`DeviceSpec`	A device specification containing an ID and device type.

`Buffer`	Device-resident buffer representation.
`DLPackArray`	Protocol for objects that exchange data through the DLPack interface.
`HostHazardError`	Raised when host code waits on a completion whose producer failed.
`SymmetricPool`	A driver-owned symmetric multicast pool shared by a set of devices.
`Usage`	Allocation-intent descriptor for `Buffer`.

`batch_inplace_copy`	Copies `srcs[i]` into `dsts[i]` in as few driver submissions as possible.
`copy_pinned_to_destinations`	Copies a host buffer into destination buffers safely.
`load_max_buffer`	Experimental method for loading serialized MAX buffers.

`accelerator_api`	Returns the API used to program the accelerator.
`accelerator_architecture_name`	Returns the architecture name of the accelerator device.
`accelerator_count`	Returns number of accelerator devices available.
`devices_exist`	Returns `True` if all specified devices exist.
`enable_all_peer_access`	Enables peer-to-peer memory access between all available GPU pairs.
`load_devices`	Initializes and returns a list of `Device` objects from a sequence of `DeviceSpec` objects.
`scan_available_devices`	Returns all available accelerators, or a CPU device if none are available.

`LaunchTraceEntry`	One operation recorded by `max.driver.begin_launch_trace`.

`begin_launch_trace`	Starts a process-global recording of enqueued device operations.
`launch_trace`	Records the device operations enqueued within the `with` block.
`take_launch_trace`	Stops the global recording and returns the recorded entries.

`calculate_virtual_device_count`	Calculates the minimum virtual device count needed for the given device specs.
`calculate_virtual_device_count_from_cli`	Calculates the minimum virtual device count from raw CLI inputs (before parsing).
`get_virtual_cpu_target`	Gets the current virtual CPU target.
`get_virtual_device_api`	Gets the current target API for virtual devices.
`get_virtual_device_count`	Gets the current virtual device count.
`get_virtual_device_target_arch`	Gets the current target GPU architecture for virtual devices.
`is_virtual_device_mode`	Checks if virtual device mode is currently enabled.
`set_virtual_cpu_target`	Sets the CPU target for host-independent kernel codegen.
`set_virtual_device_api`	Sets the target API for virtual devices in compile-only mode.
`set_virtual_device_count`	Sets the number of virtual devices for device creation.
`set_virtual_device_target_arch`	Sets the target GPU architecture for virtual devices in compile-only mode.