nvml-wrapper

Fixed

• Corrected NVML_FI_PWR_SMOOTHING_* field IDs which used CUDA 13.0U1 numbering instead of v12, causing silent data corruption on CUDA 12 hosts (#134, #137)
• Added runtime driver version detection and transparent field ID remapping for drivers >= 580.82 (CUDA 13.0 Update 1+) so callers don't need to handle the v12/v13U1 numbering difference (#137)
• Fixed APIs that need double-calls for buffer size calculation (#129)
• Fixed nvmlDeviceGetActiveVgpus argument ordering (#128)

Added

• FieldIdScheme enum and Nvml::field_id_scheme() to inspect the detected driver's field ID numbering
• Method for getting vGPU type ID (#124)

chore: Release nvml-wrapper version 0.12.0

And nvml-wrapper-sys 0.9.0

Fixes, improvements, and updates for NVML 12.8.90.

Fixed

• Include license text in the packaged crates (#66)
• Fix EventLoop elided lifetime warnings (#69)
• Use proper api version from header. (#80)

Added

• Add Gpc and Mem clock offset methods (#61)
• Add nvmlDeviceGetP2PStatus (#62)
• Add Blackwell to DeviceArchitecture enum (#70)
• Add DeviceAttributes and device::attributes(). (#71)
• Add attestation report fetch functionality to device (#73)
• Add confidential compute gpu certificate functionality (#74)
• Add checks for confidential compute (#76)
• Add fan control functions (#77)
• Add device::cpu_affinity_within_scope. (#79)
• Add FanSpeedInfo and FanSpeedRPM wrappers. (#80)
• Add device::memory_affinity(). (#82)
• Add device::mig_mode(). (#83)
• Add wrappers for clock offset and pstate functions, min_max_fan_speed (#85)
• Add device::virtualization_mode() wrapper. (#86)
• Add vgpu_driver_capabilities() (#87)
• Add device::performance_modes() (#88)
• Add device::active_vgpus() (#89)
• Add util functions for cc and multi-gpu ppcie (#90)
• Add set fan control policy function (#91)
• Add last_seen_timestamp parameter to get more accurate process utilization counts. (#92)

Changed

• Update nvmlDeviceGetMemoryInfo to version 2 to be consistent with nvidia-smi (#58)
• Set LIB_PATH to libnvidia-ml.so.1 instead of libnvidia-ml.so on Linux (#63)

And 0.8.0 of nvml-wrapper-sys

• Release 0.9.0
  And 0.7.0 of nvml-wrapper-sys

• Downgrade indexmap
  In order to downgrade hashbrown to avoid requiring Rust 2021 edition

Release 0.8.0 (#30)

Release Summary

Dynamically loading the NVML library at runtime is here! Thanks to a new bindgen feature that landed recently, nvml-wrapper-sys now has regenerated bindings that make use of the libloading crate and don't require linking to NVML at compile-time.

This means it's now possible to drop NVIDIA-related features in your code at runtime on systems that don't have relevant hardware.

Rust Version Support

The MSRV of this release is 1.42.0.

Added

• NVML struct:
  • Added methods:
    • builder
• NvmlBuilder struct
  • A builder struct that provides further flexibility in how NVML is initialized
• bitmasks
  • Added events:
    • POWER_SOURCE_CHANGE
    • MIG_CONFIG_CHANGE
  • Added init flags:
    • NO_ATTACH
• Device struct:
  • Added methods:
    • new (replaces the From<nvmlDevice_t> impl)
    • nvml
• EventSet struct:
  • Added methods:
    • new (replaces the From<nvmlEventSet_t> impl)
• EventData struct:
  • Added methods:
    • new (replaces the From<nvmlEventData_t> impl)
• Unit struct:
  • Added methods:
    • new (replaces the From<nvmlUnit_t> impl)
    • nvml
• NvmlError
  • Added variants:
    • LibloadingError
    • FailedToLoadSymbol

Changed

• InitFlags now implements Default
• NvmlError and NvmlErrorWithSource no longer implement Clone, Eq, or PartialEq
  • They can no longer implement these traits because libloading::Error doesn't implement them
  • As a result of this change, FieldValueSample no longer implements Clone or PartialEq
• NvmlError::UnexpectedVariant now contains a u32 instead of an i32 due to a change in the types of the generated bindings

Dependencies

• libloading: new dependency on 0.6.6
• static_assertions: new dependency on 1.1.x

Bump nvml-wrapper version to 0.6.0

Release Summary

A long time in the works, 0.5.0 contains the last two years of my extremely sporadic work wrapping some of the new functionality provided in NVML since version 8 alongside a handful of small fixes and improvements.

Additions

• An import library (nvml.lib) has been added that enables compilation using the MSVC toolchain on Windows.
• The basic_usage example now prints the system's CUDA driver version.
• bitmasks
  • Added throttle reasons:
    • SW_THERMAL_SLOWDOWN
    • HW_THERMAL_SLOWDOWN
    • HW_POWER_BRAKE_SLOWDOWN
    • DISPLAY_CLOCK_SETTING
  • Added device::FbcFlags
  • Added InitFlags
• enums::device
  • SampleValue
    • Added variant I64(i64)
    • The from_tag_and_union constructor has been updated to support i64 values
• enum_wrappers::device
  • MemoryLocation
    • Added variants:
      • Cbu
      • SRAM
  • TemperatureThreshold
    • Added variants:
      • MemoryMax
      • GpuMax
  • PerformancePolicy
    • Added variants:
      • BoardLimit
      • LowUtilization
      • Reliability
      • TotalAppClocks
      • TotalBaseClocks
  • SampleValueType
    • Added variant SignedLongLong
  • Brand
    • Added variant Titan
  • Added the EncoderType enum
  • Added the DetachGpuState enum
  • Added the PcieLinkState enum
  • Added the FbcSessionType enum
• Device struct:
  • Added methods:
    • cuda_compute_capability
    • encoder_capacity
    • encoder_stats
    • encoder_sessions
    • encoder_sessions_count
    • fbc_stats
    • fbc_sessions_info
    • fbc_session_count
    • process_utilization_stats
    • total_energy_consumption
    • field_values_for
    • set_gpu_locked_clocks
    • reset_gpu_locked_clocks
• error
  • Added errors:
    • InsufficientMemory
    • VgpuEccNotSupported
• NVML struct:
  • Added methods:
    • init_with_flags
    • sys_cuda_driver_version
    • blacklist_device_count
    • blacklist_device_info
• structs::device
  • Added structs:
    • EncoderStats
    • CudaComputeCapability
    • FieldId
    • RetiredPage
• struct_wrappers
  • Added struct:
    • BlacklistDeviceInfo
• struct_wrappers::device
  • Added structs:
    • EncoderSessionInfo
    • ProcessUtilizationSample
    • FieldValueSample
    • FbcStats
    • FbcSessionInfo
• lib.rs
  • Added functions:
    • cuda_driver_version_major
    • cuda_driver_version_minor

Removals

• bitmasks
  • ThrottleReasons::Unknown was removed since its counterpart in the NVML library was removed

Changes

• enum_wrappers::device
  • TopologyLevel
    • The Cpu variant was replaced by the Node variant
• The UnexpectedVariant error value is now an i32 (previously u32)
• The Device.remove() method now takes additional parameters for more removal options
• The Device.fan_speed() method now takes a fan index to allow reading the speed of different fans
• The Device.retired_pages() method now returns the timestamps for each page's retirment along with their addresses
• The NVML.sys_cuda_driver_version() method now errors if the CUDA shared library cannot be found

Fixes

• Methods that allocate i8 vectors to be passed as cstrings now do so via the vec! macro rather than simply using with_capacity, meaning the length of the vector gets set appropriately
  • This did not cause a memory leak because we were just working with primitive types that don't have Drop impls, but it's nice to have fixed regardless

Dependencies

• error-chain: 0.11.x -> 0.12.x

Also 0.3.1 of the sys crate

Release Summary

This is a small release that updates dependencies and makes a handful of changes for forward-compatibility purposes.

Rust Version Support

This release requires and supports Rust 1.20.0 or higher.

Additions

• CI has been set up
  • All it can do is build the crate (no testing of any kind), but at least it's something.

Changes

• EventData::try_from() is replaced by a From<nvmlEventData_t> impl as it can no longer error
  • This is because of the from_bits_truncate() usage described next
• Methods that deal with bitmasks now use the from_bits_truncate() constructor instead of from_bits()
  • This allows the wrapper, which is using bindings for NVML 8, to still accept bitmasks from future versions of NVML (such as NVML 9) that may have additional flags
  • *_strict() method counterparts are available for most such methods if you need them
• As a result of the bitflags update, flags are now associated constants
  • This means that, for instance, nvml_wrapper::bitmasks::event::CLOCK_CHANGE is now nvml_wrapper::bitmasks::event::EventTypes::CLOCK_CHANGE
• Imports were deglobbed in a number of places
• The basic_usage example now uses pretty-bytes instead of number_prefix to pretty-print bytes

Dependencies

• bitflags: 0.9.x -> 1.0.x
• error-chain: 0.10.x -> 0.11.x

Release Summary

The major highlight of this release is the high_level::event_loop module, an interface to NVML's event capabilities. Only available on Linux platforms, this module provides you with the boilerplate necessary to quickly and easily watch for events on any number of devices, handling both errors and the events themselves. See the event_loop example in the examples folder at the root of the repository for more.

This release also marks the point at which no nightly features are required for any reason (meaning the removal of the nightly feature flag) and the addition of a couple examples demonstrating use of the crate.

Rust Version Support

This release requires and supports Rust 1.19.0 or higher.

Additions

• Examples:

  • basic_usage
  • event_loop

• enums::event:

  • New file with the following:
    • XidError

• high_level module

  • This module will be the home of any high-level abstractions over the NVML API.
  • high_level::event_loop:
    • New file with the following:
      • Event
      • EventLoop
      • EventLoopState
      • EventLoopProvider, implemented for:
        • NVML

Removals

• The nightly feature flag has been removed as nightly features are no longer required (union has been stabilized).

Changes

• The EventData.event_data field is now an Option<XidError> instead of a u64

  • This was done to more strongly represent the field's presence via the type system (it is None for most events) and also to statically type the Unknown value.

• The UnexpectedVariant error now contains the enum value that could not be mapped to an enum variant defined in the wrapper.

• The project is now formatted via rustfmt as much as possible.

• Markdown headers now have two newlines after them, which is (to my knowledge) how they are supposed to be formatted.

Release Summary

The major highlight of this release is the NvLink struct, an interface to NVML's various NvLink-related functions. This release additionally corrects some issues / oversights in the wrapper and replaces Rust enums with numerical constants for FFI use (see rust-lang/rust#36927).

Rust Version Support

This release requires and supports Rust 1.18.0 or higher.

Additions

• NvLink struct added and fleshed out, wrapping all NvLink-related functions.

  The NvLink struct can be obtained from a Device via Device.link_wrapper_for(). It provides a convenient interface to access the various NvLink-related functions with a common link value.

  • Functions wrapped:
    • nvmlDeviceGetNvLinkState
    • nvmlDeviceGetNvLinkVersion
    • nvmlDeviceGetNvLinkCapability
    • nvmlDeviceGetNvLinkRemotePciInfo
    • nvmlDeviceGetNvLinkErrorCounter
    • nvmlDeviceResetNvLinkErrorCounters
    • nvmlDeviceSetNvLinkUtilizationControl
    • nvmlDeviceGetNvLinkUtilizationControl
    • nvmlDeviceGetNvLinkUtilizationCounter
    • nvmlDeviceFreezeNvLinkUtilizationCounter
    • nvmlDeviceResetNvLinkUtilizationCounter
  • Tests written for all new functionality
    • While I cannot personally run these, they are useful both for static analysis and for those who are able to run the tests.

• Device struct:

  • Device.link_wrapper_for()

• enums::nv_link:

  • New file with the following:
    • Counter

• struct_wrappers::nv_link:

  • New file with the following:
    • UtilizationControl

• structs::nv_link:

  • New file with the following:
    • UtilizationCounter

• No-run tests were added for all methods that modify state. They are useful to statically verify whatever is statically verifiable.

Changes

• InsufficientSize error now contains an Option<usize> instead of a usize.
  • This allows for None to be returned rather than 0 in the default case, and Some(size) to be specified explicitly if possible.
• PciInfo's pci_sub_system_id field is now Option<u32> instead of u32.
  • This change was made in order to accommodate NvLink.remote_pci_info(). NVIDIA says that the sub system ID is not set in that function call, and we represent that with None.
  • As a direct result of this change, PciInfo::try_from() now takes a boolean arg, sub_sys_id_present, which is used to specify whether the pci_sub_system_id field should be Some() or None.

Fixes

• Methods that should have taken &mut self now do so, including:

  • Device.clear_cpu_affinity()
  • Device.reset_applications_clocks()
  • Device.clear_accounting_pids()
  • Device.clear_ecc_error_counts()

• A method that should have taken &self now does:

  • NVML.are_devices_on_same_board()

• Numerical constants are now used in place of Rust enums for safety reasons

  • For more info see rust-lang/rust#36927
  • Many methods now return an UnexpectedVariant error where they did not previously

Dependencies

• bitflags: 0.8.x -> 0.9.x

Release Summary

Initial release wrapping the majority of the NVML API surface.

Rust Version Support

This release requires and supports Rust 1.17.0 or higher.