Skip to content

Release Notes

Changes in the HyperAccel Profiler (hapti). For changes across the whole stack, see the SDK release notes.


0.0.6 — 2026-11-16

Version: 0.0.6. Requires Python 3.10 or later and Runtime 0.0.6.

Migration (Breaking Changes)

Item Before 0.0.6
Counter selection environment variable HA_PROFILE_COUNTERS HAPTI_COUNTERS
The old name is no longer read

For configuration, see Environment Variables.

New Features

  • Configurable device record ring size — HAPTI_DEVICE_RING_BYTES sets the size of the host ring that receives device records. The default is 1 MiB, and the number can take a K, M, or G suffix (1024-based). If a record loss warning appears, increase this value.

Fixes

  • Fixed MPU segments not being shown on the signal logger track.
  • Fixed the hapti shared libraries not being found in wheel installations.

0.0.5 — 2026-10-14

Version: 0.0.5. Requires Python 3.10 or later and Runtime 0.0.5.

New Features

  • Output control options now take effect — -d, -o, -f, and --log-level are passed to the tool, so you can set the output directory, file name, and format. With --profile-children, the -o value becomes the prefix of the per-rank file names. See the CLI Reference.
  • Memory transfer bandwidth — H2D and D2H transfer slices of 4 KiB or more carry a measured bandwidth, and haprof traces gain a per-device bandwidth counter track.
  • Kineto plugin device timeline — The Kineto plugin collects the device kernel and memory transfer timeline, controlled by HAPTI_KINETO_DEVICE_TRACE and HAPTI_KINETO_DEVICE. See Device Timeline Collection in PyTorch Profiler Integration.
  • Device record loss warnings — When records are lost, a warning is printed at process exit. It distinguishes three cases. - Records overwritten because the reader fell behind the ring - Records the driver dropped before writing them to the ring - Signal logger executions that could not be shown because their kernel record was missing

Changes

  • Per-execution signal logger collection — The tool creates a slot pool in user space and registers it with the runtime, and the runtime reads and hands over the data for each execution.
  • Environment variable isolation — The launcher sets all HAPTI_* environment variables anew on every run, so values left in the shell do not leak into the tool's configuration.
  • Device record ring drain order — At runtime shutdown or device reset, the device record ring is drained and released before the device's resources are reclaimed.

Fixes

  • Fixed a bug where signal logger data for an execution with no kernel record was drawn over another kernel. That execution is now not shown, and a warning is printed.
  • Fixed meaningless bandwidth values shown for D2D transfers and small transfers.

0.0.4 — 2026-07-15

Version: 0.0.4rc4. Requires Python 3.10 – 3.12 and Runtime 0.0.4.

Highlights

  • Two-level device tracing — --runtime-trace collects the device kernel and transfer timeline, and --device-trace adds per-kernel hardware counters and the signal logger on top of it.
  • 6 hardware blocks, 41 counter events — Includes stall breakdown metrics that distinguish memory-bound from compute-bound.
  • PyTorch profiler integration — Through a Kineto plugin.

New Features

Area Added
Device tracing options --runtime-trace, --device-trace (passed to the tool through HAPTI_TOOL_DEVICE_TIMELINE and HAPTI_TOOL_DEVICE_COUNTERS)
Hardware counters Per-kernel-execution counter collection. Events selected with HA_PROFILE_COUNTERS
Multi-rank --profile-children for fork-based multi-rank runs
haprof merge, which merges per-rank traces into a straggler report
User annotation API annotate, range_push / range_pop, range_start / range_stop, mark
Device selection HAPTI_TOOL_DEVICE_ID (the device whose records are collected)

Changes

  • Unified counter representation — haprof and the Kineto plugin show the same counters the same way (per-execution delta, utilization, bandwidth).
  • Timeline linking — Host and device timelines are linked by correlation id instead of being compared side by side.

Known Issues

Issue Impact Workaround
Most CLI options are parsed but not applied -o, -d, -f, --log-level, and per-domain trace flags are ignored Change to the target directory before running
The profiler's own overhead is not recorded The overhead activity kind is empty Compare with a run without the profiler
Memory allocation records have no device information Allocation locations cannot be told apart on multi-device systems —
Limited default counter scope Only the 20 events of core 0 (AXIBRG, DMC, UTIL) are collected; CDMA, CGDMA, MPU, and other cores are not Specify them with HA_PROFILE_COUNTERS
One device per trace In a multi-device process, kernels of only one device are recorded Set HAPTI_TOOL_DEVICE_ID, or split ranks per device and use --profile-children
No device timeline or hapti annotations on the Kineto path torch.profiler traces show only runtime API calls and the layers below Use the haprof launcher
Configuration is fixed at process start A running process cannot be profiled, and profiling cannot be started or stopped mid-run Mark the range to analyze with annotations
Traces are written at process exit A forcibly killed process leaves no result file Exit normally with an interrupt signal instead of SIGKILL
The first execution is not a measurement target It includes kernel binary loading and has no baseline for counter deltas Analyze executions after enough warm-up
CGDMA bandwidth is not measured It is a utilization metric scaled by firmware constants Use it only for relative comparison between executions, not as an absolute rate
Utilization is aggregated by sampling Units that run in short bursts may be under-counted Check it together with the stall breakdown metrics

Version Support

Version Release date Status
0.0.6 2026-11-16 Current
0.0.5 2026-10-14 Superseded
0.0.4 2026-07-15 Superseded

hapti is in beta.

  • The trace output formats (Perfetto Protobuf, Chrome JSON) follow external standards and are stable.
  • The CLI is still being wired up, so option behavior may change between releases. In particular, the per-domain trace flags are currently parsed and ignored; do not assume they will stay ignored.
  • Entry points scheduled for deprecation (APIs, options) are announced in the release in which the deprecation is decided, and they remain for at least two subsequent releases before removal.