perf-analyzer: Add version 2.60.0 - #2350
Merged
Merged
Conversation
Build the Triton Performance Analyzer CLI from the r26.04 head of triton-inference-server/perf_analyzer (the commit the PyPI 2.60.0 binary reports) without CUDA, and package it with the repo's own pyproject.toml as a py3-none-manylinux_2_39_riscv64 wheel.
luhenry
added a commit
that referenced
this pull request
Sep 26, 2026
Contributor
|
The unit test asserts RapidJSON 1.1.0's error code for an unescaped control character (kParseErrorStringEscapeInvalid, 10); RapidJSON after Tencent/rapidjson@85500e8 reports kParseErrorStringInvalidEncoding (12), and 1.1.0 itself does not compile with GCC 14.
luhenry
force-pushed
the
perf-analyzer
branch
from
September 26, 2026 06:18
fce7935 to
3758e73
Compare
concurrency_sequence checks completed requests against a sleep-based
expectation (+/-10%), and custom_load_sequences' serial_sequences subcase
compares per-sequence counters read without the stats lock while the
worker threads are still sending. Both passed on the first run and failed
on the second (522 vs 600; 45,44,44 logged after a failed check) with the
same sources, while the deterministic sequence-to-thread split tests
("Configure threads", "Calculate thread ids") passed both times.
The wheel test found perf_analyzer silent: GCC links the non-PIE CLI at riscv64's default text base 0x10000, and the RPATH auditwheel's patchelf writes for the grafted libssl/libcrypto moved the first PT_LOAD down to 0xc000 in the built wheel, below the runner's vm.mmap_min_addr (65536), so the exec is killed before main and the Python wrapper, which ignores the return code, exits 0 with no output. Raise the base with -Ttext-segment, as build-pygraphviz.yml does. PrintVersion() also writes to std::cerr, so the version check has to capture stderr to see the commit.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
perf-analyzer2.60.0Compiles the Triton Performance Analyzer CLI (C++, with gRPC, protobuf, abseil and curl built from source) and ships it in a
py3-nonewheel. Upstream publishes no riscv64 wheel.Mirrors upstream's documented source build; upstream's wheel job runs in NVIDIA's internal CI.
Differs from upstream
TRITON_ENABLE_GPU=OFF- CUDA does not exist on riscv64; drops CUDA shared memory only-DABSL_USE_UNSCALED_CYCLECLOCK=0- abseil's riscv64rdcycletraps on Linux 6.6+ (unscaledcycleclock: remove RISC-V support abseil/abseil-cpp#1644)Testing
perf_analyzer_unit_tests, thenperf_analyzer --versionfrom the installed wheel on each interpreterinvalid jsonsubcase - it asserts RapidJSON 1.1.0's error code, which later RapidJSON changedconcurrency_sequenceand theserial_sequencessubcase ofcustom_load_sequences- they compare request counts against wall-clock sleeps (the latter reads the counters while the workers are still sending), and failed intermittently on a busy runner: both passed on the first run, then failed on the second with the same sources (522 of an expected 600 +/-10%; per-sequence counts one to two requests apart). The deterministic sequence-to-thread split is covered byConfigure threadsandCalculate thread ids, which pass.License: Wheel statically links gRPC, abseil (Apache-2.0), protobuf, re2 (BSD-3-Clause), c-ares (MIT), curl (curl) and libb64, and bundles OpenSSL (Apache-2.0); upstream ships no licence text for them, so the build adds it.
Patches
0001-forward-third-party-repo-tag.patch- To upstream [not submitted yet]. Without it the release branch builds against third_party main (gRPC 1.81.1 instead of 1.54.3). Not riscv64-specific.