[WIP] [part-2] [dep: #1430] Convert inference-perf multimodal output into v0.2.1 reports#1431
Draft
Bslabe123 wants to merge 5 commits into
Draft
[WIP] [part-2] [dep: #1430] Convert inference-perf multimodal output into v0.2.1 reports#1431Bslabe123 wants to merge 5 commits into
Bslabe123 wants to merge 5 commits into
Conversation
Bslabe123
requested review from
Vezio,
achandrasekar,
kalantar,
maugustosilva,
mengmeiye and
namasl
as code owners
Contributor
|
Unsigned commits detected! Please sign your commits. For instructions on how to set up GPG/SSH signing and verify your commits, please see GitHub Documentation. |
Bslabe123
changed the title
Bslabe123
force-pushed
the
multimodal-br-v0_2_1-converter
branch
from
e5639b1 to
1d74f2c
Compare
Additive minor revision of the v0.2 benchmark report schema. Introduces optional per-modality (image/video/audio) payload statistics on the request aggregates and new unit categories (pixels, ratio, bytes, media throughput). - base.py: add Units.PIXELS/RATIO/BYTES and images/videos/audios per-second rate units; add UNITS_RATIO and UNITS_MEDIA_THROUGHPUT lists. - schema_v0_2_1.py: MediaPayloadStats hierarchy (image/video/audio), extended AggregateRequests (request_size, multimodal) and AggregateThroughput (image/video/audio_rate), and BenchmarkReportV021. - core.py / __init__.py: register and dispatch schema version 0.2.1. - regenerated br_v0_2_1 example and JSON schema artifacts. All new fields are Optional, so any v0.2 report remains valid under v0.2.1. Refs: #989, kubernetes-sigs/inference-perf#450 Signed-off-by: Brendan Slabe <slabe@google.com>
Enforces the hard requirement that every valid v0.2 benchmark report is also valid under v0.2.1, via three angles: the committed v0.2 example validates under both models, v0.2 data dumps identically under either version, and every field v0.2.1 adds is Optional (no v0.2 field becomes required). Runs under the existing `pytest tests/` PR CI job. Signed-off-by: Brendan Slabe <slabe@google.com>
- tests/test_benchmark_report_v0_2_1_multimodal.py: a fully populated multi-modal report validates and round-trips; each per-modality unit is accepted with its correct category; mismatched units are rejected, including proof that the inherited v0.2 request_rate guardrail is not loosened. - README: document v0.2.1 as an additive superset of v0.2, listing the new request_size, per-modality multimodal block, and media-rate fields plus the new unit categories. Signed-off-by: Brendan Slabe <slabe@google.com>
Grounding the v0.2.1 schema against inference-perf's emitted lifecycle report (post #450/#477) shows video temporal extent is modeled as sampled frames, not wall-clock seconds: the native Video record carries pixels/bytes/aspect_ratio/frames and no duration. The seconds field on VideoPayloadStats therefore had no producer, so remove it; audio.seconds stays since the native Audio record emits it. Updates the schema, generated JSON schema, example, README, and the video unit tests accordingly. Signed-off-by: Brendan Slabe <slabe@google.com>
Adds native_to_br0_2_1.import_inference_perf, which reuses the v0.2 inference-perf converter for all shared content (scenario, latency, token throughput) and folds in the additive v0.2.1 fields: per-request request_size, the per-modality multimodal block (image/video/audio), and per-modality delivery rates. Native median/p0.1/p99.9 are renamed to the schema's p50/p0p1/p99p9, *_per_sec rates to *_rate, and request_size_bytes to request_size. The CLI gains a 0.2.1 target via -b/--br-version. The converter test runs against a genuine inference-perf lifecycle report (tests/fixtures/inference_perf_lifecycle.yaml), captured from inference-perf's own summarize_requests over the upstream shape test's multimodal mix, and pins the field/units mapping against the raw native values rather than hardcoded numbers. Signed-off-by: Brendan Slabe <slabe@google.com>
Bslabe123
force-pushed
the
multimodal-br-v0_2_1-converter
branch
from
1d74f2c to
f4627a0
Compare
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.
Summary
Populates a v0.2.1 report from a real inference-perf run, adds
native_to_br0_2_1.import_inference_perf, which reuses the v0.2 converter for everything shared (scenario, latency, token throughput) and folds in only the additive multimodal fields. The CLI gains a0.2.1target via-b/--br-version.Depends on #1430