feat(bench): ternary f32 gemv scenario + loaded-weight memory evidence (#1141) - #1167
Merged
Merged
Conversation
…dence engine-ternary-f32-gemv is the first ternary scenario, and deliberately benches the REAL dispatch path rather than a raw kernel loop — what a decode step pays, adapters included. Three providers for the same operands: ffm-lut (the vendored NeoGPU kernel behind the exact FP32×b1.58 key), int8 (per-call I8-absmax requantize + the portable bitnet_gemv reference), f32-reference (the Kotlin oracle on the exact key). BitNet-2B FFN dims (k=2560, n=6912 — crosses the LUT kernel's internal pthread threshold). Measured on an Apple-arm64 dev host through real dispatch: ffm-lut 51.8 GOPS vs int8 0.43 GOPS vs reference 0.30 GOPS — exact math, two orders of magnitude over the portable path (#1136 proof point 1). TernaryWeightMemoryTest is the evidence artifact for proof point 2: the same I2_S GGUF tensor loaded keep-packed vs FP32-widened, asserting the >=15x ratio (measured 15.98x: 0.25 B/weight + scale vs 4 B/weight). jvm-cpu-publish gains the backend-native-cpu dependency so the benchmark runs the bundled libskainet_kernels exactly as a consumer would. Closes #1141 together with the tutorial that landed in #1163. The OpenBenchmarking profile dir and a JMH twin stay optional follow-ups. Refs #1141, #1136 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
7 tasks
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Phase 5 of #1136 — closes #1141 (the docs half landed in #1163).
What
engine-ternary-f32-gemvinjvm-cpu-publish+ScenarioRegistryentry — the first ternary scenario. Benches the realKernelDispatchpath (adapters included), not a raw kernel loop, at BitNet-2B FFN dims (k=2560, n=6912 — crosses the LUT kernel's internal pthread threshold). Providers:ffm-lut|int8|f32-reference.TernaryWeightMemoryTest(io-gguf) — the memory evidence artifact: same I2_S tensor loaded keep-packed vs FP32-widened, asserts ≥15× (measures 15.98×).jvm-cpu-publishgains thebackend-native-cpudependency (bundledlibskainet_kernelsruns exactly as a consumer would).Measured (Apple arm64 dev host, real dispatch, warmups 3 / measured 3)
ffm-lut(vendored NeoGPU kernel, exact)int8(requantize adapter + portablebitnet_gemv)f32-reference(Kotlin oracle)Two orders of magnitude over the portable path, with exact math — #1136 proof points 1 and 2 now have their evidence artifacts. (Pi-4/A72 numbers will differ; upstream measured 6.78 GOPS there — the qemu lane and an on-device run remain the arm-Linux checks.)
Optional follow-ups not included: OpenBenchmarking profile dir, JMH twin.
🤖 Generated with Claude Code