Conversation
First bench run on RTX 4090 sm_89, median TFLOPS per shape: Size f16 TF bf16 TF bf16/f16 256x256x256 0.09 0.09 94.1% 512x512x512 0.68 0.68 99.9% 1024x1024x1024 4.74 5.22 110.1% 2048x2048x2048 26.69 26.75 100.2% 4096x4096x4096 53.83 55.59 103.3% SC-2 gate at 4096³ (worst-of-10): f16 worst-of-10: 2.33 ms / 58.96 TF bf16 worst-of-10: 2.29 ms / 59.96 TF bf16/f16 (TFLOPS): 101.70% (delta +1.70%, well inside ±5% bound) bf16 at 4096³ is 91.8% of cuBLAS sgemm — same regime as f16 (82.3% per existing perf doc). Comparison is project-local reference, not apples- to-apples; the cublasGemmEx-bf16 future reference is tracked at C8.
…ogy-evolution writeup
Breaking (kaio-core IR, pre-v1.0): PtxModule::validate now returns ValidationError::MmaSyncBf16Rejected when MmaSync is constructed with PtxType::BF16 on a_ty or b_ty. Bf16 mma emission must use MmaSyncBf16. This closes a legacy hole where the generic MmaSync path silently emitted bf16 PTX from FragmentA_F16/FragmentB_F16 operands, defeating the D2 sibling-IR design intent. Inverts the prior emit_mma_sync_m16n8k16_bf16_f32 regression test (which documented the footgun) and adds four new validation tests. Updates the 5 existing SmTooLow match sites in kaio-ops to handle the new variant.
…d min_normal class The near_denorm class name was inaccurate: inputs [1e-18, ~1.94e-18] are normal bf16 values 20 orders of magnitude above min-normal (~1.175e-38), so the class did not cover what its name promised. Rename: - patterned_bf16_near_denorm -> patterned_bf16_tiny_product - assert_bf16_close_d5_near_denorm -> assert_bf16_close_d5_relative_only - 4 tc_bf16_*_near_denorm tests -> tc_bf16_*_tiny_product New min_normal magnitude class with asymmetric magnitudes — A in bf16's min-normal exponent band [2e-38, ~3.88e-38] (just above the ~1.175e-38 normal/subnormal boundary), B large positive [1e10, ~1.94e10]. Per-element products land at ~2e-28, normal-f32 range, so the f32 accumulator does not underflow. Isolates the FTZ-on-load failure mode on the bf16 side. 4 new tc_bf16_*_min_normal tests at 32³/64³/256³/512³, symmetric with the existing magnitude grid. Total bf16 correctness suite: 21 -> 25 tests.
…README - README.md: add Tensor-core matmul (bf16) row to the feature summary, API + constraints only (perf table deferred to v0.5.0 aggregate). - kaio-ops/README.md: full refresh against current lib.rs exports. The prior README only listed scalar matmul + scalar attention; the refreshed version lists every shipped op (matmul, matmul_tc family, matmul_tc_bf16, matmul_int8, matmul_int4, attention family, qkv_project_int8/int4, auto-tuner variants) with min-SM + divisibility constraints, and defers perf numbers to docs/performance.md per the Phase 9 master-plan rule. - docs/phases.md: mark Phase 9 in-progress (9.1 shipped 2026-05-14) instead of "Planned".
SC-2 at 4096^3 on RTX 4090 sm_89: median ratio 100.72% (delta +0.72%, bound +/-3%), worst ratio 101.55% (delta +1.55%, bound +/-15%). bf16_async 59.65 TFLOPS / f16_async 59.30 TFLOPS in the per-shape table; ratios cluster tight (99.70%-101.55%) across 10 interleaved alternating-order runs.
…leg died at toolkit install)
…files; coverage badge 94.67%
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.
No description provided.