Repository navigation
gemm_ex: general GEMM domain + fp8 fallback for refused shapes (draft) - #139
Closed
aien-spark wants to merge 18 commits into
Closed
aien-spark wants to merge 18 commits into
aien-spark wants to merge 18 commits into
Conversation
…ed batch, bf16/f16/f32-via-bf16x3/fp8 scales) + fp8 GEMV/grouped GEMM fallback for refused shapes New C behind CUDAPP_MIGRATION_PENDING_GEMM_EX (default ON: only shapes the old code refused take the new path). Host tiers: gemm_ex --host-only PASS 224 checks, mutants 1-5 red. Chip: gemm_ex PASS 901 checks (receipts in the PR body; evidence dir to follow).
…; regression gemm/fp8/fp8gemm/matmul; part0of4 = first run whose one FAIL was a test defect: mutant 5 on a one-k-block shape)
This was referenced Oct 27, 2026
Merged
…neral field, omega_gpu_fp8_general_rows internal (make test-runtime-api PASS)
Collaborator
Author
|
Full make test (before this commit) reached test-runtime-api and failed there: manifest size of OmegaGpuFp8GemmInfo 80 vs 88 (new u32 general appended; EXPERIMENTAL struct, size change) and omega_gpu_fp8_general_rows undeclared. Fixed in this push; make test-runtime-api now PASS. The targets after test-runtime-api in make test: NOT_RUN on this head (full make test needs a rerun). |
…ounding The flat 2^-22 per fp32 step was below the HMMA model's worst case (DnnGpu: K=1808 adversarial input, err 5.343e-5 vs bound 2.843e-5). Now (1+c)^S - 1 with c = (g+1) 2^-25 + 2^-23, times (1 + 2^-20); g = 16 HMMA, 32 QMMA. Host: PASS 224, mutants 1-5 still red. Chip rerun under the new bound: NOT_RUN (GPU window).
…ME limits, successor #144 Manifest: header src/omega_gpu_gemm_ex.h with OmegaGemmExArgs (200 B), OmegaGemmExInfo (24 B), omega_gpu_gemm_ex, omega_gpu_gemm_ex_fp8_block_rows and 21 constants, all EXPERIMENTAL; abi 0.6.0, OMEGA_RUNTIME_API_VERSION_MINOR 6. INTERNAL: omega_gpu_gemm_ex_set_backend, omega_gpu_gemm_ex_test_set_mutant (TEST ONLY), 183 names. API.md: 0.6.0 change list (OmegaGpuFp8GemmInfo 80 -> 88 bytes, OmegaGpuFp8Info.general at 156), section 9.8 with the host-orchestration, bf16x3, fp8-output and migration-pending limits. README cuBLAS row and section 10 limits text. mk/gemm-ex.mk: the .om successor is #144.
…efore any timed run)
…ffers, one-part path without the double sum, C3 writes C in place for the bit-for-bit case ab-1 measured gemm_ex at 2.17-3.25x the C3 call it wraps. Host-only profile with a no-op core: bf16 2048x16384x2048 151.7 -> 4.4 ms of host work per call, bf16 1x8192x2048 24.6 -> 1.2 ms. Same entry, signature and results. Test summary figure is now |got - ref| / bound. Host PASS 224, sim PASS 252; scan negative control (scan never reports) FAILs 3 of 224.
…5), plan gate FAIL in 5 small cells; chip tier PASS 901 at 0d2b14a; overhead recorded in API.md 9.8
…y 0x1p+0, err above the old flat bound, inside the step_c() bound) Host: PASS 225 checks. Chip tier also runs it; chip rerun on this head NOT_RUN at commit time.
Collaborator
Author
|
Status update: CI check failed in job (). Additionally, the PR stated acceptance criteria note in 5 small cells with small-cell overhead remaining unattributed, and full unverified on the PR head. Leaving draft per acceptance and CI gate rules. |
drakestapleton
left a comment
drakestapleton
left a comment
Contributor
There was a problem hiding this comment.
Classification: needs-changes (draft).
buildcheck failed on fc591d6 (actions run 40477773654, job 122100130750).- The PR's own performance gate (r <= 1.10 in every cell) fails in 5 cells: e4m3 64x1024x2048 1.24, e4m3 64x2048x512 1.20, bf16 64x1024x2048 1.17, e4m3 16x12288x2048 1.10, e4m3 1x8192x2048 1.10.
- Full
make testis NOT_RUN on this head (per the 2026-10-27 comment).
To become merge-ready: green build, gate passing or formally re-scoped in API.md 9.8, and a full make test log on the final head. #140 is waiting on this.
drakestapleton
left a comment
drakestapleton
left a comment
Contributor
There was a problem hiding this comment.
Classification: needs-changes.
- A
buildcheck fails on fc591d6 (actions run 40477773654, job 122100130750). - The PR's own A/B gate (r <= 1.10 in every cell) fails in 5 small cells (e4m3 64x1024x2048 1.24, e4m3 64x2048x512 1.20, bf16 64x1024x2048 1.17, e4m3 16x12288x2048 1.10, e4m3 1x8192x2048 1.10). Small-cell overhead remains unattributed.
- Full
make testis NOT_RUN on this head (per the 2026-10-27 comment). - Still draft.
- To become merge-ready: green build, gate passing or formally re-scoped in API.md 9.8, a full
make testlog on the head, then mark ready. #140 waits on this PR.
— Grokbot
Signed-off-by: aien-spark <[email protected]> # Conflicts: # README.md # docs/runtime/API.md # docs/runtime/omega_runtime_api.manifest # src/omega_gpu_fp8_api.c # src/omega_gpu_gemm.c # src/omega_runtime_api.h
…drift baseline records omega's bw_warp_sim_tc.h blob The fp8 general fallback (omega_gpu_fp8_api.c) now calls omega_gpu_gemm_ex_fp8_block_rows, so test_omega_gpu_native_runtime links src/omega_gpu_gemm_ex.c, src/omega_gpu_gemm.c and the qmma codegen. tests/bw_warp_sim_tc.h: main holds omega 9920aab1's blob 59e4ad3 byte for byte; this branch adds the CUDA++-only fp16 HMMA operand path on top, so the baseline row records 59e4ad3 (precedent d2c83a1). A later Omega edit of that file still turns the drift gate red. Signed-off-by: aien-spark <[email protected]>
…M-GENERAL-00c9cf9); README points at it
Collaborator
Author
aien-spark
added a commit
that referenced
this pull request
Oct 31, 2026
gemm_ex: general GEMM domain + fp8 fallback for refused shapes (replaces #139)
Collaborator
Author
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
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.
Draft.
omega_gpu_gemm_ex: one entry with the cblas RowMajor shape contract (any M,N,K, op N/T, ld, alpha/beta, strided batch, f32 via bf16x3, bf16/f16/e4m3/e5m2, tensor or block fp8 scales), plus fp8 GEMV / grouped-GEMM fallback for refused shapes. Migration-pending C behindCUDAPP_MIGRATION_PENDING_GEMM_EX; .om successor #144.Head 7c1a708.
Done
OmegaGpuFp8GemmInfo80 -> 88 bytes.make test-runtime-apiPASS (296 functions, 46 structs, 443 fields, 229 constants; 183 INTERNAL); semver check MINOR.evidence/GEMM-GENERAL-AB-20261027/ab-2/chip/).evidence/GEMM-GENERAL-AB-20261027/:NOT_RUN
make -k testpassed on 34f1b31 (CUDA++ HOST TESTS PASS, 48 targets, EXIT 0). It has not run on the PR head; only the gemm_ex host/sim tiers and the runtime-API check ran there.