95 lines
4.7 KiB
Markdown
95 lines
4.7 KiB
Markdown
|
# Phase 2B Aligned Numerical Prototype
|
||
|
|
|
||
|
|
## Decision
|
||
|
|
|
||
|
|
VEA-B passes the isolated fixed-shape Phase 2B gate on NVIDIA GB10. The selected
|
||
|
|
kernel computes `B=1, H=1, Q=128, KV=192, D=128` from prequantized Q/K/V with
|
||
|
|
three increasing K64 epochs. It is retained as research only. Ragged tails,
|
||
|
|
preparation, complete H3 shapes, block integration, trajectories, and production
|
||
|
|
dispatch remain out of scope.
|
||
|
|
|
||
|
|
## Exactness
|
||
|
|
|
||
|
|
The four-warp checkpoint oracle reproduces the installed public SageAttention
|
||
|
|
2.2.0 output. The ten-warp VEA-B kernel matches the oracle byte-for-byte for:
|
||
|
|
|
||
|
|
- final BF16 output, SHA-256
|
||
|
|
`75d4169c9ae0692b24a463a64b538f852acba5a5567404728c8af60c16fb8e1b`;
|
||
|
|
- INT32-to-FP32 QK scores;
|
||
|
|
- per-epoch online-softmax `m` and `d`;
|
||
|
|
- FP8 E4M3 probability words and FP32 rescale values;
|
||
|
|
- FP16 PV partial words and FP32 `RO` after every epoch;
|
||
|
|
- final reciprocal values.
|
||
|
|
|
||
|
|
The selected source also completes 1,000 consecutive fast launches with the
|
||
|
|
same output hash and no deadlock.
|
||
|
|
|
||
|
|
## Selected Kernel
|
||
|
|
|
||
|
|
- 320 threads: two K/V producer warps, four QK/softmax owners, and four
|
||
|
|
PV/output owners.
|
||
|
|
- Two ready, two free, and two K-ready CUDA block-scope mbarriers.
|
||
|
|
- Two K/V and probability slots with strict epoch reuse.
|
||
|
|
- Final reciprocal publication is folded into the final score-ready handoff.
|
||
|
|
- Probability and V fragments are processed as ordered 8-column MMA pairs.
|
||
|
|
- A per-lane volatile shared retirement slot aliases scale metadata only after a
|
||
|
|
warp phase boundary. It constrains ptxas scheduling without changing
|
||
|
|
arithmetic or allocating additional shared memory.
|
||
|
|
|
||
|
|
Ptxas and runtime attributes report 168 registers/thread, zero stack frame,
|
||
|
|
zero spill stores/loads, zero local bytes/thread, 51.2 KiB dynamic shared
|
||
|
|
memory, and one resident CTA/SM. NCU independently reports zero local-memory
|
||
|
|
spilling requests, 168 registers/thread, one register/shared-memory-limited
|
||
|
|
block per SM, and ten theoretical active warps (`20.83%`).
|
||
|
|
|
||
|
|
## Synchronization And Safety
|
||
|
|
|
||
|
|
- Compute Sanitizer memcheck: `0 errors`.
|
||
|
|
- Compute Sanitizer racecheck: `0 hazards`, `0 errors`, `0 warnings`.
|
||
|
|
- QK epoch 1 overlaps PV epoch 0 by 49,317 clocks.
|
||
|
|
- QK epoch 2 overlaps PV epoch 1 by 46,408 clocks.
|
||
|
|
|
||
|
|
The retirement alias initially produced racecheck WAR warnings because lanes
|
||
|
|
could overwrite scale metadata before sibling lanes completed their reads. The
|
||
|
|
selected variant adds an explicit warp phase boundary before the first write.
|
||
|
|
|
||
|
|
## Timing
|
||
|
|
|
||
|
|
The final 2,000-sample single-launch run measures VEA-B p50 `0.016288 ms` and
|
||
|
|
public Sage2 p50 `0.019840 ms`, a prototype-only `1.218x` ratio. Single-launch
|
||
|
|
CV is `5.95%`, reflecting a bimodal microsecond-scale distribution.
|
||
|
|
|
||
|
|
The authoritative stability run uses 500 event samples with 20 identical
|
||
|
|
launches per sample and reports per-launch values. VEA-B p50 is `0.014426 ms`,
|
||
|
|
CV `0.87%`; public Sage2 p50 is `0.014871 ms`, CV `1.23%`; the ratio is `1.031x`.
|
||
|
|
This is fixed-shape screening evidence, not a complete-attention speedup claim.
|
||
|
|
|
||
|
|
## Rejected Variants
|
||
|
|
|
||
|
|
| Variant | Result |
|
||
|
|
| --- | --- |
|
||
|
|
| Initial combined ownership | Exact, 352 local bytes/thread |
|
||
|
|
| Sequential PV fragments | Exact, material spills remained |
|
||
|
|
| Union role state | Exact, 208 local bytes/thread |
|
||
|
|
| Role-separated epoch loops | Exact, 40 local bytes/thread |
|
||
|
|
| `setmaxnreg` redistribution | Compiled at 200 registers, rejected at launch for excessive resources |
|
||
|
|
| Shared persistent `RO` fragment | Final BF16 exact, FP32 `RO` diverged at epoch 1 and stack usage increased |
|
||
|
|
| Inner-pair warp fence | Exact, but increased the fast stack frame |
|
||
|
|
| Selected shared retirement alias | Exact, zero local memory; race-clean after warp phase boundary |
|
||
|
|
|
||
|
|
## Evidence
|
||
|
|
|
||
|
|
- `/home/daniel/StoryStudioAssets/H3-output/h3-blackwell-runtime/research/vortex_exact_attention/phase2b/probes/authoritative-batched-final.json`
|
||
|
|
- `/home/daniel/StoryStudioAssets/H3-output/h3-blackwell-runtime/research/vortex_exact_attention/phase2b/probes/authoritative-final.json`
|
||
|
|
- `/home/daniel/StoryStudioAssets/H3-output/h3-blackwell-runtime/research/vortex_exact_attention/phase2b/probes/memcheck-final.log`
|
||
|
|
- `/home/daniel/StoryStudioAssets/H3-output/h3-blackwell-runtime/research/vortex_exact_attention/phase2b/probes/racecheck-final.log`
|
||
|
|
- `/home/daniel/StoryStudioAssets/H3-output/h3-blackwell-runtime/research/vortex_exact_attention/phase2b/vea-b-final.ncu-rep`
|
||
|
|
- `/home/daniel/StoryStudioAssets/H3-output/h3-blackwell-runtime/research/vortex_exact_attention/phase2b/fixtures/aligned-q128-kv192-seed73021/manifest.json`
|
||
|
|
|
||
|
|
## Advancement Limit
|
||
|
|
|
||
|
|
Phase 2B authorizes no production integration. The next phase must add ragged
|
||
|
|
tails and complete-shape work distribution, then pass canonical attention,
|
||
|
|
blocks 0/24/49, trajectory, resident-service, sanitizer, and complete-block
|
||
|
|
performance gates before any dispatch proposal.
|