-
92b9221533
paper(gqa): drop committed long-ctx 4-cases derived chart PNGs
mukesh
2026-06-17 09:39:26 -07:00
-
694e0cc9b9
paper(gqa): add 6-case per-PE memory + comm summary chart
mukesh
2026-06-16 16:10:49 -07:00
-
6b6e29968a
paper: §2 platform deep-edit + §1 GQA framing + figure regen
ywkang
2026-06-16 16:00:23 -07:00
-
77eece81c4
gqa(milestone-1h-gqa): add umbrella bench + first sweep outputs
mukesh
2026-06-16 13:06:33 -07:00
-
443ede99c7
gqa(prefill-4cases): add long-context prefill 4-cases comparative study
mukesh
2026-06-16 13:06:08 -07:00
-
e45626c036
gqa: reorganize benches into gqa_helpers/ subpackage; drop legacy headline
mukesh
2026-06-16 13:05:41 -07:00
-
359a0eaa44
paper(platform): Table 1 -> table*, shrink Fig 1 row 1, move Fig 1/2 source up
ywkang
2026-06-16 11:58:48 -07:00
-
1cf8dd7868
paper(platform): widen SIP/CUBE subfigures, drop height cap
ywkang
2026-06-16 11:22:01 -07:00
-
b715002b5a
paper: §2 figure layout + Accuracy subsection + Table 1 → 2-col; §6 trim historical 4-panel content
ywkang
2026-06-16 11:16:16 -07:00
-
22f3968b51
paper(gqa): wire the 4-cases long-context decode comparison into §6
ywkang
2026-06-16 10:44:32 -07:00
-
e3f08972da
paper: §2 polish + §5 fresh all-reduce data + kernbench probe table
ywkang
2026-06-16 10:38:47 -07:00
-
7c346dec1b
paper(figures): GQA long-context decode 4-cases comparative figures
mukesh
2026-06-15 17:48:34 -07:00
-
1fbe833992
gqa(decode-4cases): wire latency + engine occupancy into sweep.json; add comparative plot script (5C.F)
mukesh
2026-06-15 17:06:13 -07:00
-
7f437a20bd
paper(platform): edit-pass through §2 KernBench Platform done
ywkang
2026-06-15 17:17:05 -07:00
-
756680f4e6
gqa: fold decode_long into the Case 4 kernel; drop 1D-chain dead code
mukesh
2026-06-15 16:32:37 -07:00
-
ddee28a499
gqa(decode-4cases): rename bench/kernels/panels with long_ctx token
mukesh
2026-06-15 16:09:21 -07:00
-
3c155be8e6
gqa(decode-4cases): give Case 4 its own thin kernel file for naming symmetry
mukesh
2026-06-15 15:58:11 -07:00
-
0ef4fde5d8
gqa(decode-4cases): Case 1 — Cube-SP × PE-TP (inter-CUBE lrab only; PE-TP B=1 waste) (5C.A)
mukesh
2026-06-15 15:30:00 -07:00
-
ae942f6959
gqa(decode-4cases): Case 3 — Cube-Repl × PE-SP (intra-CUBE AR only; 8× memory) (5C.C)
mukesh
2026-06-15 15:18:41 -07:00
-
5672c8f3ef
gqa(decode-4cases): Case 2 — Cube-Repl × PE-TP (no comm; 8× memory) (5C.B)
mukesh
2026-06-15 14:59:58 -07:00
-
65c365f858
bench(milestone-gqa-headline): drop misleading single_user_/multi_user_ panels
mukesh
2026-06-15 14:50:25 -07:00
-
c164645aee
gqa(decode-4cases): Case 4 anchor in dedicated bench (5C.D)
mukesh
2026-06-15 14:31:44 -07:00
-
5b4d9cb597
bench(milestone-gqa-headline): scale single_kv_group prefill panel T_q=S_kv=1K (scratch budget)
mukesh
2026-06-15 14:00:30 -07:00
-
9e1242039b
gqa: single-KV-group LLaMA-3.1-70B prefill milestone (Increments 1-5)
mukesh
2026-06-15 13:42:31 -07:00
-
9270d3435a
paper(platform): WIP — §2 KernBench Platform mid-edit
ywkang
2026-06-11 16:31:42 -07:00
-
d5267ec718
paper(intro): edit-pass through §1 Introduction done
ywkang
2026-06-11 15:13:01 -07:00
-
a965db7056
paper(exec-summary): edit-pass through Executive Summary done
ywkang
2026-06-11 14:44:45 -07:00
-
724a40453d
test: fix 3 stale tests from the prior pull (registry list + memory_store msg)
ywkang
2026-06-10 23:33:34 -07:00
-
1a4ecb7aac
gqa(ddd): DDD-0065 as-built content (prior commit only recorded the rename)
ywkang
2026-06-10 23:27:43 -07:00
-
30a0451481
gqa(ddd): DDD-0065 reconciled to as-built + promoted to docs/ddd/
ywkang
2026-06-10 23:27:13 -07:00
-
f0c5b9c33f
gqa(adr): ADR-0065 Proposed -> Accepted
ywkang
2026-06-10 23:16:52 -07:00
-
e9d62b908c
gqa(adr-0065): N4 — opt2 numeric parity vs opt3 in data mode
ywkang
2026-06-10 23:13:16 -07:00
-
5bbe293bea
gqa(adr-0065): N3 — GEMM-epilogue accumulate dataflow (O = O.corr + P.V)
ywkang
2026-06-10 23:06:01 -07:00
-
3d4a3d43c4
gqa(adr-0065): N2 — softmax_merge recipe MATH ops compute in data mode
ywkang
2026-06-10 22:56:25 -07:00
-
4089e18770
gqa(adr-0065): N1 — composite GEMM computes numerically in data mode
ywkang
2026-06-10 22:46:39 -07:00
-
e8d6c283d8
gqa(adr-0065): D8 — composite returns output handle + output-space DMA
ywkang
2026-06-10 22:29:10 -07:00
-
dd525bfcb7
paper: add /paper skill + 1H HW-SW codesign report (GEMM, All-Reduce, fused GQA)
ywkang
2026-06-10 22:15:14 -07:00
-
4a55ae5c0b
gqa(adr-0065): P6 — close dispatch-ratio loop (ADR-0064 Test #9 direction)
ywkang
2026-06-10 21:04:45 -07:00
-
35453cc4fe
gqa(adr-0065): P5(B) — decode opt2 two-composite kernel + dispatch measurement
ywkang
2026-06-10 21:02:27 -07:00
-
17fb94086e
gqa(adr-0065): P4 — strict-FIFO RW hazard tracker
ywkang
2026-06-10 20:36:31 -07:00
-
55f025c4b1
gqa(adr-0065): P3 — flat-ops PE_SCHEDULER plan (position-scan + prologue stages)
ywkang
2026-06-10 20:24:33 -07:00
-
184f654295
gqa(adr-0065): P2 — softmax_merge recipe + TLContext prologue lowering
ywkang
2026-06-10 19:48:29 -07:00
-
47e2c78c66
gqa(adr-0064/0065): flat-ops CompositeCmd (P1) + structural dispatch cost (ADR-0064 Rev2); promote ADR-0064
ywkang
2026-06-10 19:18:04 -07:00
-
79ddb12b42
gqa: explain why Tile-0 / bootstrap can't be folded into the tile loop
mukesh
2026-06-10 16:35:43 -07:00
-
3647e2d8f5
gqa(adr): ADR-0064/0065 second-pass review — 1024B rationale, FIFO as ordering source, robust formula-based tests, D5/D6 auto-bind alignment
ywkang
2026-06-10 16:34:34 -07:00
-
a8d04750e6
gqa(adr): ADR-0065 review fixes — KERNEL phase/repetition split, auto-bind conflict + MATH-TCM invariants, FIFO conservatism note, logical_bytes per-op summation
ywkang
2026-06-10 16:29:39 -07:00
-
95cecccd8e
gqa(adr): ADR-0064/0065 review fixes — composite size cap, type-aware logical_bytes, R recalibration
ywkang
2026-06-10 16:20:58 -07:00
-
7fad0371c5
gqa: tile-granular Ring KV (P3c) + rename to gqa_attention_* + ADR-0060/62/63/64 → Accepted
mukesh
2026-06-10 16:17:32 -07:00
-
a8c50238c6
gqa(adr): ADR-0064 Rev2 (structural dispatch cost) + ADR-0065/DDD-0065 (flat-ops composite + softmax_merge recipe)
ywkang
2026-06-10 16:11:20 -07:00
-
91cdfebb67
gqa: S_kv tile sweep for decode_long/decode_short/prefill_short (ADR-0063 §A.2)
mukesh
2026-06-10 15:33:29 -07:00
-
5a76ed4f6a
gqa: rename long-context kernels to *_long for symmetry with *_short
mukesh
2026-06-10 10:05:54 -07:00
-
d282144339
gqa: ADR-0060/0062/0063/0064 unified GQA kernels + CPU cost model
mukesh
2026-06-09 18:15:59 -07:00
-
b3730a33eb
policy(adr): ADR-0026 Revision 6 — document DPPolicy.cube_start
mukesh
2026-06-08 14:18:27 -07:00
-
ab324c155c
gqa(adr): review fixes — unify KV placement to contiguous, ring/reduce cost model, opt3 wording, rank def
ywkang
2026-06-05 00:52:38 -07:00
-
9103661098
gqa(adr): mark DDD-0060 as synced in ADR-0060 SB (EN+KO)
ywkang
2026-06-04 23:09:14 -07:00
-
c699e2ddfc
gqa(ddd): rewrite DDD-0060 to match current ADR (two-kernel composite hybrid)
ywkang
2026-06-04 23:08:02 -07:00
-
973a452b77
gqa(adr): sync ADR-0060 KO mirror to current EN (two-kernel design)
ywkang
2026-06-04 23:02:49 -07:00
-
8de13efb9d
gqa(adr): opt2 ex_composite decode needs tl.wait() before the reduction
ywkang
2026-06-04 22:49:59 -07:00
-
9c554e3f64
gqa(adr): fix prefill TL;DR pseudocode — K/V are TCM handles for the ring
ywkang
2026-06-04 22:40:12 -07:00
-
c6bc286740
gqa(adr): TL;DR — full code for 3 decode variants + prefill (per request)
ywkang
2026-06-04 22:05:15 -07:00
-
444810c6e3
gqa(adr): TL;DR — Q-replicated/M-fold wording + pointer to S5.6 decode variants
ywkang
2026-06-04 22:01:25 -07:00
-
943626d758
gqa(adr): ADR-0060 — 1 Q head/CUBE term, decode 3 variants (S5.6), shared contiguous KV layout
ywkang
2026-06-04 21:59:03 -07:00
-
8176cdf287
gqa(adr): split ADR-0060 SP into two kernels — decode=reduce, prefill=ring
ywkang
2026-06-04 21:32:08 -07:00
-
a61fed3ce0
gqa(adr): ADR-0060 hierarchical CUBE-Group SP — 2-level reduce-to-root
ywkang
2026-06-04 17:35:43 -07:00
-
f009ba7b1f
topology: add llama70b_4sip.yaml for the 8-KV-group diag harness
mukesh
2026-06-04 12:42:40 -07:00
-
39fc2e953f
attention: add 8-KV-group diag harness (CCL pattern + cube_start sub-meshes)
mukesh
2026-06-04 12:40:04 -07:00
-
4859149392
attention: add 2D row-then-col AllReduce-mlo decode kernel (C2)
mukesh
2026-06-04 12:39:48 -07:00
-
e2fe33180d
policy: add DPPolicy.cube_start for disjoint cube sub-meshes within a SIP
mukesh
2026-06-04 12:39:30 -07:00
-
80510b89a9
ccl: add missing configure_sfr_intracube_pe_ring (fix milestone-gqa import)
mukesh
2026-06-04 12:36:27 -07:00
-
d0f904c57e
attention: docstring cleanup — clarify single/multi user per GQA study
mukesh
2026-06-02 14:38:30 -07:00
-
7a9d4ec47b
gqa(adr): pivot ADR-0060 to composite hybrid + lazy tl.load; add ADR-0064 cost model
ywkang
2026-06-04 10:19:37 -07:00
-
6d24b9306f
gqa(adr/ddd): iteration-2 corrections — GQA reuse via M-fold, transpose caveat
ywkang
2026-06-03 17:51:03 -07:00
-
fb600b3cdd
gqa(ddd): add Detailed Design Document for GQA fused attention (ADR-0060)
ywkang
2026-06-03 17:44:39 -07:00
-
1742600641
gqa(adr): rewrite GQA fused-attention ADR as ADR-0060, aligned to kernbench
ywkang
2026-06-03 17:42:00 -07:00
-
f4f55b2c1c
gqa(adr): add supporting feature ADRs 0061-0063 for GQA fused attention
ywkang
2026-06-03 17:38:55 -07:00
-
189794510a
fused attention kernel design - in-progress
Yangwook Kang
2026-06-03 14:47:50 -07:00
-
b3ca532023
attention: milestone-gqa-llama70b figures + MILESTONE_FAST (sub-cycle 4c, 5/6)
mukesh
2026-06-01 22:23:28 -07:00
-
e748a62264
attention: land milestone-gqa-llama70b 4-panel sweep bench (ADR-0057 v1)
mukesh
2026-06-01 21:57:12 -07:00
-
222815d374
attention: add rank_axis kwarg to mesh kernels for multi_user cube ring
mukesh
2026-06-01 19:53:18 -07:00
-
d9e767d048
runtime_api: ctx.launch honors DPPolicy.num_cubes + adds _auto_dim_remap opt-out
mukesh
2026-06-01 19:33:40 -07:00
-
313dee503c
sim_engine: fix IPCQ slot-wrap snapshot race in Phase 2 replay
mukesh
2026-06-01 19:14:09 -07:00
-
b1d6fafd3a
eval: commit milestone bench output (track generated figures + results)
mukesh
2026-05-22 15:37:27 -07:00
-
cc1bbd0ab7
eval: fold GEMM/allreduce harnesses into self-contained milestone benches
mukesh
2026-05-22 15:19:32 -07:00
-
e33e76f2d1
adr: add INDEX.md (auto-generated by tools/generate_adr_index.py)
ywkang
2026-05-22 11:15:37 -07:00
-
bd49c93703
adr: add ADR-0050-0053 — close /report's second-pass G4 candidates
ywkang
2026-05-22 10:52:42 -07:00
-
9a02955770
adr: add ADR-0046-0049 — close G4 coverage gaps from /report
ywkang
2026-05-22 10:25:04 -07:00
-
5f8dd688f5
adr: add ADR-0045 (bench module contract — registration, dispatch, authoring)
ywkang
2026-05-21 16:29:45 -07:00
-
fd56b6cacd
adr: add ADR-0043/0044 (eval harnesses); reconcile ADR-0024/0032 for SIP w/h
sccl-distributed-allreduce
mukesh
2026-05-21 10:26:25 -07:00
-
0e346b939d
gemm: test-generated GEMM plots under tests/gemm/ + docs/diagrams/gemm_plots/
mukesh
2026-05-21 09:58:08 -07:00
-
b610cb0d9a
sccl: drive allreduce tests via torch.distributed; reorganize into tests/sccl/
mukesh
2026-05-20 22:24:43 -07:00
-
ff7d727ddd
CCL allreduce: rename to lrab_hierarchical_allreduce + descriptive plots
mukesh
2026-05-20 20:50:48 -07:00
-
e77e4a1703
types: narrow BenchResult.engine to GraphEngine, cast topology in engine_factory
ywkang
2026-05-20 14:54:18 -07:00
-
1f36baa898
ADR: add 0038-0042 (pcie_ep, pe_mmu, pe_tcm, sram, tiling)
ywkang
2026-05-20 14:43:03 -07:00
-
049e3d8bb3
benches: package as kernbench.benches, add @bench registry + list subcommand
ywkang
2026-05-20 14:42:10 -07:00
-
168b0c89f0
ADR: translate adr-ko/ to Korean, fix ADR-0013 slug, refine Status check
ywkang
2026-05-20 08:17:56 -07:00
-
a796c1d2f7
ADR: bilingual structure — EN canonical in adr/, KO mirror in adr-ko/
ywkang
2026-05-20 01:38:44 -07:00
-
687c98086d
ADR housekeeping: category prefixes, lifecycle folders, retroactive 0034-0037
ywkang
2026-05-20 01:15:55 -07:00
-
22fd0d2b9d
ADR: introduce docs/history/, merge 0011+0018, prune migration cruft
ywkang
2026-05-19 11:42:45 -07:00
-
ecc57d050d
CLAUDE.md: restructure into Part 1 (general) / Part 2 (project-specific)
ywkang
2026-05-18 12:08:10 -07:00
-
a7fe785e5f
tl.composite: fused epilogue ops with per-op scope
mukesh
2026-05-15 10:16:47 -07:00