- merge the quantize launchers into one Tiled template; extract shared cvt_fp8/publish_amax helpers and replace the dtype x format ladder with two-level template dispatch
- fold the gemm interior/generic operand loads into one kInterior template and the fast/generic async loads into load_async<kFast>; Policy carries the smem budget
- compress kernel comments to the load-bearing invariants, dropping measured-number essays; Policy signature and kernel code unchanged
Benchmark: NVIDIA L20, 1.2B model train step fwd+bwd+CE
- M=8192: fp8 532.2 -> 530.4 ms (1.26x, noise); tests/extension 65 passed, quantize layouts byte-exact, NT routing diff 0.0
- pass the crosswise operand count from gemm into plan_gemm so congruous and crosswise problems stop sharing one threshold ladder
- congruous grids past one big-cta wave pick big vs narrow by the wave cost ceil(tiles/sm) * T_tile with T_narrow ~= 0.53 * T_big, reproducing every measured crossover
- crosswise problems run the small 64x64 s3 cta up to ~1.5 waves of 128x128 tiles; the narrow cta never wins there (loses to small below the band, to big above it)
- keep the sub-wave congruous ladder and the padding rules unchanged
Benchmark: NVIDIA L20 (92 SMs, sm_89), CUDA 12.8, fp8 e4m3 -> bf16, interleaved A/B against the previous ladder
- NT Mx4096x4096: M=384 114.5 -> 134.5 TF (+17.4%), M=512 152.5 -> 171.7 (+12.6%), M=768 153.5 -> 165.4 (+7.7%); all other NT shapes unchanged
- NN/TN M=64..512 +3.2..+17.4%, 1024^3 +13.7%/+12.8%; M>=640 and 2048^3+ unchanged
- TT 1024^3 +25.5%, M=256 +25.4%; TT M=512 -4.6% at the 1.5-wave boundary that favors TN/NN
- remove the LeanRing knob: every production Policy already ran full kStages+1 rings (the lean variant measured slower, 1280³ +5..9%), so the barrier-4 branch, the kInterleave condition and the ring-depth ternaries collapse to a single kRingDepth in Fp8GemmSmem, now the single source the mainloop reads
- remove the always-true grouped field from Fp8GemmPlan: every layout canonicalize_gemm produces is grouped-raster, so plan_gemm drops the parameter; the plain-raster experiment knob stays available via launch_plan's GroupRaster template parameter
- extract load_b_frags for the duplicated B-fragment fill (initial + double-buffer next-seg sites)
- device_sm_count: fold the out-of-range branch into one cached query path
- Fp8GemmPolicy goes 12 -> 11 template parameters; fp8_test's CasePolicy follows
Benchmark: NVIDIA L20 (sm_89, 92 SMs), kernel bench and the 1204M bf16 model e2e training step both unchanged (fp8 step 503.7 -> 503.9 ms, 1.23x vs bf16; per-shape TFLOPS within +-2%); fp8_test All PASS, tests/extension/test_fp8_mma.py 36 passed.
Kernel restructured CUTLASS-style: Fp8GemmPolicy as the kernel's single template parameter (traits + operand layouts + scheduling knobs), the body split into Fp8GemmTileScheduler / Fp8CollectiveMainloop / Fp8CollectiveEpilogue collectives, and the entry split into canonicalize_gemm -> plan_gemm -> launch_plan behind fp8::gemm.
- NN (dual-N-contiguous) problems run as their transpose: the swap in canonicalize_gemm plus an out-transposed epilogue removes one kernel instantiation per (format, tile config)
- new 128x64 narrow CTA (8 warps of 32x32) serves the sub-wave band once its grid passes ~3/8 of a wave: +7..77% there (128x4096x4096 116->131T, 1024^3 131->174T, 4096x384x4096 147->242T, 8192x128x4096 131->233T); decode, the padding band and multi-wave shapes unchanged
- launch_with_smem no longer swallows cudaFuncSetAttribute failures
- fp8_test: GPU-side fp32 reference (O(m*n) compare instead of O(m*n*k) host loop), production-dispatch cases for the NN swap and the plan selection; dead transpose_layout trait removed
Device: NVIDIA RTX 6000D (sm_120, 156 SMs), CUDA 13.1, torch 2.11.0+cu130. Kernel-only bench vs CUTLASS 4.8.0 sm120 dense fp8: ahead up to 1.68x below one wave (512^3 44 vs 26T, 64x4096x4096 95 vs 62T), within ~7% in the DRAM-streaming regime (8192^3 248 vs 266T).
- a 128x128 CTA that is not exactly tiled (m or n not a multiple of 128) runs its edge tiles on the predicated generic path, and with a single in-flight wave the runtime is the slowest CTA — the edge tiles drag the whole shape down, so prefer_small_cta now also takes n and returns true when 64 divides both dims but 128 does not (measured sweep: 1088^3 big-CTA 76T vs 64x64 93T)
- the 64x64 grid tiles exactly on such shapes and overlaps waves (CTA count jumps 64->100->144 in the 1024-1536 band; the non-divisible points like 1088/1216 sit in deep sawtooth valleys that the divisibility rule lifts to the flat ~100-111T plateau)
- double-non-divisible shapes (e.g. 1000^3) stay on the big CTA: measured 64x64 36.7T vs 128x128 40.0T — both grids carry edge tiles there and the big CTA's efficiency wins
- m <= 64 or n <= 64 also takes the small CTA (a 128-wide CTA wastes more than half its columns on narrow N)
Benchmark: L20, e2e CUDA graph (GPU 7, same GPU as all comparisons): 960^3 60.2->100.1T, 1088^3 77.6->101.6T, 1216^3 69.7->103.5T, 1344^3 84.9->111.1T; 128-divisible shapes unchanged within noise (1024^3 118.4, 1152^3 150.6, 1280^3 106.8, 1536^3 153.7, 2048^3 192.2). 596 pytests pass.
- our 128x128 fast loop padded 26 NOPs between the 32 QMMAs while all four LDGSTS sat bunched at the loop tail: ptxas had no independent instructions to fill the tensor-pipe issue gaps, the exact structure the decompiled cuBLAS loop (166 i, NOP=0) and CUTLASS MmaMultistage avoid by issuing cp.async in small groups inside the MMA phase (copy_tiles_and_advance per warp-tile batch)
- the steady-state prefetch is now a loop-carried register pair per congruous operand (PrefetchCarry: swizzled stage offset + global source, constructed once from the same (r, c0) mapping as the interior loader), whose chunks ride after the first and last k_seg MMA batches — SASS: 282 -> 110 instructions, 0 branches, 0 UIMAD.WIDE magic-divisions, LDGSTS interleaved inside the QMMA range, 26 -> 19 NOPs, still 128 regs (2 CTAs/SM)
- the wait-count dispatch ladder (16 instructions of ISETP/SEL picking DEPBAR immediates) and the per-k-tile (tile % ring) * stage_bytes recomputation (UIMAD.WIDE by 0x55555555) are gone: the prologue commits unconditionally so the wait_group<kStages-1> immediate is valid for every iteration, and both read and write stage addresses advance as carried pointers with an equality wrap
- cp_async.cuh splits the emitter from its policies: one raw PTX site (cp_async_16_raw) plus wrappers for unconditional/predicated and pointer/offset destinations, and the now-unused wait_group_dispatch ladder is deleted; the dispatch flip: with the stall gone the big CTA wins the whole former dip band (1280^3 kernel-level 98.2->104.4T), so prefer_small_cta keeps only the sub-5/8-wave band and the single-wave s3 variant is retired
Benchmark: L20 (sm_89), kernel-level sweep 128s2ff: 1024^3 104.2->114.0T (cuBLAS 151.2), 1152^3 130.4->145.9T (157.2), 1280^3 98.2->104.7T (163.8), 1536^3 141.4->150.4T (182.3), 2048^3 178.1->189.6T (202.8), 4096^3 197.6->208.8T (223.7), 8192^3 207.6->219.1T (227.5). CUDA-graph e2e: 1024^3 108.2->118.4T, 1152^3 137.7->150.7T, 1536^3 143.4->153.7T, 2048^3 179.4->191.9T, 8192^3 199.0->208.9T, 1280^3 108.3->106.6T (old dip-band rule re-measured 106.1T — within noise). Four-layout C++ suite and 596 pytests pass.
- fold the two adjacent nt B fragments of each pair into one ldmatrix.x4: lanes 0-7/8-15 address rows n0..n7 chunks c/c+1, lanes 16-23/24-31 the same chunks of rows n8..n15, so {r0,r1} feed the even nt mma and {r2,r3} the odd nt — 4 x4 B loads per k-tile instead of 8 x2 (12 LDSM total, matching the decompiled cuBLAS and CUTLASS loop shapes)
- the +8-row half never reaches the XOR-swizzle source bits for kK <= 64 (row[2:1]), so the pairing rides the existing per-lane address closure with one extra term (rh16 * 8 * kK); kK=128 swizzles on row[2:0] and keeps the x2 path
- decompilation trail: nsys shows cuBLAS never split-Ks on the gap shapes (grid.z=1, no atomics; it fills waves with 64x128/64x64 tiles instead), and a CUTLASS 3.8 reference at our exact 128x128 s3 geometry reaches 196.5T at 2048^3 vs our 177.9T with NOP=0 and 12 LDSM — proving the loop shape is reachable from CUDA C++ (see perf/fp8_next_ideas.md F/C)
Benchmark: L20 (sm_89), kernel-level sweep: 2048^3 176.7->178.1T, 4096^3 196.3->197.2T, 8192^3 ->208.7T, 4096x512x4096 137.3->139.1T, 896x1152x4096 121->123.4T. CUDA-graph e2e: 512^3 54.2->55.3T, 1536^3 142.3->143.4T, 2048^3 177.9->179.4T, 8192^3 198.2->199.0T. Four-layout C++ suite, 596 pytests pass.
- dispatch 128x128 CTAs at kStages=3 when the grid fits one wave (tiles <= SM count): with no second wave to overlap the drain, latency hiding comes only from the pipeline depth
- multi-wave grids keep kStages=2 — the shorter prologue wins once retiring CTAs overlap (measured 4096^3: s2 196T vs s3 175T)
- geometry sweep across the mid band (128x64, 64x128, kK=128, 256-row CTAs) measured and rejected: all lose to the 128x128 fast loop; the remaining mid-band gap concentrates in the 1.0-1.4 wave dip (1280^3-class shapes, ~105T vs cuBLAS 186T), which is a scheduling problem (split-K), not a geometry one
Benchmark: L20 (sm_89), CUDA-graph e2e: 1024^3 106.3->107.9T, 1152^3 133.3->137.1T, others unchanged (512^3 54.2T, 2048^3 177.9T, 4096^3 192.9T, 8192^3 198.2T). C++ four-layout suite and 114 targeted pytests pass.
- re-enable kFastLoop on the 128x128 CTA for congruous layouts: the base-pair fragment addressing freed the registers the old offset tables spilled, and the predication-free interior loop now wins across the band (fast body 142 SASS instr with zero predicated fallback vs 719/136 generic; 128 regs, no spill)
- move the big/small CTA dispatch boundary from 3/4 to 5/8 wave: with the fast big-CTA loop the crossover sits between 49 and 63 tiles (63-tile rect +8%, 1024^3 now takes the big CTA)
- fuse the linear bias into the GEMM epilogue: FP8Params.bias_ptr adds in fp32 before the single bf16 rounding, replacing the separate out + bias elementwise pass; guarded loads keep N tails exact and batch broadcast falls out of the row-major layout
- resolve Python None bias in the pybind layer (py::object + cast) so ops/fp8.py and fp8.py pass the argument through untouched; drop the _empty_bias sentinel machinery
- add fused-bias tests covering odd N tails, no-bias parity and batched broadcast
Benchmark: L20 (sm_89), CUDA-graph e2e. Big-CTA fast loop + dispatch: 1024^3 102.6->106.3T, 1152^3 128.5->133.3T, 2048^3 173.8->178.2T, 3072^3 180.2->185.3T, 8192^3 196.2->197.7T. Bias fusion (with-bias GEMM vs unfused out + bias): 1024^3 90.5->106.1T (+17%), 2048^3 162.2->178.3T (+10%), 4096^3 178.2->191.1T (+7%). Fused bias differs from the split path by <=1 bf16 ulp and is closer to the fp64 reference. 596 tests pass.
- replace the a_off/b_off per-lane offset tables with two loop-invariant lane bases; ldmatrix fragments now address [base + immediate] with the k_seg step as a single XOR (0x20), mirroring cuBLAS SASS mechanism 1
- drop the 16-register offset table that pushed the kernel past the 128-reg budget and forced per-k-tile address rematerialization; hot-loop integer instructions 444 -> 359 (128x128), immediate-addressed LDSM 8 -> 13/16
- small CTA switches from the lean ring (two __syncthreads per k-tile) to the full ring (one barrier, cuBLAS's structure): s2/24KB below one 3-CTA wave, s3/32KB above
- remove the kAheadFrag cross-k-tile fragment pipeline after measurement (neutral to -8%); mechanism recorded in perf/fp8_gemm_optimization.md
Benchmark: NVIDIA L20 (92 SM, sm_89), torch 2.11.0+cu128, kernel-level event timing on one idle GPU, extension rebuilt from source before each run.
- 2048^3 171.0 -> 172.7 TF (+1.0%), 4096^3 191.0 -> 195.9 (+2.6%), 8192^3 ~190 -> 202.5 (+4.7%)
- 512^3 48.3 -> 50.2 (+3.9%), 1024^3 99.1 -> 101.4 (+2.3%), 1280^3 102.2 -> 106.9 (+4.6%)
- e2e mm_fp8 CUDA-graph: 1280^3 107.6 T, 2048^3 173.8 T, 8192^3 196.2 T
- numerics unchanged: accumulation order identical, per-shape precision equal to the committed baseline (594 pytest, 4-layout C++ suite, short-K and ragged repros all pass)
- replace the sm_count*14/3 small-shape threshold (calibrated on a 24-SM part, so 429 tiles on the 92-SM L20) with a wave-quantization-aware rule: 128x128 CTA for tiles in [3/4, 1] wave or >= 1.4 waves, 64x64 below and inside the just-past-one-wave dip where the finer grid fills the tail
- add a predication-free fast interior loop (kFastLoop) for the 64x64 small CTA: XOR-folded chunk addresses cut ~9 to ~3 instructions per loaded chunk
- delete the dead staged-B pipeline family and launcher dead branches (gemm.cuh 982 -> 801 lines), unused since 5745c2f
Benchmark: L20 (92 SM, sm_89), e2e CUDA-graph TF/s vs prior dispatch: 1152^3 103.1 -> 131.2 (+27%), 1536^3 112.1 -> 138.5 (+24%), 2048^3 123.4 -> 173.0 (+40%); 512/768/1024/1280/3072/4096 cubes unchanged within 1%; 594 pytest + C four-layout tests pass.
- mm_fp8 accepts 3D operands through the same signature: grid.z slices by batch strides, size-1 batches broadcast (stride 0), inner .t() views fold into the layout tag at zero copy
- fix _LinearFp8 backward crash on 3D [B,L,d] training inputs (flatten before mm_fp8, reduce grad_b over leading dims)
- expose kRasterGroup/kStreamOut as template knobs; drop the 64x128 mid CTA and staged crosswise-B path from dispatch (direct wins everywhere re-measured, including DRAM-streamed B)
- dispatch thresholds grounded in fresh sweeps: m<=64 -> 64x64 CTA (+27% at 64x8192x2048), small-CTA crossover at SM*14/3 total tiles (+13% at 96 tiles), threshold counts batch x per-matrix tiles (+31% at 64x512^3 bmm, +25% at 8x1024x2048)
- remove scripts/tools/bench_fp8_gemm.py (superseded by csrc/tests/fp8_sweep.cu for kernel-level tuning)
Benchmark: NVIDIA L20, E4M3, NT pre-quantized, median of 100-200 iters
- 64x8192x2048: 29.1 -> 22.8 us (94 TF/s)
- 1024x1536x2048: 67.4 -> 59.6 us (108 TF/s)
- bmm 64x512^3: 139.8 -> 106.7 us; bmm 8x1024x2048: 186 TF/s
- regression-free: 4096^3 192 TF/s, 8192^3 200 TF/s, 512^3 unchanged
- parameterize warp tile (WarpM/WarpN) in Fp8GemmTraits; MMA loops, fragment arrays and epilogue scale with kMt/kNt instead of the fixed 64x32/4x4, enabling cuBLAS-style 64x64 CTAs of 32x32 warps
- dispatch by output tiling (grid-searched via csrc/tests/fp8_sweep.cu): fewer than 48 output tiles take 64x64/32x32 with a lean ring (4 CTAs/SM fill the wave-quantization gap: 512^3 goes 16 -> 64 CTAs); larger shapes keep 128x128 with the kStages+1 ring
- kStages+1 canonic ring rotation drops the post-compute barrier on the congruous path (one __syncthreads per k-tile); LeanRing keeps the kStages ring for the small CTA; direct-crosswise operands always rotate kStages+1 (their prefetch issues right after barrier 1 and would race a lean ring - caught by the pure C layout suite)
- stage the bf16 epilogue through the reclaimed operand smem: swizzled scatter + barrier + coalesced 16B copy-out replaces 8 disjoint 16B per-warp segments (~50% write efficiency before)
- hoist per-lane ldmatrix swizzle offsets out of the mainloop (stage-relative table + ring-base add) so the innermost loop stops recomputing IMAD/LOP3 address chains
- bypass the torch.library dispatch for real CUDA tensors in quantize/mm_fp8 wrappers (~5us/call, ~40% of a 512-wide call's wall time); fake/subclass tensors keep the custom_op route
vs the previous kernel + python path, wall clock on NT squares: 512^3 52 -> 13us (4.0x, 5.2 -> 20.5 TF, now 1.36x cuBLAS _scaled_mm), 1024^3 1.05x, 2048^3 1.02x (46.9 -> 48.2 TF kernel-only); correctness: 4 layouts x 6 shapes pure C suite PASS, 588 pytest PASS
- split quantize into quantize.cuh, templated on input type (bf16/fp16/fp32)
- rename pybind entry quantize_bf16 to quantize; validate the fmt enum
- fix fp8x2 packing: one 32-bit word packs two pairs (halves were dropped)
- drop the dead OutFp8 template param; GEMM output is always bf16
- fp8_state.reset() restores recipe/format defaults too (test state leak)
- rewrite tests for the two-primitive API with fp32-domain amax references
- Rotate kStages+1 canonical buffers for direct-crosswise operands so their load issues right after barrier 1 and its LDG+PRMT latency overlaps the MMA phase, instead of stalling the post-compute inter-barrier window (ncu on the production 24L/dim1536/ffn6912 shapes: barrier 3.7-4.1 + long-scoreboard 1.6-1.8 stalls per issue before).
- Split the stage loaders into load_async (cp.async operands, committed after the post-compute barrier) and load_direct (ring-indexed by tile, not stage, since the rings differ in depth).
On the production shapes vs cuBLASLt _scaled_mm: dX 1.11-1.24x (from 1.27-1.42x), dW 1.32-1.35x (from 1.26-1.49x); generic 4096/11008 shapes also improve (dX 37.6->40.8 TF, dW 36.0->37.3 TF). fp8 train step vs bf16: ~1.2x at both 512 and 2048 tokens.
- Finalize scale rings inside the quantize kernels: a last-block epilogue (threadfence + counter elect) folds amax into hist, reduces the window and publishes the next scale on device, zero extra launches; _ScaleRing packs [hist | scale | counter] into one CUDA buffer.
- Split FP8QuantizeParams out of FP8Params so each operator owns its fields; linear_forward/backward_fp8 take optional ring arguments.
- Drop the inference weight-quantization cache; the optimizer bumps the weight version every step, so a cache would miss anyway.
- Zero amax scratch via empty + cudaMemsetAsync instead of torch::zeros, cutting a ~50us fill_ dispatch per quantize.
- Stage crosswise-B operands K-major with cp.async (contract >= 8192) and PRMT-transpose per k_seg region in smem, interleaved with the MMAs; the sync LDG + byte-scatter path it replaces was long-scoreboard bound (ncu 4.6 vs 0.4 stalls/issue).
- Load crosswise-A direct with an in-register PRMT transpose; its operands are typically L2-resident and the staging round trip measured as a net loss.
- Enable grouped rasterization for the congruous NT forward (shared B stripe keeps the weight operand hot in L2) and make the smem budget layout-aware (Fp8GemmSmem) while holding two CTAs per SM.
- Annotate ops/fp8.py return types; drop weight-cache and decorator tests, hoist their imports to module level.
e2e 12L/dim1024/B4xT512 fused AdamW: fp8 137.8ms/step vs bf16 210.3ms, 1.53x. Kernel vs cuBLASLt _scaled_mm: fwd 1.03-1.09x, dX 1.33-1.47x, dW 1.30-1.39x (from 1.10/1.42-1.49/1.52-1.56x), before the pre-transposed copies cuBLASLt needs for dX/dW. fp8 train step vs bf16: 1.34x at 2048 tokens (was 1.25x), 1.08x at 512.
- last-block epilogue (threadfence + counter elect) folds amax into hist[idx], reduces the window and publishes the next scale on device — zero extra launches per linear layer
- _ScaleRing packs [hist | scale | counter] into one CUDA buffer; the eager hist-write / max / scale-copy chain and update() are gone
- split FP8QuantizeParams out of FP8Params so each operator owns its fields; linear_forward/backward_fp8 take optional ring arguments
- e2e 12L/dim1024/B4xT512 (fused AdamW): fp8 137.8ms/step vs bf16 210.3ms, 1.53x; fwd 1.82x, bwd 1.50x
- load_operand_tile ColMajor path issued one LDG then immediately scattered 16 byte-granular shared stores, so every store waited on the preceding global load; the runs of one row group now batch into registers first (v[kPasses]) and scatter after, overlapping the LDG latencies
- hoist pass-invariant predicates: the alignment check folds to one uniform (base | ld) & 15 test since r0 is always a multiple of 16, and rows_full leaves the per-pass condition; the contract tail zero-fills without global traffic
- RowMajor path hoists the row bound and the (invariant) chunk-alignment check out of the per-chunk loop
- measured (cuda events, old/new interleaved): crosswise bwd gemms +5-9%, RowMajor and fwd NT within noise; model step unchanged (in the 563-617 ms band)
- file passed through clang-format with the new .clang-format config
- linear_forward_fp8 accepts pre-quantized w8 (matching fmt) and skips the weight quantize; amax_w returns 0 on that path since no bf16 values are seen
- bias is now fused into the GEMM epilogue for both dtypes, replacing the separate torch-level add (one elementwise kernel per linear removed)
- FP8Params.bias becomes void* with a new bias_scale slot: null scale = raw bf16 bias, non-null = fp8 storage dequantized in the epilogue after the operand scaling and before any output quantization
- ops/fp8.py relaxes the w dtype check to bf16-or-fp8 and passes bias_scale through
- regression test covers w8/b8, w8/bf16-bias and the amax_w = 0 contract vs an explicit quantization reference
- K tile 32->64 (new default): fewer barriers, more MMA per stage; generalize tile_at swizzle and load_operand_tile accordingly
- 64x128 small-M CTA for m<=64 (2x at 64x4096x4096)
- L2 rasterization for crosswise-A layouts (+6..21%)
- micro-bench: NT 4096^3 +35%; linear fwd 1.24-1.76x, bwd 1.71-2.27x vs bf16
- add csrc/tests/fp8_test.cu (single MMA demo + GEMM layouts x K-tiles vs CPU reference)
- NSDMI null/-1 defaults for AttentionParams/FP8Params pointer+flag members: partially packed structs can no longer hold garbage non-null pointers that gate optional paths (root cause class of the paged test bug); still aggregates, still trivially copyable
- move per-lane ldmatrix wrappers (ldsm_x2/x4) from fp8/gemm.cuh to common/mma.cuh as ldmatrix_x2_lane/x4_lane, next to the single-address variants
- DEVICE_FORCEINLINE macro in common/mma.cuh (matches layout_policies.cuh, internal linkage)
- frag_addr now delegates to tile_at: the swizzle math has one source
- operand layouts as CUTLASS-style RowMajor/ColMajor tags threaded from launch_fp8_gemm through the kernel to load_operand_tile; B's operand view via transpose_layout_t; call sites read <Fmt, false, RowMajor, ColMajor> instead of <Fmt, false, false, true>
- attention family -> astrai::attention; fp8 family -> astrai::fp8
- new common/reduce.cuh (warp/group reductions, atomic_max_float)
- new common/cp_async.cuh (predicated cp_async_16, commit/wait group)
- move MAX_SPLITS into attention/common.h; delete warp_utils.cuh
- .cu bindings and pure C tests open family namespaces via using
- 128x128 CTA of 8 warps x 64x32 warp tiles: 16 mma.sync per warp per K-segment (was 8)
- ldmatrix.x4/x2 with per-lane swizzled addresses replaces 36 scalar LDS per warp-tile step
- __launch_bounds__(256, 2) caps registers at 124 so two CTAs fit per SM
- fp8 linear vs bf16 cuBLAS: fwd 1.07x->2.65x, bwd 1.41x->2.65x by size, peak 31-33 TFLOPS
ViperEkura