- evaluate_mmlu: fix double few-shot injection (build_prompt no longer
adds few-shot, apply_chat handles it once)
- evaluate_humaneval: fix pass@k k-filtering to be per-problem instead
of using first problem's n globally; reuse ProcessPoolExecutor across
problems; fix closure UnboundLocalError in test_one; handle None in
report when k > n
- evaluate_ifd: remove dead code (score_plain/score_messages); add
multi-file/directory input support with --input_path/--output_dir;
add summary.json aggregation and --max_samples; add --dtype flag
- evaluate_ppl: add --device and --dtype flags (was hardcoded to cuda)
- evaluate_ifeval: fix docstring path (scripts/tools -> scripts/eval)
- analyze_weights: add --output JSON export; fix dead code filter
("_norm" not in r was always True)
- BFDSplitPacking splits over-length sequences into chunks before BFD
- All keys (loss_mask, position_ids, ...) split in lockstep for alignment
- No tokens lost vs bfd which truncates over-length sequences
- Tests: token preservation, chunk alignment, short unchanged, vs bfd
- Add --resume bool flag to train.py CLI
- --param_path always loads weights only by default
- --resume restores epoch, consumed_samples, optimizer & scheduler
- Checkpoint.load() now preserves full meta dict
- Update test_early_stopping to use new param_path/resume API
- Replace is_causal + causal_offset with unified causal_offset (-1 = off, >=0 = first Q pos)
- Causal and mask can now coexist (was mutually exclusive)
- Add stride-based addressing for Q/KV/O (layout-agnostic, zero-copy)
- Add layout param ("bhld"/"blhd") parsed in Python, passed as int to C++
- Support 2D [batch, kv_len] and 3D [batch, q_len, kv_len] mask
- Vectorize paged KV gather in Python fallback (was per-token Python loop)
- Extract shared helpers: compute_num_splits, alloc_split_partials, dispatch_head_dim
- Unify paged_decode entry via attn_pack_paged_params
- Update mma_softmax_tile for 3D mask with per-row qrow indexing
- Mask prompt tokens to 0 so their logprobs excluded from ratio/KL
- Switch to token-level ratio + PPO clipping via reduction='none'
- Slice response token logprobs from full sequence output
- Replace k3 KL estimator with non-negative k1 estimator
- Fix epsilon from finifo.eps (~1e-38) to 1e-8
- Remove unused 'reduction' param from GRPOStrategy.__init__
- Clarify offline batch semantics in docstring
- Add 11 unit tests for masking, advantage, KL, sync, clipping
- Sync training.md and architecture.md docs
- Remove MIN_BLOCKS template param and __launch_bounds__ attribute
- Profiling shows smem (not registers) is the occupancy bottleneck for D>=64, making the hint a no-op
- D=64 sees 2-4% speedup, D=128 unchanged (smem-capped at 1 block/SM)
- Update comment blocks in kernel header and both dispatch sites
- Verified correctness via standalone CUDA test (max_err ~1e-4)
- Double-buffered KV (STAGES=2) for D<=128: next tile cp.async overlaps current tile MMA compute, hiding global load latency
- Q loaded directly from global into mma A-operand registers, removing sQ staging and prologue syncwarp
- Predicated cp.async unifies full and partial tile paths, eliminating scalar fallback branch
- STAGES=1 fallback for D=256 (double-buffer would exceed smem budget)
- Applied to both contiguous and paged decode MMA kernels
- ~1.27x average speedup on L20 (sm_89), zero precision loss
- Add attn_paged_decode wrapper in ops.py with gather fallback
- Register kernel in loader.py and export from __init__.py
- Extract test_utils.cuh shared by all attention unit tests
- Rename attn_paged_vs_contiguous.cu to attn_paged_decode_test.cu
- Refactor decode/prefill tests to use common bf16 helpers and cpu ref
- Fix k_cache dim check in attn_paged_decode.cu
- PagedAttentionParams merged into attn_common.h
- Scalar variant: warp-per-query-head split-KV, resolves page table per-position for the K/V shared-memory tile load
- MMA variant (sm_80+): tensor-core head-packing with cp.async, single page-table lookup per tile (BC=32 fits within page_size>=32)
- Standalone test: 14 cases across head_dim 32/64/128/256, GQA, multi-batch, both paths verified against CPU reference
- inline decode_use_mma() into dispatch_decode()
- drop G>1 guard, MMA works correctly for G>=1
- decode is memory-bound, tensor cores + cp.async still win at G=1
- Remove tiles_total/8 min-work cap that limited splits for small workloads
- Simplify decode_num_splits to only use base_blocks and tiles_total
- Short sequences now generate more blocks, improving SM utilization
- Convert AttentionParams to a template struct supporting arbitrary types
- Rename attn_common.cuh -> attn_common.h (no CUDA-specific code remains)
- Include standard headers explicitly in each .cuh instead of via attn_common.cuh
- Allow .h files in csrc/ via .gitignore
- Replace bf16 pre-scale Q loading with direct 32-bit aligned bf16x2 reads
- Apply scale in float32 after Q@K^T, before online softmax
- Reduces causal max error from 2^-6 to 2^-8 with zero perf cost
- Rename all csrc/kernels/gqa_*.cuh/cu to attn_*, with _split_q / _split_kv
strategy suffix and optional _mma compute suffix
- Remove non-split MMA decode kernel, keep only split-KV path
- Convert scalar decode fallback to split-KV (o_part/ml_part + combine)
- Move combine kernel to attn_decode_split_kv.cuh (shared by both paths)
- Rename GQAParams to AttentionParams
- Update all C++ #include, PYBIND11, and Python extension references
- read the 8 Q elements each lane needs straight from global into the mma
A-operand layout, pre-scaled, instead of staging through shared sQ
- removes the sQ smem area (20KB->16KB) and the serialized per-warp prologue
with its WARPS __syncthreads barriers
- result vs torch SDPA: prefill 0.70-0.82x -> 0.85-1.00x (matches torch at
seq=128; 2048 1.251->1.159ms), correctness unchanged across head dims
- double-buffer K/V one tile ahead via cp.async to overlap load with tensor-core math (ncu long_scoreboard 2.12->0.53)
- reorder wait->barrier->prefetch so one __syncthreads/tile covers both cross-warp publish and buffer-reuse (was two)
- add predicated cp_async_16_pred (src-size=0 zero-fills OOB) to unify full/partial tiles, dropping the scalar fallback
- halve BC to 16 to keep 3 blocks/SM despite the doubled smem
- pack adjacent bf16 output into one 32-bit STG, removing the uncoalesced scalar-store penalty (14%->5% sectors)
- result vs torch SDPA: prefill 0.61-0.78x -> 0.70-0.82x, spills eliminated
Decode has only batch*kv_head independent tasks, so the grid was tiny (e.g. 16 blocks) leaving most SMs idle (ncu: 0.04 waves/SM, 11% DRAM).
- Partition KV across gridDim.z blocks emitting unnormalised (O, m, l) partials, reduced by a new combine kernel
- Choose split count to fill the device (~2 blocks/SM), capped by tile count and 32; fall back to single-pass direct-write when batch*kv_head already saturates the SMs
- Refactor decode dispatch into named helpers, de-duplicate scalar fallback
Result: now DRAM-bound at 63% (99->543 GB/s), 2.1-2.5x over torch SDPA in the low-parallelism regime, on par at high parallelism
Moving Qa ldmatrix into the tile loop caused warps 0-2 to read
warp 3's Q data from sQ (only the last warp's data survives the
serialized load loop). Reverted to loading Qa during the init phase
and keeping it resident; __launch_bounds__ still forces 128 regs
(33% occupancy) with spill to local memory.
- Move Qa[KD][4] into tile loop (reload from sQ per tile)
cutting ~32 resident registers for HEAD_DIM=128
- Replace extern __shared__ with static template-sized smem
(no cudaFuncSetAttribute or dynamic allocation needed)
- Add __launch_bounds__ with MIN_BLOCKS param, dispatch by HEAD_DIM
(hd=128→4, hd=64→6, hd=32→6)
- Remove dynamic smem from scalar kernel and C test
- Result: hd=128 168→128 regs, 25%→33% occupancy
- Extract SingleOutputMaskBuilder for SFT and pretrain configs
- Extract MultiOutputMaskBuilder for DPO and GRPO configs
- Keep SectionedMaskBuilder as backward-compatible facade
- Register "single" and "multi" names in MaskBuilderFactory
- Add parity and rejection tests for concrete builders
Decode MMA kernel previously used scalar global→shared loads with
LD=HEAD_DIM+8 padding and per-tile scale multiply. This commit brings it
in line with the prefill MMA kernel (which already had these optimizations):
- cp.async K/V loads (bypasses registers, halves load instructions)
- XOR swizzle: LD=HEAD_DIM instead of HEAD_DIM+8 (zero waste smem)
- Pre-scale Q during load (removes per-tile scale multiply in softmax)
- Clean up prefill MMA kernel comments (no code change)
~2x speedup on decode (0.47ms→0.24ms at seq_len=512)
- add case 32 to decode/prefill dispatch switch
- fix swiz_col out-of-bounds for HEAD_DIM=32: XOR mask now limited to chunk count (3 for 32, 7 for >=64) instead of always 7, which produced column offsets >= LD=32 and corrupted shared memory
- restructure decode dispatch to #ifndef/#else/#endif matching prefill
- split astrai/extension/__init__.py into loader.py (kernel .so discovery) and ops.py (wrapper functions + torch SDPA fallback); __init__.py now re-exports the public API
- Add gqa_decode_attn_mma.cuh for tensor-core decode path
- Add dispatch_decode<> selecting MMA vs scalar based on G and mask
- Add TORCH_CHECK for unsupported head_dim instead of silent scalar launch
- centralize CXX_FLAGS/NVCC_FLAGS in csrc/build.py as single source
- add --use_fast_math, --ptxas-options=-O3,-v, --extra-device-vectorization
- add -march=native -funroll-loops host flags
- setup.py reads shared cxx_flags/nvcc_flags from registry
- sync pure-C test build commands with new flags
ViperEkura