biondizzle 014d647ba3 fix: sink bias domain correction — add attn_sink/scale to raw logits ...

The softmax scales by scale_log2 = scale * log2(e). Adding sink_val to
raw logits causes it to be scaled too. Fix: add sink_val/scale instead,
so after scaling: (sink_val/scale) * scale_log2 = sink_val * log2(e).
This correctly multiplies attention weights by exp(sink_val).

2026-05-26 15:03:49 +00:00

..

attention

fix: sink bias domain correction — add attn_sink/scale to raw logits

2026-05-26 15:03:49 +00:00

cache

KV Cache: schema, allocator, pools, manager, append_swa kernel

2026-05-22 00:08:38 +00:00

compressor

Restructure: cutedsl/ -> dsv4/ with proper layering

2026-05-21 17:30:44 +00:00

cuda

fix quantize_nvfp4 kernel: use proven single-thread-per-CTA pattern from deinterleave_quantize.cu

2026-05-25 16:21:44 +00:00

decode

Restructure: cutedsl/ -> dsv4/ with proper layering

2026-05-21 17:30:44 +00:00

gemm

fix: add SwiGLU clamping to fused kernel (paper §4.2.3, CG-1)

2026-05-23 06:32:54 +00:00

indexer

Indexer: score+topk kernel, gather KV, compute_valid_lens

2026-05-22 01:20:39 +00:00

router

Router: Blackwell-native fused decode kernel — real CuTeDSL implementation

2026-05-21 22:04:20 +00:00

__init__.py

Restructure: cutedsl/ -> dsv4/ with proper layering

2026-05-21 17:30:44 +00:00