perf: rebuild decode gemm dispatch around shape-driven tile configs

- split-K removed entirely: tiled kernel walks K in one pass, no partials/semas workspace, no memset, single launch per call
- skinny GEMM (M<=8) dispatch table replaces the hand-written switch
- shape-driven four-family table replaces plan_gemm: wide-N (n>=4096) default {16,64,64,3,128} with BM=32 at M>16; narrow-N deep-K rings {16,32,256,2,64} while the grid fits one wave, {16,32,128,2,64} past it
- narrow-N is K-serial: widening the grid measurably does nothing (BN 64->32 ties, doubled m_tiles tie, kv at 4 blocks ties q/o at 24); deeper K chunks win until 72KB smem forces one CTA per SM and past one wave the 2-wave quantization loses to BK=128
- launch-check macros in common/launch.cuh; smem opt-in for the 72KB/60KB rings
- rename kernels/bf16_*.cu to gemm.cu/swiglu.cu; module names unchanged
- Python gate: lm_head (N>32768) falls back to cuBLAS, band narrows to M<=32
- drop the stale per-op benchmark narratives; fold the live numbers into cuda_kernels.md

Benchmark: NVIDIA L20 (sm_89, 92 SMs), CUDA 12.8, bf16, L2-thrash weight rotation, per-call medians at M=16: q/o 9.5us, kv 8.6us, gate/up 33.3us, down 33.7us (down -29% vs prior default). End-to-end 1B decode (gen 128, 3 trials, tokens/s vs cuBLAS): B=1 260 vs 252, B=8 1660 vs 1446, B=16 2464 vs 2437, B=32 3620 vs 3690. Prior split-K dispatch measured B=16 2243 / B=32 3393.

astrai/extension/backend/swiglu.py

@@ -52,16 +52,16 @@ def _swiglu_capable(x: Tensor, up_weight: Tensor, gate_weight: Tensor) -> bool:
 def swiglu(x: Tensor, up_weight: Tensor, gate_weight: Tensor) -> Tensor:
     """Apply the dense-MLP SwiGLU projection with a safe torch fallback.
 
-    ``ASTRAI_SWIGLU=0`` and ``auto`` keep the unfused linear-backend chain;
-    ``1`` forces the fused primitive for supported inputs. Auto will adopt
-    an M-banded rule mirroring the linear backend once end-to-end evidence
-    qualifies one.
+    ``ASTRAI_SWIGLU=0`` keeps the unfused linear-backend chain; ``1`` forces
+    the fused primitive for supported inputs; ``auto`` (the default) uses
+    the fused primitive for decode batches with M in ``{1, ..., 8}``. The
+    fused kernel reads x once and covers both projections plus the SiLU
+    gate-multiply in a single launch, measured 9-15% faster than the
+    unfused chain per MLP call on L20 with L2-thrashing weight rotation.
     """
-    if env_mode("ASTRAI_SWIGLU") != "1" or not _swiglu_capable(
-        x, up_weight, gate_weight
-    ):
-        return _unfused_swiglu(x, up_weight, gate_weight)
-    return _fused_swiglu(x, up_weight, gate_weight)
+    if env_mode("ASTRAI_SWIGLU") != "0" and _swiglu_capable(x, up_weight, gate_weight):
+        return _fused_swiglu(x, up_weight, gate_weight)
+    return _unfused_swiglu(x, up_weight, gate_weight)
 
 
 __all__ = ["swiglu"]