cuda-ab: ctaid_probe (CTA-coordinate + A-row reader) + GEO geometry print in transformer_mt

The vGPU GR engine's failing chains are correct for the by=0 band and
replicate one tile for rows >= 16 with the QMD geometry and the input recs
proven clean, so the remaining hypothesis is engine-side CTA-coordinate
behaviour. ctaid_probe reads it directly in two phases on the exact
(16,4,1)/(16,16,1) geometry:

  Phase 1 coord_map writes by*10000+bx*100+ty*10+tx per element (int32
  64x256) and reports per-band row0 plus a bad count - a collapsed or
  wrong CTAID.Y shows instantly.
  Phase 2 arow_read writes A[(by*16+ty)*256+(bx*16+tx)] with A[i]=i, so a
  correct engine reads each band's own rows; a collapsed row register reads
  one band's rows everywhere.

transformer_mt prints its per-chain geometry (GEO lines) for the host-side
GRID-WIT comparison. Runtime <10 s. Built with
nvcc -O2 --cudart=static -arch=sm_75 (deterministic static runtime).

1 files changed, 8 insertions, 0 deletions

diff --git a/tools/cuda-ab/transformer_mt.cu b/tools/cuda-ab/transformer_mt.cu
index ffd23aa..5cfe55e 100644
--- a/tools/cuda-ab/transformer_mt.cu
+++ b/tools/cuda-ab/transformer_mt.cu
@@ -142,6 +142,14 @@ static void launch_chain(cudaStream_t st, const float *x, struct attn_bufs b,
 	dim3 g1(DH / TILE, SEQ / TILE);
 	dim3 g2(SEQ / TILE, SEQ / TILE);
 	dim3 tb(TILE, TILE);
+	static int chain_n;
+
+	/* r226 GRID-WIT: the exact geometry every chain launch uses, so the
+	 * host-side QMD geometry capture (GRID-WIT in vgsp) can be compared
+	 * against what the client actually launched. Chains 0..2 are the
+	 * multi-stream class, 3..7 the 5-launch loop. */
+	printf("GEO chain=%d gemm_nn grid=(%u,%u,1) block=(%u,%u,1) gemm_nt grid=(%u,%u,1) softmax grid=(64,1,1) block=(64,1,1)\n",
+	       chain_n++, g1.x, g1.y, tb.x, tb.y, g2.x, g2.y);
 
 	gemm_nn<<<g1, tb, 0, st>>>(x, wq, b.q, SEQ, DH, DH);
 	gemm_nn<<<g1, tb, 0, st>>>(x, wk, b.k, SEQ, DH, DH);