mmq: compacted MoE tiling for RDNA3.5 (default on)

[roberteg16]

Summary

Compacted grid for the MMQ Mixture-of-Experts matmul on RDNA3.5, enabled by default (previously opt-in via GGML_CUDA_MMQ_MOE_COMPACT, now removed). Modeled on vLLM's moe_align_block_size + fused_moe_kernel.

Instead of the worst-case (nty, ntx, ne02) grid -- where ntx = ceil(ne12/mmq_x) assumes all tokens land on one expert, so most ne02 * ntx column-blocks belong to experts that early-exit -- this launches only the minimum number of mmq_x-wide column blocks implied by the actual per-expert token counts (padded to mmq_x).

• mmq_build_moe_block_map (single block, parallel prefix sum in shared memory +
  parallel binary-search scatter) builds from expert_bounds:
  • block_start[] -- exclusive prefix sum of ceil(tokens_e / mmq_x)
    (block_start[ne02] = total real blocks, the device-side "num_tokens_post_padded").
  • block_expert[] -- compacted column-block -> expert.
• The conventional-tiling kernel decodes blockIdx -> (it, jt, zt) via those arrays;
  padding blocks self-exit (m_block >= block_start[n_experts]).
• Grid is sized to the host-known upper bound
  ceil((ne_get_rows + ne02*(mmq_x-1)) / mmq_x) -> fixed per shape, no host readback
  (graph-safe).

Scope / safety

• Gated on expert_bounds != nullptr && GGML_CUDA_CC_IS_RDNA3_5(cc). The
  block_expert pointer is the mode flag: non-null only for this launch.
• Dense, multi-GPU split and stream-K (NVIDIA/CDNA) paths are untouched
  (they pass nullptr, nullptr, 0).
• The alternate (it, jt, zt, wt) decode only re-enumerates the tiles the legacy grid
  would have kept; wt = 0 is exact (MoE asserts ne13 == 1), zt/jt stay in the
  same ranges, padding blocks return before indexing block_expert (no OOB).
• Falls back to the legacy grid if max_m_blocks >= MMQ_MAX_GRIDDIM_Y (65535).
• mm_ids_helper is left unmodified.

Verification (gfx1151, Strix Halo)

Decode is provably unaffected (rocprofiler / GGML_ROOFLINE_OUT)

Per-op kernel traces confirm the compacted path is prefill-only:

• Decode (tg): MUL_MAT_ID dispatches mul_mat_vec_q (MMVQ) + quantize_q8_1. Zero mul_mat_q and zero mmq_build_moe_block_map dispatches (0 of 603437 kernel launches in the tg128 trace).
• Prefill (pp): MUL_MAT_ID dispatches mul_mat_q + quantize_mmq_q8_1 + mm_ids_helper + mmq_build_moe_block_map.

So enabling by default carries no decode risk: the MoE GEMM in decode never enters the MMQ path.

Prefill throughput, 4 MoE models (branch vs gfx11 integration baseline)

llama-bench -ngl 999 -n 128 -r 10 -t 8 -fa 1 --poll 50 -ctk f16 -ctv f16, uncontended via gpu-lock, ROCBLAS_USE_HIPBLASLT=1.

model	pp128	pp512	pp1024	pp4096
Qwen3-30B-A3B-Instruct-2507	+5.1%	+2.1%	+1.0%	+1.3%
Qwen3.6-35B-A3B (Q4_K_M)	+5.4%	+4.5%	+4.7%	+4.4%
Qwen3.5-35B-A3B (Q4_K_M)	+3.6%	+4.2%	+3.5%	+3.3%
gemma-4-26B-A4B-it	+2.4%	+1.9%	+3.0%	+2.2%

Decode is unaffected (same-build A/B)

Toggling the compacted path on/off in the same build leaves decode throughput at parity (tg128, -n 128 -r 10):

model	compact on	compact off	Δ
Qwen3.6-35B-A3B (Q4_K_M)	58.33 ± 0.08	58.42 ± 0.09	-0.15% (noise)
Qwen3.5-35B-A3B (Q4_K_M)	57.75 ± 0.07	57.75 ± 0.07	0.0%

Test plan

• Roofline traces: MMQ compact path active in prefill, absent in decode.
• E2E llama-bench prefill A/B across 4 MoE models (above).
• test-backend-ops -o MUL_MAT_ID with default-on: 790/790 pass (gfx1151).