perf(ds4): batch heterogeneous MoE prefill by expert

[cheese-cakee]

Summary

Accelerates DeepSeek-V4 heterogeneous prefill by batching owner-local routed experts into grouped ggml_mul_mat_id graphs instead of issuing a serial graph per active expert.

• enables the existing expert-major heterogeneous path from 32 tokens (previously 512);
• builds one grouped gate/up/down MMID graph per owner GPU;
• adds a GPU-native route histogram/prefix/scatter helper for up to 256 experts;
• preserves masked owner routes as exact zero contributions;
• retains DFLASH_MOE_GROUPED_MMID_PREFILL=0 and DFLASH_MOE_EXPERT_MAJOR_MIN_TOKENS=512 as rollback controls.

Matched performance

Measured on Lucebox6 with AMD Radeon AI PRO R9700 (gfx1201) + Strix Halo Radeon 8060S (gfx1151), the same ROCmFP2 model, cache disabled, two warmups, and five measured requests per cell. Baseline is current upstream/main at ac22a3ed; the measured implementation is retained in the final 041b59df history, followed only by masked-route correctness and regression-test hardening.

Workload	Current main	Candidate	Result
401-token prefill, chunk=512	54.16 tok/s (MAD 0.19)	184.11 tok/s (MAD 1.41)	3.40x faster (+239.9%)
MoE FFN time, 401 tokens	6,791.6 ms	1,661.2 ms	4.09x reduction (-75.5%)
2048-token prefill, chunk=2048	292.42 tok/s (MAD 10.64)	304.69 tok/s (MAD 8.70)	+4.2%

The wide-context cells use the same fixed 22,000 MiB expert budget (hot 60 experts/layer) on both heads so that both can run the single 2048-token chunk without an allocation failure. The headline is the matched 401-token production-prefill improvement

Correctness and validation

• All warmup and measured full-model responses in the retained 401- and 2048-token A/B cells produced the same SHA-256: cd5cb9fb5ac3c4f4007e8b41d117da21622439cd05c1728f3e82f90e4f869dad.
• Every retained request was cache-cold (restore=false, slot=-1, prefix_len=0, disk_hit=false).
• test_deepseek4_mmid_grouped_cuda: parity PASS across 72 cases on physical gfx1201 and gfx1151, including direct +0.0 assertions for wide masked ROCmFP2/ROCmFP3 owner routes.
• test_deepseek4_unit: PASS on physical gfx1201 and gfx1151.
• git diff --check: clean.

The masked-route regression uncovered during qualification is fixed by zero-initializing compact MMID destination lanes before dispatch; this prevents unwritten owner-masked columns from contributing stale/NaN values during weighted reduction.

[cubic-dev-ai]

All reported issues were addressed across 5 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>