[zihaomu]

Related to #671; follows #823.

Problem

For ggml.mul_mat_id with MMQ and repeated nonnegative expert IDs in one token, mm_ids_helper packed only one row per (token, expert) but counted every route when computing expert offsets. The resulting holes in ids_src1 / ids_dst could produce wrong output or an illegal GPU memory access. A direct call with 9 tokens, top-k 3 and IDs [0, 0, 2] for the first token reproduced the fault on an R9700.

#823 fixed the known hybrid-decode trigger by preserving inactive routes as -1. Normal model top-k produces distinct active IDs. This PR makes the lower-level operator safe for callers that supply repeated active IDs; it does not claim an observed production model regression after #823.

Change

Pack one record for every matching route and derive expert bounds from the same route count. The common fixed top-k path uses shared memory when it fits; larger inputs and generic top-k use a direct two-pass mapping. Document the per-route contract and add four CPU-reference regression cases for Q4_0 / IQ1_M and top-k 3 / 8.

Validation

• ROCm 10.1, Radeon AI PRO R9700 (gfx1201), PyTorch 2.14.0: FREETOKEN_DISABLE_KERNEL_CACHE=1 python -m pytest tests/kernels/test_ggml.py -q -> 65 passed. An external qualification also passed 65 combinations, including repeated active IDs, inactive -1, per-route activations and a large shared-memory fallback.
• RTX 5090 (sm_120), driver 615.71.09, CUDA / nvcc 13.0.88, PyTorch 2.14.1+cu130, apache-tvm-ffi 0.1.13.post3: the same GGML test command rebuilt the native module for sm_120 -> 65 passed, 0 skipped. The original 9-token IQ1_M shape also matched a CPU reference with shared and per-route activations (3,456 elements each; max absolute errors 4.29e-6 and 5.25e-6).
• Exploratory R9700 Q4_0/BF16 Graph replay for unique IDs (main -> patch): T=17/top-2 15.86 -> 15.78 us, T=32/top-8 16.01 -> 16.01 us, T=256/top-8 50.18 -> 50.89 us. These are operator measurements, not model throughput.

All tests used synthetic packed GGUF expert tensors; no model checkpoint or end-to-end serving run is claimed.

[jason-fxz]

No caller produces repeated active ids after #823, so there's no workload for this fix. Closing this.