ggml-hrx: MXFP4 weights (shared dequantizer, exact E8M0 scale, known-answer test)

[bong-water-water-bong]

MXFP4 (gpt-oss's expert format) as HRX weight format 39 in the shared dequantizer.

Changes

• motifs/dequant_1bit.loom (new, ours): MXFP4 decoder. 17-byte blocks (E8M0 exponent, 16 bytes of E2M1 nibbles, IQ4_NL nibble order), scaled by the exact GGML_E8M0_TO_FP32_HALF built from its f32 bits. The file is self-contained: no func.decl of the inline helpers in dequant.loom, because those declarations don't resolve at kernel link time (LINK/MATERIALIZE: unresolved exact declaration).
• dequant.loom: tile (136 B / 256 values), row (17 B per block) and f16/f32 value hooks.
• manifest.json: the new file after dequant_prism.loom in all 87 lists.
• Format enum, type mapping, config value 39. Dense input sizes are multiples of 32.
• MUL_MAT_ID now admits only input sizes that are a multiple of 256, which its kernels declare. It used to reuse the dense rule (multiples of 32 for 32-value block formats). gpt-oss-20b (input 2880) then failed in the JIT with CONFIG/INVALID instead of running its experts on the CPU.
• tests/test-hrx-mxfp4.cpp (new): bit-exact known-answer check of MXFP4 GET_ROWS on HRX0.
  • Runs with no scheduler, so no CPU fallback.
  • Requires a get_rows kernel in the HRX plan.
  • Exponents 120/127/134 and the edges 0/1/2/254; every code in both nibbles; memcmp against dequantize_row_mxfp4.
  • At e = 0 and 1 the scale itself is subnormal (2^-128, 2^-127). The GPU kernels flush it, so those rows come back as same-sign zeros (values at most 12 * 2^-127). The test accepts that for those exponents only.

Tested on HRX0, balanced power mode. Tree: 1bit/hrx-tq (these commits plus TQ1_0/TQ2_0, on d60cc4f).

• test-hrx-mxfp4: 7 rows x 256 values bit-exact; 448 values with a subnormal scale flushed. The dispatch log shows loom_libs:ggml_get_rows_f32.
• test-backend-ops: full suite 1051/1051.
  • MXFP4 MUL_MAT 13 OK / 64 not supported, MUL_MAT_ID 12 OK / 62 not supported, GET_ROWS 1 OK / 3 not supported, 0 FAIL.
  • Same not-supported pattern as Q4_0, Q8_0 and Q4_K (broadcast and batched shapes).
• loom-link --verify=true and ggml-hrx-compile-kernel on the rebased tree (e2b946a) for the kquant and get_rows recipes.

gpt-oss-20b (ggml-org gpt-oss-20b-MXFP4.gguf):

• It loads and runs on HRX0. Greedy text is correct ("red, green, and blue ... additive color mixing ...").
• Its MXFP4 tensors are all expert weights with input size 2880, which isn't a multiple of 256, so MUL_MAT_ID declines them and they run on the CPU. On this model MXFP4 doesn't reach HRX yet. That needs a MUL_MAT_ID kernel that handles input sizes in multiples of 32, which is the next step and is not in this PR.
• With that split: wikitext KLD vs CPU 0.0309, same top token 88.8% (8 chunks of 512); pp512 25.8, tg128 12.6 ± 1.6 tok/s.
• The KLD is above what I'd expect and isn't explained yet. It doesn't come from MXFP4 on HRX, which isn't used on this model, but I haven't traced it.

🤖 Generated with Claude Code

[bong-water-water-bong]

Review (PR-Agent duty): approve.

• MXFP4 in the shared dequantizer, in our own file (motifs/dequant_1bit.loom, full notice), with minimal hooks in AMD's dequant.loom, the enum/config/type maps and the manifest.
• 17 bytes per 32 values, 136-byte tiles, row bytes = blocks x 17, ggml's nibble order.
• The E8M0 half scale is built from bits exactly as GGML_E8M0_TO_FP32_HALF does. tests/test-hrx-mxfp4.cpp checks it bit for bit on HRX0, and requires an HRX dispatch plan, so a CPU fallback can't pass. The only allowed difference is a same-sign zero for the subnormal scales (e = 0/1, values about 1e-38); the GPU flushes those.
• The MUL_MAT_ID guard (input_size % 256, which the kernels declare) is a correctness fix: 2880-wide experts used to be accepted and then fail in the JIT.
• Full suite 1051/1051, MXFP4 ops 0 FAIL.

Not covered yet (stated in the PR): on gpt-oss-20b all 72 MXFP4 tensors are 2880-wide experts, so they run on the CPU. The follow-up (MUL_MAT_ID for multiples of 32) is the next MXFP4 item. Merging when the hosted jobs pass.