Pin llama.cpp d5048ad: IQ2_XXS and IQ2_XS weights on HRX0 (llama.cpp #50)

[bong-water-water-bong]

Moves third_party/llama.cpp from d1747cb to d5048ad. The only change is llama.cpp #50, which gives IQ2_XXS and IQ2_XS weights kernels on HRX0.

Why it matters. IQ2_XXS and IQ2_XS had no matmul on HRX0. llama.cpp's load-time 512-token matmul test failed, so these weights went to CPU_REPACK and each use split the graph. Unsloth's UD GGUFs below 4 bits carry both.

What #50 does

• Adds IQ2_XXS (format 24) and IQ2_XS (format 25) to the shared dequantizer, which covers prompt WMMA, generic decode and GET_ROWS.
• Adds lane functions to the K-quant decode kernels, for 1 token and for 2-8 tokens. The grids are staged in workgroup memory, like IQ3_S.
• The generators for the grid tables and decoders are in ggml/src/ggml-hrx/tools/generate_iq2_*.py, and their output is verbatim in the kernel files.

Measured on strixhalo (HRX0), Qwen3-4B requantized Q8_0 → IQ2_*

• Speed, previous (IQ2 on the CPU) → this:
  • IQ2_XXS: pp512 59 → 715 tok/s, tg128 23.0 → 39.5 tok/s.
  • IQ2_XS: pp512 87 → 268 tok/s, tg128 28.0 → 30.2 tok/s.
  • tg128 is from 2 × 6 runs with the box lock held.
• Accuracy, KLD against the Q8_0 model on the decode path (-b 1):
  • IQ2_XXS: 1.1306, against 1.141 on CPU.
  • IQ2_XS: 0.7744, against 0.786 on CPU.
• Tests: test-backend-ops -b HRX0 passes 971/971. 64 batched IQ2_XXS matmul shapes are still declined and run on the CPU, as before.

Also in this PR:

• registry/architectures.json regenerated for the pin. tools/registry_build.py --check-pins passes, and so does tools/check_pins.py origin/main (ahead).
• docs/hrx.md: the sub-4-bit note now lists IQ1_S/IQ1_M as the remaining CPU fallback and gives the IQ2 numbers.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 33215a7

[bong-water-water-bong]

Review (standing in for PR-Agent while coder-llm is paused)

• Pin: d1747cb..d5048ad is exactly fork context7.json: description under 200 characters (fixes the claim) #50: 6cbab7e, the two review follow-ups 0174b32 and a05c1b6, and the merge. I reviewed context7.json: description under 200 characters (fixes the claim) #50 on the fork: grid staging and the kquant_needs_grid guard, formats 24/25, and the inline barrier condition.
• Registry: only the llama.cpp (hrx) source line moves, as a kernel-only pin change should.
• docs/hrx.md: the numbers match context7.json: claim the Context7 library #49 and context7.json: description under 200 characters (fixes the claim) #50 as measured there. The note that some batched IQ2_XXS shapes are still declined is right to keep.

Good to merge once build passes.