IQ1_XS, IQ1_XXS, IQ1_XXXS on an upstream base so the nightly can pin them

[danielhanchen]

Same change as #61, on an upstream base so the nightly can pin it. #61 stays open for review of the design; this is the branch the prebuilds would actually merge.

Why #61 cannot be pinned as it is

resolve merges every pin onto the upstream base tag, not onto our master. Our master only shares history with upstream up to e95dae18d (2026-06-10), because upstream syncs land as squashes, most recently 53fd974b4 (Sync master with upstream, #80). #61 branches off that squash, so merging it onto b10356 drags two months of upstream work in as fresh additions:

merge-base(iq1-narrow, b10356)  = e95dae18d  2026-06-10   765 commits behind
merge-base(master,     b10356)  = e95dae18d  2026-06-10
merge-base(kimi-k3-vision-only, b10356) = 071327508  2026-08-04   86 commits behind

The result is hundreds of conflicts in files the change never touches, tools/ui/, vendor/cpp-httplib/, vendor/sheredom/, all of the "both sides add the same line(s)" shape, and additive_merge.py refuses them, correctly:

refused  tools/ui/vite.config.ts: both sides add the same line(s), so this is one change made twice
refused  vendor/cpp-httplib/httplib.cpp: merge base is not empty, so at least one side edited existing text
refused  vendor/sheredom/subprocess.h: merge base is not empty, so at least one side edited existing text
unslothai/llama.cpp#61 (36fe8e1c) does not merge cleanly onto b10356 + the PRs listed before it

That is a branch-base problem, not a code problem. #70 avoids it by sitting on an upstream commit and targeting kimi-k3-text-upstream, and this branch follows the same pattern: iq1-narrow-upstream is b10356 plus the single commit, targeting iq1-narrow-upstream-base, which is b10356 untouched.

The content is unchanged from #61: 41 files, 2334 insertions, 25 deletions, identical to what #61 reports. Cherry-picking 36fe8e1c onto b10356 applies with zero conflicts, and onto b10356 + 24423 + 25731 + 70 also with zero conflicts.

Checks on the merged nightly tree

Built the full mix (b10356 + 24423 + 25731 + 70 + this), CUDA sm_100, on a B200:

check	result
resolve merge of all four pins	clean, tag b10356-mix-e84bfed
build	completes, no errors
ctest	59 of 60 pass; the 60th is test-backend-ops hitting the 1500s ctest timeout, run separately below
test-backend-ops -o MUL_MAT -b CUDA0	1219/1219, 2/2 backends
test-backend-ops -o MUL_MAT_ID -b CUDA0	874/874, 2/2 backends
test-quant-type-selection	pass
test-llama-archs	pass

The three types register in llama-quantize at the fork-local ids and quantize and generate end to end on Llama-3.2-1B with an imatrix:

  24  or  IQ1_S   :  1.56 bpw quantization      393551840 bytes
  64  or  IQ1_XS  :  1.44 bpw quantization      380182496
  65  or  IQ1_XXS :  1.31 bpw quantization      366813152
  66  or  IQ1_XXXS :  1.19 bpw quantization     353443808

Sizes fall in the order the bpw table predicts. libggml-cpu.so exports ggml_vec_dot_iq1_xs_q8_K, ggml_vec_dot_iq1_xxs_q8_K, ggml_vec_dot_iq1_xxxs_q8_K and the three grids, and the three CUDA MMQ instances compile.

Existing quant types are untouched

Built the identical tree without this commit and requantized the same model from the same imatrix with both binaries. The GGUFs are byte-identical:

IQ1_S    IDENTICAL  02dc2c2c1a63817dfb3998acc272923d8f68ed1ac10d06dc00e87b137c2c4482
IQ2_XXS  IDENTICAL  3a461ee80e793e93ab28b742da8513a9bd2342c5270f403a9398a7871efd657c
Q4_K_M   IDENTICAL  574d6de2b69c35bcb539dd210adee181b8eba1b7ae8864f525ef20b6ef83c2b3
Q8_0     IDENTICAL  05085ca8b9697115aeb2be01ee32873d476dec5b060abaaf2367b394fae82392

DiffusionGemma, Inkling, Kimi-K3 and Muse Glimmer all survive the merge alongside it: every arch registers in llama-arch.cpp, test-llama-archs passes, and the built libraries export their symbols (libllama.so: diffusion_gemma 17, inkling 24, kimi_k3 23, muse_glimmer 17; libmtmd.so: kimivl 7, muse_glimmer 14), with all four llama-diffusion-* binaries linking.

Pinning

Depends on #90, which drops the merged 26841 pin that currently blocks the merge before anything else is reached. Once both are in, the pin url is:

https://github.com/unslothai/llama.cpp/pull/91/commits/c86ed269986f2dced6325c5c58bda966a2e2ead1