Repin ggml-org/llama.cpp#25731 (TML Inkling) at 946fc11d1 after the upstream parsers and ggml_prec refactors

[danielhanchen]

Why

The nightly on b10865 (run 34287079907) stopped in resolve: pin 2, ggml-org#25731 @ 1066edc3a, no longer merged. Two upstream changes landed between the PR's base (Sep 7) and b10865:

• chat : split specialized parsers into common/parsers ggml-org/llama.cpp#27764 (895c045fd) moved every model-specific common_chat_params_init_* out of common/chat.cpp into common/parsers/<model>.cpp. The PR inserted init_inkling into chat.cpp at the spot the refactor emptied, which is an edit/delete conflict the additive resolver refuses by design.
• ggml : update ggml_prec specification ggml-org/llama.cpp#26675 (5a6caa05f) renumbered enum ggml_prec into a ranked scale, added ggml_prec_set_acc, and deprecated ggml_mul_mat_set_prec / ggml_flash_attn_ext_set_prec. The PR adds GGML_PREC_F32_PEDANTIC and declares ggml_flash_attn_ext_banded on the same lines.

What changed on the PR branch (danielhanchen/llama.cpp:add-inkling, 1066edc -> 946fc11)

1. d3a67005d Merge upstream master. common_chat_params_init_inkling moves to common/parsers/inkling.cpp (body unchanged, static dropped), declared in parsers.h, listed in sources.cmake; chat.cpp keeps only the 7-line detection branch. GGML_PREC_F32_PEDANTIC joins the new ranked enum at 5, below GGML_PREC_F32, because it is the stricter contract (lower rank = stricter in the new scheme; 11 would have read as looser than F32). The ggml_flash_attn_ext_banded declaration stays next to the flash attention API.
2. 946fc11d1 Every deprecated setter call becomes ggml_prec_set_acc (the macOS leg builds with LLAMA_FATAL_WARNINGS=ON and no deprecation opt-out, so this is required, not cosmetic), ggml_prec_set_acc accepts GGML_OP_FLASH_ATTN_EXT_BANDED, and the backends that compared the accumulator slot with == GGML_PREC_F32 (CUDA cuBLAS compute type, CPU and spacemit flash attention, Vulkan flash attention) compare by rank so a pedantic request never falls to a lower-precision path. The CUDA mul_mat_id slice carries the acc and src precision slots, and ggml_cuda_mul_mat_id_needs_sync gets the same pedantic gate as ggml_cuda_mul_mat_id (a pedantic mmf-eligible F32 expert matmul would otherwise trip its assertion).

PR delta against its base: 64 files / +4093 -124 before, 70 files / +4120 -126 after. The six extra files are the three parser registrations plus the three backend comparison fixes.

Verification

Local, on the rebased branch:

gate	result
CPU build, LLAMA_FATAL_WARNINGS=ON, tests on	clean
test-chat (8 Inkling cases among them), test-chat-peg-parser, test-chat-template	pass
test-backend-ops test -b CUDA0 -o FLASH_ATTN_EXT_BANDED	13/13
test-backend-ops test -b CUDA0 -o MUL_MAT / MUL_MAT_ID / FLASH_ATTN_EXT	1288/1288, 880/880, 2959/2959
test-flash-attn-bias, test-flash-attn-generic-hash	pass
test-llama-archs -a inkling -s 1234	OK
test-mtmd-impl test_projector_registry	pass
unsloth/Inkling-Small-GGUF UD-IQ1_S, two prompts, temp 0, 48 tokens, CUDA and CPU, pinned build vs rebased build	byte-identical text

Mix, all 13 pins in pr-set.json order with the new pin, reproduced with the resolve job's own git commands:

base	result
b10865	merges; additive add/add only on llama-arch.{cpp,h}, llama-model-saver.cpp, test-llama-archs.cpp (ggml-org#25731) and tools/mtmd/CMakeLists.txt (#70); merge_checks OK; pin_contract all 13 intact
b10870	same

Script tests (test_additive_merge, test_pin_merge, test_merge_checks, test_sync_deletes, test_carry_vintage, test_pin_contract, test_feature_matrix) pass; every pin is covered by feature-checks.json.

[chatgpt-codex-connector]

You have reached your Codex usage limits for security reviews. Please try again later.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-09T07:00:55.261615Z	435bb22	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[danielhanchen]

Feature matrix on the 13-pin mix (base b10870, CUDA build, feature_matrix.py --gpu):

ok   diffusion-gemma: 3/3 device rows decoded and matched CPU
ok   glm5next: 2/3 device rows decoded and matched CPU; 144/144 cases matched the CPU reference
ok   inkling: 2/3 device rows decoded and matched CPU; 13/13 cases matched the CPU reference; projector registry intact over 194 assertions
ok   iq1-narrow-grids: 13/13 cases matched the CPU reference (x3)
ok   kimi-k3: 2/3 device rows decoded and matched CPU; projector registry intact over 194 assertions
ok   projector-registry: projector registry intact over 194 assertions
ok   qwen4exp-mtp: 2/3 device rows decoded and matched CPU; 4/4 cases matched the CPU reference
all 7 features demonstrated on GPU