fix(metal): register PR #2166 kernels in runtime-compile path

[ljchang]

Fixes #2168 — MISTRALRS_METAL_PRECOMPILE=0 builds (introduced by #1518) regressed in two places after recent kernel work.

Commit 1: register PR #2166's new kernels in the runtime-compile path

PR #2166 added three new top-level kernel files (rmsnorm_residual.metal, topk_logits.metal, flash_attn.metal) and registered them in mistralrs-quant/build.rs's METAL_SOURCES for the precompiled metallib path, but did not register them in the runtime-compile path's file_system map and main_files list in mistralrs-quant/src/metal_kernels/mod.rs.

When precompile is disabled (e.g. on Apple Silicon with Metal 4 where the precompiled metallib lacks function variants for the current GPU), runtime falls back to MTLDevice.newLibraryWithSource via Kernels::compile_kernels_at_runtime. Without the new files in both lists, the runtime library is built without those kernels and library.get_function("rmsnorm_residual_bf16", ...) fails on first inference:

Error while loading function: rmsnorm_residual_bf16

13 lines: 3 entries each to file_system and main_files. Each file has clean header dependencies (only <metal_stdlib> and <metal_simdgroup*>, which main_source's preamble already provides).

Commit 2: typedef bfloat16_t for runtime-compiled GDN kernels

mistralrs-core/src/metal/kernels/gdn.metal instantiates templates with bfloat16_t, the ggml/llama.cpp type-name convention inherited from the CUDA port. Apple Metal's stdlib doesn't expose bfloat16_t: it has bfloat (Metal 3.1+, macOS 14+) as the public type, and bfloat16 as a forward-only struct (__Reserved_Name__Do_not_use_bfloat16) in metal_extended_vector.

The precompiled path resolves bfloat16_t via build-time headers, but the runtime-compiled path bails with:

error: unknown type name 'bfloat16_t'; did you mean 'bfloat16'?
instantiate_conv1d_update(bfloat16_t);

Aliasing to the forward-only bfloat16 fails differently (incomplete-type errors at template instantiation). The fix is a one-line typedef bfloat bfloat16_t; at the top of gdn.metal so both compile paths see a fully-defined type. This unblocks Qwen 3.5 / 3.6 hybrid models on Apple Silicon under precompile=0; Gemma 4 and other non-hybrid models were unaffected.

This fix overlaps with #2047 (fix(metal): GDN bfloat16, PA scheduler, error handling, MLX SDPA fixes) which has been open since 2026-04-02 and contains the same typedef plus several unrelated improvements. Happy to drop this commit if #2047 is prioritized.

Validation

• cargo check --features metal on mistralrs-quant with MISTRALRS_METAL_PRECOMPILE=0 — clean.
• End-to-end: Qwen 3.6 27B AFQ4 ISQ inference on Apple Silicon / macOS 26 — both commits required for the model to load and produce tokens. Before commit 1: rmsnorm_residual_bf16 missing on first inference. After commit 1, before commit 2: GDN kernel compile fails with bfloat16_t unknown. After both: clean inference, PR feat(metal): optimize Gemma 4 prefill and decode on Apple Silicon #2166's decode perf wins intact.

Out of scope (separate work)

• f8q8.metal (Add new quant method: F8Q8 #1883) and hqq_bitpack.metal (Add blockwise fp8 quantize kernels #1586) are also in METAL_SOURCES but missing from the runtime path. They predate the regression; leaving them out to keep this PR surgical.
• Three hand-maintained mirrors of the same kernel-file list (METAL_SOURCES in build.rs + file_system + main_files in mod.rs) is the design smell that made this regression possible. Single-sourcing (e.g. build.rs emits a generated kernels.rs that mod.rs include!s) is a worthwhile follow-up refactor.

Suggestion

Adding MISTRALRS_METAL_PRECOMPILE=0 to the CI matrix would catch this class of regression before it ships. Three precompile=0 incompatibilities have surfaced in the last six weeks (the two in this PR plus an f8q8/hqq_bitpack gap that affects pre-existing kernels).

[github-actions]

Code Metrics Report

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Language              Files        Lines         Code     Comments       Blanks
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 C Header                  6          482          253          175           54
 CSS                       3          281          252            5           24
 CUDA                     63        22172        18136         1531         2505
 Dockerfile                1           38           21            8            9
 HTML                      2           27           27            0            0
 JavaScript                3          392          387            2            3
 Jinja2                    7          694          656            5           33
 JSON                     26         9360         9357            0            3
 Makefile                  1            6            5            0            1
 MDX                       1          149            0          133           16
 Metal Shading Lan|       37        14287        11284         1136         1867
 PowerShell                1          357          276           33           48
 Python                  131        10342         8515          460         1367
 Shell                     2          549          379          101           69
 Plain Text                3         3723            0         2413         1310
 TOML                     28         1359         1185           43          131
 TypeScript               11         1607         1371           66          170
 YAML                      3           25           23            2            0
─────────────────────────────────────────────────────────────────────────────────
 Jupyter Notebooks         3          122           83           23           16
 |- Markdown               1           60           30           22            8
 |- Python                 1          122          113            1            8
 (Total)                              304          226           46           32
─────────────────────────────────────────────────────────────────────────────────
 Markdown                127         9600            0         6575         3025
 |- BASH                  61          600          520           47           33
 |- Dockerfile             2            5            5            0            0
 |- JSON                  18          700          700            0            0
 |- PowerShell             3            5            5            0            0
 |- Python                25          830          722            5          103
 |- Rust                  15          437          382            1           54
 |- TOML                  10          124           98            3           23
 |- YAML                   1            9            9            0            0
 (Total)                            12310         2441         6631         3238
─────────────────────────────────────────────────────────────────────────────────
 Rust                    599       265776       235112         6354        24310
 |- Markdown             392         9302          452         7728         1122
 (Total)                           275078       235564        14082        25432
─────────────────────────────────────────────────────────────────────────────────
 Svelte                   18         1831         1696           50           85
 |- CSS                    1            4            4            0            0
 |- JavaScript            18          876          727           24          125
 (Total)                             2711         2427           74          210
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Total                  1076       356253       292785        26946        36522
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[EricLBuehler]

Hey @ljchang! Thanks for the PR.

The PR looks good broadly but I added a few very small comments.

[EricLBuehler]

This is the source of the merge conflict I think, if you remove this change (I made a fix on master), it should be resolved.

[ljchang]

Dropped that commit entirely and rebased onto master — your typedef bfloat bfloat16_t at gdn.metal:8 covers it, so the conflict is gone. The PR is now mod.rs-only.

[EricLBuehler]

Let's not reference specific PRs here, ideally this comment can be removed.

[ljchang]

Removed.

[EricLBuehler]

Same here, ideally just describe the kernel and not refernece when it was added.

[ljchang]

Done — the three entries now just describe the kernel (Flash attention DK=512 variants / Fused RMSNorm + residual / Two-stage top-k + softmax stats).

[EricLBuehler]

Thank you!