fix(metal): rename test-local fp8_nblk to avoid quant.h collision

[gouravkargwal]

Fixes #940

The test file defined its own static int fp8_nblk while quant.h now provides static inline int64_t fp8_nblk. Same name and parameter list with different return types breaks compilation of the Metal test (make metal-test): "functions that differ only in their return type cannot be overloaded".

Renamed the test-local helper to ref_fp8_nblk. Header API and production behavior are untouched.

Reproduced and verified on Apple M3, 16 GB: make metal-test passes in full.

[gouravkargwal]

Proof — reproduced and verified on Apple M3, 16 GB

Environment:

• MacBook (Apple M3, 16 GB), macOS
• Apple clang 17.0.0, Homebrew libomp
• Commit: this branch (fix/metal-test-fp8-nblk)

Before the fix — exact failure of make metal-test:

tests/../quant.h:483:23: error: functions that differ only in their return type cannot be overloaded
tests/test_backend_metal.mm:99:12: note: previous definition is here

After the fix — full run below, exits 0. Note the FP8 grouped cases (the ones exercising the renamed helper) all pass:

[gouravkargwal]

clang++ -x objective-c++ -std=gnu++17 -fobjc-arc -O3 tests/test_backend_metal.mm backend_metal.mm -framework Metal -framework Foundation -o backend_metal_test
./backend_metal_test
Metal backend kernel tests:
int8 gate/up S=1 nerr=2.36e-06 ok
int4 gate/up S=1 nerr=1.98e-06 ok
int4 down S=1 nerr=1.03e-06 ok
int2 gate/up S=1 nerr=1.72e-06 ok
f32 S=1 nerr=1.77e-06 ok
int8 gate/up S=4 nerr=2.36e-06 ok
int4 gate/up S=7 (odd) nerr=1.98e-06 ok
int4 non-mult-4 dims nerr=3.78e-06 ok
Metal fmt=4 grouped-int4 tests (coli_metal_matmul vs matmul_i4_grouped semantics):
grouped gate/up I=6144(mult64) S=1 worst_rel=1.90e-08 (I=6144 gs=64 ng=96 S=1) ok
grouped down I=2048(mult64) S=1 worst_rel=2.09e-08 (I=2048 gs=64 ng=32 S=1) ok
grouped gate/up I=6144(mult64) S=4 worst_rel=1.83e-08 (I=6144 gs=64 ng=96 S=4) ok
grouped I=200 (non-mult-64) S=2 worst_rel=1.29e-08 (I=200 gs=64 ng=4 S=2) ok
grouped I=201 (odd, non-mult-64) S=3 worst_rel=2.40e-08 (I=201 gs=64 ng=4 S=3) ok
grouped I=64 (degenerate, ng=1) S=1 worst_rel=1.15e-08 (I=64 gs=64 ng=1 S=1) ok
grouped I=40 (<gs, single partial group) S=1 worst_rel=2.53e-08 (I=40 gs=64 ng=1 S=1) ok
grouped I=512(mult64) outlier-heavy S=3 worst_rel=5.89e-08 (I=512 gs=64 ng=8 S=3) ok
grouped I=201(non-mult-64) outlier-heavy S=2 worst_rel=4.91e-08 (I=201 gs=64 ng=4 S=2) ok
Metal fmt=8 native FP8-e4m3 passthrough tests:
fp8 LUT exactness (256/256 codes via GPU kernel) 0/256 mismatches ok
fp8 gate/up-shaped O=2048 I=6144 (spec example) S=1 worst_rel=3.01e-08 (I=6144 O=2048 nblkO=16 nblkI=48 S=1) ok
fp8 down-shaped O=6144 I=2048 S=1 worst_rel=3.63e-08 (I=2048 O=6144 nblkO=48 nblkI=16 S=1) ok
fp8 gate/up-shaped O=2048 I=6144 S=4 worst_rel=4.05e-08 (I=6144 O=2048 nblkO=16 nblkI=48 S=4) ok
fp8 block edges: O,I both non-mult-128 worst_rel=5.95e-08 (I=200 O=130 nblkO=2 nblkI=2 S=2) ok
fp8 block edges: O just over 128, I exact worst_rel=4.65e-08 (I=128 O=129 nblkO=2 nblkI=1 S=1) ok
fp8 block edges: O exact, I just over 128 worst_rel=6.93e-08 (I=129 O=128 nblkO=1 nblkI=2 S=1) ok
fp8 degenerate 1x1 (single sub-block) worst_rel=5.87e-08 (I=1 O=1 nblkO=1 nblkI=1 S=1) ok
fp8 non-square block grid nblkO=3 nblkI=48 (stride audit) worst_rel=3.33e-08 (I=6144 O=384 nblkO=3 nblkI=48 S=3) ok
fp8 GEMM entry explicitly gated off (coli_metal_gemm refuses) rc=0 (expect 0/CPU-fallback) ok
fp8 MB_BUILD/moe_submit entry gated off (shared-expert fmt=8 hazard) rc=0 (expect 0/CPU-fallback) ok
Metal batched moe_block tests:
moe decode nb=8 R=8 nerr=2.43e-06 ok
moe ragged nb=6 R=16 nerr=2.49e-06 ok
Metal fmt=6 (E8/IQ3) moe_block tests:
e8 decode nb=8 R=8 nerr=4.80e-07 ok
e8 ragged nb=6 R=16 nerr=5.12e-07 ok
Metal large-batch gemm test:
gemm S=64 int4 nerr=2.85e-06 ok
gemm S=64 int4-grouped worst_rel=2.42e-08 ok
Metal fused attention tests:
attn S=1 pos=0 nerr=4.04e-06 cache=1.38e-05 ok
attn S=1 pos=37 nerr=4.13e-06 cache=1.48e-05 ok
attn S=4 pos=12 (MTP) nerr=5.37e-06 cache=1.34e-05 ok
attn S=3 pos=0 nerr=5.30e-06 cache=1.31e-05 ok
Metal fused attention tests (fmt=4 grouped q_a, proves bind_gemv gs plumbing):
attn grouped-qa S=1 pos=37 nerr=5.53e-06 cache=7.87e-06 qraw=1.94e-08 ok
attn grouped-qa S=4 pos=12 (MTP) nerr=5.00e-06 cache=1.38e-05 qraw=2.02e-08 ok
Metal negative control: fmt 1/2/3 unaffected by the fmt=4 shader branch --
see the runs above (int8/int4/int2/f32/moe/gemm/attn cases): all still ok.
Metal top-8 select serial-vs-parallel tests (exact-match contract, E=256):
top8 generic S=1 ok (serial==parallel bitwise, S=1 E=256)
top8 generic S=4 ok (serial==parallel bitwise, S=4 E=256)
top8 ALL-EQUAL ties ok (serial==parallel bitwise, S=1 E=256)
top8 massed dup ties S=4 ok (serial==parallel bitwise, S=4 E=256)
top8 3-level ties rscale ok (serial==parallel bitwise, S=2 E=256)
top8 denormal logits ok (serial==parallel bitwise, S=1 E=256)
top8 topp=0.01 (Ke=1 edge) ok (serial==parallel bitwise, S=1 E=256)
top8 topp=0.6 tied weights ok (serial==parallel bitwise, S=1 E=256)
top8 topp=0.999 S=4 ok (serial==parallel bitwise, S=4 E=256)
top8 topp on ALL-EQUAL ok (serial==parallel bitwise, S=2 E=256)
Metal top-8 select expert-count-generality tests (E!=256, REAP/#428 motivated):
top8 E=168 (REAP) generic S=1 ok (serial==parallel bitwise, S=1 E=168)
top8 E=168 (REAP) massed dup ties S=4 ok (serial==parallel bitwise, S=4 E=168)
top8 E=24 (<32 lane width) generic ok (serial==parallel bitwise, S=1 E=24)
top8 E=24 (<32 lane width) ALL-EQUAL ties ok (serial==parallel bitwise, S=1 E=24)
top8 E=200 (lane straddles E boundary) ok (serial==parallel bitwise, S=4 E=200)
top8 E=257 (>256, auto-serial-fallback) ok (serial==parallel bitwise, S=1 E=257)
metal backend tests: ok
exit: 0

[JustVugg]

Closing, with appreciation rather than rejection: this exact fix — same rename, same ref_fp8_nblk name — landed on dev six days ago as c7aaaf7, addressing what became #940. You based this on the released code, which doesn't have it yet; it ships in v1.6.0, which is imminent (#943). Independent convergence on the identical solution is the best possible confirmation the fix is right.

Your #947 is the more valuable half anyway — the reason this bug ever shipped is that no CI ran metal-test, and that PR closes the hole itself. Review over there.