[aittalam]

Description

llama.cpp/BUILD.mk used the -DNDEBUG flag with llama and common targets, but not with the ggml
objects, so plain assert() stayed live in the CPU hot path (upstream's Release build, instead, compiles them out).

-$(LLAMA_OBJS) $(COMMON_OBJS): private CCFLAGS += -DNDEBUG
+$(LLAMA_CPP_OBJS): private CCFLAGS += -DNDEBUG

LLAMA_CPP_OBJS is the existing umbrella variable (GGML + LLAMA + COMMON + GGUF).

What was running: 25 #ifndef NDEBUG blocks in ops.cpp that re-read each FFN
activation output row checking isnan/isinf (swiglu, silu, gelu, geglu…), plus
per-row asserts in the quant dot-products.

Measured (token generation, interleaved A/B, medians):

path	gain	why
CPU x86 (9B / 27B)	+1.25% / +1.19%	ggml built here — was missing the flag
CPU arm64 (9B)	+1.00%	same
CUDA	+0.34%	kernels already had it; only shared host code gained
Metal	0.00%	metal.c rebuilds ggml at runtime, already with -DNDEBUG

Binary shrinks 77–94 KB. make check passes.

Trade-off: two GGML_LOG_DEBUG diagnostics go quiet (ggml-backend.cpp:496
"slow copy", :899 "no backend supports op"), both useful when debugging offload.
Same as upstream Release.

Not covered: whisper.cpp and transcribe.cpp build their own ggml copies and
still lack the flag — separate PR. stable-diffusion.cpp reuses $(GGML_OBJS) and
benefits for free.

PR Type

• 💅 Refactor

Relevant issues

#980 - see this in particular, TY @dexhunter !

Checklist

• I understand the code I am submitting.
• I have run this code locally and verified the change.
• New and existing tests pass locally, or I have explained why tests were not run.
• I have read and followed the contribution guidelines.
• AI Usage:
  • No AI was used.
  • AI was used in an assistive capacity (to run tests for the 1-line manual change)
  • This PR includes substantial AI-generated content.