[cleanup] Gate profiling prints behind debug flag (#44) #54

Merged

sleepy merged 1 commit from refactor/44-remove-profiling-prints into main

2026-05-20 20:10:43 +02:00

sleepy commented

2026-05-20 20:02:01 +02:00

Owner

Summary

Gates decode_profile and argmax_compare prints (and their macOS-specific mach_absolute_time / mach_timebase_info calls) behind @import("builtin").mode == .Debug.

Preserves the CLI "Decode: X tokens in Ys = Z tok/s" timing in engine.zig
Removes profiling noise from release builds

Test Results

zig build compiles cleanly

Benchmarks

No performance impact expected (timing calls removed from release builds).

## Summary Gates `decode_profile` and `argmax_compare` prints (and their macOS-specific `mach_absolute_time` / `mach_timebase_info` calls) behind `@import("builtin").mode == .Debug`. - Preserves the CLI "Decode: X tokens in Ys = Z tok/s" timing in `engine.zig` - Removes profiling noise from release builds ## Test Results - `zig build` compiles cleanly ## Benchmarks No performance impact expected (timing calls removed from release builds).

sleepy added 1 commit

2026-05-20 20:02:02 +02:00

refactor(#44 ): gate profiling prints behind debug flag, keep CLI decode timing 103f6d5c2a

sleepy commented

2026-05-20 20:04:29 +02:00

Author

Owner

CHANGES_REQUESTED

Dead GPU work in release builds: dispatch.set_argmax_bf16() is dispatched unconditionally, but the GPU argmax result is never read in non-debug builds. Move the dispatch inside the if (Debug) block to avoid wasting GPU cycles per token in release mode.
Unnecessary code duplication: The copy_bf16_buffer_to_f32 + CPU argmax logic is duplicated in both branches. Restructure so the shared readback and CPU argmax happen once after the conditional debug timing, rather than duplicating ~20 lines of identical logic.

Suggested: gate only the mach_absolute_time calls and std.debug.print lines behind Debug, keep the shared readback/argmax single-path.

**CHANGES_REQUESTED** 1. **Dead GPU work in release builds**: `dispatch.set_argmax_bf16()` is dispatched unconditionally, but the GPU argmax result is never read in non-debug builds. Move the dispatch inside the `if (Debug)` block to avoid wasting GPU cycles per token in release mode. 2. **Unnecessary code duplication**: The `copy_bf16_buffer_to_f32` + CPU argmax logic is duplicated in both branches. Restructure so the shared readback and CPU argmax happen once after the conditional debug timing, rather than duplicating ~20 lines of identical logic. Suggested: gate only the `mach_absolute_time` calls and `std.debug.print` lines behind `Debug`, keep the shared readback/argmax single-path.

sleepy force-pushed refactor/44-remove-profiling-prints from 103f6d5c2a to e11677ecdc

2026-05-20 20:09:54 +02:00

Compare

sleepy merged commit 953928c4d4 into main

2026-05-20 20:10:43 +02:00

sleepy referenced this pull request from a commit

2026-05-20 20:10:43 +02:00

[cleanup] Gate profiling prints behind debug flag (#44) (#54)

sleepy deleted branch refactor/44-remove-profiling-prints

2026-05-20 20:10:43 +02:00