DEV Community: futurexindiafx-gif

Beyond seeds: which determinism flags actually mattered for reproducible fine-tunes on a T4?

futurexindiafx-gif — Mon, 26 Oct 2026 08:03:52 +0000

I set torch, random, and numpy seeds, and my small fine-tune still moves run to run on a T4. Same recipe, same data — the weights at the same step differ between runs. For those of you who actually got reproducible fine-tunes: which flags mattered in practice — math attention backend only, a cuBLAS workspace config, cudnn.deterministic plus benchmark=False, or torch.use_deterministic_algorithms as a diagnostic? And what slowdown did you accept for it? I'm trying to find out which of these is the one that actually closes the gap, rather than turning them all on and hoping.

My renderer's seed argument changes nothing — how do I prove the seed actually reaches the sampler?

futurexindiafx-gif — Mon, 26 Oct 2026 07:55:46 +0000

I'm working with a diffusion-based renderer whose inference entry accepts a seed argument. I render the same input twice with two different seeds, and the two outputs come out byte-identical — I confirmed it with a per-frame hash comparison, not just by eye. So either the path is genuinely deterministic, or the seed is accepted and then never consumed by the sampler. How do practitioners verify this? Is there a standard probe — for example, checking which RNG stream the sampler actually draws from, or a minimal seed-sensitivity test where you change one thing and expect the output to move? I want to know whether my seed path is real before I assume the renderer is deterministic. What has worked for you to tell these two cases apart?

Same fine-tune recipe, different seeds — one narrow behaviour swings wildly. What actually stabilises it?

futurexindiafx-gif — Mon, 26 Oct 2026 07:41:21 +0000

I'm fine-tuning a small adapter for a talking-face model — a short recipe, same data throughout. The only thing I change between runs is the seed. The overall score barely moves, but one narrow behaviour swings: on one seed it collapses, on another it lands fine, on a third it overshoots.

Has anyone dealt with this pattern? What actually helped in your experience — seed averaging, EMA / checkpoint soup, a longer schedule, or judging the behaviour with its own per-behaviour metric instead of the aggregate score?