Cold greedy output occasionally differs between server starts under the new decode numerics (upstream #631/#585), think prompt only #83

Symptom: with decode numerics that differ from the old ones, the cold greedy output of ~/models/bench/greedy_ident.py's think prompt (thinking on, 256 tokens) occasionally differs between server starts of the same build. The code and ctx prompts never did.

What it is not:

• perf(kernel): avoid layout conversions in nvfp4 marlin decode FlashML-org/FreeToken#631's kernel: repeated calls are bit-identical, and its wide (16-byte aligned) and narrow activation loads give the same bits (~/models/bench/integ4/check/marlin_det.py).
• Run order: the outlier run in the 1639600 set came right after another build's arm. Three back-to-back restarts of the same build all gave the majority fingerprint.

Hypothesis (unverified): some start-dependent but small numeric choice elsewhere, for example Triton autotune picking among near-equal configs at startup (FLA GDN kernels autotune). It only changes an argmax where the new numerics put the think output on a near-tie.

Why it matters: the lossless checks for scheduling and prefetch changes compare cold greedy output across restarts. Under these numerics, a rare flip would look like a regression.

Next step:

1. Log the autotune choices at startup (TRITON_PRINT_AUTOTUNING=1) for several starts, and see whether they differ when the fingerprint does.
2. If they do, pin or cache the configs.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Bu6LgoxLR4wETqb7RPR2vt