[ashhart]

On Metal, DeepSeek-V4.1 picks a row's token on the host (metal/target.zig, choose). A greedy pick scanned all 129,280 logits one at a time. A sampled pick allocated a float64 copy of the logits and an id list, about 2 MB, on every row before calling the lane core's sampling.choose. A drafted round picks one row per window row, so this adds up.

• Greedy: finds the first maximum in 16-wide vector steps, then the first index holding it. A row containing a NaN takes the old scalar scan, so ties and NaNs give exactly the id they gave before. A test compares both over 200 random rows of the release's vocabulary with ties, NaNs and all -inf rows planted.
• Sampled: reuses a float64 buffer and a constant id list kept in the target, then calls the same sampling.choose on the same values, so the draw is unchanged.

On an M5 Ultra (512 GB), release weights, with #2973, #2983 and the FP8 rows change underneath, on the five prose prompts of tools/bench_prompts.json at 512 tokens, one run each:

	Before	After
Greedy, mean tok/s	63.0	64.7
Sampled (T 1.0, seed 1234), mean tok/s	53.7	55.7

Every reply's text equals main's (10 of 10). zig build test-dsv41-metal -Dcpu=apple_m1 passes on an M5 Max, and lean reports no new problems.

• Synthena (Ash's Agent)