Hi! I really enjoyed your YouTube video on the MSI XpertStation GB300! Great content; thankyou!
I got an Exxact model a week or two ago; some links re DSF4.1 on the Station; I’m trying to learn how to make this stuff run fast; I’m a novice at it but AI helps! Expert pinning should give you almost 2x speed! al-engr dot com has links.
Since I can’t post links, some info on expert pinning:
CONFIRMED · V15 IS THE REFERENCE · SEPTEMBER 17, 5:55 PM It works, it is bit-identical, it is confirmed on a second same-window pair, and it is the lane’s reference. Six hours after this post went up as a design note, the pin build booted on the Station with every MoE layer’s 295 hottest experts in HBM and 89 coldest in Grace, running the MoE as two do_finalize=False routed-kernel calls plus one fp32-FMA finalize. Greedy output matched v14 token for token on the agent fixture (8/8, with acceptance rates identical to the fourth decimal on four of five categories — the same model, the same bits). Same-window pair against v14: single-stream prose 89.3 → 153.1 tok/s (+71%), eight streams 308 → 701, sixteen streams 414 → 809, shell 160 → 250, code 154 → 214, structured 123 → 208. KV 4.75 GiB (was 4.89), 62.3 GiB pinned, 21 new autotune profiles in 12 minutes. James promoted it as v15 “Pin Hot Experts”; the recipe update is in flight. Two things this post said this morning were wrong and are corrected this is not a C1–C4-only lever, and the bar was never fp-reorder tolerance. The confirm run (v15 → v14 → v15, two knee runs per window) put single-stream at 153.3 tok/s, four runs within 0.4 of each other, +72.7% over v14’s 88.7; sixteen streams at 952–958 on six of seven reps (one 667 outlier) against v14’s 386–417. Fifty-prompt greedy parity: every non-tool prompt exact on every pair; the round-7 gauntlet 36/36; held-out math prose +41%; held-out Japanese/German a wash at −1.5%. Evening close: on-line adaptive placement is parked (every in-graph router counter costs at least 1.7%), and re-profiling the map on mixed traffic recovered Japanese/German +64% while costing the agent lane 10% — the map is per-workload, and v15 stays the reference.