clm-0010

Source: a detailed Strix Halo write-up (HP ZBook, Ryzen AI MAX+ PRO 395, 128 GB, Windows, Vulkan then HIP) on a Qwen3.6-35B-A3B finetune in ROCmFP4 via the ciru-ai/ROCmFPX fork. MECHANISM: the speculator matches n-grams against text already in the context window, so the draft costs nothing to produce and only verify rows are paid. Reported 86% acceptance on 48-64 token copied drafts, and on fresh content the gate simply stays closed with no penalty. That asymmetry is what makes it different from a draft model. WHY IT MATTERS TO US MORE THAN MOST: our heaviest repetitive workload is exactly their best case — Claude Code doing code edits, re-emitting large chunks of a file. And the shared-pool effect (acceptance RISING with stream count, 93% to 97%) is a genuine argument for the council pattern in clm-0008, where several agents work adjacent problems. ⚠ MTP IS THROUGHPUT-NEGATIVE AT BATCH per the same source: roughly breakeven at 2 streams, -13% at 4, because every accepted draft still costs a verify row through the MoE and the expert-union already saturates the bus. We run -np 1, so we are in MTP's favourable regime — but this is a further argument against multi-slot, on top of llama.cpp #25992 and the memory cost. CAVEATS: single reporter; the 4-stream figures used IDENTICAL prompts across streams, which is maximum pool sharing and flattering (they say so). Requires the ROCmFPX fork, which would be a third carried runtime (see clm-0009). Their own updates report ngram fighting flash attention at larger contexts, and crashing when combined with MTP. TEST: ngram-mod vs MTP vs neither, on OUR code-edit workload, single stream. It is the cheapest large win available if it reproduces.