Stage M2 — Profile one decode token (old 15)
Goal: account for where one Metal decode token spends its time,
with evidence, before any fusion.
- Instruments Metal System Trace + the existing
ZYNFER_SIGNPOSTS=1
machinery, extended to per-op-family signposts on the Qwen path.
- Produce the master prompt's per-operation table (RMSNorm, QKV, RoPE,
attention, O-proj, MLP, LM head, sampling, encode/wait overhead).
- Compute and record the machine's sustainable memory bandwidth
(measured, not the marketing number) and the f32 bytes-per-token for
Qwen3-0.6B; state the roofline tok/s and the current fraction of it.
Gate: the question "where does one decode token go?" has a
documented, evidence-backed answer that names the top three costs.
Tutorial: docs/tutorials/16-profiling-a-token-on-apple.md.
Stage M2 — Profile one decode token (old 15)
Goal: account for where one Metal decode token spends its time,
with evidence, before any fusion.
ZYNFER_SIGNPOSTS=1machinery, extended to per-op-family signposts on the Qwen path.
attention, O-proj, MLP, LM head, sampling, encode/wait overhead).
(measured, not the marketing number) and the f32 bytes-per-token for
Qwen3-0.6B; state the roofline tok/s and the current fraction of it.
Gate: the question "where does one decode token go?" has a
documented, evidence-backed answer that names the top three costs.
Tutorial:
docs/tutorials/16-profiling-a-token-on-apple.md.