GLM 5.3 Flash: Metal resident-Q2 prefill −6–8% from 0e9cc2d (attention masking fix), measured on M5 Max

[nomishbhardwaj]

Setup: MacBook Pro M5 Max 128 GB, GLM-5.3-Flash-Q2.gguf (resident, --warm-weights),
ds4-server --ctx 393216 --vision GLM-5.3-Flash-Vision-Encoder.gguf --kv-disk-dir ... --kv-cache-cold-max-tokens 393216,
server chat path, cold prefill on unique text (KV purged per block), server-reported avg=.

Observation: updating from 110afdd to 6289c51 cost GLM cold prefill consistently at every depth,
while decode stayed flat. Interleaved A/B (3 rounds per condition, per-block checkout+binary+KV purge,
thermal settle + cooldowns; first round discarded as cold):

depth	110afdd	6289c51	delta
8K cold	411 t/s	381 t/s	−7.3%
16K cold	385 t/s	362 t/s	−6.0%
24K cold	372 t/s	341 t/s	−8.4%
2K delta on warm 14K	313 t/s	278 t/s	−11%
decode 0–24K	24.2–31.6 t/s	24.2–30.4 t/s	flat/noise

Attribution: direct two-commit differential, 0e9cc2d^ (=9d9e129) vs 0e9cc2d
("Fix GLM attention masking and reductions across GPU backends"), 2 interleaved rounds each,
same harness:

depth	9d9e129	0e9cc2d	delta
8K cold	415 t/s	382 t/s	−8.0%
16K cold	386 t/s	359 t/s	−7.0%
24K cold	376 t/s	355 t/s	−5.5%

That accounts for essentially the whole update delta. The commit message records the Strix
continued-prefill improvement (21→68 t/s) and unchanged official scores, but not a Metal cost,
so this may be an unnoticed side effect of the added reduction-storage synchronization /
padded-selection handling on Apple GPUs rather than an accepted trade-off.

Happy to run patches or diagnostics on this M5 Max config. (Same machine/config as the
measurements in #897/#898.)