What's Changed
- engine_view_test: the 70 KiB body is refused from its headers alone, raw socket with no body; 30 of 30 pass (#1162) by @alish-nace-23 in #1163
- engine_view_test: the viewer-cap check retries a 503 from a leftover stream; 1 failure in 300 runs under load before (#1162) by @alish-nace-23 in #1164
- Deduplicate shared grammar alternatives (Not verified on a GPU) by @chainhelen in #1167
- Metal: cut sorted-list sampling stage 2 GPU time by 49.8–64.3% by @chainhelen in #1165
- Flash-Next: batch index-cache growth copies by @chainhelen in #1160
- Fix stale chat-template content modes after reload (Not verified on a GPU) by @chainhelen in #1168
- Engine view: RAM prefix cache and SSD cache in status.engine and the Memory card (#457) by @alish-nace-23 in #1172
- Engine view: GET /models and a picker of the downloaded models for Switch (#458) by @alish-nace-23 in #1177
- KFlash: the budget cache loads only into the build that wrote it (#1150) by @alish-nace-23 in #1174
- Engine view: the log shows what the server logged before the page opened, with a warnings filter (page only, #455) by @alish-nace-23 in #1175
- tests: accept equivalent Q4_K/Q5_K NaNs (Not verified on a GPU) by @chainhelen in #1178
- Engine view: a finished request opens to its waited, read and wrote times and its error text (page only, #456) by @alish-nace-23 in #1176
- Flint: batch KV cache growth copies by @chainhelen in #1171
- KQuant: copy that sells rigor and results, without the build steps by @alish-nace-23 in #1181
- Engine view: a request refused before its prompt was read says so, not "read 12,010 tokens" (page only) by @alish-nace-23 in #1182
- Metal: compile 125 fewer duplicate entry points in the weight-format library by @chainhelen in #1180
Full Changelog: v0.2.34...v0.2.35-rc1