known-issues.md

Known issues

This is the current list. Each release's notes summarize it and say what changed; an issue leaves the list only when a release fixes it, and the release notes say which one.

Models

• GLM-5.3 (EXLR8) decodes slowly next to the other models: a one-row round on two M5 Ultras takes about 45 ms. No CUDA engine serves it, and only the home profile on M5 Ultras has run. Replies past 2,048 tokens, where the sparse indexer starts picking keys, have not been compared with a whole-model reference.
• Qwen3.5-2B with a JSON schema: on an integer field it wants to write as a decimal, the 2B can keep adding digits until max_tokens, because the grammar allows no decimal point there. Nemotron writes the same schema correctly.
• Qwen3.8-27B on an M1 to M4 Mac: the drafter's landing counts last for the server's life, so after four prose replies greedy code drafts shallower, about 4% slower on an M3 Ultra. The tokens are the same.
• Qwen3.8-27B on Metal allocates DeltaNet rollback scratch lazily for the verification slots that need it. The bank can be resized or released between completed rounds; an idle or newly lone host can release it after the previous verification has been kept. A revised physical-memory peak comparison with 1.0.12 remains unmeasured.
• Qwen3.8-27B cold prefill at 64K tokens on RTX 50-series GPUs (sm_120) measures about 2% slower than 1.1.0; 32K and shorter prompts, and GB10, are at parity or faster.
• Flash Next on Metal with explicit --parallel 2 or more retains a cache-count/performance issue: on an M3 Ultra with the MLX 4-bit checkpoint, 9 of 12 chat steps reused 17 to 312 fewer tokens than the single host, with the same output tokens, and a lone request decoded 7% slower. The default --parallel auto uses the single host and matched 1.0.12's cache counts; the separate shared-session repair is held for a later 1.1.x and is absent from the current 1.1.1 series.
• The 1.0.12 gfx1151 Qwen receipts leave the 27B cancellation failure open and the 9B/27B drafted/plain comparisons unrun. Those receipts do not establish a new 1.1.0 device pass.
• Qwen3.8-Flash-Next in speed-up mode: a drafted request right after a "draft": false one decodes slowly, and the next request is back to speed. The tokens are the same.
• Qwen3.8-Flash-Next on a Mac: a turn more than --keep-warm seconds after the last request (900 by default) takes about 2 s to its first token on an M3 Ultra, even when its prompt is cached, while macOS wires the model again.
• Qwen3.8-Flash-Next on one GB10 reads its n-gram embedding tables (29.8 GiB in the MLX 4-bit MTP checkpoint) from the checkpoint files through the page cache, so they compete with the server's own memory. With TF_FLASHNEXT_CHUNK=16384 the larger prompt scratch leaves too little room for all of them: in our runs, time to first token at 65,536 tokens ranged from 32.9 s to 49.1 s, about 1.5 times, with heavy paging in the slow runs. With the default 1,536-token chunks and the prompt cache off, three runs differed by at most 2.2%; saved prompt states (up to 16 GiB by default) use the same memory and can push table rows out at that chunk size too.
• On a Mac without tensor units (M1 to M4), Flash Next's uniform 6-bit checkpoint decodes different tokens from 1.0.9 since 1.0.10's faster products. FZ_M3_LANES=1 keeps the earlier kernels and their tokens.
• @Bizuayeu's regression run on shipped 1.1.0 loaded the whole GLM-5.3-Flash NVFP4 checkpoint on two GB10s at TP=2 in 302.1 s, kept one token hash per prompt across three runs and found a passphrase in 199,307 tokens; scored prompts and more than two ranks were not covered (#1040).
• DeepSeek-V4.1 CUDA pieces land in this release, but no release binary serves DeepSeek checkpoints until its whole-model gate (#827) passes.

Living Weights (--slide)

• Living Weights implementation and safety work is in the 1.1.1 tree, including CUDA learners, shared retention controls, lesson progress and save/restart work. Model, retention and save/restart qualification remains pending. The #1029 lesson-counter repair is excluded and remains held: rejected lessons must not shift the prompts used for later training and checks.

• Facts about the same subject can overwrite each other, a later fact about the user can disturb an earlier one's answers on a GB10, yes-or-no questions about a changed value are unreliable, facts pile up badly past a few dozen, and thinking on brings back fewer answers. 1.0.8's release notes have the measurements.

• Learned facts cost decode speed on a Mac. 1.0.9 took back about 8% on an M3 Ultra with two facts; 1.0.8's M5 Ultra figures (213 tok/s after ten facts, 248 before any) have not been measured again.

• A rollback is all or nothing: tensorfold slide rollback returns the folder to its snapshot, undoing every lesson and delete since the first. One fact can be deleted on its own (DELETE /v1/slide/fact/{id}), which silences its neurons rather than erasing them. A key keeps other clients from teaching, but anyone with the key can teach a wrong fact. Teach a copy of the checkpoint folder.

• On file systems without clones, such as ext4, the snapshot is a full copy of the files a lesson can change, 18.5 GB for Nemotron 3.5 Lightning, and the rollback copies them back.

• The folder hold that keeps a server and a rollback apart works between TensorFold processes only: other tools that write files ignore it.

Web research (--cortheon)

• The instruction filter that removes orders aimed at the model from web text is a heuristic. It reads each line on its own, so an order split across a line break is not seen whole, and a page written to steer the model could still make it put conversation text into the path of a URL it fetches. docs/cortheon.md says what leaves the machine.