Issues

gemma3n: KV-shared layers only reuse cached KV when a cache is present, so no-cache forward (LoRA training, evaluation) diverges from generation

#1674

· binaydhakal opened

on Aug 7, 2026

mlx_lm.server: [metal::malloc] Resource limit (499000) exceeded in BatchGenerator._generate wedges the server invisibly (HTTP layer stays up, generation thread dies) — recurrence of #831/#1185/#1332's signature in continuous batched serving

#1672

· athal7 opened

on Aug 7, 2026

Confusing bare KeyError when LoRA config uses "alpha" instead of "scale"

#1670

· arjunneema09 opened

on Aug 7, 2026

mlx_lm.generate: repetition_penalty has no default (unlike other sampling params), model can loop into repeated text with no warning

#1669

· arjunneema09 opened

on Aug 7, 2026

completeMemory() prepare count underflow kernel panic recurs under continuous batching despite #883's KV-cache-bound mitigations

#1666

· athal7 opened

on Aug 5, 2026

Models that discard `update_and_fetch`'s values return leak one Metal buffer per layer per generated token — deterministic crash at Metal's buffer-count limit (deepseek_v4 dies at ~11.5k generated tokens)

#1662

· aidiffuser opened

on Aug 5, 2026

LoRA trainer has no gradient clipping or non-finite guard: one bad batch loses the whole run

#1658

· axiom-of-choice opened

on Aug 5, 2026

chat prompt emits raw readline markers (\x01/\x02) because readline is never imported

#1653

· Vlor999 opened

on Aug 3, 2026

quantized_kv_start default disagrees between generate_step and the CLI

#1651

· Vlor999 opened

on Aug 3, 2026

load() pays an avoidable Hub round-trip per call even when the model is fully cached

#1649

· Vlor999 opened

on Aug 3, 2026

[metal::malloc] Resource limit (499000) exceeded on qwen3_5_moe (non-MLA), single node — #839 does not cover this path

#1644

· veghadam opened

on Jul 30, 2026

`ArraysCache.advance()` leaks Metal buffer objects in qwen3_5/qwen3_next batch decode, crashing long completions at the resource limit

#1641

· thinkroth opened

on Jul 29, 2026

Navigation Menu

Uh oh!

All issues

gemma3n: KV-shared layers only reuse cached KV when a cache is present, so no-cache forward (LoRA training, evaluation) diverges from generation

mlx_lm.server: [metal::malloc] Resource limit (499000) exceeded in BatchGenerator._generate wedges the server invisibly (HTTP layer stays up, generation thread dies) — recurrence of #831/#1185/#1332's signature in continuous batched serving

Confusing bare KeyError when LoRA config uses "alpha" instead of "scale"

mlx_lm.generate: repetition_penalty has no default (unlike other sampling params), model can loop into repeated text with no warning

completeMemory() prepare count underflow kernel panic recurs under continuous batching despite #883's KV-cache-bound mitigations

Models that discard `update_and_fetch`'s values return leak one Metal buffer per layer per generated token — deterministic crash at Metal's buffer-count limit (deepseek_v4 dies at ~11.5k generated tokens)

LoRA trainer has no gradient clipping or non-finite guard: one bad batch loses the whole run

chat prompt emits raw readline markers (\x01/\x02) because readline is never imported

quantized_kv_start default disagrees between generate_step and the CLI

load() pays an avoidable Hub round-trip per call even when the model is fully cached

[metal::malloc] Resource limit (499000) exceeded on qwen3_5_moe (non-MLA), single node — #839 does not cover this path

`ArraysCache.advance()` leaks Metal buffer objects in qwen3_5/qwen3_next batch decode, crashing long completions at the resource limit

Uh oh!

Search results