Issues

perf(decode): two hard rules for the decode path (no extra GPU op per token, no f32 between GPU ops), a per-token primitive census gate, and an audit of the test targets

Pushkinist/rMLX#733 · opened

on Oct 24, 2026

3

perf(decode): Ternary-Bonsai-2 decodes 12% slower than Bonsai 1 because each Hadamard linear adds four small GPU kernels per token

Pushkinist/rMLX#732 · opened

on Oct 24, 2026

3

test(gates): the host-memory guard of the 0.4.5 whole ci-perf peaked at 39.9 GiB of a 40 GiB bound, 7.3 GiB above the 0.4.4 gate at the same offset — attribute the peak to a test

Pushkinist/rMLX#731 · opened

on Oct 23, 2026

server(spec): `thinking_budget` has no effect on a speculative request — the server builds the force-close, and no speculative loop reads it

Pushkinist/rMLX#730 · opened

on Oct 23, 2026

refactor: three source files crossed the 1000-line guideline in 0.4.5 with no split and no LOC-exempt marker (decode_loop.rs, ssd_index.rs, anthropic/streaming.rs)

Pushkinist/rMLX#729 · opened

on Oct 23, 2026

docs: 0.4.5 added SPEC_READING_A_RUN and grew PROMPT_CACHE — cut the run records, state one argument once, and note that FFI.md is 10 bytes under the size cap

Pushkinist/rMLX#728 · opened

on Oct 23, 2026

docs(releasing): the retrospective sets PREV from the nearest tag of any name, so a local tag gives the wrong range

Pushkinist/rMLX#722 · opened

on Oct 17, 2026

docs: cut the five docs that 0.4.4 grew or added to current truth (AUDIO, AUDIO_TTS, AUDIO_WAV, SERVER_SHUTDOWN, WEIGHT_QUANTS)

Pushkinist/rMLX#721 · opened

on Oct 17, 2026

server(refactor): the OpenAI and Anthropic routes each carry their own engine-error map and token count; 0.4.4 made the same edit in both

Pushkinist/rMLX#720 · opened

on Oct 17, 2026

test(refactor): fourteen test files implement tracing::Subscriber to capture events; 0.4.4 added eight and none is shared

Pushkinist/rMLX#719 · opened

on Oct 17, 2026

ci(debt-report): the file-pair scan pairs files inside one directory only, so three per-arch generate.rs files at 57% to 73% never reach the report

Pushkinist/rMLX#718 · opened

on Oct 17, 2026

kv: eval_gpu_state leaves the iso V buffers and the rotor K ring lazy; in a chunked tail prefill the last layer's ring can chain across all chunks

Pushkinist/rMLX#715 · opened

on Oct 16, 2026

server: /v1/messages drops image and input_audio blocks and answers as if they were not sent

Pushkinist/rMLX#708 · opened

on Oct 14, 2026

tts: the codec decoder's pre-transformer ignores the 72-frame sliding window of its config

Pushkinist/rMLX#707 · opened

on Oct 14, 2026

ci(gpu-test): the shader-validation canary ran and printed no report, so the runner stopped before any test

Pushkinist/rMLX#705 · opened

on Oct 14, 2026

server: the embeddings job has no stop input and no place in the GPU queue; a remote image is fetched on an async worker

Pushkinist/rMLX#704 · opened

on Oct 14, 2026

whisper: the frames at the edges of a 30 s window read a reflection of the window; the reference reads the samples of the next window

Pushkinist/rMLX#703 · opened

on Oct 14, 2026

server: a retry attempt of a chat request runs with no GPU permit, so the queue bound and the in-flight count do not see it

Pushkinist/rMLX#702 · opened

on Oct 13, 2026

server: a remote image is fetched inside the generation job, under the GPU permit, with a 20 s timeout for each image

Pushkinist/rMLX#701 · opened

on Oct 13, 2026

server: a load evicts a model that a request still uses; two models are in memory at --max-loaded-models 1, and the context check can be skipped

Pushkinist/rMLX#700 · opened

on Oct 13, 2026

server: an unload frees the model on an async worker (the unload route, the idle timer, the embedding unload)

Pushkinist/rMLX#699 · opened

on Oct 13, 2026

server: the 503 of the chat and messages routes carries the text of an engine error; an MLX message can name a file location on the server

Pushkinist/rMLX#697 · opened

on Oct 13, 2026

models(refactor): four generate loops still copy the token-selection fork of select_token

Pushkinist/rMLX#694 · opened

on Oct 13, 2026

server: the OpenAI stream emits reasoning and tool-call deltas after a stop string matched

Pushkinist/rMLX#693 · opened

on Oct 13, 2026

server: a body that the JSON extractor refuses gets a plain-text reply that can repeat the client's value

Pushkinist/rMLX#687 · opened

on Oct 13, 2026

Navigation Menu

Repository navigation

Issues

perf(decode): two hard rules for the decode path (no extra GPU op per token, no f32 between GPU ops), a per-token primitive census gate, and an audit of the test targets

perf(decode): Ternary-Bonsai-2 decodes 12% slower than Bonsai 1 because each Hadamard linear adds four small GPU kernels per token

test(gates): the host-memory guard of the 0.4.5 whole ci-perf peaked at 39.9 GiB of a 40 GiB bound, 7.3 GiB above the 0.4.4 gate at the same offset — attribute the peak to a test

server(spec): `thinking_budget` has no effect on a speculative request — the server builds the force-close, and no speculative loop reads it

refactor: three source files crossed the 1000-line guideline in 0.4.5 with no split and no LOC-exempt marker (decode_loop.rs, ssd_index.rs, anthropic/streaming.rs)

docs: 0.4.5 added SPEC_READING_A_RUN and grew PROMPT_CACHE — cut the run records, state one argument once, and note that FFI.md is 10 bytes under the size cap

docs(releasing): the retrospective sets PREV from the nearest tag of any name, so a local tag gives the wrong range

docs: cut the five docs that 0.4.4 grew or added to current truth (AUDIO, AUDIO_TTS, AUDIO_WAV, SERVER_SHUTDOWN, WEIGHT_QUANTS)

server(refactor): the OpenAI and Anthropic routes each carry their own engine-error map and token count; 0.4.4 made the same edit in both

test(refactor): fourteen test files implement tracing::Subscriber to capture events; 0.4.4 added eight and none is shared

ci(debt-report): the file-pair scan pairs files inside one directory only, so three per-arch generate.rs files at 57% to 73% never reach the report

kv: eval_gpu_state leaves the iso V buffers and the rotor K ring lazy; in a chunked tail prefill the last layer's ring can chain across all chunks

server: /v1/messages drops image and input_audio blocks and answers as if they were not sent

tts: the codec decoder's pre-transformer ignores the 72-frame sliding window of its config

ci(gpu-test): the shader-validation canary ran and printed no report, so the runner stopped before any test

server: the embeddings job has no stop input and no place in the GPU queue; a remote image is fetched on an async worker

whisper: the frames at the edges of a 30 s window read a reflection of the window; the reference reads the samples of the next window

server: a retry attempt of a chat request runs with no GPU permit, so the queue bound and the in-flight count do not see it

server: a remote image is fetched inside the generation job, under the GPU permit, with a 20 s timeout for each image

server: a load evicts a model that a request still uses; two models are in memory at --max-loaded-models 1, and the context check can be skipped

server: an unload frees the model on an async worker (the unload route, the idle timer, the embedding unload)

server: the 503 of the chat and messages routes carries the text of an engine error; an MLX message can name a file location on the server

models(refactor): four generate loops still copy the token-selection fork of select_token

server: the OpenAI stream emits reasoning and tool-call deltas after a stop string matched

server: a body that the JSON extractor refuses gets a plain-text reply that can repeat the client's value