Morph MORPH
SIGN IN SIGN UP
dhruvhead / vmlx UNCLAIMED
0 0 80 Python

COMMITS

main
August 21, 2026
J
feat(cli): suggest SSD-only caching on low-memory Macs at session startup
Jinho Jang committed
200e6ba
J
fix(parity): controls that existed on one surface and silently died on another
Jinho Jang committed
8208b36
J
fix(ollama): send the duration fields clients compute tok/s from
Jinho Jang committed
b05b0b2
J
test(cache): pin the invariant that keeps padding off disk
Jinho Jang committed
a05109a
J
test(compat): fail when an mlx_vlm patch target disappears instead of skipping
Jinho Jang committed
ed5bbde
J
fix(lfm2-vl): load bundles whose MLP tensors carry llama-style names
Jinho Jang committed
3bbe857
J
harden(cache): clamp every truncation through the LOGICAL extent, not the buffer
Jinho Jang committed
e6776b2
J
fix(dots3): adopt the LOGICAL extent of a restored KV layer, not the padded buffer
Jinho Jang committed
9936e83
J
fix(embeddings): batch internally so a long-chunk request cannot ask for 200GB
Jinho Jang committed
1bca990
August 20, 2026
J
fix(lfm2-vl): unwrap InputEmbeddingsFeatures at the language-model boundary
Jinho Jang committed
4f2c4ca
J
fix(lfm2-vl): image requests died on an upstream arity bug
Jinho Jang committed
928139c
J
fix: four reported issues — garbled qwen3_5, embeddings cap, VLM context cap, guard diagnostics
Jinho Jang committed
c69267a
J
fix(cache): sibling-lane memory-wave discipline + idle-rederive exemption
Jinho Jang committed
e40ee6e
J
docs: changelog notes the qwen video fix
Jinho Jang committed
11b539e
J
fix(video): qwen3_5 video via SimpleEngine never expanded its placeholder
Jinho Jang committed
9b1e458
J
fix(store): disk-lane numpy view imports were the machine-killer's other half
Jinho Jang committed
b092451
J
docs: 1.6.34 changelog covers the store memory fix and DFlash2 resume fix
Jinho Jang committed
303ec9d
J
fix(store): numpy view imports dragged full base buffers into Metal
Jinho Jang committed
21fd6f9
J
fix(hybrid,dflash2): post-turn clean-prefill memory waves + resume cache math
Jinho Jang committed
e8b2b60
J
docs: changelog entries for 1.6.33 and 1.6.34
Jinho Jang committed
eac732f
J
fix(dots3): exempt positional full-latent slots from SSM companion snapshots
Jinho Jang committed
1c282ae
J
debug(census): referrer walk names the holder of the largest grown bucket
Jinho Jang committed
dce7bf8
J
debug(census): gc sweep reports per-bucket growth since previous sweep
Jinho Jang committed
eac80a7
J
debug(census): gc sweep prints total live-array count and bytes
Jinho Jang committed
bbbaeea
J
debug(census): env-gated gc sweep bucketing live mx.arrays by shape
Jinho Jang committed
b51bf95
J
feat(bundle): ship dflash + dflash-mlx (DFlash2 drafter runtime + four-row q4 verify kernel)
Jinho Jang committed
0274691
J
chore: dots3 prefill eval barriers default off (measured no effect at every-8 and every-2); mflux 0.19.0 (mlx 0.32-compatible) unblocks the bundle
Jinho Jang committed
adfb911
J
perf(dots3): prefill eval boundary every 2 layers - one full layer's DSA buffers are ~2.5GB at 60k so every-8 left the sawtooth unchanged
Jinho Jang committed
2f101e3
J
perf(dots3): bound the prefill lazy graph to eval_every layers
Jinho Jang committed
b6d14d4
J
perf(native-mtp): skip the replay forward at every depth
Jinho Jang committed
2422487