Morph MORPH
SIGN IN SIGN UP
dhruvhead / omlx UNCLAIMED
0 0 137 Python

COMMITS

main
September 2, 2026
J
fix: keep restored-prefix VLM prefill inputs off the default stream (#3305)
jundot committed
1a7ab33
J
fix: preserve Qwen4 PLE mmap admission fallback
jundot committed
33ad1ea
P
fix(engine_pool): keep the Qwen4 PLE table resident across a model swap (#3299)
pedroberaldo87 committed
1f6f4e2
J
refactor: simplify Qwen4 prefill memory accounting (#3350)
jundot committed
e1280a2
J
fix(qwen4): reject corrupt QSA prefix states (#3294)
jundot committed
9ed3be2
J
fix(qwen4): isolate gathered prefill memory history (#3350)
jundot committed
ba45eaa
J
fix(qwen4_exp): price gathered QSA prefill memory (#3351)
jundot committed
6f76565
J
fix(qwen4): preserve gathered prefill across mRoPE rebinds (#3355)
Jonathan Spangler committed
ffb5072
J
fix(qwen4): harden QSA batch cache joins (#3294)
jundot committed
e680587
A
fix(qwen4): handle mixed-rank QSA cache joins (#3369)
astro-15-alive committed
2246290
September 1, 2026
K
fix(qwen4_exp): price gathered QSA at admission, not only during forward (#3350)
KC committed
61cd558
K
fix(qwen4_exp): survive batched QSA offset and mixed-rank indexer extend (#3350)
KC committed
44e51e3
J
fix: handle multi-row VLM rope deltas (#3343)
jundot committed
ebf4f1f
K
fix(qwen4_exp): pin mRoPE equal-plane cache to the live array (#3350)
KC committed
262a2c6
K
test(qwen4_exp): treat broadcast text mRoPE as gathered-eligible on decode (#3350)
KC committed
8adac5f
K
fix(qwen4_exp): gather text mRoPE prefills and stop pricing them as dense SDPA (#3350)
KC committed
61e1027
August 31, 2026
J
fix: validate boundary snapshots before materialization
jundot committed
dc312e6
J
build: upgrade MLX to 0.32.2 with runtime compatibility (#3332)
Jun Kim committed
8b2ade3
August 29, 2026
G
formula: bump to 0.6.4
github-actions[bot] committed
e008a66
J
chore: bump version to 0.6.4
jundot committed
1d78261
J
fix(qwen4): gate wide prefill on sparse native kernel (#3244)
jundot committed
0a6ae09
J
perf(qwen4): flatten exact QSA prefill and accelerate Lightning MTP (#3244)
Jonathan Spangler committed
7467dce
J
fix(cache): initialize boundary diagnostics in snapshot test
jundot committed
8706683
J
fix(admin): validate Lightning MTP draft depth (#3280)
jundot committed
64e63cb
J
fix(cache): keep boundary diagnostics accurate
jundot committed
b798e65
C
fix(admin): persist mtp_num_draft_tokens via settings PUT (#3280)
Cold Cloud committed
dc18a18
A
fix: allow TurboQuant with Lightning MTP in macOS app (#3254)
Aka.Fido committed
4a082f4
M
feat(cache): expose hybrid boundary diagnostics (#3249)
MDX Tom committed
fe8fe7e
R
fix(cache): normalize mixed Qwen4 QSA positions (#3219)
rsnow committed
119d204
J
fix(qwen4_exp): bound QSA long-prefill memory (#3283)
jundot committed
9934b28