Morph MORPH
SIGN IN SIGN UP
dhruvhead / vmlx UNCLAIMED
0 0 81 Python

COMMITS

main
August 20, 2026
J
fix(prefix-cache): cumulative state keeps its dtype across the numpy bridge
Jinho Jang committed
15ab55c
J
fix(panel): stream-interval 8 default — interval 1 was the warm-turn collapse
Jinho Jang committed
08768f9
J
test: hybrid round-trip pins the padded-restore contract (offset logical, physical step-padded)
Jinho Jang committed
73ec6c0
J
fix(restore): pad plain-KV restored buffers to the step boundary like ZAYA/CacheList branches
Jinho Jang committed
44c20c2
August 19, 2026
J
fix(native-mtp): restore-aware depth gates
Jinho Jang committed
f303064
J
chore(prefill): slot census every 8 chunks unconditionally
Jinho Jang committed
71b43b7
J
feat(prefill): always-on per-kind cache-slot census at 32-chunk cadence
Jinho Jang committed
2cd0ffc
J
fix(dots3): adopt restored sliding slots into the latent lane
Jinho Jang committed
5ee0ba6
J
feat(admission): teach the wired-limit lever in every rejection
Jinho Jang committed
c5d3fea
J
test: refresh three source-pin audits stale since the DFlash2 merge
Jinho Jang committed
1a931a0
J
fix(block-disk): 4GB pending-write budget default
Jinho Jang committed
fbee066
J
fix(dots3): chunk-scaled prefill valve + burst-surviving L2 admission
Jinho Jang committed
ccc085a
J
fix(dflash2): cut the boundary snapshot before the assistant tag
Jinho Jang committed
2dd4528
J
feat(dflash2): prompt-boundary snapshots rescue reuse under think-strip
Jinho Jang committed
5ce794b
J
fix(dflash2): splice the draft cache across turns via a hidden gap
Jinho Jang committed
2a8d244
J
feat(dflash2): multiturn prefix reuse via in-process session store
Jinho Jang committed
309ea39
J
fix(cache,mtp): bf16 stores broken by MLX 0.32.1, plus a runtime cost gate
Jinho Jang committed
f538972
J
fix(deps): pin dflash==0.1.0 - the DFlash2 drafter import only resolved via /tmp
Jinho Jang committed
24706cb
J
Report the effective DFlash2 verifier width
Jinho Jang committed
3561cef
J
Count DFlash2 multi-token chunks correctly
Jinho Jang committed
7e8cd3c
J
Route server chat through DFlash2 verifier
Jinho Jang committed
ee58fc1
J
Keep native MTP preview deterministic
Jinho Jang committed
75ea762
J
Match DFlash2 command preview to launcher
Jinho Jang committed
33c9e9b
J
Expose matched DFlash2 in VLM settings
Jinho Jang committed
90a1d16
J
Add matched DFlash2 path for Qwen3.8
Jinho Jang committed
0c0fd5a
J
Accelerate Qwen native MTP verification on M5
Jinho Jang committed
f89c4fe
J
perf(mtp): verify prefetch - Qwen 4D 36.5 t/s, +65 percent over baseline
Jinho Jang committed
3219cff
J
feat(mtp): aligned head cache - acceptance 74.8 -> 92.0 percent, depth 2 sticks
Jinho Jang committed
21b420b
J
fix(mtp): three depth-controller defects that locked depth 2 out structurally
Jinho Jang committed
2b4d453
J
feat(mtp): skip the replay forward on rejection via rollback_state (gated, +6%)
Jinho Jang committed
d5b6673