perf(glm5next): complete B12X CKV support for DCP and MTP - #488

Draft

jackzampolin wants to merge 20 commits into

local-inference-lab:dev/jovian-judgementfrom

jackzampolin:codex/glm53-jovian-followup

Commits on Aug 28, 2026

perf(glm5next): route MTP verification through B12X decode

jackzampolin
and
codex
committed
683e51d
perf(glm5next): gather full C4 cache for DCP prefill

jackzampolin
and
codex
committed
998af63
fix(glm5next): size CKV scratch from parallel config

jackzampolin
and
codex
committed
d71b3ca
perf(glm5next): overlap DCP CKV gathers across layers

jackzampolin
and
codex
committed
665f6fd
Fix full-cache DCP query selection

jackzampolin
and
claude
committed
ea338d7
fix(spec-decode): scope CP interleave gate to MTP

jackzampolin
and
codex
committed
cff3dc5
Support DCP-replicated DFlash draft caches

jackzampolin
and
codex
committed
ed40055
Group hybrid MLA targets apart from DFlash drafts

jackzampolin
and
codex
committed
99f3b3e
Stabilize mixed-DCP slot mapping compilation

jackzampolin
and
codex
committed
daea2ce
Keep DFlash hybrid prefix caches reusable

jackzampolin
and
claude
committed
913753b
Retain replicated DFlash prefix checkpoints

jackzampolin
and
claude
committed
1529503
Retain DFlash proof blocks for prefix reuse

jackzampolin
and
claude
committed
d445358
Preserve model-specific sliding retention

jackzampolin
and
claude
committed
ff5c342
Honor per-group DFlash cache retention

jackzampolin
and
claude
committed
2836cba
Align DFlash prefix registration with lookup

jackzampolin
and
codex
committed
9ffd6ca
Clarify DCP CKV runtime contracts

jackzampolin
and
codex
committed
aefc986
Format GLM DCP follow-up

jackzampolin
and
codex
committed
a4bde82
Refresh B12X CKV test fixtures

jackzampolin
and
codex
committed
68eb9de
Keep CKV gather on bounded workspace

jackzampolin
and
codex
committed
f5ab1b1
Preallocate B12X MLA scratch before KV profiling

jackzampolin
and
codex
committed
e72447a

Navigation Menu

perf(glm5next): complete B12X CKV support for DCP and MTP - #488

perf(glm5next): complete B12X CKV support for DCP and MTP#488
jackzampolin wants to merge 20 commits into
local-inference-lab:dev/jovian-judgementfrom
jackzampolin:codex/glm53-jovian-followup

Commits on Aug 28, 2026

perf(glm5next): route MTP verification through B12X decode

perf(glm5next): gather full C4 cache for DCP prefill

fix(glm5next): size CKV scratch from parallel config

perf(glm5next): overlap DCP CKV gathers across layers

Fix full-cache DCP query selection

fix(spec-decode): scope CP interleave gate to MTP

Support DCP-replicated DFlash draft caches

Group hybrid MLA targets apart from DFlash drafts

Stabilize mixed-DCP slot mapping compilation

Keep DFlash hybrid prefix caches reusable

Retain replicated DFlash prefix checkpoints

Retain DFlash proof blocks for prefix reuse

Preserve model-specific sliding retention

Honor per-group DFlash cache retention

Align DFlash prefix registration with lookup

Clarify DCP CKV runtime contracts

Format GLM DCP follow-up

Refresh B12X CKV test fixtures

Keep CKV gather on bounded workspace

Preallocate B12X MLA scratch before KV profiling