|
|
@@ -170,7 +170,12 @@ args = [
|
|
|
# DCP4 — required to make 512K fit (KV sharded across the 4 TP ranks).
|
|
|
# Unverified for the full model; first fallback is to drop both lines.
|
|
|
"--decode-context-parallel-size", "4",
|
|
|
- "--cp-kv-cache-interleave-size", "4",
|
|
|
+ # interleave MUST be 1 here: the engine (r35) asserts
|
|
|
+ # DCP sparse indexer currently supports only cp_kv_cache_interleave_size=1
|
|
|
+ # (the DSA top-k sparse-MLA path is not the Flash dense-MLA path where the
|
|
|
+ # x4 interleave / parent-page-stride rule applied; that constraint does not
|
|
|
+ # carry over to this model - learned on the first DCP4 boot, 2026-09-13).
|
|
|
+ "--cp-kv-cache-interleave-size", "1",
|
|
|
# No --kv-cache-dtype: let the r35 fork pick its native default for
|
|
|
# GlmMoeDsa attention (the Flash lane's explicit fp8/fp8_ds_mla spellings
|
|
|
# both failed auto-selection on this build, and those were Flash-MLA
|
