Limit prompt cache reuse to text-only profiles

2026-08-25 19:26:48 +02:00
parent d921dfaf69
commit c25e57af57
6 changed files with 22 additions and 16 deletions
@@ -90,12 +90,9 @@ services:
      - q4_0
      - --cache-type-v
      - q4_0
-      # Keep the cross-chat prefix cache explicit. cache-reuse tolerates small
-      # changes after Hermes' stable system-prompt prefix without enabling the
-      # currently unreliable on-disk slot restore path.
+      # Keep the cross-chat prefix cache explicit. On-disk slot restore stays
+      # disabled until the current upstream restore regressions are fixed.
      - --cache-prompt
-      - --cache-reuse
-      - "${LLAMA_CACHE_REUSE:-256}"
      - --cache-ram
      - "${LLAMA_CACHE_RAM_MIB:-8192}"
      - --threads
@@ -172,8 +169,6 @@ services:
      - --cache-type-v
      - q4_0
      - --cache-prompt
-      - --cache-reuse
-      - "${LLAMA_CACHE_REUSE:-256}"
      - --cache-ram
      - "${LLAMA_CACHE_RAM_MIB:-8192}"
      - --threads
@@ -261,8 +256,6 @@ services:
      - --cache-type-v
      - q4_0
      - --cache-prompt
-      - --cache-reuse
-      - "${LLAMA_CACHE_REUSE:-256}"
      - --cache-ram
      - "${LLAMA_CACHE_RAM_MIB:-8192}"
      - --threads
@@ -421,8 +414,6 @@ services:
      - --cache-type-v
      - q4_0
      - --cache-prompt
-      - --cache-reuse
-      - "${LLAMA_CACHE_REUSE:-256}"
      - --cache-ram
      - "${LLAMA_CACHE_RAM_MIB:-8192}"
      - --threads