From 894e92275d700a6419eb2c18bc0a69b3c499099e Mon Sep 17 00:00:00 2001
From: Vuong Hoang <vh@phasefinal.com>
Date: Mon, 27 Apr 2026 18:05:49 -0700
Subject: [PATCH] =?UTF-8?q?llama-swap:=20drop=20qwen3.6-35-a3b=20from=20`p?=
 =?UTF-8?q?inned`=20group=20=E2=80=94=20was=20blocking=20other=20model=20l?=
 =?UTF-8?q?oads=20under=20memory=20pressure?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Symptom: qwen3.6-35-a3b refused to deload when other models needed
the VRAM, even with the model itself at ttl: 0. The pinning came from
the `pinned` group's `persistent: true` flag, which exempts members
from eviction by the scheduler regardless of memory pressure. The
model's ttl: 0 only governs idle-timeout, NOT scheduler eviction —
those are separate concerns.

Removed qwen3.6-35-a3b from the group's members. Kept ttl: 0 on the
model itself: still no idle-unload, but the scheduler CAN now evict
it when another non-coexistent model is requested. qwen3.5-9b stays
pinned (~6 GB at Q4 — cheap to hold).

Updated the inline comment + the group-header docstring to reflect
the new semantics so future-Claude doesn't undo this.
---
 stacks/llama-swap/conf/config.yaml | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/stacks/llama-swap/conf/config.yaml b/stacks/llama-swap/conf/config.yaml
index 891b3df..8bc680c 100644
--- a/stacks/llama-swap/conf/config.yaml
+++ b/stacks/llama-swap/conf/config.yaml
@@ -129,7 +129,8 @@ models:
   "qwen3.6-35-a3b":
     name: "Qwen 3.6 35B-A3B UD-Q6_K_XL"
     description: "Qwen3.6 MoE, 3B active of 35B. Stock unsloth Q6_K_XL (~29GB). Thinking on by default."
-    ttl: 0   # pinned — member of the `pinned` group, never unloads
+    ttl: 0   # never idle-unload, but evictable under memory pressure
+             # (out of `pinned` group as of 2026-04-27 — was OOM'ing other loads).
     cmd: |
       /app/llama-server
       --context-shift
@@ -546,12 +547,15 @@ groups:
 
   # Pinned general-purpose chat models. Coexist in VRAM, never unload.
   # Members also have ttl: 0 individually so idle-timeout can't drop them.
-  # VRAM budget check: qwen3.5-9b (~6 GB at Q4) + qwen3.6-35-a3b (~29 GB
-  # at Q6_K_XL) + 128K KV cache for each ≈ 50-60 GB. Size for your GPU.
+  # Currently just qwen3.5-9b — small enough (~6 GB at Q4 + KV) to be
+  # cheap to keep resident. qwen3.6-35-a3b was previously here but its
+  # ~29 GB at Q6_K_XL made other model loads OOM under memory pressure
+  # (persistent: true exempts members from eviction). It now sits with
+  # ttl: 0 outside any group — never idle-unloads, but the scheduler
+  # CAN evict it when another non-coexistent model is requested.
   "pinned":
     swap: false
     exclusive: false
     persistent: true
     members:
       - "qwen3.5-9b"
-      - "qwen3.6-35-a3b"