From 1fd39358c5aba4c86ffd944e91a453d2f37e12c3 Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Sun, 26 Jul 2026 04:12:06 +0100
Subject: [PATCH] fix(sdnq): reset dynamo caches at model unload

Dynamo tracks a lifetime recompile counter per compiled function that
freed models leave climbing while their graphs and guards die, and the
compiled dequant runs fullgraph, so crossing the accumulated limit is a
hard FailOnRecompileLimitHit instead of an eager fallback; enough model
or quant switches in one process got there. unload_model_weights now
calls reset_compile_caches when the compiled dequant is active, dropping
the dead graphs and the counters in the same sweep as the unload gc.
Raised limits only move the wall; the reset removes it.

- scoped to the model unload branch: the reset is global and must only
  run when the graphs' owner is being discarded
- regression test trips the wall under a lowered limit and recovers
  through the same helper the unload path calls
---
 modules/sd_models.py   |  2 ++
 modules/sdnq/common.py | 16 ++++++++++++++++
 2 files changed, 18 insertions(+)

diff --git a/modules/sd_models.py b/modules/sd_models.py
index 494c5bf1d..69abadc7e 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -1573,6 +1573,8 @@ def unload_model_weights(op='model'):
             disable_offload(model_data.sd_model)
             move_model(model_data.sd_model, 'meta')
         model_data.sd_model = None
+        from modules.sdnq.common import reset_compile_caches
+        reset_compile_caches() # dead compiled-dequant graphs and their lifetime recompile counters otherwise accumulate across switches
         devices.torch_gc(force=True, reason='unload')
         log.debug(f'Unload {op}: {memory_stats()} fn={fn}')
     elif (op == 'refiner') and model_data.sd_refiner:
diff --git a/modules/sdnq/common.py b/modules/sdnq/common.py
index 92801ce25..d82bb4823 100644
--- a/modules/sdnq/common.py
+++ b/modules/sdnq/common.py
@@ -368,6 +368,22 @@ else:
         return fn
 
 
+def reset_compile_caches():
+    """Drop compiled graphs and dynamo's lifetime recompile counters at model unload.
+
+    Freed layers leave their guards invalid, so the dead graphs cannot be
+    reused, but the per-frame lifetime compile counter keeps climbing and a
+    fullgraph compile hard-fails (FailOnRecompileLimitHit) once it crosses the
+    accumulated limit; enough model or quant switches in one process get there.
+    The reset is global, so call it only when the compiled graphs' owner is
+    being discarded.
+    """
+    if check_torch_compile():
+        from modules.logger import log
+        torch._dynamo.reset()
+        log.debug('SDNQ compile: dynamo reset')
+
+
 common_skip_keys = (
     ".time_embed",
     ".context_embedder",