From be995185f598c83abb990a088e9f634c4d36eb46 Mon Sep 17 00:00:00 2001
From: Jaret Burkett <jaretburkett@gmail.com>
Date: Sat, 29 Aug 2026 12:01:40 -0600
Subject: [PATCH] Fixed issue with offloading with new model loader on nvfp4
 weights.

---
 toolkit/models/v2/_mixin.py | 18 +++++++++++++++---
 version.py                  |  2 +-
 2 files changed, 16 insertions(+), 4 deletions(-)

diff --git a/toolkit/models/v2/_mixin.py b/toolkit/models/v2/_mixin.py
index 9a0f879..6be23f0 100644
--- a/toolkit/models/v2/_mixin.py
+++ b/toolkit/models/v2/_mixin.py
@@ -279,13 +279,19 @@ class OstrisModelMixin:
                 }
                 - {None}
             )
+            # a checkpoint may deliberately mix backends (minimax's TE: nvfp4
+            # LM linears + int8 embeddings). The request matches when the
+            # requested backend is among the shipped ones — the whole shipped
+            # quantization is then kept exactly as-is.
             matches = (
                 qtype is not None
                 and ara_path is None
-                and shipped == [qtype]
+                and qtype in shipped
             )
             if matches:
-                status_fn(f"Checkpoint is pre-quantized ({qtype}); keeping it")
+                status_fn(
+                    f"Checkpoint is pre-quantized ({'/'.join(shipped)}); keeping it"
+                )
             else:
                 target_is_ostris = qtype is not None and isinstance(
                     get_qtype(qtype), ostristype
@@ -364,9 +370,15 @@ class OstrisModelMixin:
         if offload and offload > 0:
             from toolkit.memory_management import MemoryManager
 
+            # the manager stages layers to the COMPUTE device. `device` is the
+            # component's parking device — "cpu" under low_vram — which must
+            # never become the manager's process device (bounces would stage
+            # to cpu against cuda activations). quantize_device carries the
+            # actual gpu in the holder flow.
+            compute_device = quantize_device if quantize_device is not None else device
             MemoryManager.attach(
                 self,
-                device,
+                torch.device(compute_device),
                 offload_percent=offload,
                 ignore_modules=list(self.get_offload_ignore_modules() or []),
             )
diff --git a/version.py b/version.py
index f5a5cb4..bec33f0 100644
--- a/version.py
+++ b/version.py
@@ -1 +1 @@
-VERSION = "0.13.1"
+VERSION = "0.13.2"