fix: apply peft<->gptqmodel AWQ shim on the inference LoRA path too

Loading a trained LoRA at inference (e.g. a fighter LoRA on an image pipeline)
crashed with "cannot import name AwqGEMMQuantLinear" because peft dispatches AWQ
for any non-bnb target when gptqmodel is installed, and gptqmodel 7.1.0 renamed
that class to AwqGEMMLinear. The alias shim existed only in the training path.

Extract it to codai/models/peft_compat.ensure_peft_awq_compat() (cached) and call
it before load_lora_weights in codai/api/images._apply_loras and the acceleration
fuse path; the trainer now delegates to the shared shim.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RdMufYvtTbtGDWsiZVoXce

codai/__init__.py

@@ -16,7 +16,7 @@
 
 # Canonical product version for CoderAI — single source of truth. Both the API
 # metadata and the admin web UI read from here.
-__version__ = "0.1.25"
+__version__ = "0.1.26"
 
 # Configure the CUDA caching allocator BEFORE torch is imported anywhere.
 # expandable_segments lets the allocator return freed pages to the driver even

codai/api/images.py

@@ -676,6 +676,11 @@ def _ensure_ip_adapter_loaded(pipeline) -> bool:
 def _apply_loras(pipeline, loras):
     """Load and activate LoRA weights on a diffusers pipeline."""
     try:
+        # peft dispatches AWQ for any non-bnb target when gptqmodel is installed;
+        # gptqmodel 7.1.0 renamed the class peft imports, which would crash
+        # load_lora_weights below. Alias it before applying any adapter.
+        from codai.models.peft_compat import ensure_peft_awq_compat
+        ensure_peft_awq_compat()
         names = []
         weights = []
         for i, lora in enumerate(loras):

codai/api/loras.py

@@ -1443,13 +1443,11 @@ def _ensure_peft_awq_compat():
     import AwqGEMMQuantLinear`, but gptqmodel 7.1.0 renamed that class to
     AwqGEMMLinear. Since peft calls dispatch_awq for ANY non-bnb target whenever
     gptqmodel is installed, the failed import crashes EVERY add_adapter() (SDXL, Wan
-    and Z-Image LoRA training alike). Alias the renamed class so the import succeeds;
-    no-op when the name already exists or gptqmodel isn't present."""
+    and Z-Image LoRA training alike). Delegates to the shared shim (also used by
+    the inference LoRA-apply path)."""
     try:
-        import importlib
-        m = importlib.import_module("gptqmodel.nn_modules.qlinear.gemm_awq")
-        if not hasattr(m, "AwqGEMMQuantLinear") and hasattr(m, "AwqGEMMLinear"):
-            m.AwqGEMMQuantLinear = m.AwqGEMMLinear
+        from codai.models.peft_compat import ensure_peft_awq_compat
+        ensure_peft_awq_compat()
     except Exception:
         pass
 

codai/models/acceleration.py

@@ -293,6 +293,14 @@ def apply_accel_to_pipeline(pipe, accel: Optional[dict]) -> None:
                     "acceleration LoRA", type(pipe).__name__)
         return
 
+    # gptqmodel 7.1.0 renamed the AWQ class peft imports during dispatch; alias
+    # it before any load_lora_weights so the import doesn't crash the fuse.
+    try:
+        from codai.models.peft_compat import ensure_peft_awq_compat
+        ensure_peft_awq_compat()
+    except Exception:
+        pass
+
     weight = float(accel.get("lora_weight") or 1.0)
 
     def _load_one(ref, into_t2: bool, adapter: str) -> bool:

codai/models/peft_compat.py

+# CoderAI - OpenAI-compatible API server
+# Copyright (C) 2026 Stefy Lanza <stefy@nexlab.net>
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program. If not, see <https://www.gnu.org/licenses/>.
+
+"""Compatibility shims for peft <-> gptqmodel version skew."""
+
+_awq_aliased = False
+
+
+def ensure_peft_awq_compat():
+    """peft's LoRA AWQ dispatcher does ``from gptqmodel.nn_modules.qlinear.gemm_awq
+    import AwqGEMMQuantLinear``, but gptqmodel 7.1.0 renamed that class to
+    AwqGEMMLinear. peft calls dispatch_awq for ANY non-bnb target whenever
+    gptqmodel is installed, so the failed import crashes EVERY add_adapter() /
+    load_lora_weights() — at LoRA *training* AND at *inference* when a trained
+    LoRA is applied to a pipeline. Alias the renamed class so the import succeeds;
+    no-op when the name already exists or gptqmodel isn't present. Cached so the
+    per-request inference path doesn't re-import gptqmodel every call."""
+    global _awq_aliased
+    if _awq_aliased:
+        return
+    try:
+        import importlib
+        m = importlib.import_module("gptqmodel.nn_modules.qlinear.gemm_awq")
+        if not hasattr(m, "AwqGEMMQuantLinear") and hasattr(m, "AwqGEMMLinear"):
+            m.AwqGEMMQuantLinear = m.AwqGEMMLinear
+        _awq_aliased = True
+    except Exception:
+        # gptqmodel not installed, or import failed — nothing to alias. Don't
+        # cache so a later call can retry once gptqmodel is importable.
+        pass