From 5978716c987a3f7ef2f03bbcb597a494e13bb103 Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Sat, 9 May 2026 16:09:43 +0100
Subject: [PATCH] feat(flux2): native GLoRA loader

Loads the four ab.weight tensors plus alpha. Fused QKV is skipped:
GLoRA's target-dependent term doesn't row-slice cleanly without
recomputing against a fused proxy weight.
---
 pipelines/flux/flux2_lora.py | 52 +++++++++++++++++++++++++++++++++++-
 1 file changed, 51 insertions(+), 1 deletion(-)

diff --git a/pipelines/flux/flux2_lora.py b/pipelines/flux/flux2_lora.py
index abfbed43c..c8f3e5d60 100644
--- a/pipelines/flux/flux2_lora.py
+++ b/pipelines/flux/flux2_lora.py
@@ -38,7 +38,7 @@ import time
 import torch
 from modules import shared, sd_models
 from modules.logger import log
-from modules.lora import network, network_lora, network_lokr, network_hada, network_oft, network_ia3, lora_convert
+from modules.lora import network, network_lora, network_lokr, network_hada, network_oft, network_ia3, network_glora, lora_convert
 from modules.lora import lora_common as l
 
 
@@ -86,12 +86,18 @@ IA3_SUFFIXES = (
     ".weight", ".on_input",
     ".alpha", ".scale",
 )
+GLORA_SUFFIXES = (
+    ".a1.weight", ".a2.weight",
+    ".b1.weight", ".b2.weight",
+    ".alpha", ".dora_scale", ".scale",
+)
 
 LORA_MARKERS = (".lora_down.weight", ".lora_up.weight", ".lora_A.weight", ".lora_B.weight")
 LOKR_MARKERS = (".lokr_w1", ".lokr_w2")
 LOHA_MARKERS = (".hada_w1_a", ".hada_w1_b", ".hada_w2_a", ".hada_w2_b")
 OFT_MARKERS = (".oft_blocks", ".oft_diag")
 IA3_MARKERS = (".on_input",)  # NOT .weight — too generic, overlaps every other family
+GLORA_MARKERS = (".a1.weight", ".a2.weight", ".b1.weight", ".b2.weight")
 
 
 # === BFL → diffusers mapping ===
@@ -523,6 +529,50 @@ def try_load_ia3(name, network_on_disk, lora_scale):
     return finalize_network(net, name, 'IA3', lora_scale, t0, unmapped=unmapped, skipped=skipped)
 
 
+def try_load_glora(name, network_on_disk, lora_scale):
+    """Load a Flux2/Klein GLoRA adapter as native modules.
+
+    GLoRA stores four low-rank components (``a1``/``a2``/``b1``/``b2``) and
+    computes ``ΔW = w2b @ w1b + (target @ w2a) @ w1a`` — the second term is
+    target-dependent. Fused QKV in double_blocks is skipped with a warning
+    because the target-dependent term doesn't slice cleanly without
+    redirecting calc_updown to a fused proxy weight, and zero real-world
+    GLoRA-on-DiT files exist.
+
+    Depends on the ``self.dim`` initialization fix in network_glora.py so
+    that alpha-based ``calc_scale`` is honored.
+    """
+    t0 = time.time()
+    state_dict = sd_models.read_state_dict(network_on_disk.filename, what='network')
+    if not has_marker(state_dict, GLORA_MARKERS):
+        return None
+
+    mapping = resolve_mapping()
+    net = new_network(name, network_on_disk)
+    groups = group_by_suffixes(state_dict, GLORA_SUFFIXES)
+
+    unmapped = 0
+    skipped = 0
+    for (prefix, base), w in groups.items():
+        if not all(k in w for k in ('a1.weight', 'a2.weight', 'b1.weight', 'b2.weight')):
+            continue
+        targets = resolve_targets(prefix, base)
+        if any(t[1] is not None for t in targets):
+            log.warning(f'Network load: type=GLoRA name="{name}" key={base} fused QKV skipped (unsupported)')
+            skipped += 1
+            continue
+        for diffusers_path, _, _ in targets:
+            network_key = "lora_transformer_" + diffusers_path.replace(".", "_")
+            sd_module = mapping.get(network_key)
+            if sd_module is None:
+                unmapped += 1
+                continue
+            nw = network.NetworkWeights(network_key=network_key, sd_key=network_key, w=w, sd_module=sd_module)
+            net.modules[network_key] = network_glora.NetworkModuleGLora(net, nw)
+
+    return finalize_network(net, name, 'GLoRA', lora_scale, t0, unmapped=unmapped, skipped=skipped)
+
+
 # === Diffusers-PEFT path helpers (used when lora_force_diffusers is on) ===