From 710b6b7ee0e2dfbb1c89cae1de752ce012c799f4 Mon Sep 17 00:00:00 2001
From: Pyro <178290668+pyros-projects@users.noreply.github.com>
Date: Tue, 4 Aug 2026 00:47:14 +0200
Subject: [PATCH] Support MiniMax attention patch callback variants

---
 comfy/ldm/minimax/model.py                    |  6 ++-
 .../comfy_test/test_minimax_attention.py      | 46 +++++++++++++++++++
 2 files changed, 51 insertions(+), 1 deletion(-)
 create mode 100644 tests-unit/comfy_test/test_minimax_attention.py

diff --git a/comfy/ldm/minimax/model.py b/comfy/ldm/minimax/model.py
index 9c0b6e352..57f6b3aec 100644
--- a/comfy/ldm/minimax/model.py
+++ b/comfy/ldm/minimax/model.py
@@ -191,7 +191,11 @@ class Attention(nn.Module):
             k = k.reshape(1, s, -1)
             v = v.reshape(1, s, -1)
             for p in patches["attn1_patch"]:
-                q, k, v = p(q, k, v, extra_options)
+                out = p(q, k, v, extra_options=extra_options)
+                if isinstance(out, dict):
+                    q, k, v = out.get("q", q), out.get("k", k), out.get("v", v)
+                else:
+                    q, k, v = out
             q = q.view(q.shape[0], q.shape[1], self.heads, self.head_dim).transpose(1, 2)
             k = k.view(k.shape[0], k.shape[1], self.heads, self.head_dim).transpose(1, 2)
             v = v.view(v.shape[0], v.shape[1], self.heads, self.head_dim).transpose(1, 2)
diff --git a/tests-unit/comfy_test/test_minimax_attention.py b/tests-unit/comfy_test/test_minimax_attention.py
new file mode 100644
index 000000000..3f010be24
--- /dev/null
+++ b/tests-unit/comfy_test/test_minimax_attention.py
@@ -0,0 +1,46 @@
+import torch
+
+import comfy.ldm.minimax.model as minimax
+
+
+class QKV(torch.nn.Module):
+    def forward(self, x):
+        return torch.cat((x, x + 1, x + 2), dim=-1)
+
+
+def test_attention_patch_accepts_tuple_and_mapping_callbacks(monkeypatch):
+    attention = minimax.Attention(4, 2, 2, 1e-6, operations=torch.nn)
+    attention.qkv_proj = QKV()
+    attention.q_norm = torch.nn.Identity()
+    attention.k_norm = torch.nn.Identity()
+    attention.out_proj = torch.nn.Identity()
+    seen = {}
+
+    def tuple_patch(q, k, v, extra_options):
+        assert extra_options["block_index"] == 3
+        return q + 1, k + 1, v + 1
+
+    def mapping_patch(q, k, v, pe=None, attn_mask=None, extra_options=None):
+        assert pe is None
+        assert attn_mask is None
+        assert extra_options["n_heads"] == 2
+        return {"q": q * 2, "v": v * 3}
+
+    def fake_attention(q, k, v, *args, **kwargs):
+        seen.update(q=q, k=k, v=v)
+        return v.transpose(1, 2).reshape(1, v.shape[2], -1)
+
+    monkeypatch.setattr(minimax, "optimized_attention", fake_attention)
+    x = torch.zeros(2, 4)
+    output = attention(
+        x,
+        transformer_options={
+            "block_index": 3,
+            "patches": {"attn1_patch": [tuple_patch, mapping_patch]},
+        },
+    )
+
+    assert torch.equal(seen["q"], torch.full((1, 2, 2, 2), 2.0))
+    assert torch.equal(seen["k"], torch.full((1, 2, 2, 2), 2.0))
+    assert torch.equal(seen["v"], torch.full((1, 2, 2, 2), 9.0))
+    assert torch.equal(output, torch.full((2, 4), 9.0))