From fb2315f11db0ebfaafa9099a5df5227dc6bb42bc Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 29 Sep 2026 21:00:09 -0700 Subject: [PATCH] Fix crash when selecting int8 or int4 cache in qwen image 2.1 (#16667) --- comfy/ldm/qwen_image21/model.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/ldm/qwen_image21/model.py b/comfy/ldm/qwen_image21/model.py index 8e2d6ad5a..b804b9eca 100644 --- a/comfy/ldm/qwen_image21/model.py +++ b/comfy/ldm/qwen_image21/model.py @@ -395,6 +395,9 @@ class QwenImage21Transformer2DModel(nn.Module): hidden_states = block(hidden_states, mod, pe, attn_fn, prefix_len, transformer_options) for p in patches.get("single_block", []): hidden_states = p({"img": hidden_states, "x": x, "block_index": i, "transformer_options": transformer_options})["img"] + if cache is not None: + # Release dequantized K/V before leaving the block's allocation scope. + del attn_fn, prefix_k, prefix_v comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, x.device, None, malloc_scope="block") comfy.model_prefetch.malloc_graph_end()