From 425e1a209e8c8ce728112256500a1baabb44fe7d Mon Sep 17 00:00:00 2001
From: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>
Date: Fri, 17 Jul 2026 14:49:16 +0800
Subject: [PATCH] [https://nvbugs/6395830][fix] Qwen-VL mRoPE: move seq-slot
 delta cache update to model device

Text-only startup of the Qwen3.6-27B dense VL checkpoint crashes during
KV-cache capacity estimation in the shared Qwen-VL mRoPE path
(`_prepare_qwen_vl_mrope_config`):

    mrope_position_deltas_cache.index_copy_(0, seq_slots, deltas)
    RuntimeError: Expected all tensors to be on the same device, but got
    source is on cpu, different from other tensors on cuda:0

Root cause: checkpoints carrying `language_model_only: false` with the
dense `Qwen3_5ForConditionalGeneration` arch are routed to
`QwenImageBenchModel`, whose `multimodal_data_device_paths` was missing
the `mrope_config.mrope_position_ids` / `mrope_config.mrope_position_deltas`
entries that its sibling `_Qwen3_5VLModel` lists. The engine's H2D move
(`MultimodalParams.to_device` with `target_keywords`) therefore skips the
mRoPE tensors, leaving the deltas CPU-resident when the GPU seq-slot cache
write (introduced in #11943) consumes them. MoE VL checkpoints
(`Qwen3_5MoeForConditionalGeneration`) never match the image-bench route
and are unaffected.

Fix both layers:
- `QwenImageBenchModel.multimodal_data_device_paths`: add the two
  `mrope_config.*` entries, matching `_Qwen3_5VLModel`, so mRoPE tensors
  ride the engine's pinned async H2D move like every other Qwen-VL model.
- `_prepare_qwen_vl_mrope_config`: normalize `deltas` onto the cache
  device before `index_copy_` as a defensive backstop for any path that
  still reaches the write branch with CPU tensors (no-op when devices
  already match).

Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>
---
 tensorrt_llm/_torch/models/modeling_qwen2vl.py          | 9 ++++++++-
 tensorrt_llm/_torch/models/modeling_qwen_image_bench.py | 6 ++++++
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/tensorrt_llm/_torch/models/modeling_qwen2vl.py b/tensorrt_llm/_torch/models/modeling_qwen2vl.py
index 0a4107bd3d9b..2079bfdf3127 100644
--- a/tensorrt_llm/_torch/models/modeling_qwen2vl.py
+++ b/tensorrt_llm/_torch/models/modeling_qwen2vl.py
@@ -154,7 +154,14 @@ def _prepare_qwen_vl_mrope_config(
         if len(delta_tensors) != num_seq_slots:
             raise RuntimeError(
                 "Missing MRoPE position deltas for seq-slot cache update")
-        deltas = torch.cat(delta_tensors, dim=0)
+        # `delta_tensors` originate from per-request `multimodal_data` and may
+        # be CPU-resident when the owning model's `multimodal_data_device_paths`
+        # does not cover `mrope_config.*` (or a path skips the engine's H2D
+        # move), while the seq-slot cache and `seq_slots` live on the model
+        # device. `index_copy_` requires all tensors on the same device.
+        deltas = torch.cat(delta_tensors,
+                           dim=0).to(device=mrope_position_deltas_cache.device,
+                                     non_blocking=True)
         mrope_position_deltas_cache.index_copy_(0, seq_slots, deltas)
 
     if position_ids is not None \
diff --git a/tensorrt_llm/_torch/models/modeling_qwen_image_bench.py b/tensorrt_llm/_torch/models/modeling_qwen_image_bench.py
index 92d3cccb5550..061d21bc0df0 100644
--- a/tensorrt_llm/_torch/models/modeling_qwen_image_bench.py
+++ b/tensorrt_llm/_torch/models/modeling_qwen_image_bench.py
@@ -40,10 +40,16 @@
 class _QwenImageBenchModelMixin:
     @property
     def multimodal_data_device_paths(self) -> List[str]:
+        # Keep the mrope_config entries in sync with `_Qwen3_5VLModel`
+        # (modeling_qwen3_5.py): the shared Qwen-VL mRoPE seq-slot cache path
+        # consumes `mrope_position_deltas` on the model device, so the engine
+        # must move them H2D along with the rest of the multimodal payload.
         return [
             "image.pixel_values",
             "video.pixel_values_videos",
             "multimodal_embedding",
+            "mrope_config.mrope_position_ids",
+            "mrope_config.mrope_position_deltas",
         ]
 
     @property