From a67a31aa81d23bbd83be9f8a921474d48d73ad36 Mon Sep 17 00:00:00 2001
From: YanbingJiang <yanbing.jiang@intel.com>
Date: Wed, 9 Sep 2026 09:22:58 +0800
Subject: [PATCH] [CPU] Support Qwen3.8 text+video: adding torchcodec, ffmpeg
 and removing pin_memory (#35492)

---
 docker/xeon.Dockerfile                             |  1 +
 python/pyproject_cpu.toml                          |  1 +
 python/sglang/srt/multimodal/processors/qwen_vl.py |  3 ++-
 python/sglang/srt/utils/video_decoder.py           | 14 +++++++++++++-
 4 files changed, 17 insertions(+), 2 deletions(-)

diff --git a/docker/xeon.Dockerfile b/docker/xeon.Dockerfile
index 7c22ebbb2..5d99806e2 100644
--- a/docker/xeon.Dockerfile
+++ b/docker/xeon.Dockerfile
@@ -8,6 +8,7 @@ RUN apt-get update && \
     apt-get full-upgrade -y && \
     DEBIAN_FRONTEND=noninteractive apt-get install --no-install-recommends -y \
     ca-certificates \
+    ffmpeg \
     git \
     curl \
     wget \
diff --git a/python/pyproject_cpu.toml b/python/pyproject_cpu.toml
index 571c89ef5..7ee090217 100644
--- a/python/pyproject_cpu.toml
+++ b/python/pyproject_cpu.toml
@@ -62,6 +62,7 @@ dependencies = [
   "timm==1.0.16",
   "torch==2.12.0",
   "torchaudio==2.11.0",
+  "torchcodec==0.12.0 ; sys_platform != 'linux' or (sys_platform == 'linux' and platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')",
   "torchvision==0.27.0",
   "tqdm",
   "transformers==5.12.1",
diff --git a/python/sglang/srt/multimodal/processors/qwen_vl.py b/python/sglang/srt/multimodal/processors/qwen_vl.py
index eba723827..e911611d5 100644
--- a/python/sglang/srt/multimodal/processors/qwen_vl.py
+++ b/python/sglang/srt/multimodal/processors/qwen_vl.py
@@ -268,7 +268,8 @@ async def preprocess_video(
         [resized_height, resized_width],
         interpolation=InterpolationMode.BILINEAR,
     )
-    video = video.pin_memory()
+    if not is_cpu():
+        video = video.pin_memory()
     video_metadata = {
         "fps": video_fps,
         "duration": total_frames / video_fps,
diff --git a/python/sglang/srt/utils/video_decoder.py b/python/sglang/srt/utils/video_decoder.py
index 3b398b58e..13cfff5cd 100644
--- a/python/sglang/srt/utils/video_decoder.py
+++ b/python/sglang/srt/utils/video_decoder.py
@@ -33,6 +33,13 @@ def _try_cuda_backend() -> bool:
     return _cuda_backend_enabled
 
 
+def _is_cpu_engine() -> bool:
+    # Lazy import to avoid circular dependency issues and unnecessary imports on module load
+    from sglang.srt.utils.common import is_cpu
+
+    return is_cpu()
+
+
 class VideoDecoderWrapper:
     """Unified video decoder that uses torchcodec when available, decord as fallback.
 
@@ -140,10 +147,13 @@ class VideoDecoderWrapper:
 
         if _BACKEND == "torchcodec":
             batch = self._decoder.get_frames_at(indices)
+            if _is_cpu_engine():
+                return batch.data
             return batch.data if batch.data.is_cuda else batch.data.pin_memory()
         else:
             arr = self._decoder.get_batch(indices).asnumpy()
-            return torch.from_numpy(arr).pin_memory()
+            output = torch.from_numpy(arr)
+            return output if _is_cpu_engine() else output.pin_memory()
 
     def _parallel_decode(self, indices, num_threads):
         """Decode frames using multiple VideoDecoder instances in parallel threads."""
@@ -177,6 +187,8 @@ class VideoDecoderWrapper:
                 results[idx] = future.result()
 
         output = torch.cat(results, dim=0)
+        if _is_cpu_engine():
+            return output
         return output if output.is_cuda else output.pin_memory()
 
     @property