From dc4380e33ac9745f2fe73dfd0910f95b6762185b Mon Sep 17 00:00:00 2001
From: Michael <13900043+michaelzhang-ai@users.noreply.github.com>
Date: Tue, 10 Mar 2026 17:04:35 -0700
Subject: [PATCH] [AMD] [DeepSeek-OCR-2 Day 0] Enable DeepSeek-OCR-2 on AMD
 GPUs and add nightly test (#19732)

---
 python/sglang/srt/models/deepseek_ocr.py      |  7 ++++---
 .../multimodal/processors/base_processor.py   |  2 ++
 .../accuracy/mi30x/test_vlms_mmmu_eval_amd.py | 19 +++++++++++++++++--
 3 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/python/sglang/srt/models/deepseek_ocr.py b/python/sglang/srt/models/deepseek_ocr.py
index fa334af7b..f87c06d32 100644
--- a/python/sglang/srt/models/deepseek_ocr.py
+++ b/python/sglang/srt/models/deepseek_ocr.py
@@ -125,8 +125,9 @@ def isin_list(
     elements: torch.Tensor,
     test_elements_list: list[int],
 ) -> torch.Tensor:
-    test_elements = torch.tensor(test_elements_list, pin_memory=True).to(
-        device=elements.device, non_blocking=True
+    use_pin = torch.cuda.is_available() and not getattr(torch.version, "hip", None)
+    test_elements = torch.tensor(test_elements_list, pin_memory=use_pin).to(
+        device=elements.device, non_blocking=use_pin
     )
 
     return torch.isin(elements, test_elements)
@@ -1685,7 +1686,7 @@ class DeepseekOCRForCausalLM(nn.Module):
 
         images_crop = (
             torch.stack([item.images_crop for item in mm_items], dim=0)
-            .type(torch.long)
+            .type(target_dtype)
             .to(device=pixel_values.device)
         )
         images_spatial_crop = (
diff --git a/python/sglang/srt/multimodal/processors/base_processor.py b/python/sglang/srt/multimodal/processors/base_processor.py
index 4cc4cfb50..36861d3a3 100644
--- a/python/sglang/srt/multimodal/processors/base_processor.py
+++ b/python/sglang/srt/multimodal/processors/base_processor.py
@@ -210,6 +210,8 @@ class BaseMultimodalProcessor(ABC):
             "image_emb_mask": Modality.IMAGE,
             "images_spatial_crop": Modality.IMAGE,
             "images_crop": Modality.IMAGE,
+            "has_local_crops": Modality.IMAGE,
+            "has_images": Modality.IMAGE,
             "tgt_size": Modality.IMAGE,
             "image_grid_hws": Modality.IMAGE,
             "aspect_ratio_ids": Modality.IMAGE,
diff --git a/test/registered/amd/accuracy/mi30x/test_vlms_mmmu_eval_amd.py b/test/registered/amd/accuracy/mi30x/test_vlms_mmmu_eval_amd.py
index e896c6c26..6cb9b8807 100644
--- a/test/registered/amd/accuracy/mi30x/test_vlms_mmmu_eval_amd.py
+++ b/test/registered/amd/accuracy/mi30x/test_vlms_mmmu_eval_amd.py
@@ -8,7 +8,7 @@ VLMs tested here:
 - Qwen VL series (Qwen2-VL-7B, Qwen2.5-VL-7B, Qwen3-VL-30B)
 - InternVL2 series (InternVL2_5-2B)
 - MiniCPM series (MiniCPM-v-2_6, MiniCPM-o-2_6)
-- DeepSeek VL series (deepseek-vl2-small, Janus-Pro-7B)
+- DeepSeek VL series (deepseek-vl2-small, Janus-Pro-7B, DeepSeek-OCR-2)
 - Kimi VL (Kimi-VL-A3B-Instruct)
 - MiMo VL (MiMo-VL-7B-RL)
 - GLM VL (GLM-4.1V-9B-Thinking)
@@ -116,13 +116,28 @@ AMD_VLM_MODELS = [
         "accuracy_threshold": 0.27,
         "extra_args": ["--trust-remote-code"],
     },
+    # DeepSeek-OCR-2 - last to avoid memory pressure on subsequent models
+    {
+        "model_path": "deepseek-ai/DeepSeek-OCR-2",
+        "tp_size": 1,
+        "accuracy_threshold": 0.25,
+        "extra_args": [
+            "--trust-remote-code",
+            "--disable-cuda-graph",
+            "--mem-fraction-static",
+            "0.70",
+            "--max-total-tokens",
+            "16384",
+        ],
+    },
 ]
 
-# Models that need special handling on AMD (MoE models)
+# Models that need triton attention on AMD (aiter pa_ragged JIT compilation OOMs)
 TRITON_ATTENTION_MODELS = {
     # "deepseek-ai/deepseek-vl2-small",
     # "Qwen/Qwen3-VL-30B-A3B-Instruct",
     # "moonshotai/Kimi-VL-A3B-Instruct",
+    "deepseek-ai/DeepSeek-OCR-2",
 }
 
 # Models known to fail on AMD - exclude from testing