[BugFix][VLM] keep Qwen3-VL MoE inference deepstack order (#34690)

python/sglang/srt/models/qwen3_vl.py

@@ -1152,6 +1152,11 @@ class Qwen3LLMModel(Qwen3Model):
         self.deepstack_embed_to_decoder_layer = range(
             len(config.vision_config.deepstack_visual_indexes)
         )
+        # Use HF deepstack order only if rl_on_policy_target is set;
+        # otherwise, retain original order for inference accuracy.
+        self.use_hf_deepstack_order = (
+            get_exec().deterministic.rl_on_policy_target is not None
+        )
 
     def get_deepstack_embeds(
         self, layer_idx: int, input_deepstack_embeds: Optional[torch.Tensor]
@@ -1196,25 +1201,43 @@ class Qwen3LLMModel(Qwen3Model):
                     hidden_states + residual if residual is not None else hidden_states
                 )
 
-            # SGLang applies residual at the START of the next layer, not at the END like HuggingFace.
+            if self.use_hf_deepstack_order:
-            # See: https://github.com/huggingface/transformers/blob/v5.0.0rc0/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L549
+                # HF-order path (RL on-policy / FSDP). SGLang applies residual at the START of the
-            # To match HF behavior, deepstack must be added AFTER residual: (hidden_states + residual) + deepstack
+                # next layer, so to match HF's (hidden_states + residual) + deepstack, deepstack for
-            # The order matters because addition with different tensors is not associative in practice.
+                # the previous layer is added after residual via post_residual_addition.
-            # Deepstack for prev_layer is applied at the start of current layer via post_residual_addition.
+                deepstack_embeds = self.get_deepstack_embeds(
-            deepstack_embeds = self.get_deepstack_embeds(
+                    layer_idx - 1, input_deepstack_embeds
-                layer_idx - 1, input_deepstack_embeds
+                )
-            )
+                hidden_states, residual = layer(
-            hidden_states, residual = layer(
+                    positions,
-                positions,
+                    hidden_states,
-                hidden_states,
+                    forward_batch,
-                forward_batch,
+                    residual,
-                residual,
+                    post_residual_addition=deepstack_embeds,
-                post_residual_addition=deepstack_embeds,
+                )
-            )
+            else:
+                # Inference path: add deepstack directly to hidden_states at the end of the layer
+                # (original, grounding-correct order).
+                hidden_states, residual = layer(
+                    positions,
+                    hidden_states,
+                    forward_batch,
+                    residual,
+                )
+                if (
+                    input_deepstack_embeds is not None
+                    and layer_idx in self.deepstack_embed_to_decoder_layer
+                ):
+                    sep = self.hidden_size * layer_idx
+                    hidden_states.add_(
+                        input_deepstack_embeds[:, sep : sep + self.hidden_size]
+                    )
 
-        # Handle deepstack for the last processed layer if it exists.
+        # Handle deepstack for the last processed layer (HF-order path only).
-        last_deepstack = self.get_deepstack_embeds(
+        last_deepstack = (
-            self.end_layer - 1, input_deepstack_embeds
+            self.get_deepstack_embeds(self.end_layer - 1, input_deepstack_embeds)
+            if self.use_hf_deepstack_order
+            else None
         )
 
         if not self.pp_group.is_last_rank:

python/sglang/srt/models/qwen3_vl_moe.py

@@ -31,6 +31,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTe
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen3_moe import Qwen3MoeDecoderLayer, Qwen3MoeModel
 from sglang.srt.models.qwen3_vl import Qwen3VLForConditionalGeneration
+from sglang.srt.runtime_context import get_exec
 from sglang.srt.utils.hf_transformers_utils import get_processor
 
 logger = logging.getLogger(__name__)
@@ -58,6 +59,11 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
         # This approach follows the original implementation.
         # TODO: make config of type Qwen3VLMoeConfig, so that we can directly obtain deepstack_visual_indexes.
         self.deepstack_embed_to_decoder_layer = range(3)
+        # Use HF deepstack order only if rl_on_policy_target is set;
+        # otherwise, retain original order for inference accuracy.
+        self.use_hf_deepstack_order = (
+            get_exec().deterministic.rl_on_policy_target is not None
+        )
 
     def get_input_embeddings(self) -> nn.Embedding:
         return self.embed_tokens
@@ -104,25 +110,43 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
                     hidden_states + residual if residual is not None else hidden_states
                 )
 
-            # SGLang applies residual at the START of the next layer, not at the END like HuggingFace.
+            if self.use_hf_deepstack_order:
-            # See: https://github.com/huggingface/transformers/blob/v5.0.0rc0/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L549
+                # HF-order path (RL on-policy / FSDP). SGLang applies residual at the START of the
-            # To match HF behavior, deepstack must be added AFTER residual: (hidden_states + residual) + deepstack
+                # next layer, so to match HF's (hidden_states + residual) + deepstack, deepstack for
-            # The order matters because addition with different tensors is not associative in practice.
+                # the previous layer is added after residual via post_residual_addition.
-            # Deepstack for prev_layer is applied at the start of current layer via post_residual_addition.
+                deepstack_embeds = self.get_deepstack_embeds(
-            deepstack_embeds = self.get_deepstack_embeds(
+                    layer_idx - 1, input_deepstack_embeds
-                layer_idx - 1, input_deepstack_embeds
+                )
-            )
+                hidden_states, residual = layer(
-            hidden_states, residual = layer(
+                    positions,
-                positions,
+                    hidden_states,
-                hidden_states,
+                    forward_batch,
-                forward_batch,
+                    residual,
-                residual,
+                    post_residual_addition=deepstack_embeds,
-                post_residual_addition=deepstack_embeds,
+                )
-            )
+            else:
+                # Inference path: add deepstack directly to hidden_states at the end of the layer
+                # (original, grounding-correct order).
+                hidden_states, residual = layer(
+                    positions,
+                    hidden_states,
+                    forward_batch,
+                    residual,
+                )
+                if (
+                    input_deepstack_embeds is not None
+                    and layer_idx in self.deepstack_embed_to_decoder_layer
+                ):
+                    sep = self.hidden_size * layer_idx
+                    hidden_states.add_(
+                        input_deepstack_embeds[:, sep : sep + self.hidden_size]
+                    )
 
-        # Handle deepstack for the last processed layer if it exists.
+        # Handle deepstack for the last processed layer (HF-order path only).
-        last_deepstack = self.get_deepstack_embeds(
+        last_deepstack = (
-            self.end_layer - 1, input_deepstack_embeds
+            self.get_deepstack_embeds(self.end_layer - 1, input_deepstack_embeds)
+            if self.use_hf_deepstack_order
+            else None
         )
 
         if not self.pp_group.is_last_rank:

test/registered/vlm/test_vision_openai_server_a.py

@@ -4,9 +4,13 @@ python3 -m unittest test_vision_openai_server.TestOpenAIVisionServer.test_mixed_
 python3 -m unittest test_vision_openai_server.TestOpenAIVisionServer.test_multi_images_chat_completion
 """
 
+import base64
+import io
+import re
 import unittest
 
 import openai
+from PIL import Image, ImageDraw
 
 from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_cuda_ci
@@ -28,6 +32,34 @@ from sglang.test.vlm_utils import (
 register_cuda_ci(est_time=560, stage="base-b", runner_config="1-gpu-large")
 
 
+# --- Qwen3-VL grounding regression (deepstack fusion) --------------------------
+# Guards Qwen3MoeLLMModel.forward: deepstack (multi-scale ViT features) injection
+# must keep its original inference order. PR #14636 rerouted it through
+# post_residual_addition (for RL on-policy / FSDP), which is FP-order-sensitive
+# and regresses FP8 visual grounding (the predicted point drifts by ~150+ px).
+_GROUNDING_IMG_SIZE = 1000
+# Target box in pixels; on a 1000x1000 canvas this equals the 0-1000 normalized
+# coordinate, so the check is robust to normalized-vs-pixel conventions.
+_GROUNDING_BOX = (620, 180, 880, 360)  # (x0, y0, x1, y1), center (750, 270)
+_GROUNDING_MARGIN = 60
+_GROUNDING_SYSTEM = (
+    "You are a UI grounding model. Treat the image as a 1000x1000 normalized "
+    "coordinate system with the top-left at (0,0) and the bottom-right at "
+    "(1000,1000). Return the geometric center of the requested element. "
+    "Output ONLY one coordinate in the form (x, y) and nothing else."
+)
+_GROUNDING_COORD_RE = re.compile(r"\(?\s*(\d{1,4})\s*,\s*(\d{1,4})\s*\)?")
+
+
+def _make_grounding_image() -> str:
+    """White canvas with a single red box at _GROUNDING_BOX; base64 data URI."""
+    img = Image.new("RGB", (_GROUNDING_IMG_SIZE, _GROUNDING_IMG_SIZE), (255, 255, 255))
+    ImageDraw.Draw(img).rectangle(_GROUNDING_BOX, fill=(220, 30, 30))
+    buf = io.BytesIO()
+    img.save(buf, format="PNG")
+    return "data:image/png;base64," + base64.b64encode(buf.getvalue()).decode("utf-8")
+
+
 class TestLlavaServer(ImageOpenAITestMixin):
     model = "lmms-lab/llava-onevision-qwen2-0.5b-ov"
 
@@ -52,6 +84,47 @@ class TestQwen3VLServer(ImageOpenAITestMixin, VideoOpenAITestMixin):
         with envs.SGLANG_MM_FEATURE_CACHE_MB.override(512):
             super().setUpClass()
 
+    def test_deepstack_grounding_hits_target_box(self):
+        # Regression guard for the Qwen3-VL MoE deepstack fusion order: the
+        # predicted point must land inside the target box; a deepstack corruption
+        # drifts it out (see PR #14636).
+        client = openai.Client(api_key=self.api_key, base_url=self.base_url)
+        response = client.chat.completions.create(
+            model="default",
+            messages=[
+                {"role": "system", "content": _GROUNDING_SYSTEM},
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image_url",
+                            "image_url": {"url": _make_grounding_image()},
+                        },
+                        {
+                            "type": "text",
+                            "text": "Point at the center of the red rectangle.",
+                        },
+                    ],
+                },
+            ],
+            temperature=0,
+            **(self.get_vision_request_kwargs()),
+        )
+        out = response.choices[0].message.content
+        match = _GROUNDING_COORD_RE.search(out or "")
+        self.assertIsNotNone(match, f"could not parse a coordinate from: {out!r}")
+        x, y = int(match.group(1)), int(match.group(2))
+        x0, y0, x1, y1 = _GROUNDING_BOX
+        inside = (x0 - _GROUNDING_MARGIN <= x <= x1 + _GROUNDING_MARGIN) and (
+            y0 - _GROUNDING_MARGIN <= y <= y1 + _GROUNDING_MARGIN
+        )
+        self.assertTrue(
+            inside,
+            f"grounding output {out!r} -> ({x}, {y}) fell outside target box "
+            f"{_GROUNDING_BOX} (margin {_GROUNDING_MARGIN}); deepstack fusion "
+            f"likely regressed grounding.",
+        )
+
 
 class TestQwen2VLContextLengthServer(CustomTestCase):
     # --context-length 300 is calibrated to this model's mm-token expansion: