diff --git a/python/sglang/srt/entrypoints/openai/encoding_dsv41.py b/python/sglang/srt/entrypoints/openai/encoding_dsv41.py
index c04d8059d8..c8eb19344e 100644
--- a/python/sglang/srt/entrypoints/openai/encoding_dsv41.py
+++ b/python/sglang/srt/entrypoints/openai/encoding_dsv41.py
@@ -196,11 +196,18 @@ def render_reasoning_effort(
def find_last_user_index(messages: List[Dict[str, Any]]) -> int:
"""Mid-conversation system messages also count as user messages here;
they trigger the assistant generation header.
+ User messages containing only tool results do not start a new turn.
"""
last_user_index = -1
for idx in range(len(messages) - 1, -1, -1):
role = messages[idx].get("role")
if role in ["user", "developer"] or (role == "system" and idx > 0):
+ content_blocks = messages[idx].get("content_blocks")
+ if content_blocks and not any(
+ block.get("type") in ("text", "image", "image_url")
+ for block in content_blocks
+ ):
+ continue
last_user_index = idx
break
return last_user_index
@@ -434,10 +441,13 @@ def merge_tool_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
if role == "tool":
# Convert tool message to a user message with tool_result block
+ content = msg.get("content", "")
+ if msg.get("content_blocks") and not isinstance(msg.get("content"), str):
+ content = msg["content_blocks"]
tool_block = {
"type": "tool_result",
"tool_use_id": msg.get("tool_call_id", ""),
- "content": msg.get("content", ""),
+ "content": content,
}
# Merge into previous message if it's already a user (merged tool)
if (
@@ -456,7 +466,12 @@ def merge_tool_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
elif role == "user":
content_blocks = msg.get("content_blocks")
if content_blocks is None:
- content_blocks = [{"type": "text", "text": msg.get("content", "")}]
+ content = msg.get("content", "")
+ content_blocks = (
+ content
+ if isinstance(content, list)
+ else [{"type": "text", "text": content}]
+ )
if (
merged
and merged[-1].get("role") == "user"
@@ -638,14 +653,6 @@ def _encode_messages_text(
) -> str:
"""Encode preprocessed (text-only) messages into the V4.1 prompt format."""
context = context if context else []
-
- # Preprocess: merge tool messages and sort tool results
- messages = merge_tool_messages(messages)
- messages = sort_tool_results_by_call_order(context + messages)[len(context) :]
- if context:
- context = merge_tool_messages(context)
- context = sort_tool_results_by_call_order(context)
-
full_messages = context + messages
prompt = bos_token if add_default_bos_token and len(context) == 0 else ""
@@ -689,7 +696,14 @@ def encode_messages(
Returns the prompt string, or ``(prompt, {"images": [...]})`` when
``return_multi_modal_data`` is set; the image records are in prompt order.
"""
- context = context or []
+ # Validate before merging can discard message-level textual fields.
+ for msg in (context or []) + messages:
+ _validate_no_image_sp_tokens(msg)
+
+ # Collect images only after tool results are in their final prompt order.
+ context = merge_tool_messages(context or [])
+ messages = merge_tool_messages(messages)
+ messages = sort_tool_results_by_call_order(context + messages)[len(context) :]
processed_context, _ = process_image_messages(context) if context else ([], [])
processed_messages, images = process_image_messages(messages)
prompt = _encode_messages_text(
diff --git a/test/registered/unit/entrypoints/openai/test_serving_chat.py b/test/registered/unit/entrypoints/openai/test_serving_chat.py
index 81b931f5f4..d28d6663c0 100644
--- a/test/registered/unit/entrypoints/openai/test_serving_chat.py
+++ b/test/registered/unit/entrypoints/openai/test_serving_chat.py
@@ -11,6 +11,7 @@ from sglang.test.test_utils import CustomTestCase, enter_override, maybe_stub_sg
maybe_stub_sgl_kernel() # must precede any import that pulls in sgl_kernel
import asyncio
+import copy
import gc
import json
import tempfile
@@ -2992,6 +2993,152 @@ class ServingChatTestCase(CustomTestCase):
with self.assertRaisesRegex(ValueError, "dsv4_reasoning_effort_profile"):
OpenAIServingChat(tm, TemplateManager())
+ def test_dsv41_rejects_image_placeholder_in_text(self):
+ from sglang.srt.entrypoints.openai import encoding_dsv41
+
+ for role in ("user", "tool"):
+ with self.subTest(role=role):
+ with self.assertRaises(ValueError):
+ encoding_dsv41.encode_messages(
+ [
+ {
+ "role": role,
+ "content": f"hi {encoding_dsv41.IMAGE_PLACEHOLDER}",
+ }
+ ],
+ thinking_mode="chat",
+ )
+
+ def test_dsv41_tool_result_images_follow_prompt_order(self):
+ """Out-of-order tool results are sorted by call order; images must follow."""
+ from sglang.srt.entrypoints.openai import encoding_dsv41
+
+ image = encoding_dsv41.IMAGE_PLACEHOLDER
+ history = [
+ {
+ "role": "user",
+ "content": [
+ {"type": "image_url", "image_url": {"url": "history.png"}},
+ ],
+ },
+ {
+ "role": "assistant",
+ "tool_calls": [
+ {
+ "id": name,
+ "type": "function",
+ "function": {"name": name, "arguments": "{}"},
+ }
+ for name in ("a", "b")
+ ],
+ },
+ ]
+ results = [
+ {
+ "type": "tool_result",
+ "tool_use_id": name,
+ "content": [
+ {"type": "text", "text": f"result {name}"},
+ {"type": "image_url", "image_url": {"url": f"{name}.png"}},
+ ],
+ }
+ for name in ("b", "a")
+ ]
+
+ for input_format in (
+ "content_blocks",
+ "tool_messages",
+ "tool_content_blocks",
+ "context",
+ ):
+ with self.subTest(input_format=input_format):
+ if input_format in ("tool_messages", "tool_content_blocks"):
+ content_field = (
+ "content_blocks"
+ if input_format == "tool_content_blocks"
+ else "content"
+ )
+ messages = history + [
+ {
+ "role": "tool",
+ "tool_call_id": block["tool_use_id"],
+ content_field: block["content"],
+ }
+ for block in results
+ ]
+ context = []
+ elif input_format == "context":
+ messages = [{"role": "user", "content_blocks": results}]
+ context = history
+ else:
+ messages = history + [{"role": "user", "content_blocks": results}]
+ context = []
+ original = copy.deepcopy((messages, context))
+
+ prompt, media = encoding_dsv41.encode_messages(
+ messages,
+ thinking_mode="chat",
+ context=context,
+ return_multi_modal_data=True,
+ )
+
+ self.assertTrue(
+ prompt.endswith(
+ f"<|User|>result a\n\n{image}\n\n"
+ f"result b\n\n{image}"
+ "<|Assistant|>"
+ ),
+ prompt,
+ )
+ expected_urls = (
+ ["a.png", "b.png"] if context else ["history.png", "a.png", "b.png"]
+ )
+ self.assertEqual(
+ media,
+ {"images": [{"url": url} for url in expected_urls]},
+ )
+ self.assertEqual((messages, context), original)
+
+ def test_dsv41_drop_thinking_without_tools(self):
+ """Tool results do not start a new turn, so current-turn reasoning is kept."""
+ from sglang.srt.entrypoints.openai import encoding_dsv41
+
+ messages = [
+ {"role": "user", "content": "q1"},
+ {"role": "assistant", "content": "a1", "reasoning_content": "r1"},
+ {"role": "user", "content": "q2"},
+ {
+ "role": "assistant",
+ "reasoning_content": "r2",
+ "tool_calls": [
+ {
+ "id": "c1",
+ "type": "function",
+ "function": {"name": "lookup", "arguments": "{}"},
+ }
+ ],
+ },
+ {"role": "tool", "tool_call_id": "c1", "content": "value"},
+ ]
+ prompt = encoding_dsv41.encode_messages(
+ messages, thinking_mode="thinking", drop_thinking=True
+ )
+ self.assertIn(
+ "<|User|>q1<|Assistant|>a1<|end▁of▁sentence|>", prompt
+ )
+ self.assertNotIn("r1", prompt)
+ self.assertIn("r2", prompt)
+ self.assertTrue(
+ prompt.endswith("value<|Assistant|>")
+ )
+
+ next_turn = encoding_dsv41.encode_messages(
+ messages + [{"role": "user", "content": "q3"}],
+ thinking_mode="thinking",
+ )
+ self.assertNotIn("r2", next_turn)
+ self.assertTrue(next_turn.endswith("q3<|Assistant|>"))
+
def test_streaming_abort_yields_error(self):
"""Test that an abort finish reason during streaming correctly yields an error and stops."""
err_msg = "Aborted by scheduler"