[Fix] dsv41: Preserve tool-result image order and current-turn thinking (#44751)
This commit is contained in:
@@ -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(
|
||||
|
||||
Reference in New Issue
Block a user