scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-10-24 05:45:11 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 72

Merge 2e9850f860 into 24123269cc

Browse source
This commit is contained in:
Xuetao Song 2026-07-27 02:03:39 +08:00 • committed by GitHub
commit 19b768ff13
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 60 additions and 27 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

83
litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py
View file

@ -28,7 +28,7 @@ from litellm.types.llms.anthropic import (
UsageDelta,
UsageIteration,
)
from litellm.types.utils import AdapterCompletionStreamWrapper
from litellm.types.utils import AdapterCompletionStreamWrapper, StreamingChoices
if TYPE_CHECKING:
from litellm.types.utils import ModelResponseStream
@ -393,26 +393,28 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
if compaction_event is not None:
return compaction_event
if self.sent_content_block_start is False:
self.sent_content_block_start = True
self.sent_content_block_finish = False
self.chunk_queue.append(
{
"type": "content_block_start",
"index": self.current_content_block_index,
"content_block": {"type": "text", "text": ""},
}
)
return self.chunk_queue.popleft()
for chunk in self.completion_stream:
if chunk == "None" or chunk is None:
raise Exception
should_start_new_block = self._should_start_new_content_block(chunk)
if should_start_new_block:
if should_start_new_block and self.sent_content_block_start is True:
self._increment_content_block_index()
if self.sent_content_block_start is False:
if not self._chunk_has_substantial_content(chunk):
continue
self.sent_content_block_start = True
self.sent_content_block_finish = False
self.chunk_queue.append(
{
"type": "content_block_start",
"index": self.current_content_block_index,
"content_block": self.current_content_block_start,
}
)
return self.chunk_queue.popleft()
# applied_edits only needs to flow to the final message_delta
# (when finish_reason is set); skip threading it through every
# intermediate chunk. For the hold-and-merge path below,
@ -616,16 +618,7 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
return compaction_event
if self.sent_content_block_start is False:
self.sent_content_block_start = True
self.sent_content_block_finish = False
self.chunk_queue.append(
{
"type": "content_block_start",
"index": self.current_content_block_index,
"content_block": {"type": "text", "text": ""},
}
)
return self.chunk_queue.popleft()
pass
async for chunk in self.completion_stream:
if chunk == "None" or chunk is None:
@ -633,9 +626,23 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
# Check if we need to start a new content block
should_start_new_block = self._should_start_new_content_block(chunk)
if should_start_new_block:
if should_start_new_block and self.sent_content_block_start is True:
self._increment_content_block_index()
if self.sent_content_block_start is False:
if not self._chunk_has_substantial_content(chunk):
continue
self.sent_content_block_start = True
self.sent_content_block_finish = False
self.chunk_queue.append(
{
"type": "content_block_start",
"index": self.current_content_block_index,
"content_block": self.current_content_block_start,
}
)
return self.chunk_queue.popleft()
# applied_edits only needs to flow to the final message_delta
# (when finish_reason is set); skip threading it through every
# intermediate chunk. For the hold-and-merge path below,
@ -914,6 +921,12 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
tool_block["name"] = original_name
if block_type != self.current_content_block_type:
if (
block_type == "text"
and self.current_content_block_type == "thinking"
and not self._chunk_has_substantial_content(chunk)
):
return False
self.current_content_block_type = block_type
self.current_content_block_start = content_block_start
return True
@ -932,3 +945,23 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
return True
return False
@staticmethod
def _chunk_has_substantial_content(chunk: "ModelResponseStream") -> bool:
"""Return True when the chunk carries content that should determine
the initial content block type (non-empty text, tool_calls,
thinking_blocks, or reasoning_content). Role-only chunks and empty
deltas return False."""
for choice in chunk.choices:
if not hasattr(choice, "delta"):
continue
if choice.delta.tool_calls is not None and len(choice.delta.tool_calls) > 0:
return True
if getattr(choice.delta, "content", None) and len(choice.delta.content) > 0:
return True
if isinstance(choice, StreamingChoices):
if hasattr(choice.delta, "thinking_blocks") and choice.delta.thinking_blocks:
return True
if getattr(choice.delta, "reasoning_content", None):
return True
return False

4
litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py
View file

@ -1396,7 +1396,7 @@ class LiteLLMAnthropicMessagesAdapter:
thinking_blocks = choice.delta.thinking_blocks or []
if len(thinking_blocks) > 0:
thinking_block = thinking_blocks[0]
if thinking_block["type"] == "thinking":
if thinking_block["type"] in ("thinking", "redacted_thinking"):
thinking = thinking_block.get("thinking") or ""
signature = thinking_block.get("signature") or ""
@ -1443,7 +1443,7 @@ class LiteLLMAnthropicMessagesAdapter:
thinking_blocks = choice.delta.thinking_blocks or []
if len(thinking_blocks) > 0:
for thinking_block in thinking_blocks:
if thinking_block["type"] == "thinking":
if thinking_block["type"] in ("thinking", "redacted_thinking"):
thinking = thinking_block.get("thinking") or ""
signature = thinking_block.get("signature") or ""