Skip to content

[OAI Server Refactor] [ChatCompletions & Completions] Support Return Hidden States #7260

@CatherineSue

Description

@CatherineSue

Points: 2-3 days

Description: Support return hidden states in serving_chat.py and serving_completions.py

Deliverables:

  • Complete task below
  • Add UTs

Task: Missing hidden states **

Current streaming logic in adapter.py#L691-699:

hidden_states = None
if isinstance(request, list) and request[idx].return_hidden_states:
hidden_states = ret_item["meta_info"].get("hidden_states", None)
elif (not isinstance(request, list)) and request.return_hidden_states:
hidden_states = ret_item["meta_info"].get("hidden_states", None)
if hidden_states is not None:
hidden_states = (
hidden_states[-1] if hidden_states and len(hidden_states) > 1 else []
)

Current logic in adapter.py#L1370-L1382:

if isinstance(request, list) and request[idx].return_hidden_states:
include_hidden_states = True
elif not isinstance(request, list) and request.return_hidden_states:
include_hidden_states = True
else:
include_hidden_states = False
if include_hidden_states and ret_item["meta_info"].get("hidden_states", None):
hidden_states = ret_item["meta_info"]["hidden_states"]
hidden_states = (
hidden_states[-1] if hidden_states and len(hidden_states) > 1 else []
)
else:
hidden_states = None

  • Need to support it in serving_chat.py and serving_completions.py .

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions