mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2026-06-17 21:39:45 +08:00
85a12d0a83
Adversarial review caught that a LIFO stack keyed by ``prompt_id`` still
mis-attributes events when queue execution order differs from registration
order: a second submission with the same ``prompt_id`` lands on top of the
stack, so the first prompt's events read the wrong workflow_id while it
runs, and the first's ``unregister`` then pops the second prompt's entry.
Replace the stack with an internal monotonic token. ``post_prompt``
registers metadata and stashes the returned token on ``extra_data`` under
``PROMPT_METADATA_TOKEN_KEY``. ``main.py``'s queue worker pulls the token
out, pins it on ``PromptServer.active_prompt_metadata_token`` for the
prompt's execution, and clears + unregisters in ``finally``. The merge in
``send_sync`` reads the active token, so each prompt's events are merged
with its own metadata regardless of ``prompt_id`` collisions.
- comfy_execution/metadata.py: ``merge_prompt_metadata`` now takes an
active token; registry is ``dict[int, PromptMetadata]``; new
``PROMPT_METADATA_TOKEN_KEY`` constant for the extra_data carrier.
- server.py: ``register_prompt_metadata`` returns a token (or ``None``
when no metadata applies); ``unregister`` takes a token;
``get_active_prompt_metadata`` snapshots the pinned entry.
- main.py: pops the token from extra_data, pins on the server, clears
after the terminal "executing: {node: None}" send.
- execution.py ``PromptQueue``: wipe_queue / delete_queue_item now
unregister by token extracted from each item's extra_data.
- comfy_execution/progress.py: reads workflow_id via
``get_active_prompt_metadata`` rather than per-prompt_id lookup.
- tests: unit tests updated for the token signature, plus a real E2E
test (test_prompt_metadata_e2e.py) that instantiates the actual
PromptServer and verifies same-prompt_id-different-workflow_id
submissions don't cross-attribute.
Verified end-to-end against a live ComfyUI server: two submissions with
identical client-supplied prompt_id but different workflow_id each emit
their full execution event stream (execution_start, execution_cached,
executing, executed, execution_success, progress_state, terminal executing)
with the correct workflow_id top-level. 68 / 68 tests pass.
|
||
|---|---|---|
| .. | ||
| test_cache_control.py | ||
| test_prompt_metadata_e2e.py | ||
| test_prompt_metadata.py | ||