fix: unwedge Optimized Cold Start on ComfyUI v0.3.68+

[KarrixLee]

Problem

Runs on machines with Optimized Cold Start (optimized_runner) enabled hang forever in running — no error, no traceback, no node output. Reported log ends at:

Executing prompt: 3702ebf1-e065-46d0-ad64-8d9240b448c2
Created global client session with optimized connection pooling
Status: running

Root cause

ComfyUI v0.3.68 changed PromptExecutor(server, cache_type, cache_size) → cache_args (a dict), and execute_async began dereferencing self.cache_args["ram"]. Our forked prompt_worker constructed PromptExecutor(server) with no cache config, so cache_args stayed None.

From v0.20.0 that dereference sits above execute_async's try:

742  self.add_message("execution_start", ...)          <- last thing the user sees
745  ram_headroom = int(self.cache_args["ram"] * ...)  <- TypeError
750  try:

So it raised right after execution_start, killed the daemon worker thread, and q.task_done() was never called. The traceback went to bare stderr via threading.excepthook instead of through DualHandler — which is why nothing showed up in machine logs.

Broke at v0.3.68 (2025-11-04), not v0.30. On v0.3.68–v0.4.x the same deref sits at the end of the node loop, so it died after the first node; v0.20.0 hoisted it above the try, turning it into "dies instantly".

Note: ComfyUI renumbered releases — v0.3.77 → v0.4.0 → … → v0.30.2. v0.30.0 is ~9 months newer than v0.3.68.

Blast radius

Optimized runner only. prompt_worker lives on _ComfyDeployRunner, whose sole subclass is ComfyDeployRunnerOptimizedImports. ComfyDeployRunner extends BaseComfyDeployRunner and runs ComfyUI's real main.py as a subprocess, so it never drifted.

Changes

my_app.py

• build_prompt_executor() introspects PromptExecutor.__init__, passing cache_type=CLASSIC + complete cache_args on v0.3.68+ and falling back to the bare call on ≤v0.3.67. CLASSIC is pinned deliberately rather than adopting v0.30's RAM_PRESSURE default — CLASSIC is what this runner always effectively used, and RAM-pressure eviction is untested against the model caches held across Modal memory snapshots.
• Guard e.execute() so an executor exception can never again kill the worker: log via logging, emit execution_error so the run is marked FAILED, always call task_done.
• Merge item[5] (sensitive-keys slot added in v0.3.68) back into extra_data + pass process_item to strip it from /history. Without this, comfy.org API-node auth was silently dropped on optimized runners.

session.py

• SESSION_OPTIMIZED_RUNNER as single source of truth. Creation hardcoded False (disabled in 1efd895 when GPU memory snapshotting shipped) while increase_timeout read the real DB flag — so on a flagged machine the session ran on ComfyDeployRunner but the timeout extension was spawned against ComfyDeployRunnerOptimizedImports, a different Modal class, so the container was never extended and the session died early. Sessions stay pinned to non-optimized; this only stops the two call sites disagreeing.

Tests

tests/api/modal/test_prompt_worker.py — 6 passing. AST-extracts the real functions (my_app.py can't be imported: needs modal + generated cd_config) and drives them against fakes mirroring v0.30.x contracts. Includes a test reproducing the original wedge: thread dies, item stuck in currently_running, no error to client. The constructor logic was also checked against the real PromptExecutor signature at every tag v0.3.64 → v0.30.2.

Deploy note

my_app.py is re-copied from the API's current source on every machine deploy, so existing machines need a redeploy to pick this up — deploying the API alone won't unstick them.

🤖 Generated with Claude Code