All LoRA sampler checkpoints permanently hang on sample_async() — base model works fine

[xiayang851023]

Bug Description

Calling sample_async() with any LoRA checkpoint hangs indefinitely (times out after 60s), while base model sampling works normally (~2.4s response time).

Environment

• Tinker SDK: 0.18.2 (latest from PyPI)
• Base model: Qwen/Qwen3-4B-Instruct-2507
• LoRA rank: 16
• 5 training runs, 31 checkpoints total, all show "Never" expiry in console

Reproduction

from tinker import ServiceClient, ModelInput, SamplingParams
import asyncio

async def test():
    sc = ServiceClient()
    
    # Base model — works, ~2.4s
    sampler = await sc.create_sampling_client_async(base_model="Qwen/Qwen3-4B-Instruct-2507")
    
    # Any LoRA checkpoint — permanently hangs, 60s timeout
    sampler = await sc.create_sampling_client_async(
        model_path="tinker://0ac02541-f609-540f-860e-0bf885b7292e:train:0/sampler_weights/dpo_final"
    )
    # create_sampling_client_async() returns OK (0.5s)
    # but sample_async() never returns — times out after 60s with CancelledError

Tested checkpoints (all hang)

• tinker://0ac02541-f609-540f-860e-0bf885b7292e:train:0/sampler_weights/dpo_final (created 23 min ago, 136 MB)
• tinker://a735527d-a478-59b1-b510-e3b0211f4989:train:0/sampler_weights/dpo_final (created 2 weeks ago, 136 MB)
• All show "Never" expiry in the Tinker console

Call stack at timeout

sample_async() → _sample_async_impl() → _APIFuture.result_async()
  → asyncio.wait_for(future, timeout=60) → CancelledError

No server response is returned — not a 400/404/500 error. The request appears to be queued server-side and never processed.

Related Issues

• GitHub Issue #234 ("Stuck when sampling") was closed on 2026-03-27 by @YujiaBao with comment "Closing for now — if still experiencing, please reopen." The issue described the same 7200s permanent hang with sample_async() in a CUA RL workflow. That fix does not appear to have resolved this issue.

Impact

All inference via LoRA checkpoints is blocked. DPO/SFT training works fine and checkpoints save successfully, but sampling from trained checkpoints is completely unusable.