Fix loading configured self-play checkpoints

[bagel786]

Summary

• Load base.load_model_path before training starts.
• Seed the initial self-play pool checkpoint from the loaded policy instead of random initialization.
• Synchronize the asynchronous rollout actor after loading the primary policy.

Fixes #618.

Related: #632 was an earlier closed attempt from another contributor. This PR independently applies the same narrow load-order fix to the current 5c head.

Validation

• git diff --check passed.
• Static inspection verifies that the configured policy load precedes the initial self-play checkpoint save.
• Static inspection verifies that asynchronous actor parameters are copied from the loaded primary policy.
• CUDA build/runtime validation was not available on this machine and should run in CI or on a CUDA host before merge.

---

Continues #673 — the fork repo backing that PR was accidentally deleted from this account, and GitHub blocks reopening a PR whose submitting repository was deleted (a restore request is being filed with GitHub Support). This PR resumes the identical work from the same commit (f251da22072a72a6a8238db12606312b7fa857a2); review discussion continues on the original PR.

[Copilot]

🔵 Needs a closer look

CUDA build/runtime validation was unavailable and requires final CI or human verification.

Pull request overview

Fixes self-play initialization by loading configured checkpoints before training and synchronizing asynchronous rollout actors.

Changes:

• Loads base.load_model_path before self-play setup.
• Seeds the initial self-play checkpoint from loaded weights.
• Synchronizes async actor parameters after policy loading.

File summaries

File	Description
src/pufferl.cu	Implements checkpoint loading and asynchronous actor synchronization.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.