[Cherith1222]

Why

vectorize=True calls the log probability once for the whole ensemble. A return of (log_prob, blob) was parsed as two walker results, so the log probability had length 2 and sampling raised incompatible input dimensions. The ensemble return is now split into one value per walker before that parser runs.

What changed

• compute_log_prob splits a vectorized (log_prob, *blobs) tuple into per-walker returns.
• Tests cover the example from Trying to use Blobs with the vectorize=True keyword returns a ValueError("incompatible input dimensions") #516, two same-shaped blobs, and a short return.

Scope

• Covers vectorize=True when the function returns a tuple.
• A vectorized function that returns only the log probability is unchanged, as is the non-vectorized parser.

Blast Radius

Only the vectorized path that returns blobs. A bare array of log probabilities still takes the previous path.

Verification

• Before the fix, the Trying to use Blobs with the vectorize=True keyword returns a ValueError("incompatible input dimensions") #516 example raised ValueError: incompatible input dimensions. compute_log_prob returned log-probability shape (2,) and blob shape (2, 31).
• After the fix, that example finishes and prints chain shape (10, 32, 5) and blob shape (10, 32).
• python -m pytest src/emcee/tests/unit passed 66 tests and skipped 6. The skips are HDF5 backends, because h5py is not installed here.

Fixes #516

Drafted with Cursor.

[Cherith1222]

Closing this. #463 already edits compute_log_prob, so this change to the same function should wait.

Drafted with Cursor.