wandb_gql import fails with W&B SDK >=0.27.1 - ModuleNotFoundError

[TM23-sanji]

Bug: wandb_gql import fails with W&B SDK >=0.27.1

Package: prime-rl
File: prime_rl/utils/monitor/wandb.py
Commit: 256809b (latest main)

Description

prime_rl/utils/monitor/wandb.py imports from wandb_gql import gql — a private, vendored module that was bundled inside the W&B Python SDK. W&B SDK v0.27.1 removed this vendored module as part of PR #11818 (routing GraphQL through wandb-core).

This means a fresh pip install prime-rl resolves to wandb>=0.27.1 and fails at import time:

$ python3 -c "from prime_rl.utils.monitor.wandb import WandbMonitor"
Traceback (most recent call last):
  ...
  File "prime_rl/utils/monitor/wandb.py", line 18, in <module>
    from wandb_gql import gql
ModuleNotFoundError: No module named 'wandb_gql'

This also breaks the orchestrator and rl entrypoints since the monitor is imported at startup:

File "/home/ubuntu/wordle/.venv/bin/orchestrator", line 10, in <module>
    sys.exit(main())
  ...
  File "prime_rl/utils/monitor/wandb.py", line 18, in <module>
    from wandb_gql import gql
ModuleNotFoundError: No module named 'wandb_gql'

Root Cause

The file uses two private/internal APIs from the wandb SDK:

1. from wandb_gql import gql — wandb_gql was a vendored copy of the gql GraphQL library inside wandb/vendor/gql-0.2.0/wandb_gql/. W&B SDK v0.27.1 deleted it.
2. wandb.Api().client.execute() — wandb.Api().client was the gql.Client instance. The modern replacement is wandb.Api()._service_api.execute_graphql().

Both were never public API — they were internal implementation details that happened to be importable.

Fix

Replacement API: wandb.Api()._service_api.execute_graphql() accepts a raw GraphQL query string (no gql() parsing needed) and returns a plain dict — same shape as before.

Changes (3 lines in one file)

File: prime_rl/utils/monitor/wandb.py

1. Remove the import (line 18):

-from wandb_gql import gql

2. Replace gql() parsing + client.execute() with raw string + execute_graphql():

-def list_views(entity: str, project: str) -> list[tuple[str, str]]:
-    query = gql(
-        """
+def list_views(entity: str, project: str) -> list[tuple[str, str]]:
+    query = """
         query Views($entity: String!, $project: String!) {
           project(name: $project, entityName: $entity) {
             allViews(viewType: "project-view") { edges { node { name displayName } } }
           }
         }
-        """
-    )
-    res = wandb.Api().client.execute(query, variable_values={"entity": entity, "project": project})
+    """
+    res = wandb.Api()._service_api.execute_graphql(query, {"entity": entity, "project": project})

That's it. The _service_api attribute exists in wandb SDK v0.27.0+ (earlier versions had client). Since v0.27.1 is required for the vendored wandb_gql removal to be a problem, this is always available.

Verified

• Python 3.12, wandb 0.28.1 (latest)
• WandbMonitor imports successfully
• list_views imports successfully
• Full rl @ config.toml --dry-run passes
• No dependency pins or workarounds needed

Workaround (temporary)

Until this fix lands upstream, pin wandb<0.27.1:

uv pip install "wandb<0.27.1"

Context

This is the same issue that affected wandb-workspaces (fixed in v0.4.2) and the W&B MCP server (temporarily worked around with wandb<0.27.1 cap, then resolved by upgrading wandb-workspaces).

References

• wandb PR #11818: Route Public API GraphQL through wandb-core
• wandb-workspaces#173: Importing reports fails with wandb>=0.27.1
• wandb-mcp-server#93: fix: avoid broken wandb-workspaces resolver combo