Metric Logging updates 4/N - better actor name #351
Conversation
…estamp_logging_diff2
meta-cla
bot
added
the
CLA Signed
…estamp_logging_diff3
|
could you add some more comments in the description about what this PR is enabling? |
yes! sorry. I should have marked as a draft. I am doing a 2.5/4.0 before i ask you to review it |
felipemello1
changed the title
felipemello1
changed the title
| self.timestamp = datetime.now(pytz.UTC).timestamp() | ||
|
|
||
|
|
||
| def get_actor_name_with_rank() -> str: |
There was a problem hiding this comment.
moved to observability/utils.py
| if process_name is None: | ||
| process_name = detect_actor_name_from_call_stack() |
There was a problem hiding this comment.
get name here and pass it to
local_fetcher_actor = proc.spawn(
"local_fetcher_actor", LocalFetcherActor, global_logger, process_name
)
this function is called in provisioner.py, and thats how we get the process_name for every wandb run
src/forge/observability/utils.py
Outdated
| logger = logging.getLogger(__name__) | ||
|
|
||
|
|
||
| def detect_actor_name_from_call_stack() -> str: |
There was a problem hiding this comment.
main file to review
|
in the near future the mesh might hold a name. When this happens, we can delete the function to use the call stack and just get it from the mesh. The rest of the PR stands. |
…estamp_logging_diff3
felipemello1
changed the title
| # Spawn services first (triggers registrations via provisioner hook) | ||
| trainer = await TrainActor.options(**service_config).as_service() | ||
| generator = await GeneratorActor.options(**service_config).as_service() | ||
| trainer = await TrainActor.options( |
There was a problem hiding this comment.
Dumb q but what are we doing with this toy_rl stuff anyways? (Like is there some reason we're still keeping it around?)
There was a problem hiding this comment.
i still find it useful, e.g. i dont need to have gpus available to test it. Eventually this should become an integration test.
src/forge/observability/metrics.py
Outdated
| Timestamp is automatically set to current EST time if not provided. | ||
| Args: |
There was a problem hiding this comment.
nit: give an actual docstring here (otherwise i can just read this info 5 lines below)
| if process_name is None: | ||
| ctx = context() | ||
| process_name = ctx.actor_instance.actor_id.actor_name |
There was a problem hiding this comment.
a small thing but it's not immediately clear why we do this here vs get_proc_name_with_rank in other places. (after looking at the code i think it's a global vs local thing, but imo this could be more clearly documented)
There was a problem hiding this comment.
get_proc_name_with_rank returns it with replica id and rank. Here i just want the process name, so i would have to parse it :/
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## main #351 +/- ##
==========================================
- Coverage 73.68% 65.14% -8.54%
==========================================
Files 81 82 +1
Lines 7729 7850 +121
==========================================
- Hits 5695 5114 -581
- Misses 2034 2736 +702 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
Co-authored-by: Felipe Mello <[email protected]>
)" This reverts commit 1f45470.
…#351)" (meta-pytorch#429) This reverts commit 633b219.
…#351)" (meta-pytorch#429) This reverts commit 633b219.
Co-authored-by: Felipe Mello <[email protected]>
Co-authored-by: Felipe Mello <[email protected]>
5 participants
When logging per rank, we need good naming to make it easier to debug. Now provisioner.py::get_proc_mesh provides a mesh_name (#378), so i am just using it.
User can also pass a process name as input. Thats how we get the
Controllername.The process_name then goes:
LocalFetcherActor->MetricCollector->backend.init-->wandb.init(name)