Merge pull request #151 from macrocosm-os/dev

Release 4.1.1

constants/__init__.py

@@ -64,7 +64,7 @@
 
 # Minimum percent of weight on a vali for a miner to be considered a top miner.
 # Since there can be multiple competitions at different reward percentages we can't just check biggest.
-WEIGHT_SYNC_MINER_MIN_PERCENT = 0.10
+WEIGHT_SYNC_MINER_MIN_PERCENT = 0.05
 
 # A mapping of block numbers to the supported model types as of that block.
 ALLOWED_MODEL_TYPES_1 = {
@@ -181,8 +181,7 @@
                 CompetitionId.B7_MODEL,
                 MODEL_CONSTRAINTS_BY_COMPETITION_ID[CompetitionId.B7_MODEL],
                 0.57,
-            )
-
+            ),
         ],
     ),
 ]
@@ -215,7 +214,7 @@
 sample_unpack_block = BLOCK_3B_7BSTAR_UNPACK
 
 # validators number of pages to eval over miners on each step.
-pages_per_eval_unpack = 5 # With sample unpacking
+pages_per_eval_unpack = 5  # With sample unpacking
 pages_per_eval_pack = 18
 
 timestamp_epsilon_experiment_start_block = BLOCK_3B_7BSTAR_UNPACK
@@ -227,9 +226,8 @@
 # validator eval batch min to keep for next loop.
 sample_min = 5
 # Max number of uids that can be either pending eval or currently being evaluated.
-updated_models_limit = 15
-# validator incentive threshold to prioritize updates. All incentives add up to 1.
-update_priority_incentive_threshold = 0.01
+# We allow the sample_min per competition + 10 additional models to be held at any one time.
+updated_models_limit = sample_min * len(MODEL_CONSTRAINTS_BY_COMPETITION_ID) + 10
 # time required between updates to the chain.
 chain_update_cadence = dt.timedelta(minutes=20)
 # time required between retrying evaluation of a stale model. (First retry will be immediate).

neurons/validator.py

@@ -911,11 +911,13 @@ async def run_step(self):
         # Prioritize models for keeping up to the sample_min for the next eval loop.
         # If the model has any significant weight, prioritize by weight with greater weights being kept first.
         # Then for the unweighted models, prioritize by win_rate.
+        # Use the competition weights from the tracker which also handles moving averages.
+        tracker_competition_weights = self.competition_tracker.get_competition_weights(competition.id)
         model_prioritization = {
             uid: (
                 # Add 1 to ensure it is always greater than a win rate.
-                1 + self.weights[uid].item()
-                if self.weights[uid].item() >= 0.001
+                1 + tracker_competition_weights[uid].item()
+                if tracker_competition_weights[uid].item() >= 0.001
                 else wr
             )
             for uid, wr in win_rate.items()