Merge pull request #769 from macrocosm-os/staging

v2.19.8:
  - Disable tqdm for logits.
  - Fix weight synchronization.
  - Fix rewards averaging.
  - Make MSRv2 rewards linear.
  - Add weight setter unit tests.

Author: dbobrenko

.gitignore

@@ -8,6 +8,9 @@ __pycache__/
 *.npy
 *.npz
 prompting/storage/
+validator_rewards.jsonl
+test_validator_rewards.jsonl
+uid_tracker.sqlite*
 
 # C extensions
 *.so

gpu_container/vllm/reproducible_vllm.py

@@ -3,8 +3,11 @@
 
 import numpy as np
 import torch
+from tqdm import tqdm
 from vllm import LLM, SamplingParams
 
+tqdm.disable = True
+
 
 class ReproducibleVLLM:
     def __init__(

prompting/__init__.py

@@ -4,7 +4,7 @@
 
 
 def _version_to_int(version_str: str) -> int:
-    version_split = version_str.split(".") + ["0", "0"]  # in case a verson doesn't have third element, e.g. 3.0
+    version_split = version_str.split(".") + ["0", "0"]  # in case a version doesn't have third element, e.g. 3.0
     major = int(version_split[0])
     minor = int(version_split[1])
     patch = int(version_split[2])

prompting/llms/vllm_llm.py

@@ -143,7 +143,7 @@ async def generate_logits(
             continue_final_message=continue_last_message,
         )
 
-        outputs = self.model.generate(prompt, vllm_params)
+        outputs = self.model.generate(prompt, vllm_params, use_tqdm=False)
 
         if not outputs or not outputs[0].outputs[0].logprobs:
             return {}

prompting/weight_setting/weight_setter.py

@@ -5,7 +5,6 @@
 import numpy as np
 from loguru import logger
 
-from shared.constants import WHITELISTED_VALIDATORS_UIDS
 from shared.epistula import create_header_hook
 
 
@@ -16,25 +15,23 @@ def __init__(self, metagraph: bt.Metagraph, wallet: bt.Wallet, weight_dict: dict
         self.wallet = wallet
         self.current_hotkey = wallet.hotkey.ss58_address
         self.uid = metagraph.hotkeys.index(self.current_hotkey)
-        # Handle testnet case.
-        validator_uids = set([uid for uid in WHITELISTED_VALIDATORS_UIDS if uid < metagraph.n.item()])
+        self.validator_uids = np.where(np.array(metagraph.validator_permit))[0].tolist()
 
-        self.weight_matrix = np.zeros((len(validator_uids), metagraph.n.item()))
-        self.stake_matrix = np.array([metagraph.S[uid] for uid in validator_uids])
+        self.weight_matrix = np.zeros((len(self.validator_uids), metagraph.n.item()))
+        self.stake_matrix = np.array([metagraph.S[uid] for uid in self.validator_uids])
 
-        self.validator_uids = np.array(validator_uids)
-        self.validator_hotkeys = np.array([metagraph.hotkeys[uid] for uid in validator_uids])
+        self.validator_hotkeys = np.array([metagraph.hotkeys[uid] for uid in self.validator_uids])
         self.validator_addresses = np.array(
             [
                 f"{metagraph.axons[uid].ip}:{metagraph.axons[uid].port}"
-                for uid in validator_uids
+                for uid in self.validator_uids
                 if uid < metagraph.n.item()
             ]
         )
 
         self.weight_dict = weight_dict
 
-        self.request_tracker = np.zeros(len(validator_uids))
+        self.request_tracker = np.zeros(len(self.validator_uids))
 
     async def make_epistula_request(self, weight_matrix: np.ndarray, validator_address: str, validator_hotkey: str):
         """Make an epistula request to the validator at the given address."""
@@ -83,13 +80,9 @@ async def send_weight_matrixes(self, weight_matrix: np.ndarray):
 
     async def process_weight_dict(self):
         for uid, weights in self.weight_dict.items():
-            # Verify uid is in whitelist
-            validator_indices = np.where(self.validator_uids == uid)[0]
-            if len(validator_indices) == 0:
-                logger.error(f"Invalid validator UID {uid}, not in whitelist")
-                continue
-
-            # Update the weight matrix for this validator
-            validator_idx = validator_indices[0]
-            self.weight_matrix[validator_idx] = weights
-            self.request_tracker[validator_idx] = 1
+            if uid in self.validator_uids:
+                validator_index = self.validator_uids.index(uid)
+                self.weight_matrix[validator_index] = weights
+                self.request_tracker[validator_index] = 1
+            else:
+                logger.warning(f"UID {uid} is not a validator, skipping")

prompting/weight_setting/weight_synchronizer.py

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "prompting"
-version = "2.19.7"
+version = "2.19.8"
 description = "Subnetwork 1 runs on Bittensor and is maintained by Macrocosmos. It's an effort to create decentralised AI"
 authors = ["Kalei Brady, Dmytro Bobrenko, Felix Quinque, Steffen Cruz, Richard Wardle"]
 readme = "README.md"

pyproject.toml

@@ -175,7 +175,7 @@ async def query_miners(
         logger.info(
             f"Responses success: {responses_valid}/{len(uids)}. "
             f"Responses exception: {responses_exception}/{len(uids)}. "
-            f"Reponses invalid: {responses_error}/{len(uids)}"
+            f"Responses invalid: {responses_error}/{len(uids)}"
         )
         return results
     except Exception as e:

shared/epistula.py

@@ -263,6 +263,10 @@ def SUBTENSOR(self) -> Subtensor:
         self._subtensor = Subtensor(network=subtensor_network)
         return self._subtensor
 
+    def metagraph_force_sync(self) -> Metagraph:
+        self._last_update_time = 0
+        return self.METAGRAPH
+
     @property
     def METAGRAPH(self) -> Metagraph:
         if time.time() - self._last_update_time > 1200:

shared/settings.py

@@ -80,7 +80,7 @@ def get_random_uids(k: int | None = 10**6, exclude: list[int] = None, own_uid: i
         if exclude is None or uid not in exclude:
             candidate_uids.append(uid)
 
-    # Check if candidate_uids contain enough for querying, if not grab all avaliable uids
+    # Check if candidate_uids contain enough for querying, if not grab all available uids
     if 0 < len(candidate_uids) < k:
         logger.warning(
             f"Requested {k} uids but only {len(candidate_uids)} were available. To disable this warning reduce the sample size (--neuron.sample_size)"