Merge pull request #44 from RaoFoundation/dev

Release 2.1.4

constants/__init__.py

@@ -4,7 +4,7 @@
 # Project Constants.
 # ---------------------------------
 
-__version__ = "2.1.0"
+__version__ = "2.1.4"
 version_split = __version__.split(".")
 __spec_version__ = (
     (1000 * int(version_split[0]))
@@ -21,7 +21,7 @@
 # The maximum bytes for the hugging face repo (1 Gigabyte).
 MAX_HUGGING_FACE_BYTES = 1 * 1024 * 1024 * 1024
 # The maximum parameter size allowed for models.
-MAX_MODEL_PARAMETER_SIZE = 122268040
+MAX_MODEL_PARAMETER_SIZE = 186_000_000
 # The number of run steps to log to single wandb run.
 MAX_RUN_STEPS_PER_WANDB_RUN = 100
 

docs/miner.md

@@ -121,8 +121,8 @@ python scripts/upload_model.py --load_model_dir <path to model> --hf_repo_id my-
 
 ## Running a custom Miner
 
-As of Jan 1st, 2024 the subnet works with any model supported by [AutoModelForCausalLM](https://huggingface.co/docs/transformers/model_doc/auto#transformers.AutoModelForCausalLM) subject to the following constraints:
-1. Has less than 122268040 parameters.
+As of Jan 15th, 2024 the subnet works with any model supported by [AutoModelForCausalLM](https://huggingface.co/docs/transformers/model_doc/auto#transformers.AutoModelForCausalLM) subject to the following constraints:
+1. Has less than 186,000,000 parameters.
 2. Total size of the repo is less than 1 Gigabyte.
 3. Models sequence_length parameter must be 1024 tokens.
 

pretrain/mining.py

@@ -141,6 +141,24 @@ async def push(self, model: PreTrainedModel, retry_delay_secs: int = 60):
                     self.wallet.hotkey.ss58_address, model_id
                 )
 
+                bt.logging.info(
+                    "Wrote model metadata to the chain. Checking we can read it back..."
+                )
+
+                model_metadata = (
+                    await self.model_metadata_store.retrieve_model_metadata(
+                        self.wallet.hotkey.ss58_address
+                    )
+                )
+
+                if not model_metadata or model_metadata.id != model_id:
+                    bt.logging.error(
+                        f"Failed to read back model metadata from the chain. Expected: {model_id}, got: {model_metadata}"
+                    )
+                    raise ValueError(
+                        f"Failed to read back model metadata from the chain. Expected: {model_id}, got: {model_metadata}"
+                    )
+
                 bt.logging.success("Committed model to the chain.")
                 break
             except Exception as e:

setup.py

@@ -53,7 +53,7 @@ def read_requirements(path):
 
 # loading version from setup.py
 with codecs.open(
-    os.path.join(here, "pretrain/__init__.py"), encoding="utf-8"
+    os.path.join(here, "constants/__init__.py"), encoding="utf-8"
 ) as init_file:
     version_match = re.search(
         r"^__version__ = ['\"]([^'\"]*)['\"]", init_file.read(), re.M

tests/model/storage/fake_model_metadata_store.py

@@ -1,4 +1,4 @@
-from collections import deque
+from collections import defaultdict, deque
 from typing import List, Optional
 from model.data import ModelId, ModelMetadata
 from model.storage.model_metadata_store import ModelMetadataStore
@@ -11,6 +11,7 @@ def __init__(self):
         self.current_block = 1
         self.metadata = dict()
         self.store_errors = deque()
+        self.injected_metadata = defaultdict(deque)
 
     async def store_model_metadata(self, hotkey: str, model_id: ModelId):
         """Fake stores model metadata for a specific hotkey."""
@@ -38,8 +39,15 @@ async def store_model_metadata_exact(
     async def retrieve_model_metadata(self, hotkey: str) -> Optional[ModelMetadata]:
         """Retrieves model metadata for a specific hotkey"""
 
+        if len(self.injected_metadata[hotkey]) > 0:
+            return self.injected_metadata[hotkey].popleft()
+
         return self.metadata[hotkey] if hotkey in self.metadata else None
 
+    def inject_model_metadata(self, hotkey: str, metadata: ModelMetadata):
+        """Injects a metadata for hotkey to be returned instead of the expected response."""
+        self.injected_metadata[hotkey].append(metadata)
+
     def inject_store_errors(self, errors: List[Exception]):
         """Injects a list of errors to be raised on the next N calls to store_model_metadata."""
         self.store_errors.extend(errors)

tests/pretrain/test_mining.py

@@ -43,7 +43,7 @@ def test_model_to_disk_roundtrip(self):
 
         assert_model_equality(self, self.tiny_model, model)
 
-    def _test_push(self):
+    def _test_push(self, min_expected_block: int = 1):
         asyncio.run(self.actions.push(model=self.tiny_model, retry_delay_secs=1))
 
         # Check that the model was uploaded to hugging face.
@@ -54,7 +54,7 @@ def _test_push(self):
         model_metadata = asyncio.run(
             self.metadata_store.retrieve_model_metadata(self.wallet.hotkey.ss58_address)
         )
-        self.assertGreaterEqual(model_metadata.block, 1)
+        self.assertGreaterEqual(model_metadata.block, min_expected_block)
         self.assertEqual(model_metadata.id, model.id)
 
         self.metadata_store.reset()
@@ -73,6 +73,16 @@ def test_push_model_chain_failure(self):
 
         self._test_push()
 
+    def test_push_metadata_read_is_old(self):
+        """Tests that pushing a model to the chain is successful even if the metadata read back is stale."""
+
+        # Inject an empty response when push tries to read back the metadata commit.
+        self.metadata_store.inject_model_metadata(
+            self.wallet.hotkey.ss58_address, metadata=None
+        )
+
+        self._test_push(min_expected_block=2)
+
 
 if __name__ == "__main__":
     unittest.main()