Merge branch 'rtfd' of github.com:bigscience-workshop/petals into rtfd

.github/workflows/run-tests.yaml

@@ -65,14 +65,6 @@ jobs:
           python -m pip install --upgrade pip
           pip install -r requirements.txt
           pip install -r requirements-dev.txt
-      - name: Build bitsandbytes cpuonly
-        run: |
-          git clone https://github.com/TimDettmers/bitsandbytes.git
-          cd bitsandbytes
-          git checkout main
-          make cpuonly
-          pip install .
-          cd -
       - name: Test
         run: |
           export HF_TAG=$(python -c "import os; print(os.environ.get('GITHUB_HEAD_REF') or os.environ.get('GITHUB_REF_NAME'))")
@@ -80,8 +72,8 @@ jobs:
           export REF_NAME=bigscience/bloom-560m
 
           python -m cli.run_server --converted_model_name_or_path $MODEL_NAME --block_indices 0:12 \
-            --torch_dtype float32 --identity tests/test.id --host_maddrs /ip4/127.0.0.1/tcp/31337 \
-            --throughput 1 --attn_cache_size 0.2GiB &> server1.log &
+            --identity tests/test.id --host_maddrs /ip4/127.0.0.1/tcp/31337 --throughput 1 \
+            --torch_dtype float32 --compression NONE --attn_cache_size 0.2GiB &> server1.log &
           SERVER1_PID=$!
           
           sleep 5  # wait for the first server to initialize DHT
@@ -90,13 +82,13 @@ jobs:
           # ^-- server 1 multiaddr is determined by --identity and --host_maddrs
           
           python -m cli.run_server --converted_model_name_or_path $MODEL_NAME --block_indices 12:22 \
-            --torch_dtype float32 --initial_peers $INITIAL_PEERS --throughput 1 &> server2.log &
+            --initial_peers $INITIAL_PEERS --throughput 1 --torch_dtype float32 &> server2.log &
           SERVER2_PID=$!
 
           sleep 10 # wait for initial servers to declare blocks, then let server decide which blocks to serve
 
           python -m cli.run_server --converted_model_name_or_path $MODEL_NAME --block_indices 0:6 \
-            --torch_dtype float32 --initial_peers $INITIAL_PEERS --throughput 1 &> server3.log &
+            --initial_peers $INITIAL_PEERS --throughput 1 --torch_dtype float32 &> server3.log &
           SERVER3_PID=$!
 
           python -m cli.run_server --converted_model_name_or_path $MODEL_NAME --block_indices 4:16 \
@@ -104,7 +96,7 @@ jobs:
           SERVER4_PID=$!
           
           python -m cli.run_server --converted_model_name_or_path $MODEL_NAME --num_blocks 3 \
-            --torch_dtype float32 --initial_peers $INITIAL_PEERS --throughput 1 &> server5.log &
+            --initial_peers $INITIAL_PEERS --throughput 1 --torch_dtype float32 &> server5.log &
           SERVER5_PID=$!
           
           tail -n 100 -f server*.log &

cli/run_server.py

@@ -86,7 +86,7 @@ def main():
     if args.pop("increase_file_limit"):
         increase_file_limit()
 
-    compression_type = args.pop("compression")
+    compression_type = args.pop("compression").upper()
     compression = getattr(CompressionType, compression_type)
 
     attn_cache_size = args.pop("attn_cache_size")

requirements.txt

@@ -1,6 +1,7 @@
 torch==1.12.0
+bitsandbytes==0.33.0
 accelerate==0.10.0
 huggingface-hub==0.7.0
+transformers==4.21.3
+https://github.com/learning-at-home/hivemind/archive/131f82c97ea67510d552bb7a68138ad27cbfa5d4.zip
 humanfriendly
-https://github.com/learning-at-home/hivemind/archive/20b3b3d5f225ed525515a5383a008a8f9fad8173.zip
-https://github.com/huggingface/transformers/archive/6589e510fa4e6c442059de2fab84752535de9b23.zip

src/server/server.py

@@ -73,7 +73,8 @@ class Server(threading.Thread):
         logger.info(f"Serving {len(self.module_backends)} blocks:")
         for block_name, backend in self.module_backends.items():
             num_parameters = sum(p.numel() for p in backend.module.parameters() if p.requires_grad)
-            logger.info(f"{block_name}: {backend.module.__class__.__name__}, {num_parameters} parameters")
+            parameter_msg = f"{num_parameters} trainable parameters" if num_parameters else "frozen"
+            logger.info(f"{block_name}: {backend.module.__class__.__name__}, {parameter_msg}")
 
         if not self.dht.is_alive():
             self.dht.run_in_background(await_ready=True)

src/server/task_pool.py

@@ -170,6 +170,3 @@ class PrioritizedTaskPool(TaskPoolBase):
         assert len(item) == 2
         self._priority.value = float(item[0])
         self._oldest_undispatched_timestamp.value = float(item[1])
-
-    def iterate_minibatches(self, *args, **kwargs) -> Generator[List[Task], None, None]:
-        raise NotImplementedError()