Attempt at fixing setup.

Author: finbarrtimbers

open_instruct/vllm_utils3.py

@@ -361,9 +361,9 @@ def create_vllm_engines(
     use_hybrid_engine = pg is not None
     num_gpus = int(tensor_parallel_size == 1)
     if use_hybrid_engine and tensor_parallel_size == 1 and single_gpu_mode:
-        # every worker will use 0.5 GPU, so that we can schedule
-        # 2 instances on the same GPUs.
-        num_gpus = 0.5
+        # every worker will use 0.5/num_engines GPU, so that we can schedule
+        # multiple instances on the same GPU while leaving 0.5 for the learner.
+        num_gpus = 0.5 / num_engines
 
     print(f"num_gpus: {num_gpus}")
 
@@ -381,7 +381,7 @@ def create_vllm_engines(
         scheduling_strategy = PlacementGroupSchedulingStrategy(
             placement_group=pg,
             placement_group_capture_child_tasks=True,
-            placement_group_bundle_index=i * tensor_parallel_size,
+            placement_group_bundle_index=0 if single_gpu_mode else i * tensor_parallel_size,
         )
 
         additional_kwargs = {}

scripts/train/build_image_and_launch.sh

@@ -30,7 +30,7 @@ fi
 
 # Install Python dependencies
 echo "Installing dependencies with uv..."
-uv sync --only-group dev
+uv sync 
 
 # Run the provided script
 bash $1 "$beaker_user/$image_name"

scripts/train/debug/single_gpu_integration_test.sh

@@ -34,7 +34,7 @@ uv run python mason.py \
     --per_device_train_batch_size 1 \
     --num_unique_prompts_rollout 8 \
     --num_samples_per_prompt_rollout 4 \
-    --model_name_or_path Qwen/Qwen3-1.7B \
+    --model_name_or_path EleutherAI/pythia-14m \
     --stop_strings "</answer>" \
     --apply_r1_style_format_reward \
     --apply_verifiable_reward true \
@@ -46,6 +46,7 @@ uv run python mason.py \
     --deepspeed_stage 2 \
     --num_epochs 1 \
     --num_learners_per_node 1 \
+    --vllm_num_engines 2 \
     --vllm_tensor_parallel_size 1 \
     --beta 0.01 \
     --seed 3 \

scripts/train/debug/single_gpu_on_beaker.sh

@@ -47,7 +47,6 @@ uv run python mason.py \
     --with_tracking \
     --num_epochs 1 \
     --num_learners_per_node 1 \
-    --vllm_num_engines 2 \
     --vllm_tensor_parallel_size 1 \
     --beta 0.01 \
     --seed 3 \