feat: try with 2 GPUs (agent & training separated) | use better model

Author: ndrpp

autoresearch/Dockerfile

@@ -19,7 +19,7 @@ RUN uv pip install --system --no-cache-dir \
     torch==2.9.1 --index-url https://download.pytorch.org/whl/cu128
 
 RUN uv pip install --system --no-cache-dir \
-    vllm \
+    "vllm>=0.17.0" \
     kernels>=0.11.7 \
     rustbpe>=0.1.0 \
     tiktoken>=0.11.0 \

autoresearch/algo.py

@@ -1,9 +1,9 @@
 """
 Autonomous autoresearch agent loop for ocean network.
 
-Runs inside a Docker container on a remote GPU node.
-Uses a local open-source LLM (Qwen3-32B-AWQ via vLLM) to iteratively
-improve train.py, measuring val_bpb as the optimization target.
+Runs inside a Docker container on a 2×H200 GPU node.
+GPU 0: dedicated to the agent LLM (Qwen3.5-27B via vLLM, unquantized bf16)
+GPU 1: dedicated to training (full 141GB VRAM)
 """
 
 import json
@@ -14,15 +14,23 @@
 import time
 from datetime import datetime, timezone
 
+# ---------------------------------------------------------------------------
+# GPU isolation — must be set before any CUDA imports
+# ---------------------------------------------------------------------------
+
+AGENT_GPU = "0"
+TRAINING_GPU = "1"
+os.environ["CUDA_VISIBLE_DEVICES"] = AGENT_GPU  # vLLM only sees GPU 0
+
 # ---------------------------------------------------------------------------
 # Configuration
 # ---------------------------------------------------------------------------
 
-AGENT_MODEL = "Qwen/Qwen3-32B-AWQ"
-GPU_MEMORY_UTILIZATION = 0.25  # ~35GB for weights+KV cache, rest for training
+AGENT_MODEL = "Qwen/Qwen3.5-27B"
+GPU_MEMORY_UTILIZATION = 0.90  # dedicated GPU — use most of it
 MAX_ITERATIONS = 200
 TRAINING_TIMEOUT = 600  # 10 minutes
-MAX_MODEL_LEN = 40960   # total context window (input + output)
+MAX_MODEL_LEN = 65536   # larger context — dedicated GPU has plenty of room
 MAX_OUTPUT_TOKENS = 16384  # max tokens for LLM output (enough for full train.py)
 TEMPERATURE = 0.7
 STAGNATION_THRESHOLD = 5  # consecutive non-improvements before nudge
@@ -138,13 +146,18 @@ def run_training(train_py_content):
     """
     write_file(TRAIN_PY_PATH, train_py_content)
 
+    # Run training on the dedicated training GPU
+    train_env = os.environ.copy()
+    train_env["CUDA_VISIBLE_DEVICES"] = TRAINING_GPU
+
     try:
         result = subprocess.run(
             [sys.executable, TRAIN_PY_PATH],
             capture_output=True,
             text=True,
             timeout=TRAINING_TIMEOUT,
             cwd="/app",
+            env=train_env,
         )
     except subprocess.TimeoutExpired as e:
         stderr_text = e.stderr if isinstance(e.stderr, str) else (e.stderr.decode() if e.stderr else "")
@@ -419,6 +432,7 @@ def main():
         max_model_len=MAX_MODEL_LEN,
         dtype="auto",
         trust_remote_code=True,
+        enforce_eager=True,  # avoid DeltaNet compilation issues with Qwen3.5
     )
     sampling_params = SamplingParams(
         max_tokens=MAX_OUTPUT_TOKENS,