add Laguna XS-2.1 SWE example

examples/advanced/laguna-xs-2.1/swe.toml

@@ -0,0 +1,159 @@
+# Laguna XS-2.1 on ScaleSWE with the standard bash harness and pass@1
+# evaluation on SWE-Bench Verified.
+
+max_steps = 1000
+seq_len = 131072
+
+[env_vars]
+TRITON_CACHE_DIR = "/tmp/.triton-cache"
+
+[slurm]
+job_name = "laguna-xs21-swe"
+partition = "all"
+pre_run_command = 'timeout 120 prime sandbox delete --label "laguna-xs21-swe" -y || true'
+
+[deployment]
+type = "multi_node"
+num_train_nodes = 2
+num_infer_nodes = 1
+num_infer_replicas = 4
+
+[wandb]
+project = "laguna-xs21-swe"
+name = "laguna-xs21-scaleswe-dep8"
+
+[weight_broadcast]
+type = "nccl"
+timeout = 3600
+
+[ckpt]
+interval = 25
+keep_last = 1
+resume_step = -1
+
+[model]
+name = "poolside/Laguna-XS-2.1"
+
+[trainer]
+dist_timeout_seconds = 3600
+enable_router_replay = true
+
+[trainer.model]
+impl = "custom"
+attn = "flash_attention_3"
+cp = 4
+cp_style = "ulysses"
+fused_lm_head_token_chunk_size = 1024
+optim_cpu_offload = true
+
+[trainer.model.compile]
+
+[trainer.model.ac]
+
+[trainer.model.ac_offloading]
+max_inflight_activations = 1
+
+[trainer.ckpt]
+skip_gather_master_weights = true
+skip_optimizer = true
+
+[trainer.optim]
+type = "muon"
+lr = 1e-6
+
+[orchestrator]
+batch_size = 256
+group_size = 16
+max_off_policy_steps = 32
+max_inflight_episodes = 768
+
+[orchestrator.algo]
+type = "grpo"
+
+[orchestrator.algo.length_penalty]
+type = "linear"
+num_output_tokens_weight = 0.0
+num_input_tokens_weight = 0.1
+num_turns_weight = 0.1
+
+[orchestrator.renderer]
+name = "laguna-xs-2.1"
+enable_thinking = true
+thinking_retention = "all"
+
+[orchestrator.train.sampling]
+temperature = 1.0
+max_completion_tokens = 32768
+extra_body = { top_k = 20, min_p = 0.0 }
+
+[[orchestrator.train.source]]
+name = "scaleswe"
+
+[orchestrator.train.source.env.taskset]
+id = "scaleswe-v1"
+
+[orchestrator.train.source.env.agent.harness]
+id = "bash"
+
+[orchestrator.train.source.env.agent.runtime]
+type = "prime"
+vm = true
+cpu = 1.0
+memory = 2.0
+disk = 5.0
+idle_timeout = "None"
+labels = ["laguna-xs21-swe"]
+
+[orchestrator.prime_monitor]
+
+[orchestrator.eval]
+interval = 20
+group_size = 1
+
+[orchestrator.eval.sampling]
+temperature = 1.0
+top_p = 1.0
+top_k = 20
+min_p = 0.0
+max_completion_tokens = 32768
+
+[[orchestrator.eval.source]]
+name = "swebench-verified"
+
+[orchestrator.eval.source.env.taskset]
+id = "swebench-verified-v1"
+
+[orchestrator.eval.source.env.agent.harness]
+id = "bash"
+
+[orchestrator.eval.source.env.agent.runtime]
+type = "prime"
+vm = true
+cpu = 1.0
+memory = 2.0
+disk = 5.0
+idle_timeout = "None"
+labels = ["laguna-xs21-swe"]
+
+[orchestrator.eval.source.env.agent.timeout]
+rollout = 3600
+
+[inference]
+enable_expert_parallel = true
+enable_return_routed_experts = true
+
+[inference.env_vars]
+VLLM_CACHE_ROOT = "/tmp/.vllm-cache"
+FLASHINFER_WORKSPACE_BASE = "/tmp/.flashinfer-cache"
+
+[inference.model]
+max_model_len = 131072
+tool_call_parser = "poolside_v1"
+reasoning_parser = "poolside_v1"
+
+[inference.parallel]
+tp = 1
+dp = 8
+
+[inference.vllm_extra]
+default_chat_template_kwargs = { enable_thinking = true }