# User-authorized four-task, two-update async integration run.
max_steps = 2
seq_len = 50000
output_dir = "../runs/prime-training"
[model]
name = "Qwen/Qwen3.5-9B"
[deployment]
type = "single_node"
gpus_per_node = 2
num_train_gpus = 1
num_infer_gpus = 1
[trainer]
[trainer.model]
# Use Prime RL default CompileConfig (fullgraph=false, mode=null).
# Native FSDP forward/backward compute is BF16; no trainer FP8 quantization.
optimization_dtype = "float32"
reduce_dtype = "float32"
[trainer.model.lora]
rank = 16
alpha = 32
dropout = 0.0
[trainer.optim]
lr = 1e-5
[orchestrator]
batch_size = 16
max_off_policy_steps = 1
[orchestrator.concurrency]
initial_inflight = 16
min_inflight = 16
max_inflight = 16
[orchestrator.renderer]
name = "qwen3.5"
enable_thinking = true
[orchestrator.train]
group_size = 4
[orchestrator.train.sampling]
temperature = 1.0
top_p = 0.95
top_k = 20
max_completion_tokens = 8192
[orchestrator.train.sampling.extra_body]
min_p = 0.0
chat_template_kwargs = { enable_thinking = true }
[[orchestrator.train.source]]
name = "workspace-four"
[orchestrator.train.source.env.taskset]
id = "harbor"
dataset = "workspace-four@1"
repo = "ssh://git@github.com/SUSINDRAREDDY/workspace-agent-rl-tasks.git@4e536677eb3da39855df3642b9cbbbce253d77f1"
registry_path = "registry.json"
tasks = ["ds2-task_00687_c4c7ea", "ds2-task_12106_bd7f8d", "task_00002_aa2b6f", "task_00004_88fabf"]
ignore_timeouts = false
[orchestrator.train.source.env.retries]
max_retries = 0
[orchestrator.train.source.env.agent]
max_turns = 50
[orchestrator.train.source.env.agent.timeout]
setup = 600
[orchestrator.train.source.env.agent.retries]
max_retries = 0
[orchestrator.train.source.env.agent.runtime]
type = "modal"
# Baseline requests; scalar values permit Modal bursting, without explicit caps.
cpu = 0.125
memory = 0.125 # GiB: passed to Modal as 128 MiB.
[orchestrator.train.source.env.agent.harness]
id = "bash"
edit = true
search = false
[orchestrator.train.source.env.agent.harness.compaction]
summarize_at_tokens = 30000
[orchestrator.train.source.env.verifier]
retries = 0
[inference.vllm]
dtype = "bfloat16"
quantization = "fp8_per_block"
reasoning_parser = "qwen3"
tool_call_parser = "qwen3_coder"
enable_auto_tool_choice = true
max_model_len = 50000
max_num_seqs = 16
gpu_memory_utilization = 0.8
language_model_only = true
revision = "c202236235762e1c871ad0ccb60c8ee5ba337b9a"
[ckpt]

[orchestrator.eval]
interval = 2
skip_first_step = true
group_size = 1
[orchestrator.eval.select]
limit = 4
shuffle = false
[orchestrator.eval.sampling]
temperature = 1.0
top_k = 20
top_p = 0.95
min_p = 0.0
max_tokens = 8192