[aniravich]

Adds examples/advanced/glm-5.3/swe-rebench-v2.toml, the config of a 16-node GLM-5.3 RL experiment (job 1862) kept exactly as it ran, and lists it in the GLM-5.3 README.

What it runs

• Data: PrimeIntellect/SWE-rebench-V2-Filtered-Verified, train split (6,272 tasks from 1,890 repos). Repos with sum(repo.encode()) % 20 == 13 are held out (95 repos, 241 tasks). Training uses the other 6,025 tasks, after also dropping 3 train-side aliases of held-out repos (6 tasks). An eval of 100 held-out tasks × 4 rollouts runs at step 0 and every 50 steps.
• Trainer: 8 nodes / 64 H200. CP8 × EP8 with FSDP over 8 at 131k seq_len, Muon with streamed CPU state (fix(optim): stream Muon gradients and state for 8-node GLM-5.3 SFT #3719), router replay (R3), lr 1e-6.
• Inference: one 4-node prefill instance (DP32/EP32, DeepEP high-throughput) and one 4-node decode instance (DP32/EP32, DeepEP low-latency), with NIXL KV transfer and one vllm-router (sticky_least_loaded). The BF16 checkpoint is quantized to per-block FP8 online, KV cache is FP8, DeepGEMM MoE runs on both roles, and prefill DBO is off.
• Orchestrator: batch 256 = 16 groups × 16, bash harness, up to 40 turns, and adaptive concurrency from 512 to 2048 in flight (starting at 1024). max_off_policy_steps = 16.
• No [ckpt], because one save is ~6 TB.

Not runnable on main

The config is kept as an experiment record, not ported:

• It targets a prime-rl checkout from before feat: select taskset tasks per train and eval source #3703: sources still take shuffle / num_examples, and group_size / algo sit on [orchestrator]. It uses trainer.model.optim_cpu_offload (since renamed by refactor(configs)!: one model.offload field for CPU offloading #3988) and optim.stream_cpu_offload from the still-open fix(optim): stream Muon gradients and state for 8-node GLM-5.3 SFT #3719.
• Out-of-tree patches: solver_network on the swerebench-v2 taskset (deps/prime-envs), bash_timeout on the bash harness (deps/verifiers), vllm-router PR 58, a vLLM 0.29 overlay for prefill, and a local R3 diagnostics patch. [slurm] pre_run_command checks each of them read-only and aborts the job if one is missing.
• Paths point at the cluster it ran on: output_dir, [slurm] project_dir / template_path, the /tmp/ani caches, and the decode VLLM_TUNED_CONFIG_FOLDER.

Cleanup relative to the run's config

The cleanup only touched comments and layout. The committed file parses (tomllib) to the same 151 values as the config that ran:

• Dropped the version-history notes and kept the reasoning behind each setting.
• Fixed stale comments: decode gpu_memory_utilization is 0.65 (not 0.60), max_num_seqs × 32 ranks now matches max_inflight 2048, and the prefill/decode sizing notes describe the 4 + 4 node layout.
• Rewrote the inline compilation_config table as a [...compilation_config] section (same value).

🤖 Generated with Claude Code