fix(sglang): tighten rollout argument guidance

examples/ar/qwen3_5_moe_grpo_35b_a3b_geo3k_mc_sglang.yaml

@@ -102,7 +102,8 @@ rollout:
       enable_thinking: false
     engine_kwargs:
       # Colocate MoE + geo3k: cuda graph OFF and mem_fraction_static 0.15
-      # (4B dense preset is 0.3 + cuda graph ON @ bs 16).
+      # (4B dense baseline is 0.3 + cuda graph ON @ bs 16). Unlike that
+      # baseline, attention_backend is not pinned; SGLang selects it by platform.
       # See unirl/rollout/README.md (SGLang AR knobs).
       mem_fraction_static: 0.15
       disable_cuda_graph: true

examples/ar/qwen3_dppo_4b_base_dapo_sglang.yaml

@@ -133,8 +133,8 @@ rollout:
       # FULL fine-tuning: lower the SRT KV reservation to give FSDP headroom for
       # the per-rollout full-weight all-gather. No LoRA pool — full dense weights
       # pushed via TensorWeightSync. CUDA graph ON for decode speed, but capped at
-      # cuda_graph_max_bs=16 so its captured buffers stay small (the default 256
-      # would balloon server memory + compete with the weight push).
+      # cuda_graph_max_bs=16 so its captured buffers stay small (the auto-tuned
+      # default can be much larger and compete with the weight push).
       mem_fraction_static: 0.3
       skip_server_warmup: true
       attention_backend: triton

examples/ar/qwen3_drgrpo_4b_base_dapo_sglang.yaml

@@ -125,8 +125,8 @@ rollout:
       # FULL fine-tuning: lower the SRT KV reservation to give FSDP headroom for
       # the per-rollout full-weight all-gather. No LoRA pool — full dense weights
       # pushed via TensorWeightSync. CUDA graph ON for decode speed, but capped at
-      # cuda_graph_max_bs=16 so its captured buffers stay small (the default 256
-      # would balloon server memory + compete with the weight push).
+      # cuda_graph_max_bs=16 so its captured buffers stay small (the auto-tuned
+      # default can be much larger and compete with the weight push).
       mem_fraction_static: 0.3
       skip_server_warmup: true
       attention_backend: triton

examples/ar/qwen3_drpo_4b_base_dapo_sglang.yaml

@@ -143,8 +143,8 @@ rollout:
       # FULL fine-tuning: lower the SRT KV reservation to give FSDP headroom for
       # the per-rollout full-weight all-gather. No LoRA pool — full dense weights
       # pushed via TensorWeightSync. CUDA graph ON for decode speed, but capped at
-      # cuda_graph_max_bs=16 so its captured buffers stay small (the default 256
-      # would balloon server memory + compete with the weight push).
+      # cuda_graph_max_bs=16 so its captured buffers stay small (the auto-tuned
+      # default can be much larger and compete with the weight push).
       mem_fraction_static: 0.3
       skip_server_warmup: true
       attention_backend: triton

examples/ar/qwen3_drpo_4b_veomni_sp_sglang.yaml

@@ -135,8 +135,8 @@ rollout:
       # FULL fine-tuning: lower the SRT KV reservation to give FSDP headroom for
       # the per-rollout full-weight all-gather. No LoRA pool — full dense weights
       # pushed via TensorWeightSync. CUDA graph ON for decode speed, but capped at
-      # cuda_graph_max_bs=16 so its captured buffers stay small (the default 256
-      # would balloon server memory + compete with the weight push).
+      # cuda_graph_max_bs=16 so its captured buffers stay small (the auto-tuned
+      # default can be much larger and compete with the weight push).
       mem_fraction_static: 0.3
       skip_server_warmup: true
       attention_backend: triton

examples/ar/qwen3_grpo_4b_base_dapo_sglang.yaml

@@ -132,8 +132,8 @@ rollout:
       # FULL fine-tuning: lower the SRT KV reservation to give FSDP headroom for
       # the per-rollout full-weight all-gather. No LoRA pool — full dense weights
       # pushed via TensorWeightSync. CUDA graph ON for decode speed, but capped at
-      # cuda_graph_max_bs=16 so its captured buffers stay small (the default 256
-      # would balloon server memory + compete with the weight push).
+      # cuda_graph_max_bs=16 so its captured buffers stay small (the auto-tuned
+      # default can be much larger and compete with the weight push).
       mem_fraction_static: 0.3
       skip_server_warmup: true
       attention_backend: triton

examples/ar/qwen3_grpo_4b_veomni_sp_sglang.yaml

@@ -136,8 +136,8 @@ rollout:
       # FULL fine-tuning: lower the SRT KV reservation to give FSDP headroom for
       # the per-rollout full-weight all-gather. No LoRA pool — full dense weights
       # pushed via TensorWeightSync. CUDA graph ON for decode speed, but capped at
-      # cuda_graph_max_bs=16 so its captured buffers stay small (the default 256
-      # would balloon server memory + compete with the weight push).
+      # cuda_graph_max_bs=16 so its captured buffers stay small (the auto-tuned
+      # default can be much larger and compete with the weight push).
       mem_fraction_static: 0.3
       skip_server_warmup: true
       attention_backend: triton

examples/ar/qwen3_grpo_vanilla_4b_base_dapo_sglang.yaml

@@ -134,8 +134,8 @@ rollout:
       # FULL fine-tuning: lower the SRT KV reservation to give FSDP headroom for
       # the per-rollout full-weight all-gather. No LoRA pool — full dense weights
       # pushed via TensorWeightSync. CUDA graph ON for decode speed, but capped at
-      # cuda_graph_max_bs=16 so its captured buffers stay small (the default 256
-      # would balloon server memory + compete with the weight push).
+      # cuda_graph_max_bs=16 so its captured buffers stay small (the auto-tuned
+      # default can be much larger and compete with the weight push).
       mem_fraction_static: 0.3
       skip_server_warmup: true
       attention_backend: triton

examples/ar/qwen3_ppo_4b_base_dapo_sglang.yaml

@@ -92,7 +92,7 @@ rollout:
       # Colocate full-FT preset (unirl/rollout/README.md): lower the SRT KV
       # reservation so FSDP can all-gather full dense weights. enable_lora must
       # stay false for TensorWeightSync. CUDA graph ON, cuda_graph_max_bs=16
-      # (SGLang default 256 would balloon server memory + compete with the push).
+      # (the auto-tuned default can be much larger and compete with the push).
       mem_fraction_static: 0.3
       skip_server_warmup: true
       attention_backend: triton

unirl/rollout/README.md

@@ -113,35 +113,47 @@ change surface:
 
 ## SGLang AR knobs
 
-Qwen3 AR recipes (`examples/ar/qwen3_*_sglang*.yaml`) share one **colocate
-full-FT** `engine_kwargs` preset. Async/separate and larger-model recipes keep
-the same keys and change the memory numbers. Typed `SGLangEngineConfig` fields
-overlay `engine_kwargs`; reserved ports always win. Keys that are not live
-SGLang `ServerArgs` fields warn at boot (or raise if
-`UNIRL_SGLANG_STRICT_SERVER_ARGS=1`) — see [`engine/README.md`](engine/README.md).
+The 4B dense Qwen3 AR recipes (`examples/ar/qwen3_*_sglang*.yaml`) share a
+**colocate full-FT** `engine_kwargs` baseline. Async/separate, larger-model,
+MoE, and LoRA recipes deliberately deviate as described below and in their
+inline comments. Typed `SGLangEngineConfig` fields overlay `engine_kwargs`;
+reserved ports always win. Keys that are not live SGLang `ServerArgs` fields
+warn at boot (or raise if `UNIRL_SGLANG_STRICT_SERVER_ARGS=1`) — see
+[`engine/README.md`](engine/README.md).
 
-**Colocate full-FT preset** (FSDP train shard time-shares the GPU; TensorWeightSync):
+**4B dense colocate full-FT baseline** (FSDP train shard time-shares the GPU;
+TensorWeightSync):
 
 | Knob | Preset | Why |
 |---|---|---|
 | `mem_fraction_static` | `0.3` | Lower SRT KV reservation so FSDP can all-gather full dense weights. `server_intent()` defaults to `0.88` if omitted — too high for colocate. |
 | `enable_lora` | `false` | TensorWeightSync pushes full dense weights; a LoRA pool would be the wrong receive path. |
-| `cuda_graph_max_bs` | `16` | CUDA graph stays on (`disable_cuda_graph: false`); SGLang's default `256` captures buffers that fight the weight push. |
-| `skip_server_warmup` | `true` | Skip SRT warmup on every colocated boot/wake. |
+| `cuda_graph_max_bs` | `16` | CUDA graph stays on (`disable_cuda_graph: false`); SGLang otherwise auto-tunes the capture cap from GPU memory and TP size, often reserving much larger buffers that fight the weight push. |
+| `skip_server_warmup` | `true` | Skip SRT startup warmup on the 4B colocate path; this knob does not control `wake_up()`. |
 | `attention_backend` | `triton` | Matches the in-tree 4B full-FT recipes. |
 
+**Larger / MoE colocate:** memory and graph settings are recipe-specific
+(`mem_fraction_static` ranges from `0.15` to `0.5`; graph caps include `64` and
+`128`, and one VLM MoE recipe disables graphs). Several MoE/CPPO recipes set the
+typed `skip_server_warmup: false`, which wins over `engine_kwargs`, so startup
+warmup runs. The VLM MoE geo3k recipe also leaves `attention_backend` to
+SGLang's platform selection.
+
 **Async / separate** (`*_sglang_async.yaml`): the engine owns the GPU (no colocated
-FSDP shard). Raise `mem_fraction_static` to `0.8` and leave the other preset keys
-as-is; keep headroom for NCCL weight-receive buffers. Sync is `NCCLWeightSync`.
+FSDP shard). Raise `mem_fraction_static` to `0.8` and leave the other 4B baseline
+keys as-is; keep headroom for NCCL weight-receive buffers. Sync is
+`NCCLWeightSync`.
 
 **LoRA colocate** (`*_sglang_lora.yaml`): `enable_lora: true` plus the SGLang LoRA
 pool knobs. This is the exception to the full-FT `enable_lora: false` receive path
 (`LocalLoraWeightSync` instead of TensorWeightSync). Memory numbers are recipe-specific.
 
-**Reserved ports:** `SGLangPorts.reserve()` binds HTTP `port` and `nccl_port` on
-the engine's node. Server port is capped at 35535 because SGLang derives
-`grpc_port = port + 30000`. Do not set `port` / `nccl_port` in `engine_kwargs` —
-the reserved sockets overwrite them.
+**Reserved ports:** `SGLangPorts.reserve()` selects candidate HTTP `port` and
+`nccl_port` values by binding temporary sockets on the engine's node, then
+closes those sockets so SGLang can bind (the usual bind-to-zero TOCTOU gap
+remains). With pinned SGLang 0.5.12.post1, the server port is capped at 55535
+because an enabled gRPC endpoint defaults to `grpc_port = port + 10000`. Do not
+set `port` / `nccl_port` in `engine_kwargs` — the selected values overwrite them.
 
 ## Gotchas
 

unirl/rollout/engine/sglang/backends/base.py

@@ -28,15 +28,6 @@
 )
 
 
-def _strict_dropped_server_args() -> bool:
-    return os.environ.get(_STRICT_SERVER_ARGS_ENV, "").strip().lower() in {"1", "true", "yes", "on"}
-
-
-def _unknown_server_arg_keys(server_intent: Dict[str, Any], allowed: set[str]) -> List[str]:
-    """Intent keys that are neither live ``ServerArgs`` fields nor UniRL-only."""
-    return sorted(key for key in server_intent if key not in allowed and key not in _UNIRL_ONLY_INTENT_KEYS)
-
-
 def _filter_server_args_or_raise(
     server_intent: Dict[str, Any],
     *,
@@ -56,14 +47,15 @@ def _filter_server_args_or_raise(
             "Upgrade SGLang to a build that supports these fields, or remove the explicit UniRL "
             "rollout config that depends on them."
         )
-    dropped = _unknown_server_arg_keys(server_intent, allowed)
+    dropped = sorted(key for key in server_intent if key not in allowed and key not in _UNIRL_ONLY_INTENT_KEYS)
     if dropped:
         message = (
             f"SGLang {backend_name} backend dropping unknown ServerArgs keys: {dropped}. "
             "They are not fields on the installed SGLang ServerArgs (typo or version skew). "
             f"Set {_STRICT_SERVER_ARGS_ENV}=1 to fail closed."
         )
-        if _strict_dropped_server_args():
+        strict = os.environ.get(_STRICT_SERVER_ARGS_ENV, "").strip().lower() in {"1", "true", "yes", "on"}
+        if strict:
             raise RuntimeError(message)
         logger.warning(message)
     return {k: v for k, v in server_intent.items() if k != _REQUIRED_SERVER_ARGS_METADATA_KEY and k in allowed}

unirl/rollout/engine/sglang/config.py

@@ -11,7 +11,7 @@
 from unirl.rollout.engine.base import BaseEngineConfig
 from unirl.rollout.engine.ports import ReservedPorts
 
-_SGLANG_GRPC_PORT_OFFSET = 30000
+_SGLANG_GRPC_PORT_OFFSET = 10000
 _SGLANG_MAX_DERIVED_GRPC_BASE_PORT = 65535 - _SGLANG_GRPC_PORT_OFFSET
 _SGLANG_SAFE_SERVER_PORT_MIN = 1024
 _REQUIRED_SERVER_ARGS_METADATA_KEY = "_unirl_required_server_args"