[tools]: Add sync script and CI check for Miles server args sync

.github/workflows/param-docs-sync-check.yml

@@ -0,0 +1,33 @@
+name: Param Docs Sync Check
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - "docs/en/advanced/miles_server_args.md"
+      - "miles/utils/arguments.py"
+  pull_request:
+    types: [opened, synchronize, reopened, ready_for_review]
+    paths:
+      - "docs/en/advanced/miles_server_args.md"
+      - "miles/utils/arguments.py"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  check-param-doc-sync:
+    name: Check Param Docs Sync
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Validate docs and arguments sync
+        run: python tools/sync_param_docs.py --check

docs/en/advanced/miles_server_args.md

@@ -58,7 +58,7 @@ Arguments for configuring Ray cluster resources and GPU allocation.
 | `--rollout-num-gpus` | Total number of GPUs required for rollout (inference). In `--colocate` mode, this is ignored and set to `actor-num-gpus-per-node * actor-num-nodes` (plus critic GPUs if enabled). | `None` | Type: int | Miles Native |
 | `--rollout-num-gpus-per-engine` | Number of GPUs per inference engine, same as `tp_size` in SGLang. For multi-node serving, this should be the total GPU count / `tp_size` for each SGLang instance. | `1` | Type: int | Miles Native |
 | `--num-gpus-per-node` | Total GPUs per node on the physical machine. This informs the Ray scheduler of the hardware capacity. In **Colocate mode**, it is required that the machine has fewer than 8 GPUs to calculate correct VRAM offsets. In **Disaggregated mode**, it ensures SGLang engines are distributed correctly across nodes without exceeding per-node GPU limits. | `8` | Type: int | Miles Native |
-| `--colocate` | Deploy training and rollout on the same GPUs. This mode automatically enables `--offload-train` and `--offload-rollout` to facilitate weight-swapping between the training actor and inference engine. **Note:** The offload parameters are currently only used for AMD GPUs and will be removed soon. **Memory Tip:** When colocating, it is highly recommended to set `--sglang-mem-fraction-static` to **0.8** (especially on **NVIDIA Blackwell B200/B300** GPUs). This leaves sufficient VRAM (~20%) for Megatron to initialize its structures before the first weight offload to CPU occurs. On GB200/GB300, values up to 0.75 are safer for long-running jobs to prevent potential OOMs. #TODO: Verify optimal fraction for Blackwell in production | `False` | bool flag (set to enable) | Miles Native |
+| `--colocate` | Deploy training and rollout on the same GPUs. This mode automatically enables `--offload-train` and `--offload-rollout` to facilitate weight-swapping between the training actor and inference engine. **Note:** The offload parameters are currently only used for AMD GPUs and will be removed soon. **Memory Tip:** When colocating, it is highly recommended to set `--sglang-mem-fraction-static` to **0.8** (especially on **NVIDIA Blackwell B200/B300** GPUs). This leaves sufficient VRAM (~20%) for Megatron to initialize its structures before the first weight offload to CPU occurs. On GB200/GB300, values up to 0.75 are safer for long-running jobs to prevent potential OOMs. #TODO: Verify optimaltest fraction for Blackwell in production | `False` | bool flag (set to enable) | Miles Native |
 | `--prefill-num-servers` | Number of dedicated prefill servers for PD disaggregation. | `None` | Type: int | Miles Native |
 | `--distributed-backend` | Backend for distributed communication. | `nccl` | `nccl`, `gloo` | Megatron-LM (Reset by Miles) |
 | `--distributed-timeout-minutes` | Timeout for distributed operations in minutes. | `10` | Type: int | Megatron-LM (Reset by Miles) |