GLM-5.3 sampling example

examples/api/README.md

@@ -9,6 +9,6 @@ These JSON files demonstrate individual Cortex Training request bodies:
 | `rl.json` | Create colocated training and sampling sub-jobs |
 | `fwd-bwd.json` | Submit a readable forward/backward batch |
 | `generate.json` | Submit prompts for generation |
-| `glm-sampling.json` | Create a GLM sampling sub-job |
+| `glm-5.3-sampling.json` | Create a single-node GLM-5.3 sampling sub-job on 8 H200 GPUs with an FP8 MLA KV cache and a reasoning parser |
 
 They are wire-format examples, not end-to-end training recipes.

examples/api/glm-5.3-sampling.json

@@ -0,0 +1,19 @@
+{
+  "sub_job_configs": [
+    {
+      "inference_config": {
+        "max_seq_len": 720896,
+        "n_gpus": 8,
+        "vllm_config": {
+          "enforce_eager": true,
+          "gpu_memory_utilization": 0.95,
+          "kv_cache_dtype": "fp8_ds_mla",
+          "reasoning_parser": "glm45",
+          "tensor_parallel_size": 8
+        }
+      },
+      "job_type": "sampling",
+      "model_name": "zai-org/GLM-5.3"
+    }
+  ]
+}