glm53-flash-fp4-local-2x: complete the DGX Spark x2 card params — add --dtype bfloat16, --quantization modelopt_mixed, --no-enable-flashinfer-autotune, --max-num-seqs 4, --max-num-batched-tokens 4096 (my first transcription grepped a truncated serve block and dropped them); add the card's vllm/flashinfer/triton/b12x cache mounts. Verified against the full '2 x NVIDIA DGX Spark' section (not the RTX PRO 6000 one)

recipes/glm53-flash-fp4-local-2x/recipe.toml

@@ -26,16 +26,21 @@ required_nodes = 2
 tensor_parallel_size = 2
 args = [
     "--served-model-name", "glm53-fp4-local",
+    "--dtype", "bfloat16",
     "--mamba-cache-mode", "align",
     "--enable-prefix-caching",
     "--enable-chunked-prefill",
     "--kv-cache-dtype", "fp8",
+    "--quantization", "modelopt_mixed",
     "--attention-backend", "B12X",
     "--block-size", "256",
     "--moe-backend", "b12x",
     "--linear-backend", "b12x",
+    "--no-enable-flashinfer-autotune",
     "--load-format", "b12x",
     "--max-model-len", "500000",
+    "--max-num-seqs", "4",
+    "--max-num-batched-tokens", "4096",
     "--reasoning-parser", "glm45",
     "--tool-call-parser", "glm47",
     "--enable-auto-tool-choice",
@@ -64,3 +69,21 @@ VLLM_USE_AOT_COMPILE = "1"
 VLLM_USE_MEGA_AOT_ARTIFACT = "1"
 VLLM_USE_V2_MODEL_RUNNER = "1"
 VLLM_ENGINE_READY_TIMEOUT_S = "3600"
+
+# JIT / AOT / b12x-autotune caches the card mounts on every node (speeds boot +
+# re-serve; docker creates the host dir if missing). Card: mkdir -p + mount.
+[[extra_mounts]]
+host = "~/.cache/vllm"
+container = "/root/.cache/vllm"
+
+[[extra_mounts]]
+host = "~/.cache/flashinfer"
+container = "/root/.cache/flashinfer"
+
+[[extra_mounts]]
+host = "~/.triton"
+container = "/root/.triton"
+
+[[extra_mounts]]
+host = "~/.cache/b12x"
+container = "/root/.cache/b12x"