mimo-vllm: drop sparkring 20GiB KV pin — GB10 UMA auto-sizing keeps ~54GiB pool; pin would strand 34GiB

recipes/mimo26-flash-mopd-4x-vllm/recipe.toml

@@ -65,9 +65,10 @@ args = [
     "--max-model-len", "524288",
     "--max-num-seqs", "16",
     "--max-num-batched-tokens", "8192",
-    # SparkRing tp4 contract: deterministic 20 GiB KV pin (not auto-sizing),
-    # bounce-buffer loader, c1/c8/c16 concurrency-friendly scheduler bounds.
-    "--kv-cache-memory-bytes", "21474836480",
+    # SparkRing tp4 contract, concurrency-friendly scheduler bounds +
+    # bounce-buffer loader. The 20 GiB KV pin is NOT adopted: on GB10's UMA
+    # with GMU 0.85 the auto-sizing profiler already claims ~54 GiB free
+    # (8.5M-token pool in boot logs), so pinning 20 GiB would strand ~34 GiB.
     "--model-loader-extra-config", """
 '{"read_mode":"bounce","io_threads":8}'""",
     "--enable-chunked-prefill",