brutus/vllm-xpu: chat to maxNumSeqs=3, maxModelLen=131072

Cuts the REQUEST_CONSTANT pool ((max_num_seqs * (1+num_spec) + 1) *
~3.1 MiB * 48 GDN layers) from 6.68 GiB at (15, 65536) to 1.45 GiB at
(3, 131072); doubles the TOKEN_PROPORTIONAL pool to 1.80 GiB. Total
KV budget ~3.25 GiB, clears the post-#41495 startup gate even on
worst-case ~1.3 GiB cold-boot accounting.

max_num_seqs=3 covers a 1+1+1 agentic loop (chat + tool call +
background). With MTP-2 each request occupies (1 + num_speculative
_tokens) = 3 query tokens per step, so effective batch dim peaks at 9
(matches existing cudagraphCaptureSizes [3 9]).

Author: jasonboukheir

hosts/brutus/services/vllm-xpu.nix

@@ -52,8 +52,8 @@ in {
       dtype = "bfloat16";
       quantization = "inc";
       kvCacheDtype = "turboquant_k3v4_nc";
-      maxModelLen = 65536;
-      maxNumSeqs = 15;
+      maxModelLen = 131072;
+      maxNumSeqs = 3;
       gpuMemoryUtilization = 0.75;
       speculativeConfig = {
         method = "mtp";