H100 K2.5: add --kv-cache-dtype fp8

The H100 INT4 launcher was running with default bf16 KV cache while the
B300 launcher already uses fp8. With max_num_seqs * max_model_len (256K)
KV pool sizing, bf16 KV pinned too much memory inside gpu-mem-util's
budget, leaving no runtime headroom for fused_marlin_moe's 896 MiB
intermediate buffer. fp8 halves the KV pool footprint, freeing tens of
GB per GPU for runtime allocations.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: cquil11

benchmarks/single_node/agentic/kimik2.5_int4_h100.sh

@@ -59,6 +59,7 @@ vllm serve $MODEL \
 --reasoning-parser kimi_k2 \
 --tool-call-parser kimi_k2 \
 --compilation_config.pass_config.fuse_allreduce_rms true \
+--kv-cache-dtype fp8 \
 --trust-remote-code \
 $OFFLOAD_ARGS > "$SERVER_LOG" 2>&1 &
 SERVER_PID=$!