perf: optimize Qwen3.5-122B SGLang serving for H200 (#5)

* perf: optimize Qwen3.5-122B SGLang serving for H200

Add performance flags to improve throughput and memory efficiency:
- FP8 KV cache quantization for ~2x concurrent requests
- Chunked prefill to prevent OOM on 128K context
- FlashInfer attention backend
- Aggressive scheduling for better batching

Made-with: Cursor

* chore: increase context length to 262K (native max)

Made-with: Cursor

* fix: use H200 default chunked-prefill-size of 8192

4096 was below SGLang's auto-detected default for H200 GPUs (<160GB),
which unnecessarily limited prefill throughput.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: PierreLeGuen

Qwen3.5-122B.yaml

@@ -51,8 +51,12 @@ x-sglang-qwen35-122b-common: &sglang-qwen35-122b-common
     sglang serve
     --model-path Qwen/Qwen3.5-122B-A10B
     --tp 4
-    --mem-fraction-static 0.80
-    --context-length 131072
+    --mem-fraction-static 0.88
+    --context-length 262144
+    --kv-cache-dtype fp8_e4m3
+    --chunked-prefill-size 8192
+    --attention-backend flashinfer
+    --schedule-conservativeness 0.5
     --reasoning-parser qwen3
     --tool-call-parser qwen3_coder
     --log-requests-level 0