--- spark.py.orig 2026-11-05 12:40:00.417941408 +0900 +++ spark.py 2026-11-05 13:18:45.255779819 +0900 @@ -139,7 +139,7 @@ spec_args = [ "--speculative-draft-model-path", str(draft), "--speculative-algorithm", "DFLASH", "--speculative-draft-model-quantization", "unquant", - "--speculative-num-draft-tokens", "32", "--speculative-draft-window-size", "16384", + "--speculative-num-draft-tokens", os.environ.get("RIPPLE_DRAFT_TOKENS", "16"), "--speculative-draft-window-size", "16384", "--speculative-draft-attention-backend", "triton", ] if speculative else [] args = prepare_server_args( @@ -154,13 +154,13 @@ --attention-backend flashinfer --fp4-gemm-backend flashinfer_cutlass --flashinfer-autotune-skip-ops fp8_gemm --context-length 131072 --max-total-tokens 139264 --mem-fraction-static 0.78 - --max-running-requests 1 --cuda-graph-bs-decode 1 - --cuda-graph-backend-prefill disabled --disable-radix-cache + --max-running-requests __MRR__ --cuda-graph-bs-decode __BS__ + --cuda-graph-backend-prefill disabled __RADIX__ --weight-loader-disable-mmap --weight-loader-drop-cache-after-load --model-loader-extra-config '{"enable_multithread_load": false}' --random-seed 42 --watchdog-timeout 1800 --log-level warning --tool-call-parser ripple --reasoning-parser ripple - """) + """.replace("__RADIX__", "" if os.environ.get("RIPPLE_RADIX", "1") == "1" else "--disable-radix-cache").replace("__MRR__", os.environ.get("RIPPLE_MAX_RUNNING", "3")).replace("__BS__", " ".join(str(b) for b in range(1, int(os.environ.get("RIPPLE_MAX_RUNNING", "3")) + 1)))) ) super().__init__(server_args=args) self._chat = OpenAIServingChat(self.tokenizer_manager, self.template_manager)