first commit

examples/command-serial.json

@@ -0,0 +1,40 @@
+{
+  "args": [
+    "vllm",
+    "serve",
+    "/model",
+    "--trust-remote-code",
+    "--kv-cache-dtype",
+    "fp8",
+    "--block-size",
+    "256",
+    "--tensor-parallel-size",
+    "8",
+    "--data-parallel-size",
+    "2",
+    "--enable-expert-parallel",
+    "--no-enable-flashinfer-autotune",
+    "--tokenizer-mode",
+    "deepseek_v4",
+    "--tool-call-parser",
+    "deepseek_v4",
+    "--enable-auto-tool-choice",
+    "--reasoning-parser",
+    "deepseek_v4",
+    "--max-model-len",
+    "4096",
+    "--max-num-batched-tokens",
+    "1024",
+    "--enforce-eager",
+    "--port",
+    "18000"
+  ],
+  "env": {
+    "VLLM_DISABLED_KERNELS": "PPUCutlassFp8BlockScaledMMKernel",
+    "VLLM_PPU_BF16_O_PROJ": "1",
+    "CUDA_VISIBLE_DEVICES": "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15",
+    "VLLM_MULTI_STREAM_GEMM_TOKEN_THRESHOLD": "0",
+    "VLLM_PPU_SERIAL_ATTENTION": "1",
+    "VLLM_ENGINE_READY_TIMEOUT_S": "3600"
+  }
+}

examples/launch.py

@@ -0,0 +1,20 @@
+import os,sys,json,resource,subprocess,signal,time
+from pathlib import Path
+base=Path(__file__).parent
+resource.setrlimit(resource.RLIMIT_MEMLOCK,(resource.RLIM_INFINITY,resource.RLIM_INFINITY))
+cfg=json.loads((base/"command-serial.json").read_text())
+env=os.environ.copy();env.update(cfg["env"])
+log=base/(sys.argv[1] if len(sys.argv)>1 else "serve-fixed.log")
+with log.open("w") as f:
+ p=subprocess.Popen(cfg["args"],env=env,stdout=f,stderr=subprocess.STDOUT,start_new_session=True,cwd="/workspace/pytorch")
+(base/"serve.pid").write_text(str(p.pid))
+print("SERVER_PID",p.pid,"LOG",log,flush=True)
+rc=p.wait()
+print("SERVER_EXIT",rc,flush=True)
+if rc:
+ try: os.killpg(p.pid,signal.SIGTERM)
+ except ProcessLookupError: pass
+ time.sleep(3)
+ try: os.killpg(p.pid,signal.SIGKILL)
+ except ProcessLookupError: pass
+sys.exit(rc)