[Core][SM70] Split Qwen3.8 prefill and decode compilation

[yangzhuxinyzx]

Purpose

Keep the established Qwen3.8 large-prefill graph and no-MTP decode fast graph in one TP4 service without duplicating parameters or KV/state caches. Prefill uses async disk-mmap PLE plus the original FlashQLA-SM70 TileLang GDN path; FULL decode uses rank-local pinned PLE and the M=1 graph operators.

Test Plan

• Focused CPU config/phase and delayed-spawn tests
• Exact TP4 NVFP4 no-MTP repeated 8K and 256K prefill
• Exact TP4 513-token pure decode throughput
• Greedy arithmetic, Chinese response, and thinking-parser quality gates
• Authenticated local and 1cattunnel API probes

Test Result

• Focused dual-graph tests: 5 passed
• Delayed PLE spawn/executor tests: 3 passed
• Ruff, py_compile, and git diff --check: passed
• Startup: zero systemd restarts; mmap worker retained 47.684 GiB in file-backed mappings with about 1.3 GiB RSS
• Graph audit: (1,8193) contains ple_offload_wait and no decode-only M=1 operators; (1,2) contains local pinned PLE gather plus FP16 GEMV/fused HC/fused GDN
• 8K exact hash 8aab945ad780...: 6,878 and 6,984 prompt tok/s warm (historical matched result: 7,026 tok/s)
• No-MTP decode: 513 tokens / 5.9602 s = 86.07 tok/s
• 256K gate: 262,143 prompt + 1 output; 51.8063 s pure prefill = 5,060 tok/s
• Prefix caching disabled; maximum model length 262,144; V2 runner; TP4 V100; FP16 KV; NVFP4 checkpoint
• Quality: 173*29 -> 5017, 41*37 -> 1517, accurate two-sentence Chinese lunar-phase answer; thinking exposed in message.reasoning
• Authenticated local and public /v1/models: HTTP 200