[Kernel][SM70] Default safe FP8 prescaled M1 decode

[yangzhuxinyzx]

Purpose

Supersede #309 with a latest-main rebuild based on 1b1c9faeb82c30bf3d81158279404cd884bbfb19.

Absorb the exact E4M3 power-of-two exponent factor into reversible FP16 UE8M0 scales at load time for the measured SM70 M=1 tensor/runtime contract. Admission uses hardware, block layout, operator role, tensor shape, TP/PP topology, scheduler state, and actual scale values; it never reads model/checkpoint/type/architecture identity.

The route is default-on when its exact contract is available and remains behind the faster accepted QPN8 route. Missing old-extension capability or unsafe scale values fall back to ordinary TurboMind; explicit VLLM_SM70_FP8_PRESCALED_M1_DECODE=1 fails closed, and =0 is the rollback. A distinct M=1 operator schema prevents old 8K-prefill-only extensions from receiving the new shapes.

Test Plan

• Reuse the recorded real-weight same-process CUDA Graph A/B/B/A operator screen.
• Run focused admission, scale-reversibility, M=1 dispatch, and environment-default tests.
• Run the complete repository pre-commit suite.
• No end-to-end rerun, per current audit scope and because all V100s are occupied.

Test Result

• Recorded real layer-0 TP4-rank0 screen: 320/320 changing-input comparisons bitwise equal across five M=1 shapes.
• Fused WQA/WKV: 63.330 us to 20.577 us, 3.078x.
• Five-shape weighted projection: -1.867 ms/token warm and -1.821 ms/token after cache scrub; production admission is intentionally narrower and does not claim endpoint TPOT.
• pytest -q tests/quantization/test_sm70_fp8_prefill_exact_dense.py -k prescaled: 4 passed.
• pytest -q tests/test_envs.py -k sm70_concurrency_tuning_envs: 1 passed.
• pre-commit run -a: passed.
• git diff --check origin/main: passed.

Evidence: /data/models/v100-dsv4-0731-pp2tp4-fp8-prescaled-decode-screen-20260826-r1/.