feat(deepseek-v4): enable SM120 serving

[lucifer1004]

Summary

This is the draft integration PR for DeepSeek V4 serving on NVIDIA SM120. It combines the independently reviewable prerequisites with the remaining runtime and kernel wiring so the complete TP2 path can be reviewed and qualified in one place.

The prerequisite PRs remain independent:

• fix(kernel): avoid mixing CUDA toolkit headers #644: isolate CUDA toolkit headers from wheel-provided CUDA headers (merged; inherited from main)
• feat(kernel): add SM120 FlashInfer MXFP4 MoE #645: add the FlashInfer MXFP4 MoE solution for SM120
• fix(kernel): use Triton FP8 quantization on SM120 #646: select Triton FP8 quantization on SM120 (merged; inherited from main)
• fix(pd): correct DeepSeek V4 layerwise cache handoff #660: correct DeepSeek V4 layerwise PD cache handoff (independent PD correctness follow-up)

Integration scope

• expose mHC through tokenspeed-kernel and use the DeepGEMM nv_dev prenorm kernels
• preserve the existing DeepGEMM warmup flow, including mHC warmup
• select FlashInfer 0.6.14 sparse MLA for SM120 while retaining the existing FlashMLA path on SM90/SM100
• wire the SM120 Lightning Indexer kernels into the DeepSeek V4 runtime
• gate unsupported Triton RSAG/symmetric-memory paths by platform while keeping the existing fast paths on SM90/SM100
• align apache-tvm-ffi==0.1.12 across build, runtime, and release inputs

Dependencies

The branch now inherits tokenspeed-deepgemm==2.5.0.post20260629 from main. That wheel was published for lightseekorg/whl#2 from the required deepseek-ai/DeepGEMM nv_dev@a6b593d2826719dcf4892609af7b84ee23aaf32a revision, so the temporary VCS dependency and its merge blocker have been removed.

Validation

• source-built TokenSpeed stack check through Inferlab/Pixi: passed
• focused runtime tests: 21 passed
• kernel/API/compatibility tests: 41 passed, 77 platform/solution skips
• new sparse MLA, mHC, and DeepGEMM warmup tests: 13 passed
• prek run --all-files: passed
• PEP 517 wheel build and install-from-wheel smoke test: passed

After rebasing onto main at 08b1153, the source-built stack check and prek passed again; the focused runtime suite passed 21/21 and the focused kernel/API/call-site/warmup suite passed 37 tests with 55 expected platform/solution skips.

The SM120 stack plus #660 was qualified through Inferlab with TP2/EP2 and one prefill plus one decode replica. Full GSM8K reached 93.63% flexible exact match over 1,319 samples; random 8K input / 1K output completed 252/252 requests through concurrency 32, reaching 648.5 output tok/s at c32.

Before ready for review

• merge/rebase feat(kernel): add SM120 FlashInfer MXFP4 MoE #645 and remove its duplicated commit from this integration branch
• rerun upstream CI and final TP2 qualification on the resulting dependency set