blackwell: phase 6 — manual end-to-end validation on B200

[Andrewxu313]

Run DeepSeek-R1 + one GQA model end-to-end on the B200 pod with MMLU Δ ≤ 0.5 pp vs. the H20 baseline. Manual validation (no self-hosted runner).

Parent: #169.
Status: blocked by the deferred Phase 2 kernel port.

[drunkcoding]

sm100 / B200 — back-test pending GPU availability

The sm120 (consumer Blackwell) decode path is now implemented and merged (#253; see status comment on #153 for file locations + commits). That work is sm120-specific and does not cover sm100/B200.

sm100 (B200) end-to-end validation in this issue is blocked on hardware — to be back-tested when a B200 node is available. Until then:

• The BUILD_ARCH=sm100 path builds SM80-class fused ops retargeted to sm_100; WGMMA-only kernels fall back at runtime, and FA3/FlashMLA is the Blackwell attention path (per blackwell: phase 2 — batchgen_kernels SM100 build path #204).
• The real +sm100 WGMMA→UMMA kernel port (blackwell: phase 2 — batchgen_kernels SM100 build path #204, [blackwell] sm100: setup.py _sm100_extensions infra #224–[blackwell] sm100: Fused INT4 grouped + SiLU Triton kernel (K2.5 MoE FFN) #231) remains deferred to the dedicated follow-up project.

Action: hold this e2e validation open; run the B200 smoke + MMLU-Pro accuracy pass once a B200 host is provisioned.