B300 NVFP4 port P1 stage-1: advance flashinfer_trtllm NVFP4 MoE + modelopt_fp4 path to upstream HEAD
Port the flashinfer-series NVFP4 path (GLM-5.2-NVFP4 + --moe-runner-backend flashinfer_trtllm
+ --fp4-gemm-backend flashinfer_trtllm) from upstream HEAD. b300-exp's flashinfer files were
byte-identical to fork-base (2d288ba8c9), so wholesale-replace == replaying the commit chain
without dragging baseline files forward (cherry-pick rejected: avg 17-54 files/commit entanglement).
Wholesale-replace (advanced to HEAD): modelopt_quant.py, moe_runner/flashinfer_trtllm.py,
flashinfer_trtllm_moe.py, fp4_utils.py, moe_runner/base.py. Brings split-w13 gate/up scales
(#27588), deferred-finalize symm-output (#27720), 4over6 + per-token activation (default-off).
ADD: marlin_utils_fp4.py, mxfp4_flashinfer_trtllm_moe.py (PackTopkIds), nvfp4_online.py (C1).
Surgical deps: environ +5 flags, common.alias_or_bind_derived_param, fp8_utils.apply_fp8_linear_bmm_flashinfer,
pynccl_allocator.is_tensor_in_symmetric_mempool, moe/utils.is_flashinfer_cutedsl_v1_path.
D1: gate modelopt-fp8 enable_flashinfer_bmm OFF by default (SGLANG_MODELOPT_FP8_FLASHINFER_BMM)
to keep the GLM-5.1-FP8 baseline byte-identical (upstream #28333 defaults it on for sm100).
Attention backend classes deferred (NSA bypasses them, verified 0.6.12-compatible).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
leavelet
b0baea52ef