replayssm_read#
Per-request, per-head NVFP4 read refresh with ModelOpt’s E4M3/E2M1 rounding.
Request seeds and logical element offsets keep the random stream independent of cache slot placement and CUDA graph replay.
Per-request, per-head NVFP4 read refresh with ModelOpt’s E4M3/E2M1 rounding.
Request seeds and logical element offsets keep the random stream independent of cache slot placement and CUDA graph replay.