Update Nemotron 3 Super B200 BF16 config

[zuriz-nv]

Update Nemotron3 Super B200 BF16 Config

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[malay-nagda]

/claude review

[claude]

Light review — LGTM

Small, self-contained perf-config update for Nemotron 3 Super on B200 BF16. No correctness issues found.

Observations (non-blocking):

• Removing expert_model_parallel_size=64 from BASE_NEMOTRON_3_SUPER_CONFIG_B200 is a no-op: the parent BASE_NEMOTRON_3_SUPER_CONFIG already sets EP=64, so B200_FP8_MX_V1 (= base) and B200_NVFP4_V1 (replace from base, TP=2/EP=64) are functionally unchanged. Good cleanup.
• B200_BF16_V1 now recomputes core_attn while attn is in cuda_graph_scope. Other configs in this file that recompute core_attn keep attn out of the graph scope (e.g. NANO H100 FP8 uses scope ["mamba"]). Worth confirming the measured B200 run captured cleanly with attn graphed and core_attn recomputed, since it differs from the sibling pattern.
• cuda_graph_scope order here is ["mamba", "attn", ...] vs ["attn", "mamba", ...] elsewhere — cosmetic only, order is not functional.

Suggested test cases

• nemotron_3_super_64gpu_b200_bf16_perf
• nemotron_3_super_64gpu_b200_fp8_mx_perf
• nemotron_3_super_64gpu_b200_nvfp4_perf

(The base BASE_NEMOTRON_3_SUPER_CONFIG_B200 was edited, so all three B200 precisions derived from it are in scope; BF16 changed materially, FP8-MX/NVFP4 should be net-neutral.)

[malay-nagda]

/ok to test 2e28c81