fix(perf): halve MBS on b300/gb300 MoE perf configs to fit HybridEP IB QP cap

[ko3n1g]

Claude summary

What

Halve micro_batch_size on the b300 / gb300 MoE perf base configs for nemotron_3_nano, qwen3_30b_a3b, and qwen3_vl_30b_a3b.

Why

NVIDIA/Megatron-LM PR #4094 (commit a08e259f32, 2026-05-08) added a deliberate Python-side guardrail in megatron/core/transformer/moe/fused_a2a.py:

# DeepEP calculates tx_depth = 3 * num_tokens + 1.
# InfiniBand strictly asserts tx_depth < 65536.
tx_depth = 3 * num_tokens + 1
if tx_depth >= 65536:
    raise ValueError(
        f"HybridEP RDMA Queue Pair depth ({tx_depth}) exceeds the InfiniBand "
        f"hardware limit of 65535. ..."
    )

This means num_tokens ≤ 21844 per rank is now a hard constraint ((65536-1) // 3 = 21845).

Nine perf tests in our nightly fall on the wrong side of that cap — all with num_tokens = 32768 (tx_depth = 98305):

Config	MBS (current)	seq_len	TP	CP	num_tokens	After fix
qwen3_30b_a3b_64gpu_b300_{bf16,fp8_mx}	8	4096	1	1	32768	16384
qwen3_30b_a3b_64gpu_gb300_{bf16,fp8_mx}	8	4096	1	1	32768	16384
qwen3_vl_30b_a3b_8gpu_gb300_bf16	8	4096	1	1	32768	16384
nemotron_3_nano_8gpu_b300_{bf16,fp8_mx}	4	8192	1	1	32768	16384
nemotron_3_nano_8gpu_gb300_{bf16,fp8_mx}	4	8192	1	1	32768	16384

The corresponding gb200 / b200 variants of these models already use the lower MBS and are unaffected by the new check — that's the proof point that the lower-MBS geometry works for these recipes.

Alternatives considered

Option	After fix num_tokens	Why not
Halve MBS ✅ (this PR)	16384	Smallest diff; preserves the EP=8/TP=1 perf-tuning point chosen for these configs; matches what the gb200 variants already do
Set TP=2	16384	Crosses NVLink domain only above 8 GPUs (fits here single-node), but shifts the workload geometry; less appropriate for an MoE recipe deliberately tuned at TP=1
Set CP=2	16384	Pure sequence-axis split, but less-validated path for these models; risk of regression

Test plan

• Run tests label triggers MBridge unit tests
• Re-run the affected internal nightly tests against this branch (via --mbridge-ref); confirm they no longer trip the IB cap
• Re-baseline golden values for the 9 affected tests in nemo-ci (separate MR there)

Affected presets

nemotron_3_workload_base_configs.py:
  NEMOTRON_3_NANO_PRETRAIN_CONFIG_GB300_BF16_V1     MBS 4 -> 2
  NEMOTRON_3_NANO_PRETRAIN_CONFIG_GB300_FP8_MX_V1   (alias)
  NEMOTRON_3_NANO_PRETRAIN_CONFIG_GB300_NVFP4_V1    (alias)
  NEMOTRON_3_NANO_PRETRAIN_CONFIG_B300_BF16_V1      MBS 4 -> 2
  NEMOTRON_3_NANO_PRETRAIN_CONFIG_B300_FP8_MX_V1    (alias)
  NEMOTRON_3_NANO_PRETRAIN_CONFIG_B300_NVFP4_V1     (alias)

qwen3_workload_base_configs.py:
  QWEN3_30B_A3B_PRETRAIN_CONFIG_GB300_BF16_V1       MBS 8 -> 4
  QWEN3_30B_A3B_PRETRAIN_CONFIG_GB300_FP8_CS_V1     MBS 8 -> 4
  QWEN3_30B_A3B_PRETRAIN_CONFIG_GB300_FP8_MX_V1     (alias)
  QWEN3_30B_A3B_PRETRAIN_CONFIG_B300_BF16_V1        MBS 8 -> 4
  QWEN3_30B_A3B_PRETRAIN_CONFIG_B300_FP8_CS_V1      MBS 8 -> 4
  QWEN3_30B_A3B_PRETRAIN_CONFIG_B300_FP8_MX_V1      (alias)

qwen3_vl_workload_base_configs.py:
  QWEN3_VL_30B_A3B_PRETRAIN_CONFIG_GB300_BF16       MBS 8 -> 4
  QWEN3_VL_30B_A3B_PRETRAIN_CONFIG_GB300_FP8_CS     MBS 8 -> 4
  QWEN3_VL_30B_A3B_PRETRAIN_CONFIG_GB300_FP8_MX     (alias)

Related

• Upstream: NVIDIA/Megatron-LM #4094 (the guardrail)
• Internal triage MR: dl/joc/nemo-ci!2324
• Internal bucket issue: dl/joc/nemo-ci#4037

[claude]

Light Review

LGTM — straightforward MBS reduction to stay under the HybridEP IB QP depth cap.

Verified:

• All alias configs (FP8_MX, NVFP4, FP8_CS) correctly inherit the new MBS via assignment — no stale values.
• GB200/B200 variants already use MBS=2 (nemotron_3_nano) and MBS=4 (qwen3_30b_a3b), confirming the target values are valid.
• Post-fix num_tokens = MBS * seq_len = 16384 gives tx_depth = 49153, well under the 65535 cap.

Minor nit (PR description only, not code): The table in the PR body lists qwen3_30b_a3b_64gpu_b300 and qwen3_30b_a3b_64gpu_gb300, but the actual configs have num_gpus=8. Looks like a copy-paste from the 235B configs.

Suggested test cases:

All affected configs need golden-value re-baselining since MBS changed:

• nemotron_3_nano_8gpu_gb300_bf16_perf
• nemotron_3_nano_8gpu_gb300_fp8_mx_perf
• nemotron_3_nano_8gpu_gb300_nvfp4_perf
• nemotron_3_nano_8gpu_b300_bf16_perf
• nemotron_3_nano_8gpu_b300_fp8_mx_perf
• nemotron_3_nano_8gpu_b300_nvfp4_perf
• qwen3_30b_a3b_8gpu_gb300_bf16_perf
• qwen3_30b_a3b_8gpu_gb300_fp8_cs_perf
• qwen3_30b_a3b_8gpu_gb300_fp8_mx_perf
• qwen3_30b_a3b_8gpu_b300_bf16_perf
• qwen3_30b_a3b_8gpu_b300_fp8_cs_perf
• qwen3_30b_a3b_8gpu_b300_fp8_mx_perf
• qwen3_vl_30b_a3b_8gpu_gb300_bf16_perf
• qwen3_vl_30b_a3b_8gpu_gb300_fp8_cs_perf
• qwen3_vl_30b_a3b_8gpu_gb300_fp8_mx_perf