[Performance] nemotron-hindi-4b latency is 17.0% above reference

This was generated by AI during triage.

Summary

nemotron-hindi-4b completes aligned BF16 generation and matches exact token IDs, but TRTMC latency is 17.0% above the reference.

Observed behavior

• Precision: TRTMC BF16, reference BF16
• Output contract: exact token IDs (matched)
• Request: batch 1, 20 new tokens
• Measurement: 3 warmups followed by 10 iterations
• Candidate p50: 717.17 ms
• Reference p50: 612.85 ms
• Regression: +17.0% (allowed margin: 5%)

The ten measured iterations are stable enough that this is not a single-sample outlier.

Expected behavior

The generation path should remain within the registered performance margin after warmup.

Acceptance criteria

• Profile prefill, decode, KV-cache access, and synchronization.
• Identify and fix the measured regression.
• Preserve exact-token output and the existing performance gate.

Reproduction platform

• Hardware: NVIDIA Thor X
• Architecture: aarch64
• TensorRT: 11.0
• GPU carveout: 48 GiB

No source file or test is named. Profile prefill, decode, KV-cache access, and synchronization on NVIDIA Thor X using the stated BF16 batch-1 workload, comparing warmed iterations with the reference. Done means identifying and correcting the regression while preserving exact token IDs and the existing performance gate.