fix(recipes): pin transformers in deepseek-v32-fp4 perf jobs (cherry-pick #8690)

[dagil-nvidia]

Summary

Cherry-pick of #8690 to release/1.1.0. Pins transformers==4.57.6 in the deepseek-v32-fp4 perf recipes so the aiperf install step does not silently upgrade transformers to 5.x, which lacks model_type=deepseek_v32 support and breaks tokenizer load.

Fixes DYN-2878
Fixes https://nvbugs/6109655

Conflict resolution

release/1.1.0 already shipped aiperf==0.7.0 (vs. main's 0.6.0), so the cherry-pick conflicted on the install line in both agg-round-robin/perf.yaml and disagg-kv-router/perf.yaml. Resolved by keeping aiperf==0.7.0 from the release branch and adding the transformers==4.57.6 pin from the upstream fix. Final line in both files:

pip install "aiperf==0.7.0" "transformers==4.57.6"

The transformers pin is the actual fix; aiperf version is unchanged from what release/1.1.0 already had.

Original PR

• Main PR: fix(recipes): pin transformers==4.57.6 in deepseek-v32-fp4 perf jobs #8690
• Merge commit: 11077f26f787e8dc8c2ea409caf31a792fb24d1e

Test plan

• CI passes
• QA re-runs recipes/deepseek-v32-fp4/trtllm/disagg-kv-router/perf.yaml and agg-round-robin/perf.yaml against the rc9 build and confirms perf job exits 0

Made with Cursor