[Bug]: fused_cat_fp8 didn't do rotate_activation for DSA

System Info

System Info
TensorRT-LLM version: v1.3.0rc11
Model: deepseek v3.2

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

python examples/llm-api/quickstart_advanced.py
--model_dir DeepSeek-V3.2/
--attention_backend TRTLLM
--moe_backend WIDEEP
--kv_cache_fraction 0.15
--disable_overlap_scheduler
--disable_kv_cache_reuse
--tokens_per_block 64
--max_seq_len 5000
--max_num_tokens 4096
--max_tokens 200
--enable_chunked_prefill
--tp 8
--prompt "$(cat long_context.txt)"

long_context.txt contains a single prompt about 4k tokens

Expected behavior

Do rotate_activation after cat [qk_pe, qk_nope], then do quant

actual behavior

Do cat then quant, but no rotate_activation

additional notes

Although deepseek v3.2 is very robust, I think rotate_activation is still needed. @kaiyux

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

Start with examples/llm-api/quickstart_advanced.py and reproduce using the command and DeepSeek-V3.2 settings given in the issue. Trace the fused_cat_fp8 path for DSA and compare the qk_pe/qk_nope handling with the expected rotate_activation-before-quantization order. Done means the supported reproduction follows that order without regressing the FP8 path.