| # Gemma 4 31B runs that need special serving on Hopper (SM90): | |
| # FlashInfer has no mm_prefix (bidirectional image attention) support on SM90, and NVFP4 KV exists only | |
| # in FlashInfer, so BF16-FI and FP4 are served text-only (--language-model-only): RULER + MRCR, no DocVQA. | |
| # FP8 KV with images needs the TRITON_FLASH_ATTN composite (image prefills -> Triton, text -> FA3/FA4); | |
| # BF16-TFA reruns DocVQA on the same composite as FP8's matched DocVQA baseline. | |
| # Usage: gemma_rest.sh fa (GPU 0, after after_tb.sh finishes there) | |
| # gemma_rest.sh fi (GPU 1) | |
| cd ~/bench | |
| export KVQ_MODEL=~/models/gemma-4-31B-it KVQ_NAME=google/gemma-4-31B-it KVQ_REASONING_PARSER=gemma4 \ | |
| KVQ_MAX_LEN=262144 KVQ_EXTRA_BODY='{"chat_template_kwargs":{"enable_thinking":true}}' \ | |
| KVQ_DATA=data_gemma KVQ_EXP_OUT=exp2_gemma KVQ_MRCR_OUT=mrcr_gemma | |
| case $1 in | |
| fa) | |
| export GPU=0 PORT=8000 | |
| until grep -q "AFTER_TB_DONE GPU0" ~/bench/after_tb_gpu0.log; do sleep 120; done | |
| KVQ_BACKEND=TRITON_FLASH_ATTN MRCR=1 ./exp2.sh FP8 FP8 | |
| KVQ_BACKEND=TRITON_FLASH_ATTN LENS="" ./exp2.sh BF16-TFA BF16 | |
| ;; | |
| fi) | |
| export GPU=1 PORT=8001 KVQ_SERVE_ARGS=--language-model-only KVQ_DOCVQA=0 MRCR=1 | |
| ./exp2.sh BF16-FI BF16-FI | |
| ./exp2.sh FP4 FP4 | |
| ;; | |
| esac | |
| echo "GEMMA_REST_DONE $1 $(date -u)" | |