[TRTLLM-6780][fix] Add multimodal data to dummy requests during memory profiling#7539

Merged

Funatiq merged 24 commits intoNVIDIA:mainfrom

johncalesp:add-mm-dummy-req

Oct 16, 2025

Commits on Oct 15, 2025

add multimodal dummy request to profiling - drafting stage
johncalesp
committed
guard against multimodal_input being None
johncalesp
committed
rebase and change logic
johncalesp
committed
check attribute original_arch is present
johncalesp
committed
rebase from main and add fix for only-text request for qwen2.5 and add unit test
johncalesp
committed
fix max_seq_len is less than max_num_tokens during profiling
johncalesp
committed
address yechang comments - p1
johncalesp
committed
fix rebase to 80dd8fe1973323eb8f01060788c0d5485a0ce0f8
johncalesp
committed
check for new function name and fix TextPrompt attribute for qwen
johncalesp
committed
address comments - change design to use default_multimodal_input_loader
johncalesp
committed
add additional arguments for mm data
johncalesp
committed
address comments: change unit test, and add more asserts
johncalesp
committed
fix rebase to commit b1738c3f189560a857ea1adcfdfb8e68c571c81d
johncalesp
committed
address code rabbit comments && remove mrope_config.mrope_position_ids from modeling qwen
johncalesp
committed
integrate latest feedback
johncalesp
committed
add unit test to test-db
johncalesp
committed
fix unit test by adding chunked prefill parameter
johncalesp
committed
fix test L40S-PyTorch-2.test_e2e.test_ptp_quickstart_multimodal[NVILA-8B-FP16-vila/NVILA-8B-video-False]
johncalesp
committed
fix tests A10-PyTorch-1.test_e2e.test_openai_chat_multimodal_example and A10-PyTorch-1.test_e2e.test_trtllm_serve_multimodal_example
johncalesp
committed
include flag to check chunked prefill flag during profiling
johncalesp
committed
change logic to get initial input_seq_len
johncalesp
committed
fix latest rebase from 7291cdc42287297bf72015e7201fede7985edeae
johncalesp
committed
Fix rebase to 1cdb0b6
johncalesp
committed
fix format file _util.py
johncalesp
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[TRTLLM-6780][fix] Add multimodal data to dummy requests during memory profiling#7539

[TRTLLM-6780][fix] Add multimodal data to dummy requests during memory profiling#7539
Funatiq merged 24 commits intoNVIDIA:mainfrom
johncalesp:add-mm-dummy-req

Commits on Oct 15, 2025

add multimodal dummy request to profiling - drafting stage

guard against multimodal_input being None

rebase and change logic

check attribute original_arch is present

rebase from main and add fix for only-text request for qwen2.5 and add unit test

fix max_seq_len is less than max_num_tokens during profiling

address yechang comments - p1

fix rebase to 80dd8fe1973323eb8f01060788c0d5485a0ce0f8

check for new function name and fix TextPrompt attribute for qwen

address comments - change design to use default_multimodal_input_loader

add additional arguments for mm data

address comments: change unit test, and add more asserts

fix rebase to commit b1738c3f189560a857ea1adcfdfb8e68c571c81d

address code rabbit comments && remove mrope_config.mrope_position_ids from modeling qwen

integrate latest feedback

add unit test to test-db

fix unit test by adding chunked prefill parameter

fix test L40S-PyTorch-2.test_e2e.test_ptp_quickstart_multimodal[NVILA-8B-FP16-vila/NVILA-8B-video-False]

fix tests A10-PyTorch-1.test_e2e.test_openai_chat_multimodal_example and A10-PyTorch-1.test_e2e.test_trtllm_serve_multimodal_example

include flag to check chunked prefill flag during profiling

change logic to get initial input_seq_len

fix latest rebase from 7291cdc42287297bf72015e7201fede7985edeae

Fix rebase to 1cdb0b6

fix format file _util.py