Repository navigation

[MD-TRT] Llama and Qwen export TP examples with KV cache - #4784

Open

apbose wants to merge 11 commits into

mainfrom

abose/llama_qwen_tp_export_kvcache_refresh

Commits on Oct 8, 2026
- Llama export TP example with KV cache
  apbose
  committed
  b67a611
- adding qwen model
  apbose
  committed
  626fc2c
- fix: allocate static KV cache on the input device
  apbose
  committed
  6511f63
- fix: preserve scalar input bounds in serialized exported programs
  apbose
  committed
  617869e
- fix: serialize the actual calling interface of cached TP models
  apbose
  committed
  90a6cae
- feat: add FP32 TP export and saved-engine accuracy checks
  apbose
  committed
  c653db5
- style: clean up TP export helper lint
  apbose
  committed
  85d33c6
- docs: inline TP accuracy checks and simplify precision findings
  apbose
  committed
  b22468e
- refactor: rename TP LLM example and document both cache variants
  apbose
  committed
  dfaecdc
- docs: update the distributed TP example reference
  apbose
  committed
  8cf3fa1
- Fix TP LLM benchmarks and document experiment results
  apbose
  committed
  a1fffae