-
Notifications
You must be signed in to change notification settings - Fork 0
Open
Labels
lbRoute bringup to lbRoute bringup to lbrun_testsRun model testsRun model testsrun_tests_t3000Run model tests on t3000Run model tests on t3000
Description
Port this t3000 model from a 2x4 mesh to a 1x8 mesh (TP=8).
Target: models/microsoft/Phi-3-mini-128k-instruct/t3000/functional
Tasks:
- Update models/microsoft/Phi-3-mini-128k-instruct/t3000/functional/model.py to set MESH_SHAPE = (1, 8) and adjust any mesh-axis/sharding assumptions for a 1x8 mesh.
- Keep architecture, dtypes, and cache behavior unchanged.
- Run demo + long eval:
python demo.py models/microsoft/Phi-3-mini-128k-instruct/t3000/functional/model.py
python eval.py models/microsoft/Phi-3-mini-128k-instruct/t3000/functional/model.py --model microsoft/Phi-3-mini-128k-instruct --prompt_file prompts/bringup_eval_long.txt --max_new_tokens 100 --max_seq_len - Update MODELS.md for the t3000 row and save demo.log/eval.log under models/microsoft/Phi-3-mini-128k-instruct/t3000/functional/.
Notes:
- Keep paged attention / paged KV cache behavior unchanged.
- If TT metal cache error: set TT_METAL_CACHE=/tmp/tt-metal-cache and TT_METAL_RUNTIME_ROOT=/proj_sw/user_dev/moconnor/tt-runtime-root.
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
lbRoute bringup to lbRoute bringup to lbrun_testsRun model testsRun model testsrun_tests_t3000Run model tests on t3000Run model tests on t3000
Projects
Status
Ready