#3227 [Other] FP4-Local setup dies at step 6: ModuleNotFoundError: No module named 'transformers'

Descripción

**Area**: `[Other]` (setup.py, the source installer)

**Desktop app, Terminal or source install**: source install (git clone, SETUP.bat) on Windows

**What happened**
Choosing the NVIDIA FP4-Local (QAD) model in SETUP.bat on a Python env that lacks the converter's packages dies at step 6. `tools/nvfp4_local_pack.py prepare` converts the n-gram table (~20 minutes of the 102 GB checkpoint), then reaches the dense weights, where `install()` imports llama.cpp's converter (`third_party/llama.cpp/conversion/base.py`: `from transformers import AutoConfig`) and the whole prepare exits with `ModuleNotFoundError: No module named 'transformers'`. `setup.py`'s package step only installs `requirements.txt`'s eight runtime packages; torch, transformers and safetensors appear nowhere in it. The user must find and install the three packages by hand and re-run.

**Smallest steps or command that reproduces it**
1. fresh clone (or any clone whose `.venv` lacks torch/transformers; check with `.venv\Scripts\python -m pip list`)
2. `SETUP.bat`, pick Qwen3.8-Flash-Next, pick the NVFP4 FP4-Local (NVIDIA QAD) model
3. step 6 prints the PLE shard progress, then the traceback above, exit 1

**Reproduction status**: reproduced on this PC (0.1.45, full timestamped log; the first prepare attempt of 2026-10-19). Static check on e88ee39d (v0.2.0-beta-12): `requirements.txt` and `setup.py` still do not mention torch/transformers, so the fresh-env path still dies.

**Environment**: Strata source install @ 477ac13c (v0.1.45), engine package 0.1.45, Microsoft Windows 11 Pro, Python 3.12.0 venv, RTX 5090 + RTX 3070, 96 GB RAM, model `nvidia/Qwen3.8-Flash-Next-FP4-Local` revision b9ce18a

**Workaround that worked here**: `.venv` `pip install torch transformers safetensors`, re-run SETUP: the prepare completed (dense.gguf 6.0 GB, experts.bin 63.28 GiB, ple-nvfp4.gguf 320001536 x 90 B marked `nvfp4x16`, mtp/rt) and the model serves.

**Fix**: the linked pull request installs a pinned `requirements-nvfp4-local.txt` through the existing `pip_install` immediately before a prepare that actually runs, and makes the tool name the missing package when it is run by hand.