Add tenstorrent/README.md — Tenstorrent N300s training port

tenstorrent/README.md

@@ -0,0 +1,78 @@
+# Tenstorrent port notes
+
+## What changed
+
+- Added `--backend auto|cuda|tt|cpu`
+- Added Tenstorrent runtime setup through TT-XLA / PJRT
+- Training path uses XLA-style optimizer stepping on TT (`xm.optimizer_step`)
+- Checkpoints are always saved with CPU tensors so they can move between CUDA and TT
+- TT inference avoids dynamic KV-cache assumptions and uses a static-shape path for robustness
+- Added TT tuning flags:
+  - `--tt_dtype fp32|bf16`
+  - `--tt_bfp8`
+  - `--tt_weight_bfp8`
+  - `--tt_optimization_level`
+  - `--tt_trace`
+  - `--tt_spmd` (experimental)
+
+## Koyeb setup sketch
+
+```bash
+python3 -m venv .xla-venv
+source .xla-venv/bin/activate
+pip install pjrt-plugin-tt --extra-index-url https://pypi.eng.aws.tenstorrent.com/
+pip install torch datasets transformers sentencepiece safetensors
+```
+
+## Training example
+
+```bash
+python n_tenstorrent_port.py train \
+  --backend tt \
+  --preset nano_3x \
+  --steps 10000 \
+  --batch_size 4 \
+  --block 576 \
+  --save_dir /workspace/ckpts_expansion_tt \
+  --tt_dtype bf16 \
+  --tt_optimization_level 1
+```
+
+## Warm-start from NVIDIA checkpoint and continue training on TT
+
+```bash
+python n_tenstorrent_port.py train \
+  --backend tt \
+  --preset nano_3x \
+  --warmstart_from /workspace/ckpts_expansion/final.pt \
+  --steps 10000 \
+  --batch_size 4 \
+  --block 576 \
+  --save_dir /workspace/ckpts_tt_resume \
+  --tt_dtype bf16
+```
+
+## Inference from NVIDIA-trained checkpoint on TT
+
+```bash
+python n_tenstorrent_port.py infer \
+  --backend tt \
+  --mode ar \
+  --ckpt /workspace/ckpts_expansion/final.pt \
+  --prompt "The capital of France is" \
+  --max_new 64 \
+  --tt_dtype bf16
+```
+
+## Experimental two-chip attempt on N300
+
+```bash
+python n_tenstorrent_port.py train \
+  --backend tt \
+  --tt_spmd \
+  --batch_size 8 \
+  --block 576 \
+  --steps 10000
+```
+
+Use the SPMD flag carefully. It is intentionally marked experimental in the script.