Initial K3 snapshot: 0.5B KDA/MLA/MoE train path

Standalone tree split from LLMRL/projects/kda. Includes Triton dt_bias backward fix, train_k3 --preset 0.5b, SFT, Docker runtime, and tests.
2026-08-25 14:43:17 +08:00
commit 584f7e9e73
140 changed files with 21592 additions and 0 deletions
@@ -0,0 +1,43 @@
+"""Default container command: print usage and refuse to start a silent toy run."""
+
+from __future__ import annotations
+
+HELP = """kda image: one GPU runtime for train / monitor / eval.
+
+Do not docker run without a command — toy overfit is not the default.
+
+Volumes (host → container):
+  ./ckpts            → /workspace/kda/ckpts
+  pretrain corpus    → /data/pretrain   (ro)
+  SFT bitext         → /data/sft        (ro)
+  frozen eval        → /data/eval       (ro)
+  HF tokenizer cache → /cache/huggingface
+  SWANLAB_API_KEY    via -e  (never COPY into the image)
+
+Examples:
+  docker run --rm --gpus all kda:<tag> python train.py
+  docker run --rm --gpus all -e SWANLAB_API_KEY \\
+    -v "$PWD/ckpts:/workspace/kda/ckpts" \\
+    kda:<tag> python train_k3.py --preset toy --attnres off
+  docker run --rm --gpus all kda:<tag> swanlab ping
+  docker run --rm --gpus all \\
+    -v "$PWD/ckpts:/workspace/kda/ckpts" -v "$PWD/data/eval:/data/eval:ro" \\
+    kda:<tag> python -m kda.training.eval_mt \\
+      --ckpt /workspace/kda/ckpts/k3_wiki.pt \\
+      --src /data/eval/zh2en.src.txt --ref /data/eval/zh2en.ref.txt \\
+      --target-lang en
+  docker run --rm --gpus all \\
+    -v "$PWD/ckpts:/workspace/kda/ckpts" -v "$PWD/data/sft:/data/sft:ro" \\
+    kda:<tag> python train_sft.py --ckpt /workspace/kda/ckpts/k3_wiki.pt \\
+      --data /data/sft/toy.jsonl
+  docker run --rm --gpus all kda:<tag> python -m pytest -q
+"""
+
+
+def main() -> None:
+    print(HELP)
+    raise SystemExit(2)
+
+
+if __name__ == "__main__":
+    main()