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 @@
+[project]
+name = "kda"
+version = "0.0.1"
+description = "Hand-written KDA implementation from naive recurrent to fused Triton + training + inference"
+requires-python = ">=3.10"
+
+# kda 是独立 uv 项目 (自带 uv.lock + .venv), 自声明运行依赖, 不依赖仓库根环境
+dependencies = [
+    "torch>=2.9.0",      # uv 解析最新满足版 (含 CUDA 构建); 与根环境 2.9.0+cu128 仅下限一致
+    "einops>=0.7.0",     # kda/ops/reference/chunkwise.py 的 rearrange
+    "packaging>=23.0",   # vendored FLA utils version checks
+    "sentencepiece>=0.2.0",  # toy SentencePiece
+    "datasets>=3.0.0",       # 中文 wiki 语料加载
+    "transformers>=4.51.0",  # Qwen3 tokenizer for 0.5b preset
+    "swanlab>=0.9.7",
+]
+
+# 训练机 / Docker 镜像: uv sync --extra train
+[project.optional-dependencies]
+train = [
+    "swanlab>=0.6.0",
+    "sacrebleu>=2.4.0",
+    "langdetect>=1.0.9",
+]
+
+# uv run / uv sync 默认安装 dev group
+[dependency-groups]
+dev = [
+    "pytest>=7.0",
+    "torchlens>=2.34",   # 计算图展开集成测试; 未装时测试模块自动 skip
+    "tensorlens>=0.0.3", # Flask viewer 集成测试; 未装时测试模块自动 skip
+]
+
+[tool.uv]
+# 公共入口: `from kda import CausalLM, KDAConfig, K3Config, chunk_kda`
+package = true
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["kda"]