{
	"schemaVersion": 1,
	"slug": "p2-inference-systems",
	"title": "LLM 推理机制 Lab",
	"description": "从 KV Cache、Batched Decode 到 MHA/GQA 的 NumPy 实验 Starter 与可复查证据",
	"parent": {
		"href": "/learning/phase-2/",
		"label": "阶段 2 实验路线"
	},
	"usage": {
		"mode": "starter",
		"readme": "README.md",
		"currentTask": "assignments/m2-c-cached-mha.md",
		"currentCheckoff": "checkoffs/m2-c-cached-mha.md",
		"publicSource": "src",
		"privateSource": ".work",
		"publicationPolicy": "task-archive"
	},
	"publish": [
		"lab.json",
		"README.md",
		"Makefile",
		"grade-lab",
		"pyproject.toml",
		"uv.lock",
		"assignments/m2-b-causal-attention.md",
		"assignments/m2-c-cached-mha.md",
		"checkoffs/p0-numpy-semantics.md",
		"checkoffs/m0-kv-cache.md",
		"checkoffs/m1-batch-size-scan.md",
		"checkoffs/m2-a1-split-heads.md",
		"checkoffs/m2-a2-qkv-projection.md",
		"checkoffs/m2-b-causal-attention.md",
		"checkoffs/m2-c-cached-mha.md",
		"src/inference_lab/__init__.py",
		"src/inference_lab/multi_head_attention.py",
		"tests/test_multi_head_attention.py",
		"grader_tests/test_m2_c_cached_mha.py"
	],
	"notes": [
		"The published src directory is the intentionally unfinished Starter, not the maintainer's local implementation.",
		"Completed work lives in the ignored .work Git repository and is never published.",
		"Current and archived assignments plus blank checkoff templates are published; filled answers remain private.",
		"Markdown opens in reading mode; code, data, and other artifacts open as published files."
	],
	"checks": {
		"required": [
			"README.md",
			"assignments/m2-b-causal-attention.md",
			"assignments/m2-c-cached-mha.md",
			"Makefile",
			"grade-lab",
			"grader_tests/test_m2_c_cached_mha.py",
			"checkoffs/p0-numpy-semantics.md",
			"checkoffs/m0-kv-cache.md",
			"checkoffs/m1-batch-size-scan.md",
			"checkoffs/m2-a1-split-heads.md",
			"checkoffs/m2-a2-qkv-projection.md",
			"checkoffs/m2-b-causal-attention.md",
			"checkoffs/m2-c-cached-mha.md",
			"src/inference_lab/multi_head_attention.py",
			"tests/test_multi_head_attention.py"
		],
		"starterContains": {
			"src/inference_lab/multi_head_attention.py": [
				"M2-C TODO: append K/V and compute cached attention",
				"raise NotImplementedError"
			]
		},
		"immutableTemplates": {
			"checkoffs/p0-numpy-semantics.md": "be6ca16bead04e24bb1ca05c0524381e822d23b02097dc4ea6a8f2c6c06ef5e6",
			"checkoffs/m0-kv-cache.md": "ae119c3f24f87946e34ed42a42e81c8bb7f3c9308aa69e5a8e281cc88347ac99",
			"checkoffs/m1-batch-size-scan.md": "c5c9ab200bed849605c34aad52c549de886ef5040b3fd225620cbf7091199e89",
			"checkoffs/m2-a1-split-heads.md": "5812bd143074ab5ed27dd72c20833ec626a752ef5b2d82e5e2b43d4affa4a3a8",
			"checkoffs/m2-a2-qkv-projection.md": "a70b0ad2d14d3020fb0277f422538d8f7f4b1d5db4990fca524ffb637bfe578f",
			"checkoffs/m2-b-causal-attention.md": "a207b6d6fab292a2710d5c592c3e4a3af29b6541867e6a16706fdfd458522c86",
			"checkoffs/m2-c-cached-mha.md": "cf7aafd6ab79773ca1d2ed94bb789ca94209c8b571be70cc77ed3adb7d7bdd22"
		}
	}
}