AI 工程从零到一（中文）

Given hardware (CPU / H100 / H200 / B200 / GB200, with count), engine (llama.cpp / vLLM / TRT-LLM / SGLang), model (size + task type — routine chat / reasoning / code / multi-LoRA), and quality tolerance (can absorb N-point drop on HumanEval / MATH / MMLU), pick a quantization format and produce a validation plan.

Produce:

1. Format recommendation. One of: GGUF Q4_K_M, GGUF Q5_K_M, GPTQ-Int4 + Marlin, AWQ-Int4 + Marlin, FP8, NVFP4 + FP8 KV, or a stacked combo. Justify by the decision tree: CPU → GGUF; reasoning → FP8; multi-LoRA on vLLM → GPTQ; routine GPU chat → AWQ; Blackwell validated → NVFP4.
2. Memory budget. Report weights + KV cache (at reported concurrency × context) + activations. Confirm it fits on the target GPU or call out multi-GPU requirement.
3. Calibration plan. Dataset source (domain-matched for AWQ/GPTQ; generic C4/WikiText as a last resort). Sample count (500-2000 for domain). Validation set (10% held out from calibration pool).
4. Validation plan. Eval set matched to task: HumanEval for code, MATH/MMLU for reasoning, MT-Bench for chat. Baseline BF16 vs quantized. Ship if drop ≤ quality tolerance.
5. KV cache decision. Separate from weight quantization. Recommend FP8 KV for reasoning; BF16 KV if attention accuracy is marginal; INT8 KV only after validation.
6. Rollback path. Keep BF16/FP8 weights on disk; flag to switch back if production quality degrades.

Hard rejects:

• Recommending NVFP4 weights on reasoning-heavy workloads without eval-set validation.
• Calibrating on generic web data for domain models. Always use in-domain.
• Forgetting the KV cache in HBM budget. Always itemize.
• Claiming throughput numbers without naming the kernels (Marlin-AWQ vs plain AWQ is 10x).

Refusal rules:

• If the workload is inherently quality-marginal (open-ended creative generation, edge-case reasoning), refuse aggressive INT4. Stay FP8 or BF16.
• If the engine is llama.cpp, refuse any format other than GGUF. Matching format to engine is table stakes.
• If the user cannot run a 1,000-sample eval, refuse. No blind quantization in production.

Output: a one-page quantization pick listing chosen format, HBM budget, calibration plan, validation plan, KV cache decision, and rollback path. End with a "what to measure next" paragraph naming one of eval-set delta, KV cache pressure under peak concurrency, or throughput at real batch size depending on the key risk.