Add Ministral-3-14B-Instruct-2512 (Leanstral) contrib model

[jimburtoft]

Summary

NxDI contrib for Ministral-3-14B-Instruct-2512 (Leanstral) — a 14B dense GQA text decoder running on trn2.3xlarge (TP=4, LNC=2) with SDK 2.29.

Key results:

• 252 tok/s aggregate at BS=4 (1.8x H100 FP8)
• Multi-KV-head TKG NKI kernel with virtual-batch approach — matches baseline TPOT at BS=4, only 5-9% overhead at BS=8
• FP8 E4M3 → BF16 text-only checkpoint extraction from VL model
• Full vLLM 0.16 benchmark suite (4 workloads × 2 concurrency levels × 2 batch sizes)

What's included

• src/extract_text_model.py — FP8→BF16 text extraction, strips vision keys
• src/setup_patches.py — 6 runtime patches for SDK 2.29 compatibility
• src/attention_block_tkg_multi_kv.py — Multi-KV-head TKG kernel (NKI 0.3.0)
• src/multi_kv_adapter.py — Adapter for attention_base.py
• src/fix_nki030.py — NKI 0.3.0 compatibility
• bench.py — Async streaming benchmark (TTFT, TPOT, tok/s)
• Integration test

Architecture

• 40 layers, hidden=5120, 32 Q / 8 KV heads, d_head=128
• At TP=4: q_heads_per_rank=8, kv_heads_per_rank=2
• Uses LlamaForCausalLM code path via --hf-overrides (vLLM 0.16 auto-promotes Mistral to Pixtral)

Performance (trn2.3xlarge, TP=4, LNC=2, SDK 2.29)

BS=4 — TKG kernel matches baseline

Workload	Baseline TPOT	TKG TPOT	Overhead
short-short (128/128)	15.8ms	15.7ms	-0.6%
long-long (2048/512)	17.3ms	17.0ms	-1.7%

BS=8 — TKG kernel with modest overhead

Workload	Baseline TPOT	TKG TPOT	Overhead
short-short (128/128)	16.7ms	17.5ms	+4.8%
long-long (2048/512)	18.7ms	20.1ms	+7.5%

NxDI gaps identified

1. TKG kernel hardcodes kv_heads=1 — this contrib adds multi-KV-head support
2. convert_state_dict_to_fused_qkv assumes standard Llama head ratios
3. RMS norm epsilon not passed from config to model base