Add TKG-optimized contribs for 4 dense Mistral models

[jimburtoft]

Summary

• Add NKI TKG fused attention kernel optimization for 4 dense Mistral-family models
• Each model directory is self-contained with setup scripts, TKG kernel, NKI 0.3.0 compat fixes, and integration tests
• All benchmarked on SDK 2.29 (DLAMI 20260410), vLLM 0.16.0, trn2.3xlarge

Models

Model	tok/s (TKG)	tok/s (baseline)	Improvement	TKG Mode	GPU Comparison
Mistral-7B-Instruct-v0.3	115	97	+19%	multi-KV	1.47x GPU, 15% cheaper
Ministral-8B-Instruct-2410	99	82	+21-29%	multi-KV	1.32x GPU, 24% cheaper
Ministral-3-3B-Instruct-2512-BF16	158	134	+18-27%	multi-KV	GPU blocked
Mistral-Small-3.1-24B-Instruct-2503	48	45	+6-14%	stock	1.27x GPU, 26% cheaper

Key Technical Details

TKG kernel: Fused attention NKI kernel derived from Leanstral (Ministral-3-14B contrib). Two variants:

• Multi-KV: For models with >1 KV head per TP rank (7B, 8B, 3B at TP=4). Uses virtual batch approach.
• Stock: For models with 1 KV head per TP rank (24B at TP=8). No custom kernel needed, just config flags.

Limitations:

• TKG requires max_num_seqs=1 (BIR verification failure at BS>1 on neuronx-cc 2.24)
• --no-enable-prefix-caching required (block KV cache corruption)
• Ministral 3B and 24B require text extraction from multimodal checkpoint (LlamaForCausalLM path)
• Ministral 8B requires config patch (sliding_window=null, layer_types removed)

Files Changed

• New: Mistral-7B-Instruct-v0.3/ (10 files)
• New: Ministral-8B-Instruct-2410/ (12 files, includes patch_config.py)
• New: Ministral-3-3B-Instruct-2512-BF16/ (11 files, includes extract_text_model.py)
• Updated: Mistral-Small-3.1-24B-Instruct-2503/ (8 new files + README update, includes extract_text_model.py)