b70-optimization-lab

MiniMax M2.7 REAP AutoRound W4A16 on vLLM/XPU

Target model:

• HF repo: MJPansa/MiniMax-M2.7-REAP-172B-A10B-AutoRound-W4A16
• Local path: /mnt/fast-ai/llm-models/minimax-m2.7-reap-autoround-w4a16
• Active vLLM tree: /home/steve/src/vllm
• Baseline lane to compare against: /mnt/fast-ai/llm-models/minimax-m2.7-int4-autoround

Why this lane exists:

• It keeps the known-good MiniMax M2.7 production recipe intact.
• The REAP checkpoint is materially smaller than the Lasimeri AutoRound model.
• It uses the same MiniMaxM2ForCausalLM architecture and AutoRound W4A16 quantization path, so it should exercise the existing XPU MiniMax/vLLM work before any new DeepSeek or STEP implementation work.

Initial fit estimate:

• HF repo size: 91,544,661,302 bytes total, 91,512,175,232 bytes of safetensors.
• Decimal/GiB: 91.54 GB / 85.26 GiB total.
• TP4 floor: about 21.3 GiB of safetensors per B70 before runtime overhead, KV cache, graph capture, and fragmentation.
• This is a much cleaner 4x32GB fit than the 120.7 GB Lasimeri checkpoint already running in production.

Run order:

1. scripts/check-hf-metadata.sh
2. scripts/download-model.sh
3. scripts/quality-smoke.sh
4. scripts/bench-decode.sh

Do not submit LocalMaxxing results from this lane until quality gates pass against the existing MiniMax canaries and repeatability is clean.

This site is open source. Improve this page.