b70-optimization-lab

MiniMax M2.7 INT4 AutoRound On B70

This packet is the maintained index for the MiniMax M2.7 INT4 AutoRound work on four Intel Arc Pro B70 GPUs. It links the established recipes and evidence without relocating the path-sensitive historical artifacts.

MiniMax results in this repository cover different tasks and must not be collapsed into one score:

Lane	Shape / purpose	Result	Quality and status	Primary pointer
Fresh Ubuntu 24 deployable endpoint	TP4, p512/n1536 comparison at 2K; service supports 32K	83.172 output tok/s, 110.896 total tok/s	Exact token hashes, semantic, arithmetic, and extended checks passed; maintained deployment starting point	110 tok/s-class deployment repro
Historical strict speed baseline	TP4, p512/n1536, ctx2048, batch 1	89.314195 output tok/s, 119.085594 total tok/s	Four clean repeats after strict quality gates; historical source/runtime packet	89 tok/s strict repro
Constrained structured HTML	TP4, short constrained output	94.406 effective accepted output tok/s	30/30 accepted, zero rejects; valid only for the declared constrained task	structured-lane note
Long-context service observation	TP4, prompt 32264 / output 64	63.91 output tok/s after TTFT	Approved service/capacity observation, not comparable to p512/n1536 speed rows	production service guide

Reproduction And Operations

• Fresh Ubuntu deployment: repro/minimax-m27-b70-110tps-ubuntu24-20260523.
• Older strict-speed reproduction: repro/minimax-m27-b70-89tps-20260520.
• Production c1 service: docs/minimax-production-c1-service.md.
• Session-cache and long-context research: experiments/minimax_xpu_kv_offload.
• ReAP/AutoRound source experiments: experiments/minimax-m27-reap-autoround-vllm.

Interpretation Rules

• Keep constrained output, unconstrained generation, short decode, prefill, long-context, and concurrent-service results in separate comparison classes.
• AutoRound INT4, GGUF IQ4, FP16/BF16 activation, and compressed-KV lanes are separate quality identities.
• A faster graph result is not promoted unless the corresponding exact-token, semantic, arithmetic, and practical-task gates pass.
• Treat old accepted LocalMaxxing rows as historical evidence when later quality work changes their classification.
• Preserve failed patches and negative screens. Many of the best MiniMax leads came from ruling out launch flags and locating collective or fused-boundary costs instead.

Historical Detail

The former top-level README chronology is preserved verbatim in the 2026-07-11 archive. Detailed chronological evidence remains in ../../notes/, structured results in ../../data/, and patch outcomes in ../../patches/; those established paths are intentionally not reorganized.

This site is open source. Improve this page.