feat: add MiniMax-M3-MXFP8 recipe (vLLM, 4 nodes, TP=4)

MiniMax-M3-MXFP8 is the MXFP8 quantized variant of MiniMax M3
(~428B params / ~23B active), a native multimodal MoE with MSA
sparse attention supporting 1M context.

Key recipe parameters (from vLLM Recipes page):
  - Mandatory: --block-size 128 (MSA sparse/index cache alignment)
  - Tool/reasoning: minimax_m3 parsers + auto tool choice
  - MXFP8 auto-detected from model weights (no --quantization flag)
  - --language-model-only to skip vision encoder (text-only saves VRAM)
  - --enable-expert-parallel for MoE expert sharding
  - --kv-cache-dtype fp8 for memory-efficient KV pool

Uses the same sparkrun vLLM image as deepseek-v4-flash and 4 nodes
with TP=4 (one GPU per node, matching existing minimax-m3 NVFP4
recipe).

recipes/minimax-m3-mxfp8/recipe.toml

@@ -0,0 +1,32 @@
+engine = "vllm"
+name = "MiniMax M3 MXFP8 (vLLM)"
+docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-v2"
+model = "MiniMaxAI/MiniMax-M3-MXFP8"
+required_nodes = 4
+tensor_parallel_size = 4
+pipeline_parallel_size = 1
+distributed_backend = "mp"
+
+# MXFP8 quantized MiniMax M3 (~440B params, ~23B active).
+# 4 nodes × 1 GPU (DGX Spark GB10 Blackwell), TP=4 across nodes.
+# MSA sparse attention requires --block-size 128.
+# Text-only mode skips the vision encoder to save VRAM.
+
+entrypoint = ["vllm", "serve"]
+
+args = [
+    "--block-size", "128",
+    "--kv-cache-dtype", "fp8",
+    "--gpu-memory-utilization", "0.85",
+    "--tool-call-parser", "minimax_m3",
+    "--reasoning-parser", "minimax_m3",
+    "--enable-auto-tool-choice",
+    "--language-model-only",
+    "--enable-expert-parallel",
+    "--trust-remote-code",
+    "--served-model-name", "minimax-m3-mxfp8", "default",
+]
+
+[env]
+# M3 supports thinking via chat_template_kwargs.thinking_mode
+# enabled / adaptive / disabled — pass per-request in extra_body