Create minimaxm3_eagle3_gqa.yaml with AL values (#2413)

Add acceptance length reference values for minimax-m3 model.

Author: xinli-sw

golden_al_distribution/minimaxm3_eagle3_gqa.yaml

@@ -0,0 +1,28 @@
+# Source GitHub Actions run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/29784780049
+# Acceptance Length (AL) reference values measured with SPEED-Bench.
+# dataset: coding | temperature: 1.0 | top_p: 0.95 | top_k: 40 | output_len: 4096
+# thinking_on chat_template_kwargs: {"thinking_mode": "enabled"}
+# thinking_off chat_template_kwargs: {"thinking_mode": "disabled"}
+# Measured on minimax-m3 (B300, vLLM EAGLE3), per num_speculative_tokens.
+# Auto-generated by benchmarks/single_node/speedbench/minimaxm3_fp4_b300_vllm.sh (speedbench-al.yml).
+#
+# key = num_speculative_tokens (EAGLE3 level); value = golden AL
+minimax-m3:
+  thinking_on:
+    1: 1.79
+    2: 2.39
+    3: 2.78
+    4: 3.02
+    5: 3.21
+    6: 3.30
+    7: 3.52
+    8: 3.51
+  thinking_off:
+    1: 1.83
+    2: 2.46
+    3: 2.97
+    4: 3.29
+    5: 3.72
+    6: 3.76
+    7: 3.75
+    8: 3.93