[None][feat] Add AD custom model for MiniMax-M2 family (#243)

* [None][feat] Add AD custom model for MiniMax-M2 family

Replace the existing MiniMax-M2 MoE patch with a full custom model
implementation using AD canonical ops. Covers both MiniMaxAI/MiniMax-M2
and MiniMaxAI/MiniMax-M2.5 (same architecture, model_type: minimax_m2).

Key architecture features:
- MoE with 256 experts, top-8, sigmoid routing + e_score_correction_bias
- GQA (48 Q heads, 8 KV heads, head_dim=128)
- Partial RoPE (rotary_dim=64 out of head_dim=128)
- Per-layer QK normalization (RMSNorm on full Q/K before reshape)
- FP8 block-wise quantized checkpoint

Canonical ops used: torch_rmsnorm, torch_rope_with_explicit_cos_sin,
torch_attention (GQA-native, no repeat_kv), torch_moe.

Signed-off-by: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com>

* [None][fix] Disable fuse_finegrained_fp8_moe for MiniMax-M2 in registry config

The trtllm fused MoE kernel fails with NVRTC compilation error for
MiniMax-M2's MoE configuration (256 experts, block-wise FP8). Add
transform disablement and torch-simple compile backend to the model
registry config so --use-registry works out of the box.

Signed-off-by: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com>

---------

Signed-off-by: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com>

Author: lucaslie

examples/auto_deploy/model_registry/configs/minimax_m2.yaml

@@ -1,4 +1,12 @@
 # MiniMax-M2 - override model dtype and attention backend for AutoDeploy
 attn_backend: flashinfer
+# Disable fuse_finegrained_fp8_moe: the trtllm fused MoE kernel fails with
+# NVRTC compilation error for MiniMax-M2's MoE config (256 experts, block-wise FP8).
+# Use torch-simple compile backend since the Triton MoE fallback is not
+# CUDA-graph-capturable.
+compile_backend: torch-simple
+transforms:
+  fuse_finegrained_fp8_moe:
+    enabled: false
 model_kwargs:
   torch_dtype: bfloat16

examples/auto_deploy/model_registry/models.yaml

@@ -229,7 +229,7 @@ models:
   yaml_extra: ['dashboard_default.yaml', 'world_size_8.yaml']
 # --- MiniMax-M2.5 (Feb 2026) ---
 - name: MiniMaxAI/MiniMax-M2.5
-  yaml_extra: ['dashboard_default.yaml', 'world_size_8.yaml']
+  yaml_extra: ['dashboard_default.yaml', 'world_size_8.yaml', 'minimax_m2.yaml']
 # --- MiMo-V2-Flash (Feb 2026) ---
 - name: XiaomiMiMo/MiMo-V2-Flash
   yaml_extra: ['dashboard_default.yaml', 'world_size_8.yaml']

tensorrt_llm/_torch/auto_deploy/models/custom/__init__.py

@@ -14,6 +14,7 @@
 from .modeling_internlm3 import InternLM3ForCausalLM
 from .modeling_kimi_k2 import KimiK2ForCausalLM, KimiK25ForConditionalGeneration
 from .modeling_llama3 import Llama3ForCausalLM
+from .modeling_minimax_m2 import MiniMaxM2ForCausalLM
 from .modeling_mistral import MistralForCausalLM
 from .modeling_mistral3 import Mistral3ForConditionalGeneration, Mistral3TextForCausalLM
 from .modeling_nemotron_flash import NemotronFlashForCausalLM, NemotronFlashPreTrainedTokenizerFast
@@ -48,6 +49,7 @@
     "KimiK2ForCausalLM",
     "KimiK25ForConditionalGeneration",
     "Llama3ForCausalLM",
+    "MiniMaxM2ForCausalLM",
     "MistralForCausalLM",
     "Mistral3ForConditionalGeneration",
     "Mistral3TextForCausalLM",