[None][feat] Add AutoDeploy custom model for OpenELM family (#198)

Onboard the OpenELM architecture (apple/OpenELM-270M/1_1B/3B-Instruct)
as a custom AutoDeploy model. This is a heterogeneous transformer with:
- Per-layer varying query/KV head counts (GQA)
- Per-layer varying FFN intermediate sizes
- Fused QKV projection with Q/K normalization
- Shared input/output embeddings (no separate lm_head)
- GLU-style FFN (proj_1 = fused gate+up, proj_2 = down)

Uses canonical AD IR ops: torch_rmsnorm, torch_rope_with_explicit_cos_sin,
torch_attention. Config loaded from checkpoint via trust_remote_code=True.

Updated openelm.yaml with attn_backend=flashinfer (trtllm backend produces
degenerate output for OpenELM). Works with torch-cudagraph, default batch
settings from dashboard_default.yaml.

All 3 variants produce coherent generation via build_and_run_ad.py.

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

Author: lucaslie

examples/auto_deploy/model_registry/configs/openelm.yaml

@@ -1,3 +1,7 @@
-# Configuration for Apple OpenELM models
-# These models require Llama-2 tokenizer
+# Configuration for Apple OpenELM models (270M, 1.1B, 3B)
+# These models use the Llama-2 tokenizer (confirmed by Apple's CoreNet docs).
 tokenizer: meta-llama/Llama-2-7b-hf
+
+# Override dashboard_default's attn_backend=trtllm which produces degenerate
+# output for OpenELM. Use flashinfer which works with torch-cudagraph.
+attn_backend: flashinfer

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

@@ -22,6 +22,7 @@
 from .modeling_nemotron_flash import NemotronFlashForCausalLM, NemotronFlashPreTrainedTokenizerFast
 from .modeling_nemotron_h import NemotronHForCausalLM
 from .modeling_olmo3 import Olmo3ForCausalLM
+from .modeling_openelm import OpenELMForCausalLM
 from .modeling_qwen2 import Qwen2ForCausalLM
 from .modeling_qwen3_5_moe import Qwen3_5MoeForCausalLM, Qwen3_5MoeForConditionalGeneration
 from .modeling_qwen3_moe import Qwen3MoeForCausalLM
@@ -60,6 +61,7 @@
     "NemotronFlashPreTrainedTokenizerFast",
     "NemotronHForCausalLM",
     "Olmo3ForCausalLM",
+    "OpenELMForCausalLM",
     "Phi4ForCausalLM",
     "Phi4FlashForCausalLM",
     "Phi4MMForCausalLM",