Feature: Add HuggingFaceTB/SmolLM3-3B native TensorRT support

[ruiling-smartbear]

Request type

Model onboarding

Proposed model

Add a native TensorRT-Model-Connect family for
HuggingFaceTB/SmolLM3-3B,
pinned to an immutable Hugging Face revision.

SmolLM3 is an Apache-2.0, publicly downloadable 3B decoder-only text model.
Its published configuration declares model_type: "smollm3" and
SmolLM3ForCausalLM, so this is not an exact Qwen or Llama checkpoint alias.

Why a new family

The model is a compact long-context, multilingual reasoning model with GQA and
a documented 3:1 NoPE/RoPE layout. It is not eligible for the current native
Llama path, which deliberately accepts only model_type == "llama",
LlamaForCausalLM, and full rotary embeddings. A model-owned family keeps
the checkpoint mapping, positional encoding, build graph, native runtime, and
validation contract isolated rather than weakening those established guards.

Proposed initial scope

1. Create owned Python, native-runtime, and E2E roots named smollm3.
2. Implement the smallest BF16, single-GPU text-generation path; no quantized,
   tensor-parallel, tool-calling, or extended-context support in the first PR.
3. Add one exact checkpoint/revision E2E manifest with a short, deterministic
   chat prompt and a Hugging Face reference comparison.
4. Publish support only after a successful target-hardware build and retained
   E2E/parity evidence.

Explicit non-goals

• No generic decoder abstraction or changes to existing Llama/Qwen ownership.
• No claim that the initial profile supports the model's 64K/128K context,
  YaRN extrapolation, or every chat-template/tool mode.
• No relaxation of an oracle, threshold, or test criterion to obtain a pass.

Maintainer input requested

Before implementation, please confirm the preferred initial GPU class,
TensorRT/CUDA cohort, cache length, and reference-validation acceptance
criteria. I will keep the first PR scoped to that agreed profile and split
performance qualification into a follow-up if needed.

Initial reconnaissance

• No open issue or pull request for SmolLM3 was found before filing.
• The checkpoint is public and the upstream card documents Transformers
  support from v4.53.0.
• The local repository's model-onboarding guide requires a new model family to
  carry Python builder, native runtime, and model-owned E2E descriptors, plus
  target-hardware evidence before calling the checkpoint supported.

[yifeif-nv]

Hi @ruiling-smartbear thanks for rising this issue! Looks like a good model request and a good first issue! LMK if you want to work on this or you are open to some other community member to take this task

[Moviw]

I’d like to take this on. I’ll keep the first contribution family-local and scoped to the BF16 single-GPU path described here, following the existing model onboarding and validation contracts. I’ll post the exact pinned checkpoint/revision and validation evidence with the PR.

[ruiling-smartbear]

@yifeif-nv sorry for the slow reply — yes, I'd like to work on this, and have been since filing.

@Moviw apologies for the crossed wires, and thanks for offering to pick it up. That's on me for not answering sooner.

Where it stands on my side — the three-sided unit is implemented and locally validated:

• families/smollm3 with the per-layer NoPE schedule from no_rope_layers, falling back to no_rope_layer_interval. Verified against SmolLM3Attention.use_rope on all 36 layers, and cross-checked by driving one RoPE and one NoPE layer in isolation with identical hidden states but different relative spacing.
• src/runtime/models/smollm3 with strategy smollm3_decoder_kv_cache. SmolLM3's chat template always emits a system block and does not close it with <|im_end|>, so it needed its own format rather than the generic ChatML path; asserted byte-for-byte against apply_chat_template() in both reasoning modes.
• tests/e2e/models/smollm3 with one manifest pinned to a07cc9a04f16550a088caea529712d1d335b0ac1, bf16, hf_transformers oracle, continuation parity.
• YaRN for the 128k configuration the model card documents, checked against an exact float64 reference. The native KV path still declines scaled RoPE and routes to the standard decoder.
• No shared infrastructure touched; the tokenizer needed no work, since the shared native BPE pre-tokenizer already classifies SmolLM3's Llama-3 style regex correctly. tools/model_ci.py validate passes at 89 families.

What is left is the engine build and the parity run, which I still need a TensorRT box for. I can open a draft PR so it is visible and reviewable rather than sitting on my fork.

@Moviw if you would rather not duplicate effort — happy to split. I would take the base continuation contract, and there is plenty left over: the chat-instruct E2E case, tensor parallel, or the long-context YaRN qualification. Your call.

[chaofengw-nv]

@Moviw @ruiling-smartbear
Hi, thanks for actively contributing PRs to TRTMC. I’ve assigned the issue to you. Feel free to add the code for this model if you’d like, we’re glad to help out.