Feature: Add ornith-ai/Ornith-1.5-9B checkpoint support

This issue was drafted by AI at a user's request.

Request type

New model or checkpoint support

Problem and use case

Ornith-1.5-9B is a public,
MIT-licensed, dense 9B reasoning model derived from Qwen3.5. Its documented use
cases include coding, agentic workflows, long-context chat, reasoning output,
and tool calling. The checkpoint is small enough for single-GPU deployment, but
TensorRT-Model-Connect does not currently list or validate it as a supported
checkpoint.

TensorRT-Model-Connect already has a qwen3_5 family and native runtime path
for Qwen/Qwen3.5-9B. Users should be able to build and run the Ornith
checkpoint through the same public Model Connect workflow when its architecture
is compatible, with any checkpoint-specific differences handled and validated
explicitly.

Proposed outcome

Add supported, tested ornith-ai/Ornith-1.5-9B checkpoint coverage. The initial
scope can focus on the model card's primary text-generation/chat path:

• Build a bundle from the pinned public checkpoint through the normal trtmc
  builder interface.
• Run the bundle through the native runtime path on a single supported GPU.
• Reuse the existing Qwen3.5 family where the checkpoint is architecture- and
  weight-layout-compatible; record and handle any real differences rather than
  adding checkpoint-name-only behavior.
• Preserve the checkpoint's chat-template behavior, including thinking output
  and XML tool-call serialization, within the capabilities of the public
  runtime interface.
• Add focused upstream parity validation and list the checkpoint in the public
  supported-model documentation.

The checkpoint also contains a Qwen3.5 vision configuration and is tagged for
image-text-to-text use. If the first implementation supports only text input,
document that boundary clearly and track image/video input support separately.

Acceptance criteria

• The exact checkpoint and an immutable revision are registered in the
  model manifest/test inventory.
• An FP16 engine builds from the BF16 source checkpoint through the public
  Model Connect builder without manual checkpoint rewriting.
• The resulting bundle runs through the normal native runtime on an NVIDIA
  H100 80GB for a representative chat prompt.
• A normal reasoning/chat response and a tool-calling prompt are validated
  against the pinned upstream Transformers behavior using meaningful,
  behavior-aligned criteria.
• The implementation records whether the existing qwen3_5 family is
  reused unchanged or which checkpoint-specific config, weights, tokenizer,
  or chat-template differences require support.
• The supported-model documentation states the supported precision, task,
  GPU/runtime path, context-length boundary, and multimodal boundary.
• No test threshold, passing criterion, or validator is weakened to make
  the checkpoint pass.

Model and target details

• Model ID: ornith-ai/Ornith-1.5-9B
• Revision: 489cb97981b8654bcfcf30ce1f94ed1b62e07b53
• License: MIT
• Architecture: Qwen3_5ForConditionalGeneration (model_type: qwen3_5)
• Size/source dtype: approximately 9.65B parameters, BF16
• Initial task: reasoning-oriented text generation/chat and tool-call
  serialization
• Initial engine precision: FP16
• Initial target: NVIDIA H100 80GB, single GPU
• Expected path: public trtmc builder -> bundle -> native runtime
• Declared context window: 262,144 tokens; initial validation may use a smaller,
  explicitly documented bounded context

Alternatives considered

• Serve the checkpoint with vLLM, SGLang, or llama.cpp. These are documented
  upstream alternatives, but they do not produce or validate a Model Connect
  bundle for the native runtime.
• Use the already-supported Qwen/Qwen3.5-9B checkpoint. That exercises a
  related architecture but does not provide Ornith's trained weights and
  behavior.
• Point the existing qwen3_5 family at Ornith without adding a manifest and
  parity coverage. This would leave compatibility and regressions unsupported.

Additional context

• Model card: https://huggingface.co/ornith-ai/Ornith-1.5-9B
• Ornith-1.5 technical blog: https://ornith.ai/ornith_1_5.html

Submission checks

• I searched open and closed issues and found no duplicate request.
• I removed secrets, private/internal evidence, personal paths, and
  restricted artifacts.