feat(model): add native Inkling text bridge and LoRA support

Signed-off-by: Neil Kale <263453039+kalectory@users.noreply.github.com>

Author: kalectory

docs/index.md

@@ -43,6 +43,7 @@ models/gemma/index.md
 models/glm/index.md
 models/gpt_oss/index.md
 models/hy_v3/hy-v3.md
+models/inkling/inkling.md
 models/kimi/index.md
 models/llama/index.md
 models/minimax/index.md

docs/models/inkling/inkling.md

@@ -0,0 +1,47 @@
+# Inkling
+
+The Inkling bridge converts the text backbone of the unquantized
+[`thinkingmachines/Inkling-Small`](https://huggingface.co/thinkingmachines/Inkling-Small)
+checkpoint (276B total / 12B active parameters). It uses native Megatron tensor
+and expert parallel projections, Transformer Engine normalization and grouped
+experts, and PyTorch FlexAttention for the learned relative-position bias.
+
+```python
+from megatron.bridge import AutoBridge
+
+bridge = AutoBridge.from_hf_pretrained("/shared/models/Inkling-Small")
+provider = bridge.to_megatron_provider()
+provider.expert_model_parallel_size = 8
+provider.expert_tensor_parallel_size = 1
+provider.finalize()
+model = provider.provide_distributed_model(wrap_with_ddp=False)
+```
+
+Download a fixed checkpoint revision to the shared path before launching all
+ranks. Conversion reads the published `model.llm.*` safetensors layout directly.
+Transformers' materialized model uses different names and is not a conversion
+source. Dense, routed, and shared `w13` weights contain interleaved gate/up rows.
+Full checkpoint export preserves the source audio, vision, and auxiliary MTP
+tensors unchanged; these components are not trained.
+
+For attention-only LoRA, target `linear_qkv`, `linear_r`, and `linear_proj`.
+Native adapter export produces Inkling's published `wq_du`, `wk_dv`, `wv_dv`,
+`wr_du`, and `wo_ud` names. No downstream adapter-name translation is required.
+
+Supported execution is unpacked text training and full-sequence scoring with
+zero dropout and context parallel size 1. Packed sequences, cached generation,
+CUDA graphs, quantized checkpoint import, and multimodal/MTP training are not
+implemented. FlexAttention backward through the relative bias is tested on
+PyTorch 2.13; older releases are not qualified. Use a serving engine for cached
+generation.
+
+The two-layer FP32 functional suite passes at TP1/EP1, TP2/EP1, and TP1/EP2
+on H200 with full-layer recomputation. It checks all 42 checkpoint tensors,
+forward/backward parity against Transformers, nonzero LoRA export (20 tensors),
+merged export, and native base/adapter checkpoint restore. Operator tests also
+compare relative-attention and convolution gradients against Transformers.
+FP32 internal router/convolution weights retain the imported BF16 values but
+export in FP32 unless an export dtype is requested.
+
+Full Inkling-Small training, serving parity, and capacity qualification remain
+pending; the tiny-model checks do not qualify the 276B checkpoint.

src/megatron/bridge/models/__init__.py

@@ -104,6 +104,7 @@
     HybridModelConfig,
     HybridModelProvider,
 )
+from megatron.bridge.models.inkling import InklingBridge, InklingModelProvider
 from megatron.bridge.models.kimi import (
     KimiK2Bridge,
     KimiK3Bridge,
@@ -257,6 +258,8 @@
     "HybridModelBuilder",
     "HybridModelConfig",
     "HybridModelProvider",
+    "InklingBridge",
+    "InklingModelProvider",
     "KimiK2Bridge",
     "KimiK3Bridge",
     "KimiK3ModelProvider",

src/megatron/bridge/models/inkling/__init__.py

@@ -0,0 +1,18 @@
+# Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from megatron.bridge.models.inkling.inkling_bridge import InklingBridge, InklingModelProvider
+
+
+__all__ = ["InklingBridge", "InklingModelProvider"]