[bridge] feat: Add ERNIE 4.5 text-only MoE and VL bridges for Megatro…

examples/conversion/hf_megatron_roundtrip_multi_gpu.py

@@ -70,8 +70,10 @@
     # MiniMax-M2: QK norms stored as bf16 in HF, loaded as fp32 by Megatron config.params_dtype
     "q_norm.weight",
     "k_norm.weight",
-    # MiniMax-M2: router gate stored as fp32 in HF, loaded as bf16 via autocast_dtype
+    # MoE router gate stored as fp32 in Megatron, may be bf16 in HF
     "block_sparse_moe.gate.weight",
+    "mlp.gate.weight",
+    "moe.gate.weight",
 ]
 
 # FP8 dtypes whose dequantisation is inherently lossy — allclose is meaningless.

examples/models/vlm/ernie_vl/README.md

@@ -0,0 +1,71 @@
+# ERNIE 4.5 VL Examples
+
+This directory contains example scripts for ERNIE 4.5 Vision-Language (VL) MoE models.
+
+## Supported Models
+
+| Model | Parameters | Active Parameters | Type |
+|-------|-----------|-------------------|------|
+| ERNIE-4.5-VL-28B-A3B-Instruct | 28B | 3B | VL MoE |
+| ERNIE-4.5-VL-28B-A3B-Thinking | 28B | 3B | VL MoE |
+
+## Prerequisites
+
+- `--trust-remote-code` is required for the custom HuggingFace model class.
+- All scripts use a `WORKSPACE` environment variable for checkpoints. Default: `/workspace`.
+
+```bash
+export WORKSPACE=/your/custom/path
+```
+
+## Checkpoint Conversion
+
+### Import HF → Megatron
+
+```bash
+uv run python examples/conversion/convert_checkpoints.py import \
+    --hf-model baidu/ERNIE-4.5-VL-28B-A3B-Instruct \
+    --megatron-path ${WORKSPACE}/ERNIE-4.5-VL-28B-A3B-Instruct \
+    --torch-dtype bfloat16 \
+    --trust-remote-code
+```
+
+### Export Megatron → HF
+
+```bash
+uv run python examples/conversion/convert_checkpoints.py export \
+    --hf-model baidu/ERNIE-4.5-VL-28B-A3B-Instruct \
+    --megatron-path ${WORKSPACE}/ERNIE-4.5-VL-28B-A3B-Instruct/iter_0000000 \
+    --hf-path ${WORKSPACE}/ERNIE-4.5-VL-28B-A3B-Instruct-hf-export \
+    --trust-remote-code
+```
+
+See [conversion.sh](conversion.sh) for the full pipeline including multi-GPU round-trip validation.
+
+## Inference
+
+ERNIE 4.5 VL uses a processor API that differs from other VLMs (e.g., Qwen), so a
+dedicated inference script is provided instead of the generic `hf_to_megatron_generate_vlm.py`.
+
+### Run Inference from HF Checkpoint
+
+```bash
+uv run python -m torch.distributed.run --nproc_per_node=8 \
+    examples/models/vlm/ernie_vl/hf_to_megatron_generate_ernie_vl.py \
+    --hf_model_path baidu/ERNIE-4.5-VL-28B-A3B-Instruct \
+    --image_path "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/resolve/main/images/table.png" \
+    --prompt "Describe this image." \
+    --max_new_tokens 100 \
+    --tp 2 --pp 1 --ep 4 \
+    --trust_remote_code
+```
+
+See [inference.sh](inference.sh) for a ready-to-use launch script.
+
+## Architecture Notes
+
+ERNIE 4.5 VL uses a **dual-pool MoE** architecture:
+- Text and vision experts reside in separate pools within each MoE layer.
+- Each pool has its own router and routes tokens independently.
+- This design uses `SequentialMLP` (per-expert execution) rather than `GroupedMLP`
+  (batched GEMM), since the two pools cannot be merged into a single expert group.

examples/models/vlm/ernie_vl/conversion.sh

@@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+# 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.
+set -e
+
+# Workspace directory for checkpoints and results
+WORKSPACE=${WORKSPACE:-/workspace}
+# Supported model variants:
+#   ERNIE-4.5-VL-28B-A3B-Instruct, ERNIE-4.5-VL-28B-A3B-Thinking
+MODEL_NAME=ERNIE-4.5-VL-28B-A3B-Instruct
+
+EP=4
+TP=2
+PP=1
+
+# Import HF -> Megatron
+uv run python examples/conversion/convert_checkpoints.py import \
+    --hf-model baidu/${MODEL_NAME} \
+    --megatron-path ${WORKSPACE}/${MODEL_NAME} \
+    --torch-dtype bfloat16 \
+    --trust-remote-code
+
+# HF and Megatron models logits comparison validation
+uv run python -m torch.distributed.run --nproc_per_node=8 examples/conversion/compare_hf_and_megatron/compare.py \
+    --hf_model_path baidu/${MODEL_NAME} \
+    --megatron_model_path ${WORKSPACE}/${MODEL_NAME} \
+    --model_class "Ernie4_5_VLMoeForConditionalGeneration" \
+    --image_path "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/resolve/main/images/table.png" \
+    --prompt "Describe this image." \
+    --tp ${TP} --pp ${PP} --ep ${EP} \
+    --trust-remote-code
+
+# Export Megatron -> HF
+uv run python examples/conversion/convert_checkpoints.py export \
+    --hf-model baidu/${MODEL_NAME} \
+    --megatron-path ${WORKSPACE}/${MODEL_NAME}/iter_0000000 \
+    --hf-path ${WORKSPACE}/${MODEL_NAME}-hf-export \
+    --trust-remote-code
+
+# Round-trip validation
+uv run python -m torch.distributed.run --nproc_per_node=8 examples/conversion/hf_megatron_roundtrip_multi_gpu.py \
+    --hf-model-id baidu/${MODEL_NAME} --tp ${TP} --pp ${PP} --ep ${EP} --trust-remote-code