Merge branch 'main' into mchochowski/heterogeneous_moe

Author: chochowski

.dev.commit

@@ -1 +1 @@
-cf081d5df0b34b665d214ff936f27489d7396876
+bfa33263ca06e6974410d0ea871b25e21c5aee85

.github/workflows/claude-review.yml

@@ -44,11 +44,15 @@ jobs:
 
       - name: Run Claude Code Review
         uses: anthropics/claude-code-action@v1
+        env:
+          ANTHROPIC_BASE_URL: ${{ secrets.NVIDIA_INFERENCE_URL }}
+          CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS: "1"
+          DISABLE_PROMPT_CACHING: "1"
         with:
-          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
+          anthropic_api_key: ${{ secrets.NVIDIA_INFERENCE_KEY }}
           claude_args: |
             --allowedTools "mcp__github_inline_comment__create_inline_comment,Bash(gh pr comment:*),Bash(gh pr diff:*),Bash(gh pr view:*),Bash(gh pr review:*),Read"
-            --model "claude-opus-4-6"
+            --model "${{ vars.CLAUDE_MODEL }}"
           prompt: |
             REPO: ${{ github.repository }}
             PR NUMBER: ${{ github.event.pull_request.number }}
@@ -113,8 +117,9 @@ jobs:
 
   manual-review:
     if: github.event_name == 'issue_comment' && github.event.issue.pull_request != null && contains(fromJson('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.comment.author_association)
-    uses: NVIDIA-NeMo/FW-CI-templates/.github/workflows/_claude_review.yml@v0.88.0
+    uses: NVIDIA-NeMo/FW-CI-templates/.github/workflows/_claude_review.yml@v1.7.0
     with:
+      model: ${{ vars.CLAUDE_MODEL }}
       prompt: |
         Mandatory workflow — never skip or reorder:
         1. Read the PR diff first.
@@ -174,4 +179,5 @@ jobs:
         mcp__github_inline_comment__create_inline_comment are optional —
         use them only for specific code suggestions.
     secrets:
-      ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+      NVIDIA_INFERENCE_URL: ${{ secrets.NVIDIA_INFERENCE_URL }}
+      NVIDIA_INFERENCE_KEY: ${{ secrets.NVIDIA_INFERENCE_KEY }}

.main.commit

@@ -1 +1 @@
-25f6a09440c70ca6135ca44bc4dfbea4726c136a
+0823c731ed7d793aef047b6a64f2dbbf32bf6e2c

3rdparty/Megatron-LM

@@ -13,7 +13,7 @@
 ## 📣 News
 - [06/22/2026] **Megatron Bridge 0.5.0 released!** Highlights include expanded LLM and multimodal support (Qwen3.5, DeepSeek V4, Ernie 4.5, GLM-5/4.7, StepFun Step-3.5/3.7, MiMo-V2, Gemma 4, Falcon H1, Ling MoE V2, Nemotron-3 Nano Omni, Qwen3-Omni, Qwen3-ASR, and Nemotron Diffusion), MegatronMIMO and Energon v7 training updates, evaluator backend integration, eval-time context parallelism, deterministic recipes, quantized FP8/MXFP4 export, CUDA graph/performance improvements, and Megatron Inference/tokenizer unification with Megatron-LM. Huge thanks to our community contributors: [@HowardZorn](https://github.com/HowardZorn), [@hy2826](https://github.com/hy2826), [@bo-ke](https://github.com/bo-ke), [@beccohov](https://github.com/beccohov), [@dhiaEddineRhaiem](https://github.com/dhiaEddineRhaiem), [@pavelgein](https://github.com/pavelgein), [@ccclyu](https://github.com/ccclyu), [@hbhflw2000](https://github.com/hbhflw2000), and [@HollowMan6](https://github.com/HollowMan6)! See the [full release notes](https://github.com/NVIDIA-NeMo/Megatron-Bridge/releases/tag/v0.5.0).
 
-- [06/16/2026] NVIDIA topped [MLPerf Training v6.0](https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/) across every benchmark, including the new DeepSeek-V3 and GPT-OSS MoE training workloads. Megatron Bridge serves as the packaging layer for the NeMo 26.06 training stack that integrates full-iteration CUDA graphs, HybridEP/router optimizations, all-to-all overlap, MXFP8 attention, and pipeline-layout balancing; the blog highlights [DeepSeek-V3 training](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/scripts/performance/configs/deepseek) at **1,648 TFLOPS/GPU** (**6,338 tokens/sec/GPU**) on GB300. To try related runs, start from the [performance recipes](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/scripts/performance); the corresponding container is expected with the NeMo 26.06 release soon.
+- [06/16/2026] NVIDIA topped [MLPerf Training v6.0](https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/) across every benchmark, including the new DeepSeek-V3 and GPT-OSS MoE training workloads. Megatron Bridge serves as the packaging layer for the NeMo 26.06 training stack that integrates full-iteration CUDA graphs, HybridEP/router optimizations, all-to-all overlap, MXFP8 attention, and pipeline-layout balancing; the blog highlights [DeepSeek-V3 training](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/src/megatron/bridge/perf_recipes/deepseek) at **1,648 TFLOPS/GPU** (**6,338 tokens/sec/GPU**) on GB300. To try related runs, start from the [performance recipes](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/scripts/performance); the corresponding container is expected with the NeMo 26.06 release soon.
 
 - [06/04/2026] [**NVIDIA Nemotron 3 Ultra**](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16) is now public! Day-0 support for the 550B-A55B hybrid Mamba-Transformer MoE model is available on the [`nemotron_3_ultra`](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/nemotron_3_ultra) branch, including checkpoint conversion, inference, SFT, PEFT (LoRA), and pretraining examples. Read the [NVIDIA Technical Blog](https://developer.nvidia.com/blog/nvidia-nemotron-3-ultra-powers-faster-more-efficient-reasoning-for-long-running-agents/) and see the [examples README](https://github.com/NVIDIA-NeMo/Megatron-Bridge/blob/nemotron_3_ultra/examples/models/nemotron/nemotron_3/ultra/README.md) for the full walkthrough.
 
@@ -141,14 +141,14 @@ Training quickstart using pre-configured recipes:
 
 ```python
 from megatron.bridge import AutoBridge
-from megatron.bridge.recipes.llama import llama32_1b_pretrain_config
+from megatron.bridge.recipes.llama.h100 import llama32_1b_pretrain_1gpu_h100_bf16_config
 from megatron.bridge.training.gpt_step import forward_step
 from megatron.bridge.training.pretrain import pretrain
 
 if __name__ == "__main__":
     # The recipe uses the Llama 3.2 1B architecture from Hugging Face.
     # This is random-init pretraining and does not require a converted Megatron checkpoint.
-    cfg = llama32_1b_pretrain_config()
+    cfg = llama32_1b_pretrain_1gpu_h100_bf16_config()
 
     # The recipe already sets cfg.model internally using this pattern.
     # Override cfg.model to choose a different Hugging Face model ID as the architecture source.

README.md

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:26.04-py3
+ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:26.06-py3
 FROM ${BASE_IMAGE} AS megatron_bridge
 WORKDIR /opt/Megatron-Bridge
 

docker/Dockerfile.ci

@@ -96,32 +96,6 @@ RUN pip install "starlette==0.49.1" \
 # The actual onnx Python package is managed above via pip/uv; this stale copy triggers CVE scanners.
 RUN rm -rf /opt/pytorch/pytorch/third_party/onnx
 
-# Patch jackson-core CVE inside Ray's bundled fat-jar (cannot upgrade Ray itself)
-RUN JACKSON_VERSION=2.18.6 && \
-    TMPDIR=$(mktemp -d) && \
-    wget -q "https://repo1.maven.org/maven2/com/fasterxml/jackson/core/jackson-core/${JACKSON_VERSION}/jackson-core-${JACKSON_VERSION}.jar" \
-    -O "${TMPDIR}/jackson-core-fixed.jar" && \
-    unzip -q "${TMPDIR}/jackson-core-fixed.jar" -d "${TMPDIR}/jackson_fixed" && \
-    apt-get update && apt-get install -y --no-install-recommends zip && apt-get clean && \
-    for RAY_JAR in \
-    /usr/local/lib/python3.12/dist-packages/ray/jars/ray_dist.jar \
-    /opt/venv/lib/python3.12/site-packages/ray/jars/ray_dist.jar; \
-    do \
-    rm -rf "${TMPDIR}/ray_dist" && \
-    unzip -q "${RAY_JAR}" -d "${TMPDIR}/ray_dist" && \
-    find "${TMPDIR}/ray_dist" -type d -name "jackson" | while read JACKSON_DIR; do \
-        rm -rf "${JACKSON_DIR}/core"; \
-        cp -r "${TMPDIR}/jackson_fixed/com/fasterxml/jackson/core" "${JACKSON_DIR}/core"; \
-    done && \
-    find "${TMPDIR}/ray_dist/META-INF" -name "pom.properties" 2>/dev/null | \
-        xargs grep -l "jackson-core" 2>/dev/null | \
-        xargs -I{} sed -i "s/^version=.*/version=${JACKSON_VERSION}/" {} ; \
-    mv "${RAY_JAR}" "${RAY_JAR}.bak" && \
-    (cd "${TMPDIR}/ray_dist" && zip -qr "${RAY_JAR}" .) && \
-    rm -f "${RAY_JAR}.bak"; \
-    done && \
-    rm -rf "${TMPDIR}"
-
 # Patch wandb-core: bump google.golang.org/grpc from 1.79.2 to 1.79.3 (CVE fix)
 ARG TARGETARCH
 RUN GRPC_VERSION=1.79.3 && \

docker/Dockerfile.fw_final

@@ -1,12 +1,12 @@
 diff --git a/utils/torch_utils.py b/utils/torch_utils.py
-index 1eb9306ed..0e0008aab 100644
+index 9269fbb44..77693bfd6 100644
 --- a/utils/torch_utils.py
 +++ b/utils/torch_utils.py
-@@ -788,6 +788,9 @@ def is_torch_equal_or_newer(target: str) -> bool:
+@@ -828,6 +828,9 @@ def is_torch_equal_or_newer(target: str) -> bool:
      Returns:
          Whether the condition meets.
      """
-+    if target == "2.12.0.dev":
++    if target == "2.13.0.dev":
 +        return False
 +
      try:

docker/patches/vllm.patch

@@ -30,11 +30,11 @@ uv run python scripts/translate_mlm_to_bridge.py \
 ```bash
 # From a Bridge recipe name (defaults exported as MLM args)
 uv run python scripts/translate_mlm_to_bridge.py --reverse \
-    --recipe llama32_1b_pretrain_config
+    --recipe llama32_1b_pretrain_1gpu_h100_bf16_config
 
 # From a recipe plus inline overrides
 uv run python scripts/translate_mlm_to_bridge.py --reverse \
-    --recipe llama32_1b_pretrain_config \
+    --recipe llama32_1b_pretrain_1gpu_h100_bf16_config \
     --args "train.train_iters=1000 model.tensor_model_parallel_size=2"
 
 # From Bridge overrides only (no recipe)
@@ -99,7 +99,7 @@ Run the generic recipe launcher and override config keys directly:
 
 ```bash
 uv run python scripts/training/run_recipe.py \
-  --recipe llama3_8b_pretrain_config \
+  --recipe llama3_8b_pretrain_2gpu_h100_bf16_config \
   --dataset llm-pretrain \
   train.micro_batch_size=2 \
   train.global_batch_size=128 \

docs/fern/versions/nightly/pages/megatron-lm-to-megatron-bridge.mdx

@@ -166,7 +166,7 @@ config = gemma2_27b_pretrain_config(
 
 #### Gemma 2 2B
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=8 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/gemma-2-2b \
 --recipe gemma2_2b_sft_config \
 train.global_batch_size=64 \
@@ -188,7 +188,7 @@ config = gemma2_2b_sft_config(
 
 #### Gemma 2 9B
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=8 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/gemma-2-9b \
 --recipe gemma2_9b_sft_config \
 train.global_batch_size=64 \
@@ -198,7 +198,7 @@ checkpoint.save=$SAVE_DIR/gemma2_9b_finetune
 
 #### Gemma 2 27B
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=16 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=16 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/gemma-2-27b \
 --recipe gemma2_27b_sft_config \
 train.global_batch_size=64 \
@@ -210,7 +210,7 @@ checkpoint.save=$SAVE_DIR/gemma2_27b_finetune
 
 #### Gemma 2 2B
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=8 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/gemma-2-2b \
 --recipe gemma2_2b_peft_config \
 --peft_scheme lora \