Merge branch 'main' into nvegesna/llama3-8b-gb200-nvfp4-fp8-dpa

Signed-off-by: nvegesna-netizen <nvegesna@nvidia.com>

Author: nvegesna-netizen

.dev.commit

@@ -1 +1 @@
-af91b6b5cc5aec01458b8e6d24ccb0f4717e49e3
+cae55c1f0991abd3e391d0029b15bcb35723d3f6

.github/scripts/test_docker_dependency_layers.sh

@@ -99,6 +99,13 @@ ancestor_sha=$(git -C "$revision_worktree" rev-parse HEAD)
 printf 'main\n' >>"$revision_worktree/file"
 git -C "$revision_worktree" commit -qam main
 main_sha=$(git -C "$revision_worktree" rev-parse HEAD)
+git -C "$revision_worktree" switch -q -c dev
+printf 'dev\n' >>"$revision_worktree/file"
+git -C "$revision_worktree" commit -qam dev
+dev_ancestor_sha=$(git -C "$revision_worktree" rev-parse HEAD)
+printf 'dev tip\n' >>"$revision_worktree/file"
+git -C "$revision_worktree" commit -qam dev-tip
+dev_sha=$(git -C "$revision_worktree" rev-parse HEAD)
 git -C "$revision_worktree" branch contributor "$ancestor_sha"
 git -C "$revision_worktree" switch -q contributor
 printf 'approved\n' >>"$revision_worktree/file"
@@ -122,6 +129,7 @@ git -C "$revision_worktree" add unapproved
 git -C "$revision_worktree" commit -q -m unapproved
 unapproved_sha=$(git -C "$revision_worktree" rev-parse HEAD)
 git -C "$revision_worktree" push -q "$revision_repo" "$main_sha:refs/heads/main"
+git -C "$revision_worktree" push -q "$revision_repo" "$dev_sha:refs/heads/dev"
 git -C "$revision_worktree" push -q "$revision_repo" "$mirror_sha:refs/heads/pull-request/123"
 git -C "$revision_worktree" push -q "$revision_repo" "$merge_sha:refs/pull/123/merge"
 git -C "$revision_worktree" push -q \
@@ -140,6 +148,8 @@ sed "s#\.github/scripts/validate_mcore_repo\.sh#$temporary_dir/revision-bin/repo
   "$revision_validator" >"$temporary_dir/revision-validator"
 chmod +x "$temporary_dir/revision-validator"
 "$temporary_dir/revision-validator" "$revision_repo" "$ancestor_sha"
+"$temporary_dir/revision-validator" "$revision_repo" "$dev_ancestor_sha"
+"$temporary_dir/revision-validator" "$revision_repo" "$dev_sha"
 "$temporary_dir/revision-validator" "$revision_repo" "$mirror_sha"
 "$temporary_dir/revision-validator" "$revision_repo" "$merge_sha"
 "$temporary_dir/revision-validator" "$revision_repo" "$queue_sha"
@@ -222,6 +232,12 @@ if "$temporary_dir/revision-validator" "$revision_repo" not-a-full-sha; then
   exit 1
 fi
 
+if grep -q 'transformer-engine @ git+https://github.com/NVIDIA/TransformerEngine.git@' pyproject.toml || \
+  grep -q '^name = "transformer-engine"$' pyproject.toml; then
+  echo "Bridge must inherit the TransformerEngine source and metadata from the selected MCore ref" >&2
+  exit 1
+fi
+
 composite_action=".github/actions/test-template/action.yml"
 if grep -qE 'mcore_(commit|ref)|MCORE_COMMIT|uv sync --all-extras --all-groups' "$composite_action"; then
   echo "Test template must use the already-validated MCore in the built image" >&2

.github/scripts/validate_mcore_revision.sh

@@ -15,6 +15,7 @@ fi
 refs=$(git ls-remote \
   "$repo" \
   "refs/heads/main" \
+  "refs/heads/dev" \
   "refs/heads/pull-request/*" \
   "refs/heads/gh-readonly-queue/*/pr-*" \
   "refs/pull/*/merge")
@@ -23,18 +24,22 @@ trap 'rm -rf "$object_store"' EXIT
 git -C "$object_store" init --quiet
 
 main_sha=$(awk '$2 == "refs/heads/main" {print $1}' <<<"$refs")
-if [[ -n "$main_sha" ]] && \
-  git -C "$object_store" fetch --quiet --filter=blob:none --no-tags "$repo" "$main_sha" "$revision"; then
-  if git -C "$object_store" merge-base --is-ancestor "$revision" "$main_sha"; then
-    exit 0
+dev_sha=$(awk '$2 == "refs/heads/dev" {print $1}' <<<"$refs")
+for protected_sha in "$main_sha" "$dev_sha"; do
+  if [[ -n "$protected_sha" ]] && \
+    git -C "$object_store" fetch --quiet --filter=blob:none --no-tags "$repo" "$protected_sha" "$revision"; then
+    if git -C "$object_store" merge-base --is-ancestor "$revision" "$protected_sha"; then
+      exit 0
+    fi
   fi
-fi
+done
 
 while IFS=$'\t' read -r sha ref; do
   if [[ "$sha" != "$revision" ]]; then
     continue
   fi
-  if [[ "$ref" =~ ^refs/heads/pull-request/[0-9]+$ ]] || \
+  if [[ "$ref" == "refs/heads/dev" ]] || \
+    [[ "$ref" =~ ^refs/heads/pull-request/[0-9]+$ ]] || \
     [[ "$ref" =~ ^refs/heads/gh-readonly-queue/(main|dev)/pr-[0-9]+-[0-9a-f]{40}$ ]]; then
     exit 0
   fi

.main.commit

@@ -1 +1 @@
-6513e3e23d6b5eda6a1c934990b15e804237732b
+24bad8e677d22625d86ef2a54c9506b6e4992c93

3rdparty/Megatron-LM

@@ -11,6 +11,8 @@
 </div>
 
 ## 📣 News
+- [08/11/2026] **NVIDIA Nemotron 3.5 Lightning is released!** Read the technical blog [here](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightning-delivers-fast-accurate-specialized-task-execution-for-long-running-agents/). See the [Nemotron 3.5 Lightning README](https://github.com/NVIDIA-NeMo/Megatron-Bridge/blob/main/examples/models/nemotron/nemotron_3/lightning/README.md) for details.
+
 - [07/31/2026] **[K-EXAONE-2](https://huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B) is now supported**! Day-0 support for LG AI Research's K-EXAONE-2 (750B-A37B MoE) is now available on the [k-exaone-2 branch](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/k-exaone-2). Users can convert checkpoints, run inference, fine-tune with SFT or PEFT (LoRA), and use the NVFP4 inference-optimized checkpoint released alongside the model. See the branch documentation for setup instructions and examples.
 
 - [07/17/2026] [**LongStraw (MinT-2M)**](https://github.com/MindLab-Research/longstraw) is a long-context RL research system from MindLab Research that uses Megatron Bridge's modeling and LoRA capabilities. LongStraw explores resident-prefix, response-only GRPO by capturing a shared prompt once and replaying only the trainable response branches, and reports a 2.1M-token GLM-5.2 execution path across 32 NVIDIA H20 GPUs. The broader MinT system behind this work was also used to post-train [**Macaron-V1-Preview**](https://macaron.im/mindlab/research/macaron-v1-preview), MindLab's 749B Mixture-of-LoRA agent model derived from GLM-5.1. Learn more in the [paper](https://arxiv.org/abs/2607.14952).

README.md

@@ -290,6 +290,11 @@ Or with a helper script:
 ./scripts/conversion/convert.sh export --hf-model meta-llama/Llama-3.2-1B --megatron-path ./megatron_checkpoints/llama32_1b --hf-path ./hf_exports/llama32_1b
 ```
 
+Megatron-LM training checkpoints do not normally contain the Bridge
+`run_config.yaml` required by this export command. See
+[Export Megatron-LM checkpoints without a Bridge run config](megatron-lm-to-megatron-bridge.md#export-megatron-lm-checkpoints-without-a-bridge-run-config)
+for the configuration generation and validation workflow.
+
 ### Create Megatron models and run locally
 
 ```bash

docs/bridge-guide.md

@@ -155,7 +155,7 @@ from megatron.bridge.training.state import GlobalState
 # Minimal bootstrap
 state = GlobalState()
 state.cfg = megatron_cfg
-initialize_megatron(cfg=megatron_cfg)
+pg_collection = initialize_megatron(cfg=megatron_cfg)
 set_jit_fusion_options(megatron_cfg.model, megatron_cfg.train.micro_batch_size)
 
 ckpt_ctx = init_checkpointing_context(megatron_cfg.checkpoint)
@@ -165,6 +165,7 @@ model_list = get_model(
     use_torch_fsdp2=megatron_cfg.dist.use_torch_fsdp2,
     overlap_param_gather_with_optimizer_step=megatron_cfg.optimizer.overlap_param_gather_with_optimizer_step,
     data_parallel_random_init=megatron_cfg.rng.data_parallel_random_init,
+    pg_collection=pg_collection,
 )
 optimizer, scheduler = setup_optimizer(
     optimizer_config=megatron_cfg.optimizer,
@@ -440,11 +441,12 @@ class MegatronBridgeAdapter:
         from megatron.bridge.training.initialize import initialize_megatron, set_jit_fusion_options
         from megatron.bridge.models.model_provider import get_model
         from megatron.bridge.training.optim import setup_optimizer
-        initialize_megatron(cfg=self.megatron_cfg)
+        pg_collection = initialize_megatron(cfg=self.megatron_cfg)
         set_jit_fusion_options(self.megatron_cfg.model, self.megatron_cfg.train.micro_batch_size)
         self.model_list = get_model(self.megatron_cfg.model, self.megatron_cfg.ddp,
                                     use_torch_fsdp2=self.megatron_cfg.dist.use_torch_fsdp2,
-                                    overlap_param_gather_with_optimizer_step=self.megatron_cfg.optimizer.overlap_param_gather_with_optimizer_step)
+                                    overlap_param_gather_with_optimizer_step=self.megatron_cfg.optimizer.overlap_param_gather_with_optimizer_step,
+                                    pg_collection=pg_collection)
         self.model = self.model_list[0]
         self.optimizer, self.scheduler = setup_optimizer(self.megatron_cfg.optimizer, self.megatron_cfg.scheduler, self.model_list,
                                                          use_gloo_process_groups=self.megatron_cfg.dist.use_gloo_process_groups)

docs/bridge-rl-integration.md

@@ -151,7 +151,7 @@ from megatron.bridge.training.state import GlobalState
 # Minimal bootstrap
 state = GlobalState()
 state.cfg = megatron_cfg
-initialize_megatron(cfg=megatron_cfg)
+pg_collection = initialize_megatron(cfg=megatron_cfg)
 set_jit_fusion_options(megatron_cfg.model, megatron_cfg.train.micro_batch_size)
 
 ckpt_ctx = init_checkpointing_context(megatron_cfg.checkpoint)
@@ -161,6 +161,7 @@ model_list = get_model(
     use_torch_fsdp2=megatron_cfg.dist.use_torch_fsdp2,
     overlap_param_gather_with_optimizer_step=megatron_cfg.optimizer.overlap_param_gather_with_optimizer_step,
     data_parallel_random_init=megatron_cfg.rng.data_parallel_random_init,
+    pg_collection=pg_collection,
 )
 optimizer, scheduler = setup_optimizer(
     optimizer_config=megatron_cfg.optimizer,
@@ -431,11 +432,12 @@ class MegatronBridgeAdapter:
         from megatron.bridge.training.initialize import initialize_megatron, set_jit_fusion_options
         from megatron.bridge.models.model_provider import get_model
         from megatron.bridge.training.optim import setup_optimizer
-        initialize_megatron(cfg=self.megatron_cfg)
+        pg_collection = initialize_megatron(cfg=self.megatron_cfg)
         set_jit_fusion_options(self.megatron_cfg.model, self.megatron_cfg.train.micro_batch_size)
         self.model_list = get_model(self.megatron_cfg.model, self.megatron_cfg.ddp,
                                     use_torch_fsdp2=self.megatron_cfg.dist.use_torch_fsdp2,
-                                    overlap_param_gather_with_optimizer_step=self.megatron_cfg.optimizer.overlap_param_gather_with_optimizer_step)
+                                    overlap_param_gather_with_optimizer_step=self.megatron_cfg.optimizer.overlap_param_gather_with_optimizer_step,
+                                    pg_collection=pg_collection)
         self.model = self.model_list[0]
         self.optimizer, self.scheduler = setup_optimizer(self.megatron_cfg.optimizer, self.megatron_cfg.scheduler, self.model_list,
                                                          use_gloo_process_groups=self.megatron_cfg.dist.use_gloo_process_groups)

docs/fern/versions/0.4.2/pages/bridge-rl-integration.mdx

@@ -258,6 +258,11 @@ Or with a helper script:
 ./scripts/conversion/convert.sh export --hf-model meta-llama/Llama-3.2-1B --megatron-path ./megatron_checkpoints/llama32_1b --hf-path ./hf_exports/llama32_1b
 ```
 
+Megatron-LM training checkpoints do not normally contain the Bridge
+`run_config.yaml` required by this export command. See
+[Export Megatron-LM checkpoints without a Bridge run config](megatron-lm-to-megatron-bridge.md#export-megatron-lm-checkpoints-without-a-bridge-run-config)
+for the configuration generation and validation workflow.
+
 ### Create Megatron models and run locally
 
 ```bash