mxfp8 support

.gitmodules

@@ -1,8 +1,3 @@
-[submodule "3rdparty/Megatron-LM"]
-	path = 3rdparty/Megatron-LM-workspace/Megatron-LM
-	url = https://github.com/NVIDIA/Megatron-LM.git
-	branch = main
-	shallow = true
 [submodule "3rdparty/Megatron-Bridge"]
 	path = 3rdparty/Megatron-Bridge-workspace/Megatron-Bridge
 	url = https://github.com/NVIDIA-NeMo/Megatron-Bridge.git

3rdparty/Megatron-Bridge-workspace/setup.py

@@ -25,6 +25,7 @@
 bridge_src_dir = "Megatron-Bridge/src/megatron/bridge"
 bridge_package_name = "megatron.bridge"
 
+# Default dependencies from pyproject.toml
 CACHED_DEPENDENCIES = [
     "transformers>=5.0.0,<=5.3.0",
     "peft>=0.18.1",
@@ -48,14 +49,14 @@
     "megatron-core[dev,mlm]",
     "qwen-vl-utils",
     # TODO(https://github.com/NVIDIA-NeMo/RL/issues/2111): upgrade to core_cu13 when we move to CUDA 13 base container
-    "transformer-engine[pytorch,core_cu12]",
+    "transformer-engine[pytorch,core_cu13]",
     "mamba-ssm",
-    "nvidia-resiliency-ext~=0.5.0",
+    "nvidia-resiliency-ext",
     "causal-conv1d",
     "flash-linear-attention",
     "timm",
     "open-clip-torch>=3.2.0",
-    "mlflow>=3.5.0",
+    "mlflow>=3.9.0",
     "comet-ml>=3.50.0",
     "torch>=2.6.0",
 ]
@@ -73,20 +74,9 @@
     project = data["project"]
     deps_list = project["dependencies"]
     submodule_deps = set(str(d).strip() for d in deps_list)
-
-    # Normalize the transformer-engine CUDA variant extra (core_cu12 vs core_cu13)
-    # so our CUDA 12 override doesn't trip the consistency check against the
-    # submodule's CUDA 13 default.
-    # TODO(https://github.com/NVIDIA-NeMo/RL/issues/2111): remove this when we upgrade to CUDA 13
-    def _normalize_te_cuda(dep):
-        if dep.startswith("transformer-engine") or dep.startswith("transformer_engine"):
-            return dep.replace("core_cu13", "core_cu12")
-        return dep
-
-    normalized_submodule = set(_normalize_te_cuda(d) for d in submodule_deps)
-    normalized_cached = set(_normalize_te_cuda(d) for d in CACHED_DEPENDENCIES)
-    missing_in_cached = normalized_submodule - normalized_cached
-    extra_in_cached = normalized_cached - normalized_submodule
+    cached_deps = set(CACHED_DEPENDENCIES)
+    missing_in_cached = submodule_deps - cached_deps
+    extra_in_cached = cached_deps - submodule_deps
 
     if missing_in_cached or extra_in_cached:
         print(